Skip to content

Release 1.3.0 - #884

Merged
zlareb1 merged 42 commits into
mainfrom
prep/release-1.3.0
Nov 19, 2026
Merged

zlareb1 merged 42 commits into
mainfrom
prep/release-1.3.0

Conversation

@zlareb1

@zlareb1 zlareb1 commented Nov 18, 2026 •

Copy link
Copy Markdown
Member

Related: #956
Related: #957
Related: #933
Related: #940
Related: #941
Related: #942
Related: #943
Related: #945
Related: #947
Related: #950
Related: #952
Related: #953
Related: #955
Related: #931
Related: #935
Related: #901
Related: #927
Related: #936
Related: #937
Related: #938
Related: #932
Related: #928
Related: #934
Related: #855
Related: #856
Related: #924
Related: #809
Related: #845
Related: #888
Related: #907
Related: #908
Related: #911
Related: #913
Related: #872
Related: #875
Related: #890
Related: #902

Changelog category (leave one):

  • Not for changelog (changelog entry is not required)

Why

Prepare version 1.3.0 with the included features, client changes and deployment steps.

Scope and risk

The server, primary Python and TypeScript SDKs, frontend, root packages and skills use 1.3.0. The chart uses version 0.3.0 and appVersion 1.3.0. Both coding-agent plugins use 1.2.39 and require server 1.3.0.

This composition includes current main, canonical-parent eligibility for conversation chunks, integrity-locked native tooling and outside SDK acceptance checks. The source proposals for the chunk and tooling changes remain open and unmodified. Their corrected source is incorporated and qualified in this release pull request; no separate proposal merge is claimed. The chunk tests describe the canonical memory gate, and the complete n8n manifest, lock and patch retain the dependency mitigation already on main.

Against the included main snapshot, the diff has 55 paths: 24 release-version, minimum-server and outside-SDK paths, seven chunk-parent paths and 24 additional tooling paths. Every included change requires this exact composition's independent whole-source and complete public-content review and all applicable CI checks before merge. The outside SDK matrix compares published metadata with the checked release tag, then installs each of the 15 Python extras and eight npm rows separately. npm rows disable lifecycle scripts and check registry signatures before importing SDK code.

The existing preparation history is retained. Candidate publication, runtime qualification, the real two-day soak, registry publication and outside checks follow the qualified release merge.

Compatibility, rollout and rollback

  • Version pins add no further wire field, setting, migration or feature activation. The release's complete client and operator changes are listed below, including the optional git-proxy identity fields and additive migration 112.
  • Upgrade the server before relying on the plugins' server calls and connected-source guidance. Update plugins through each client's plugin manager.
  • Keep mirrored-source creation, query-history drafting, gap drafting and source push disabled unless their documented requirements are met. Native analyst provisioning and ordinary-event compatibility retain their supported runtime proof.
  • Rolling upgrades and rollback retain accounting, erasure, durable session authority, deployment protection and the optional proxy restrictions in the notes. Reverting version pins does not reverse migrations or remove durable state.
  • Native compiler tooling uses the installer and adjacent locked inputs from one checkout, with Node for the target platform and optional dependencies present. Command names and the install prefix remain stable.

Tests

  • The maintained version and chart checks, Ruff, public-prose and diff checks pass on this committed composition. Its 55-path comparison retains the complete current-main n8n tree.
  • The maintained prepare.py notes 1.3.0 command completes on this composition. All 45 generated wire lines are retained verbatim: 375 operations, 14 added operations, 26 added request fields and 29 added response fields since v1.2.0.
  • Selected local regression runs remain unrun after focused supervisors refused admission before execution. No alternate runner or higher resource bound was used. Their refusal is retained independently of hosted results.
  • Every applicable hosted check passes on this exact composition, including the metadata regressions, analyzer and mixed-version gates, both Postgres authority jobs and all eight native platform legs. An earlier dry run with the same executable source passed both architecture quickstarts; this snapshot also includes the two documentation walkthroughs.
  • The locked-tooling source has successful hosted platform and image checks on its published head. Its bounded local native control remains incomplete after a memory refusal and supplies no passing release qualification.
  • Candidate artifacts, ordinary runtime acceptance, the real required 48-hour soak, registry publication, signatures, installers and published SDK verification remain separate release gates.

AI assistance in the code (leave one; writing this description with AI is fine):

  • AI-generated: Codex prepared version reconciliation, source composition and release notes. Independent source and complete public-content review is required.

  • I can explain each authored change and its interaction with the included source.

Checklist

  • Every applicable remote CI check passes on the exact composition, including format, lint, types, unit, integration, schema, SDK, frontend, plugin and security checks.
  • Version pins, minimum-server values and source comparison match this composition.
  • Complete release notes preserve client, operator and rollback requirements and the freshly generated wire changes.
  • Fresh independent whole-source and complete public-content review has no findings or nits.
  • Commits are signed off.

Release notes for v1.3.0

New features

Improvements and fixes

Compatibility and deployment

  • A Postgres-backed process reads cell state before serving. A process without LOOM_CELL refuses an enforced or partial cell; a failed initial Postgres read fails boot, except while migrations are pending, when the process remains unready. (A server process with LOOM_CELL unset is gated by its deployment's cell state #586)

  • An enforced cell checks tenancy on writes, including the tenancy row itself. LOOM_PG_DSN must connect directly to Postgres. Older pods on an already prepared later cell must be rolled in one step once migrating boot installs the triggers. loom doctor detects missing, altered or disabled triggers and lost session markers. (A cell's Postgres refuses an unheld org's row from a session that ran no tenancy check #599, A cell refuses a tenancy row from a session that ran no tenancy check #676)

  • CLI writes check the deployment and require the grants in upgrade step 15. Without LOOM_CELL, they refuse a cell past recording. Commands configured for a cell refuse another cell's stores; revocations that cannot restore access retain their documented outage behavior. New org names are bounded at 2048 UTF-8 bytes. (An org-declaring CLI with LOOM_CELL unset is refused on a cell past recording #668, A cell records and the router places no org name longer than 2048 bytes #671, An org-free CLI command says whether it reads, writes or revokes, and a write on another cell's stores refuses #673)

  • memory.promote for a definition requires approved_sha256 and promotes the namespace-wide row by the token's actor. Sensitive values, empty terms, conflicting terms or aliases and the shared-definition cap refuse promotion. (A definition is promoted from its namespace-wide row with the hash an admin approves, under a name no other shared definition holds #667)

  • Native HTTPS uses LOOM_TLS_CERT_FILE and LOOM_TLS_KEY_FILE, with certificate reload in each web process. The chart's tls.secretName configures the certificate mount and HTTPS probes. The image launcher refuses UVICORN_* variables; remove those variables before upgrading and use the documented Loom settings. (The web port serves HTTPS from a certificate and key it reloads when they change #756)

  • Each metric retains up to otel_metrics_cardinality_limit series per process, default 20,000; additional attribute sets accumulate in an overflow series. (OpenTelemetry 1.45.2: each metric keeps up to otel_metrics_cardinality_limit series, 20000 by default #648)

  • Query-history and gap drafting introduce private definition candidates. Before downgrading to an older reader, use a maintenance window: disable both drafting features, quiesce drafting/review/registration/worker traffic, drain queued/model work, settle uncertain sends and retained leases, remove all versions of private candidate kinds and all registrations, and verify zero on every readable storage replica. Keep private producers disabled throughout the downgrade. Published definitions retain their existing lifecycle. (Draft definitions from query history and search gaps #768)

  • memory.search and memory.search_trace accept at most 64 raw entity_filter entries, including repeated names. Canonical alias expansion is bounded at 256 names. Entity-summary retrieval separately bounds the case-insensitive distinct union of the canonical filter, query handles and known vocabulary tokens at 256 names. Overflow of either 256-name bound returns 422 VECTOR_INPUT_TOO_LARGE through retries, memory.ask widening and composed search, without a partial ranking. Narrow the query or filter before retrying. (Entity filters refuse oversized alias expansion #800, Entity summary reads reject oversized seed sets without partial fallback #808)

  • POST /v1/me/personal_tokens.create issues the explicit personal-agent token kind only from the person's own primary sign-in session. The optional signed reads_as_creator claim defaults to false and ordinary delegation does not copy it. No operation-transport RPC is added by this route. (Checked document reads enforce current source permissions #799)

  • Source membership and listing evidence must be present and current for checked reads. LOOM_KNOWLEDGE_SOURCE_ACL_STALE_SECONDS defaults to 3600; no backfill infers domain membership from an email address. Before rolling back after mirrored writes, stop those writers, settle outstanding work and follow the documented mirror erasure and telemetry sweep. Keep mirrored creation disabled on an incomplete fleet. (Checked document reads enforce current source permissions #799)

  • Ordinary event inserts and signed, namespace-scoped analysis remain available across serving builds. agent_events retains ordinary columns and values; a permanent ordinary_events_only CHECK rejects marked events and reserved mirror namespaces. Marked requests and mirror workers write the separate agent_events_private store, which has no analyst grant or analyst-visible physical metadata. (Ordinary event analysis remains available across serving builds #825)

  • The signed analysis_agent_events catalog view unions ordinary and private events and masks protected fields before caller predicates and grouping. Ordinary agent_events reads retain their signed namespace scope; the analyst cannot read the private table. Trusted cost, budget and history-policy reads cover both stores, and user erasure scrubs both through agent_events.scrub. Private events use the ordinary ledger's retention and tiering policy. (Marked document requests keep accounting behind their privacy boundary #803, Ordinary event analysis remains available across serving builds #825)

  • Analyst administration uses loom ch-setup --apply-analyst-sql with explicitly configured existing access-management credentials. It applies only the analyst boundary under the live proof and lease, without rebuilding reader credentials. --print-analyst-sql refuses with a nonzero result and prints no SQL or secrets. General --print-sql remains available for database and reader preparation; print and apply modes are mutually exclusive. (Ordinary event analysis remains available across serving builds #825)

  • Session-content operations require exactly one X-Loom-Session-Protection: v1 header once deployment protection is set. Missing, repeated or unknown declarations grant no promise. The gate includes dual-use memory.set, memory.bulk_set, memory.search and code/context operations used by automatic clients. Missing protection or unverifiable deployment state returns nonretryable 409 SESSION_PROTECTION_REQUIRED. HTTP checks before reading the body; MCP checks before content validation, identity defaults, replay or implicit session runs. (Session-content admission uses durable deployment protection #819)

  • The optional LOOM_KNOWLEDGE_MIRRORED_SOURCES_ENABLED setting now exists, defaults false and requires restart. Setting it true persists a deployment-wide latch before startup admits clients or workers; the heartbeat repairs that protection. Setting it false does not clear the latch. With the switch false, an undeclared client still needs a fresh service-principal Keeper absence proof; a declared client does not exercise that default path. This setting does not make public mirrored-source creation available. (Session-content admission uses durable deployment protection #819)

  • Named-user budget status withholds exact spend and its percentage except for a verified platform admin without namespace-role restrictions. Admission and routing still count all spend, including marked requests. User warnings always withhold exact spend, including webhooks and operator-triggered delivery. Namespace and org totals retain their existing numeric form. (Named users have daily cost caps and inbox warnings #777)

  • Before a downgrade after budget warnings were produced, stop request/background budget-check producers and status polling, drain scheduled deliveries, settle uncertain writes and writer leases, then remove only budget.warning events and verify completion and zero warnings on every storage replica. Save the global and namespace policies first: older namespace writes reset the new cap/inbox fields. Keep the feature disabled on older builds and reapply saved policy only after the full fleet supports it again. (Named users have daily cost caps and inbox warnings #777)

  • Tracker feeds support a single self-hosted server or shared storage with ReplacingMergeTree or SharedReplacingMergeTree; self-hosted replicated tables are refused. Feed connections retain the deployment's storage-authority, organization and cell checks. The dedicated writer must have verified limited direct and effective grants and remain separate from the serving reader. (A native GitHub feed supplies team-owned tracker records #786)

  • Assistant mirror admission uses a separate strict server-issued answer_mirrors Boolean. Missing admission is false, and mirror discovery also requires the caller's separate session_protection promise. An older verifier accepting the signed claim does not enforce it. Keep mirrored-source creation disabled until every serving build and consumer enforces the documented privacy protocol. No migration or backfill is added by this protocol change. (Assistant answers bind mirror admission and private accounting #835)

  • Applications using SDK session protection must provision an independent conversation family's durable authority before content, share it across participating clients, credentials and servers, and reopen the same saved pointer on restore. Missing, damaged or inaccessible authority suppresses automatic work; restore cannot silently initialize clean state. Capsules contain control state rather than conversation text, have no expiry, reset or repair path, and must survive rollback. Namespace erasure and credential changes do not authorize resetting this application-owned state. (SDK clients retain durable protection for mirrored session content #840)

  • Python accepts optional session_protection and supplies a private POSIX capsule shared with the Node primitive. Native Windows refuses without a separately qualified durable ACL-backed authority. Python automatic callbacks use with family.automatic():; captured contributors survive authentication, body reads, credential replay and deferred writes. Every contributor must be durably protected before mirrored output returns. Interrupted or malformed observations protect known families and preserve the primary refusal, network error or cancellation. (SDK clients retain durable protection for mirrored session content #840)

  • TypeScript accepts optional sessionProtection with an application-owned SessionProtectionStore. The store rejects missing or damaged state and commits monotonic protection before mark() resolves. An in-memory store is insufficient. Every concurrent same-process participant in a family must reuse the same handle, including its pending-mark latch, and restore the same durable record. SessionCapture controls belong in SDK call options, not operation arguments. (SDK clients retain durable protection for mirrored session content #840)

  • Bound Python SSE observes each complete frame before releasing its original bytes. Bound TypeScript SSE routes default readers, standard body methods and ordinary clones through the same observer, preserving frame bytes and response metadata. Split line endings and UTF-8, multiline data, comments and an initial byte-order mark are supported. Invalid frames, interrupted or partial EOF, unsupported media or compression refuse and protect contributors. BYOB readers are unsupported. Unfinished frames are bounded to 4 MiB; that limit does not bound caller-created native clone queues. Ordinary unbound responses retain their behavior. (SDK clients retain durable protection for mirrored session content #840, TypeScript streams and sessions retain contributor protection #849)

  • TypeScript assistant calls accept optional capture controls; explicit calls can use observation-only observe: new SessionCapture(roots). The typed assistant requires [DONE] before completing an answer and keeps draining for later metadata or protection errors. Cancellation and read failures protect contributors and join bounded owned cleanup while retaining the source error identity. Observation complements the existing provisional-output and required-accounting boundary; it does not replace it. (TypeScript streams and sessions retain contributor protection #849)

  • Python and TypeScript Session buffers retain all contributing authorities. Python deferred producers and coalescer windows also retain their original run context; a manual flusher cannot donate a clean family. TypeScript uses immutable capture controls and retains its contributor union across buffer boundaries. Protection discards pending and carry-over text instead of requeueing it. Synchronous buffered state knows local taint; asynchronous boundaries check durable state before dispatch. Clean extraction retries preserve older-before-newer ordering after another state check; zero carry-over does not resend an older batch. Explicit offload, fetch and note remain available with observation and cannot relabel automatic work to evade suppression. (SDK clients retain durable protection for mirrored session content #840, TypeScript streams and sessions retain contributor protection #849)

  • Use one Agent Framework LoomContextProvider or Pipecat LoomMemoryService per logical conversation family, with its complete existing application-owned authority bound before construction. Restore reopens that authority; native user and session ids only route requests. Both check before later extraction batches. Suppression after a model or tool ran cannot undo its answer or effects. Pipecat retains contributors before awaiting native frame processing, suppresses Loom automatic memory work while frames continue, and discards only service-owned retry and write tracking rather than application history. Durable observation failures and cancellation still propagate; unsent tails are not acknowledged. (Native memory callbacks retain their conversation protection #850)

  • Use one native conversation-history instance or LangGraph store per logical conversation family, with complete existing authority bound before construction and the same capsule reopened on restore. Native history reads and additions run automatically under the retained union, including after credential refresh; a marked write stops before the next item. Explicit pop and clear observe replies without clearing an enclosing automatic context. Store namespace, key, thread and user ids do not establish protection authority. A past automatic store invocation does not make later direct calls automatic. Separate Python and TypeScript history formats remain separate. (Native conversation history retains contributor protection #859, Native stores retain contributor authority per invocation #865)

  • These SDK controls add no server capability declaration, server migration or mirror activation. The streaming, callback, history and store changes preserve the capsule format introduced by the SDK durable-authority foundation. Concrete TypeScript storage backends, native Windows authority, plugin hook lineage, restored-host lifecycle and whole-fleet qualification remain separate work. The two native callback integrations and store/history changes do not qualify every framework or an installed framework merely because structural tests pass. Retain durable records on rollback; older adapters cannot declare protection support or reset authority. Existing fleet drain, private-data erasure and deployment-latch cleanup still govern returning older readers after mirror enablement. (SDK clients retain durable protection for mirrored session content #840, TypeScript streams and sessions retain contributor protection #849, Native memory callbacks retain their conversation protection #850, Native conversation history retains contributor protection #859, Native stores retain contributor authority per invocation #865)

  • The provider OAuth HTTP helper supplies consent, code exchange, identity metadata, refresh and revocation requests for the personal-account backend. Its non-live bounds remain LOOM_KNOWLEDGE_CONNECTED_OAUTH_TIMEOUT_S=30.0 and LOOM_KNOWLEDGE_CONNECTED_OAUTH_RESPONSE_MAX_BYTES=65536, with a minimum response bound of 1024 bytes. Keeping providers unconfigured and connected creation disabled requires no provider registration or live credential. (Connected-account design follows provider token lifecycles #860, Document connections use fixed provider OAuth clients #863, Personal source connections use canonical account ownership #883)

  • OAuth helper requests use fixed origins, configured callback validation, explicit scopes, browser-state input and S256 where documented. They refuse redirects, do not retry, bound metadata before parsing and share the caller's remaining deadline. Credential failures are redacted and provider error bodies remain unread. Only Google can retain an omitted replacement refresh token; Notion and Linear require replacement material. Optional email or expiry is not fabricated, and provider identifiers, email and opaque scope text confer no Loom owner or source authority. The connection backend binds the helper to consumed browser state, canonical authorization, serialized account generations, sealed refresh storage, local inactivation before remote revocation and canonical-owner erasure. The HTTP helper itself stores no durable state; the backend adds the separate account registry and migration described below. (Document connections use fixed provider OAuth clients #863, Personal source connections use canonical account ownership #883)

  • When using native memory capture, construct separate storage and native memory instances for each user and logical conversation family, with the complete existing capture supplied before construction. Reopen the same durable authority on restore. Supplying capture makes native callbacks automatic; omitting it leaves explicitly owned CRUD observation-only and must not be used to evade automatic protection. A marked recall returns only after durable observation; later automatic work stops. Writes committed before a refusal can remain stored, with no transaction or automatic rollback. Keep earlier batch effects and protection records during recovery and do not blindly retry a refused callback. (Mastra native memory retains conversation protection #874)

  • Private proof channels preserve the original shared-client generation, endpoint, principal, backend attestation, checked-session, Keeper and nonce checks. Query retries on the private channel are zero, but driver-internal network retries remain; a socket reopening or upstream rerouting still requires validation. Missing or invalid checked sessions refuse rather than being silently recreated. Ordinary traffic retains its shared driver. (Session protection proofs own their HTTP channel #876)

  • Four streaming units bound live proof channels. Lifetime-capacity waiting uses the caller's remaining original five-second proof deadline; each factory and statement retains ordinary pool admission, including its default 500 ms cutoff and 64-unit pool. Failed or cancelled proofs force-close their channel; successful close precedes capacity release. No setting or credential change is added. Rollback needs no data migration but can restore pooled-session discontinuity risk; it does not authorize recreating checked sessions, weakening timeouts or disabling the latch. Existing durable authority and fail-closed privacy requirements remain. (Session protection proofs own their HTTP channel #876)

  • LlamaIndex aflush retains each Loom block's contributors before reading the native buffer and through dispatch and archive. A refused flush raises without archiving its active window or acknowledging an unsent extraction; started blocks finish before their exceptions propagate. Earlier successful sends can remain stored, and native overflow archival is not rolled back. Preserving the active window does not authorize retrying protected work or provide a transaction. Wire, schema, settings and dependencies are unchanged; no operator setup is added. Rollback reverts the adapter, guide and sender census together while retaining the application's durable protection authority. (Retain protection across native memory callbacks #905)

  • Browser consent uses a primary signed-in human subject, one-use browser-bound state, an HttpOnly cookie and PKCE where supported. Delegated callers and service identities cannot connect as a person. Authenticated POST /v1/connect/{kind}/start checks cell admission; browser-bound GET /v1/connect/{kind}/callback checks the pending binding and current placement before consuming state or calling a provider. These routes are always excluded from automatic HTTP tracing before request attributes or headers are captured; external proxy and error-log policies remain the operator's responsibility. (Personal source connections use canonical account ownership #883)

  • connection.list and connection.disconnect address only the authenticated caller's own accounts. knowledge.source.add accepts credential_mode=shared|connected, default shared; the connected form defaults credential to empty, while shared registration still requires its credential. Source entries add credential_mode, access and a restricted owner_subject projection. A connected owner with namespace writer standing can create, update, check, remove or request synchronization of that source. Shared credentials, another person's source and credential rotation retain administrator requirements. Administrator checks of another owner's connection do not send native item requests or disclose item existence. (Personal source connections use canonical account ownership #883)

  • The account registry uses the canonical sign-in identity (org, subject, kind) across namespaces. Namespace, end-user, agent and session ids do not confer account ownership. Credentials and optional display email are sealed and absent from source metadata. Checked provider access retains the exact source, owner and connection generation across waits and body reads; redirects, partial permission listings or traversal outside the registered scope refuse. Document permissions still govern independently authorized readers. Revalidation can remove stale read-plan evidence but cannot introduce replacement sources. (Personal source connections use canonical account ownership #883)

  • Account and namespace leases protect final publication, with provider and embedding waits outside those leases. Successful provider token rotation is retained before a later source or presenter refusal; uncertain in-flight rotation fails closed. Disconnect and deactivation fence the generation, clear credentials and queue exact owned source ids for retirement. Canonical-owner cleanup removes historical account and source rows with residue checks; namespace or end-user erasure does not substitute for that owner key, and queued source retirement does not certify mirror-content erasure. Root-secret rotation also reseals the account registry. (Personal source connections use canonical account ownership #883)

  • New provider registrations use LOOM_CONNECT_GDRIVE_CLIENT_ID/LOOM_CONNECT_GDRIVE_CLIENT_SECRET, the corresponding LOOM_CONNECT_NOTION_* and LOOM_CONNECT_LINEAR_* pairs. Empty client ids disable their providers. LOOM_CONNECT_NOTION_READ_CONTENT_ONLY defaults false and may be set true only after verifying that registration's read-content-only capability. No provider registration, usable credential or live provider acceptance is supplied by the release. (Personal source connections use canonical account ownership #883)

  • Connection lifecycle bounds default to LOOM_KNOWLEDGE_CONNECTED_STOP_TIMEOUT_S=20.0, LOOM_KNOWLEDGE_CONNECTED_REVOKE_TIMEOUT_S=5.0, LOOM_KNOWLEDGE_CONNECTED_LEASE_TTL_S=60.0 and LOOM_KNOWLEDGE_CONNECTED_LEASE_WAIT_S=5.0. The complete owner scan limit defaults to 4096 and resealing page size to 256 through LOOM_KNOWLEDGE_CONNECTED_REGISTRY_SCAN_LIMIT and LOOM_KNOWLEDGE_CONNECTED_REGISTRY_PAGE_SIZE; an oversized owner inventory refuses instead of stopping a subset. Live connected-source limits are separately LOOM_KNOWLEDGE_CONNECTED_SOURCES_MAX_PER_NAMESPACE=64 and LOOM_KNOWLEDGE_CONNECTED_SOURCES_MAX_PER_PERSON=8, the latter per person in one namespace. LOOM_KNOWLEDGE_SOURCE_ACL_MAX_ENTRIES=10000 bounds complete project permission listings; exceeding it refuses publication. Review the configuration reference before optional use. (Personal source connections use canonical account ownership #883)

  • dispatch_seeds no longer emits the unsupported virtual_by_name result source. Indexed overrides and the bounded inheritance walk remain. This query-time correction changes no operation or field type and requires no migration, backfill or reindex. (Require indexed relationships for dispatch expansion #907)

  • SDK code lookup uses an existing readable code index and application-bound scope. Supplying code_namespace or codeNamespace enables it; explicitly selecting it without a namespace refuses construction. The code result character budget defaults to 8000 and must be an integer of at least 256 when enabled. Symbol and search requests use a fixed 2500-token budget; search returns at most three results, callers allows depths 1–3 with at most 20 edges, and impact allows depths 1–2 with at most 20 symbols per hop. Whole bodies or rows may be omitted with client_truncation; server exactness cannot certify client-omitted evidence. Insufficient metadata is disclosed rather than sliced. (SDK model tools support scoped code lookup #908)

  • Memory-tool result objects add receipt and trust fields. Ordinary risk and confidence scores apply to non-governed hits; governed definitions retain their separate authority and term-leg status. Direct files and matching repository association do not enter the compact memory-tool schema. No wire field or storage migration is introduced by exposing the existing Python argument. Reverting the SDK restores its former projections and convenience signature without removing stored facts or application protection records. (SDK memory tools preserve receipts and recall trust #911)

  • Use separate callback instances for unrelated conversation families and bind the existing application-owned durable protection before construction. Native user, session and framework identifiers do not establish protection authority. LangChain retrievers, CrewAI memory, Google ADK content methods and the structural store use automatic protection when the application already owns automatic context; otherwise direct calls remain explicit and observe responses. No new capsule format, capability declaration, server migration or dependency is introduced. Retain the same authority through restore and rollback. (Retain conversation protection in native LangChain retrieval #904, CrewAI memory callbacks retain session protection #893, Google ADK memory callbacks retain session protection #898, Retain conversation protection across memory-store callbacks #910)

  • AG2 policy assembly and AutoGen update_context are automatic callback boundaries. AutoGen retains contributors before awaiting native messages and observes a marked recall before injecting a system message. Direct AutoGen query and add, and model-selected AG2 tools, retain their explicit behavior outside automatic context. (AG2 recall policies retain session protection across callbacks #882, AutoGen memory retains session protection across native callbacks #889)

  • Google ADK refuses the callback when a marked memory result suppresses its optional knowledge request; it does not manufacture a partial memory response. Interrupted extraction raises without acknowledging unsent events; earlier successful sends can remain stored. Exact-reference deletion retains observation without becoming automatic content capture. (Google ADK memory callbacks retain session protection #898)

  • The structural memory-store correction covers its store methods and real Loom SDK transport boundary. It does not qualify an external native MemoryManager or Agent that may handle store errors itself. The existing extraction default is unchanged; structural conformance is not complete native framework qualification. (Retain conversation protection across memory-store callbacks #910)

  • Proof diagnostics do not change public errors, security predicates, checked sessions, deadlines, channel limits, writer coverage, insert draining or census requirements. Stage labels and numeric timing show the last observed work, not a diagnosed transport or database cause. They do not establish deployment recovery, admission success or privacy readiness. (Event storage proof failures identify their stage #885, Session proof refusals report elapsed phases and deadline expiry #892)

  • Plain Gemini CLI setup merges the MCP entry while retaining other valid JSON settings and servers. A commented JSON configuration is refused rather than rewritten without comments; its manual-entry message redacts the token. Without an explicit client flag, Gemini CLI is detected from its settings file or executable, not its directory alone. Reverting support leaves an entry already written in the client configuration. (loom install --client gemini writes Gemini CLI's MCP entry #744)

  • Unsandboxed repository self-service requires LOOM_CODE_SELF_SERVICE_PLATFORM_ADMIN_ONLY=true, LOOM_CODE_SELF_SERVICE_GIT_UNSANDBOXED=true, effective authenticated operation, a non-open MCP auth mode and LOOM_CODE_SANDBOX_REQUIRED=false. Repository registration and remote-branch operations require the caller's own flat platform-admin token; namespace roles or an impersonated caller do not suffice. The git-only exception covers remote listing, bare cloning and fetching, not compiler-tier execution. (A deployment without the repository-job sandbox can index repositories its platform admins add #912)

  • This opt-in exposes server network access and the server user's SSH credentials to the permitted git remotes. It supplies no private-address or redirect filtering and no sandbox CPU, memory, file or process limits; existing git timeouts and disk bounds still apply. Existing active registry entries resume too. The dashboard continues to suppress repository addition while the unsandboxed warning is present. Older builds keep refusing or skipping this work without the sandbox; registry rows survive rollback. (A deployment without the repository-job sandbox can index repositories its platform admins add #912)

  • The Lua lane is outside the extractor stamp, so this addition does not trigger global re-extraction. Existing Lua files need a targeted path backfill to gain the new rows. A rollback returns extraction to text-only behavior; stored Lua symbols and edges remain until reindexing. No tree-sitter dependency or compiler tier is added. (The code index reads Lua files with a tree-less lane #743)

  • The Rust client requires Rust 1.84 or later and Tokio. Search crosses agents by default unless narrowed explicitly. A write reuses its idempotency key only within that call's retries; starting another call after an uncertain result can create another write, and cancelling a call does not undo a committed write. Rust source packaging checks do not establish registry publication or complete SDK capability coverage. (feat: A minimal Rust client for the memory ops: set, get, search and delete #842)

  • The Langflow component has a default result limit of five and requires a namespace plus a key for saving memory. Memory calls accept the configured user identity; document searches use the configured namespace. The component source and stubbed tests do not establish a live native Langflow session, automatic conversation capture, or complete session-protection capability qualification. (A Langflow component saves and searches Loom memory and knowledge documents #832)

  • Browser consent that explicitly prefers HTML returns to the fixed same-origin Connected accounts page; API clients retain the JSON callback result. Both responses disable caching, suppress the referrer, vary on Accept and clear the binding cookie. The dashboard return URL contains no OAuth code, state, token or caller-supplied destination. Dashboard controls do not activate mirrored sources or replace provider registration, fleet qualification or session protection. (Settings manages personal connected accounts #897)

  • Haystack client factories retain application authority instead of serializing protection state with connection settings. Retrievers retain both store and callback contributors through search, get and download. Agno databases retain constructor and later callback families while native user and store ids only partition data. Use separate instances for unrelated conversation families and reopen the same durable authority after restore or rollback. Earlier successful effects are not rolled back by a later refusal. (Retain session protection across document-store clients #909, Agno database callbacks retain session contributors #887)

  • Coverage claim measured details add freshness, per-file commits and selected symbol identities, the index watermark, unverified and unmeasured paths, and run_current. A mixed-commit result has an empty commit. Historical rows use coverage_history and the existing file_reference tier. An unchanged file can remain conservatively unverified when its last parsing commit differs from the measurement; other evidence can still establish coverage, while stale or partial measurements cannot establish nonexecution. Stored evidence remains readable and no schema migration is added. (Coverage claims verify measured source identities #875)

  • Live connected-source search requires authenticated operation, an existing protected client session, mirrored-source enablement and current checked account/source authority. It searches only the first page, without tags, with a query of at most 1000 characters. A single deadline of at most five seconds covers the live leg, with at most three registrations, ten hits, four Drive items and three Notion ancestor lookups. Provider responses are capped at 256 KiB or the smaller configured source limit; there are no retries or pagination. The knowledge_live_search_* settings can reduce these ceilings. (Search personal connected sources without storing live hits #890)

  • Live results retain account generation, owner, registered corpus and checked native membership through response publication. Drive folder searches cover direct children, not recursive ingestion; explicit file items do not widen native search. Notion title search requires registered page or database membership. Missing or timed-out live output is not evidence that no document exists. Marked HTTP, SDK and MCP replies retain existing protection and accounting; bodies=none omits live excerpts. Live results do not become persisted gaps, traces, snapshots, offloads or model prompts. No new content table or erasure substrate is created. (Search personal connected sources without storing live hits #890)

  • LOOM_MCP_CODE_UPLOAD_TOOLS is restart-only and defaults false. When enabled, the repository-entry requirement applies to uploads over HTTP as well as MCP; a namespace without an entry returns RESOURCE_NOT_FOUND. Unread repository state refuses, and self-service entries follow the existing registry refresh interval. Both operations are callable by name outside the assistant toolset; loom_op on readonly routes only status. With LOOM_MCP_CODE_TOOLS=true, all lists both and readonly lists status; curated rosters still list neither. Capability and generated operation catalogs continue to describe the default hidden surface. Older builds keep answering OP_NOT_ON_MCP. Reverting hides the tools while uploaded parts retain their TTL and complete batches retain their existing apply behavior. (A deployment can serve code.scip_ingest and code.scip_status over MCP, for its repositories only #916)

  • Restorations matching the current blacklist return 422 BLACKLISTED. Undo checks before recording acceptance; namespace restore checks before namespace/grant creation and rechecks before insertion; rollback checks every planned live reinstatement before its first batch. Postgres commits retain captured policy generations and refuse a policy change. A refused consolidated summary neither embeds nor retires its sources. These corrections add no migration, backfill, reindex, setting or wire field and do not replace the existing restore, rollback, erasure or deployment qualifications. (Restoration and consolidation enforce the current blacklist #923)

  • The chunk-parent correction changes eligibility without adding a wire field, migration, reindex, backfill, setting or operator setup. Reverting restores the earlier eligibility behavior and does not repair historical parent identities. It does not change ranking weights or establish a query-cost improvement. (Memory search excludes chunks whose known parent is not served #924)

  • The composed locked-tooling change requires a native compiler installer and its adjacent code-toolchains directory from the same checkout. Use Node for the target platform and retain optional dependencies; missing or mismatched native inputs stop installation. Relative executable links retain the existing prefix and copied-image layout. Rollback reinstalls the earlier script and matching inputs without a database change. The included source and matching artifacts are qualified in this release composition. (Node tooling installs from integrity-locked dependency trees #856)

Operator steps

  1. Follow the upgrade checklist and apply the additive migrations through the normal migrating boot. This release's cell and shared-memory changes have additional steps below.

  2. After every older replica is gone, follow upgrade step 14 to identify promoted rows left by earlier namespace erasures. Use the ledger's census cutoff and namespace when applying the documented tombstone or purge mutation, including hidden versions. Only a namespace-list erasure authorizes a purge. Erasures beyond the ledger's retention cannot be identified by that query. (A namespace erasure tombstones the namespace's promoted rows in loom.shared #572)

  3. Run loom secrets reseal with this build so document-source secrets and any existing source-push registry secrets are covered. No synthetic push setup is needed for an empty registry. (A document source reads up to 300 items given as links or ids, and update adds and removes them #562, Provider notifications request document source syncs #784)

  4. Before migrating reserved mirror namespaces, resolve any migration refusal with the preceding build: erase the affected namespace contents, remove its code repository and delete the namespace, then retry. (No token reaches a namespace named acl- and 24 hex digits, and no org-wide read lists one #665)

  5. Apply upgrade step 15's CLI grants, including ClickHouse SELECT on loom.storage_authority, the Postgres cell-state grants, connection limit and bucket settings. Complete step 16 before first-cell attach. (An org-declaring CLI with LOOM_CELL unset is refused on a cell past recording #668, An org-free CLI command says whether it reads, writes or revokes, and a write on another cell's stores refuses #673)

  6. Upgrade every router replica before attaching cells or upgrading cell processes. The cell processes read router state on each heartbeat; older routers do not have the required per-replica cache. Validate readiness and session tenancy markers after rollout. (The router serves the first cell's attach and removes an unattached cell's row #681, A cell's heartbeat reports serves, marked and the router's answer beside its fingerprint #693)

  7. Keep the document-source limit at or below 50,000 while any older replica serves. After rollout, the live setting may be raised. (A document source's document bound is a setting for every kind, and its stored copies are read in pages #662)

  8. After all older replicas are gone, have an admin promote each existing shared definition again so its approved hash and token-stamped actor are recorded. The new live cap memory_shared_definitions_max defaults to 500. (A definition is promoted from its namespace-wide row with the hash an admin approves, under a name no other shared definition holds #667)

  9. For local assistant models, configure the model and endpoint, mount the assistant secret and add explicit network-policy egress to the model server's port. (LOOM_LLM_PROVIDER=local runs Loom's model jobs and the assistant on a model server on the machine, with no API key, and the installer offers a running Ollama #634)

  10. Before enabling TLS, mount a certificate/key pair and change the upstream proxy to HTTPS in the same deployment window. Deployments using launcher-level UVICORN_* variables must remove them even when TLS remains off. (The web port serves HTTPS from a certificate and key it reloads when they change #756)

  11. A desktop-launched coding client must find Node on its startup PATH; configure the client's environment when Node is installed only through a version manager. Upgrade Loom to 1.3.0 before relying on plugin 1.2.39's new server calls. (The Claude Code plugin's hooks and header helper run a script file by a literal plugin path #573, Code answers carry the team's memories about the files they return #638)

  12. Install an SDK extra or optional peer for each framework integration used. Separate plugin/community packages require their own installation and publication prerequisites; the release does not imply catalog approval. (Google ADK memory services use Loom in both SDKs #696, The Python SDK ships a Pipecat memory service #750, The n8n package provides Loom memory and search tools #754, Dify workflows use native Loom memory and knowledge tools #763, Haystack pipelines store and retrieve Loom documents #765, Agno user memories use Loom through its native database interface #766)

  13. Before analyst provisioning, use a ClickHouse release that supplies system.constraints, as the pinned database release does. Missing CHECK metadata refuses the cutover. The operator must verify this on the actual database before the candidate starts. (Ordinary event analysis remains available across serving builds #825)

  14. Apply the additive query-history registration migration and upgrade every server and worker before registering a source or creating private candidates. knowledge_query_history_enabled defaults off; gap drafting also requires knowledge_gaps_enabled. Review the bounded read/pass/source/group/draft settings before enabling either feature. (Draft definitions from query history and search gaps #768)

  15. For the private cell probe, configure Postgres and authentication, choose cellProbe.port and finish both web and worker rollouts before preparation. Use the Service origin for the probe URL. Configure this port only through the chart value, not env, extraEnv, settingsFile or secrets.extra. Set zero to disable it. (Expose a private cell probe through the Helm chart #806)

  16. Apply the additive source-ACL schema before serving checked reads. Keep public mirrored creation disabled until the documented fleet, request, session and enablement protections are complete. (Checked document reads enforce current source permissions #799)

  17. Apply migrations 103–106 in order for checked reads and marked accounting. Migration 106 pauses and waits for the three local refresh schedulers, changes their definitions and resumes them; complete the cutover on every refresh-owning replica. Existing ordinary aggregate types are preserved and prior input is not replayed. Keep mirrored creation disabled. (Checked document reads enforce current source permissions #799, Marked document requests keep accounting behind their privacy boundary #803)

  18. Apply migration 107 before configuring named-user budgets. While any serving build is older or unknown, keep LOOM_DEFAULT_DAILY_USER_COST_CAP_USD=0, every namespace's daily_user_cost_cap_usd=0 and budget_alert_inbox=''. The last value disables warning configuration; it does not require emptying an inbox and does not drain existing sends. The global user cap and LOOM_COST_BUDGET_ALERT_SHARE are live; the share defaults to 0.8. Do not rely on user caps during overlap. (Named users have daily cost caps and inbox warnings #777)

  19. Apply migration 108 and preserve its deployment protection state. Keep the mirrored-source switch and both private-drafting switches off on every server and worker through the mixed fleet. Verify ordinary undeclared client calls through the supported route, including the required Keeper proof. Missing metadata, backend identity changes or uncertain proof writes refuse admission; adding a promise header is not a repair. Only the documented complete mirror cleanup can clear protection, last; tenant erasure and a local switch change cannot. (Session-content admission uses durable deployment protection #819)

  20. Apply migration 109. Keep LOOM_KNOWLEDGE_SOURCE_PUSH_ENABLED=false unless opting into source push; this switch is separate from mirrored-source protection. Later setup needs a fleet supporting the registry's erasure, an explicit HTTPS LOOM_PUBLIC_BASE_URL origin and provider-specific confirmation. Source credential/item changes retire setup. Before rollback after use, disable push and complete the documented fenced registry cleanup, including historical versions, while preserving unrelated namespace data. (Provider notifications request document source syncs #784)

  21. Documentation publication uses the existing site workflow. Automatic translation generation requires the documented restricted runner, initialized budget ledger and explicit cumulative allowance, which defaults to zero. This release does not configure that allowance or authorize generation. (Documentation publishes and maintains five translation locales #764)

  22. Apply migration 110 after migration 109. It creates private event storage, adds the ordinary-table CHECK and updates union readers without moving, erasing or replaying existing rows or aggregate history. Complete its pause/wait/update/resume of both refresh definitions on every refresh-owning replica. Before grants, analyst provisioning verifies exact CHECK/table identities on every writer, freezes and drains the finite set of prepared INSERTs, then requires a zero contamination census. Existing contaminated rows remain untouched for explicit operator resolution. ALTER completion or client disconnection alone does not settle a prepared write. (Ordinary event analysis remains available across serving builds #825)

  23. LOOM_CH_ANALYSIS_OBSERVER_URLS is a restart-only list of additional trusted writer HTTP origins, default []. The configured service origin is included and the existing service credentials are reused. Every writer UUID needs exactly one attested route. Complete native policy-catalog visibility, principal/role identity and non-replaying POST semantics are required; missing/duplicate writers, hidden metadata, redirects, HTTPS downgrade, identity drift or incomplete replies refuse provisioning. Intermediaries must preserve those semantics. Use the guarded analyst apply command for administrative setup. (Ordinary event analysis remains available across serving builds #825)

  24. Before an older build returns after mirrored enablement, settle outstanding writes and complete the documented mirror cleanup, including agent_events_private, marked receipts/audit figures and the cost rollup, before clearing deployment protection. Disabling a feature or downgrading schema alone does not establish privacy. Keep mirrored creation disabled until the separate fleet, client, restore and final erasure requirements are complete. (Marked document requests keep accounting behind their privacy boundary #803, Session-content admission uses durable deployment protection #819, Ordinary event analysis remains available across serving builds #825)

  25. To opt into tracker feeds, create the dedicated record with separate writer/reader credentials, set the tracker organization, initialize the record and register its read-only contract views, then configure and schedule each allowlisted repository. Attest the supplied token's read-only permissions: a successful API read does not certify its grants. Renew installation tokens externally and configure each replacement before expiry. (A native GitHub feed supplies team-owned tracker records #786)

  26. Tracker visibility observations expire after 15 minutes and are not an instant permission mirror. Turns are bounded to 120 seconds, 100 items per page and 2–100 pages per stream. Incremental polling does not detect deletions. Stop every writer before erasing raw rows, comments and checkpoints or retiring a repository. Namespace deletion removes Loom's derived state and grants, not the external record or its backups. Rollback stops the scheduled writer; observation expiry hides serving rows without erasing the record. (A native GitHub feed supplies team-owned tracker records #786)

  27. When opting into bound SDK protection, provision the application's durable conversation authority before constructing its participating clients, native history or store. Restore the same authority and retain its records through rollback. These optional APIs do not declare capability or enable mirrors. They require no connected-account credential setup or provider call. Keep the existing fleet, privacy and native analyst checks in force. (SDK clients retain durable protection for mirrored session content #840, TypeScript streams and sessions retain contributor protection #849, Native memory callbacks retain their conversation protection #850, Native conversation history retains contributor protection #859, Document connections use fixed provider OAuth clients #863, Native stores retain contributor authority per invocation #865, Mastra native memory retains conversation protection #874)

  28. On the first start with the system-log retention definition, ClickHouse can rename existing tables to the first available query_log_N and text_log_N names, including suffixes beyond _0. Those tables retain their rows and old retention policy; the configuration does not delete them. Before deliberately removing any exact old table, preserve the diagnostics you need and follow the deployment guide's Log retention section. An old table without a TTL can retain rows indefinitely. (The bundled ClickHouse keeps query_log and text_log for 30 days #828)

  29. An external self-managed ClickHouse receives this policy only after its operator installs the retention configuration and restarts it. If text_log is disabled, retain only the query_log block: adding the text_log block enables trace-level logging. Removing the configuration can rename current logs again and leaves replacement logs without this TTL; retained suffixed tables remain separate. A Loom server-image upgrade alone does not apply this policy to an external database. (The bundled ClickHouse keeps query_log and text_log for 30 days #828)

  30. Apply additive migration 111 before the new build. It adds connected_accounts and shared-preserving source defaults. Connected registrations leave legacy provider-kind and credential fields empty so older workers cannot treat them as shared credentials. Keep connected and mirrored creation disabled until the existing fleet, session and privacy requirements are qualified; applying the migration does not activate the feature. (Personal source connections use canonical account ownership #883)

  31. Before optional connected-account use, configure the provider's registration and verified callback origin, confirm its documented permissions and review the bounded lifecycle/source/permission settings. Keep existing source synchronization and mirror enablement requirements in force. A successful source-head test or a provider's metadata is not live account acceptance or proof of operator registration capabilities. (Personal source connections use canonical account ownership #883)

  32. Stop every server and worker before a restore. Under Postgres authority, loom reconcile --apply clears restored connection authority before publishing held documents for either restored store; cleanup failure leaves publication held, and a live override does not skip the offline check. On ClickHouse-only deployments, restore the configured stores, run python dev/restore_connected_accounts.py as a dry run, then run it with --apply. Use normal deployment configuration without arbitrary host/database remapping; recent worker activity refuses cleanup. (Personal source connections use canonical account ownership #883)

  33. Before restored serving or rollback to an older reader, settle writers and complete canonical-owner account cleanup plus the separate mirror-content cleanup. Account cleanup fences historical connection keys, queues exact source ids and verifies physical deletion of account/source history; it does not prove erasure of every document, chunk, head or original. Failure leaves serving or publication held. Local cleanup precedes bounded best-effort provider revocation; owners establish connections again afterward. Reverting code leaves additive schema and stored rows and cannot substitute for cleanup. (Personal source connections use canonical account ownership #883)

  34. For SDK code tools, bind the existing code namespace explicitly and use a token already authorized to read it. The application fixes scope and budgets; authorization failures do not retry another namespace. Honor graph lower bounds, abstention and client omission markers. Memory tools retain their separate namespace and the existing session-protection requirements. (SDK model tools support scoped code lookup #908, SDK memory tools preserve receipts and recall trust #911, Code reads receive their authorized organization #913)

  35. Deploy the server update for the optional-organization read correction. Explicit and required identity fields keep their rules; writes retain their validated organization and stable adjacent-build replay hashes. These read and dispatch corrections require no new credential, migration or reindex. (Require indexed relationships for dispatch expansion #907, Code reads receive their authorized organization #913)

  36. To configure plain Gemini CLI MCP, use loom install --client gemini with the normal endpoint and scoped token, or use the dashboard token dialog, then restart the client. The configuration uses httpUrl for streamable HTTP. If installation refuses a commented settings file, preserve it and add the redacted example deliberately with the actual scoped token; do not discard unrelated settings or comments. Plain MCP setup and the existing extension remain distinct installation choices. (loom install --client gemini writes Gemini CLI's MCP entry #744)

  37. Before opting into unsandboxed repository self-service, review all active repository registry rows, including entries outside the current caller's visible list, and remove untrusted remotes through code.repo.remove. Authenticate every caller, restrict registration to platform admins, review the server user's SSH access and network reach, set both new flags with the required auth and sandbox settings, then restart. Use the platform admin's own flat token for the direct repository operations. Leaving the flags false retains the existing behavior. This opt-in is not a release requirement or an instruction to bypass an existing refusal. (A deployment without the repository-job sandbox can index repositories its platform admins add #912)

  38. To index existing Lua source with the new lane, supply the intended .lua paths to loom code backfill --paths-file; do not request a global re-extraction solely for this lane. Review unresolved dynamic imports and receiver calls as coverage limits. (The code index reads Lua files with a tree-less lane #743)

  39. After deploying the coverage import correction, run loom code coverage-import --namespace <namespace> --dry-run against the intended measured source, inspect unmapped test names, then import without --dry-run to rebuild pairs. Updating the server alone does not rebuild stored coverage. Do not reinterpret historical ranges as current execution merely because a repository watermark advances. (Coverage imports map pytest contexts to live test files #872, Coverage claims verify measured source identities #875)

  40. Before using Settings → Connected accounts, complete the existing provider registration, callback-origin and personal-session prerequisites. Confirm the account and current organization before connecting or disconnecting; a disconnect stops that connection's document sources. The UI does not replace migration 111, canonical-owner cleanup or the separate mirror-content and fleet requirements in steps 30–33. (Settings manages personal connected accounts #897, Personal source connections use canonical account ownership #883)

  41. Leave live search unrequested and mirrored-source creation disabled until the documented fleet, client, authority and runtime privacy requirements are complete. Only then opt into search.query(live=true) with the existing qualified client and inspect the separate live section and its coverage limits. Reduce knowledge_live_search_* bounds where needed. Do not change a promise, header, latch or credential to repair an admission refusal; rollback after mirrored use retains the existing writer settlement and cleanup requirements. (Search personal connected sources without storing live hits #890)

  42. Leave LOOM_MCP_CODE_UPLOAD_TOOLS=false unless an uploader needs the MCP route. To opt in, configure the repository entry, give the uploader a writer token naming its exact repository namespace, set the flag and restart. Do not give an agent an organization-wide wildcard writer token. The existing LOOM_CODE_SCIP_APPLY_ENABLED switch separately controls the apply worker; this MCP flag does not start it. (A deployment can serve code.scip_ingest and code.scip_status over MCP, for its repositories only #916)

  43. The software-factory example requires its documented indexed public repository, separate scoped service principals and an inbox policy that permits the review handoff. Configure those application resources before running the example. (Document memory between coding-agent delivery stages #931, Add a software-factory memory handoff example #935, Link software-factory guidance to architecture and evidence references #945)

  44. Install the integration directory in the Langflow environment and restart Langflow to discover the LFX extension. Configure the reachable Loom endpoint and its ordinary scoped credentials. The core SDK release does not separately publish this extension package. (Package the Langflow component as an LFX extension #943)

  45. Install the Pipecat extra and configure the text example with its scoped application credentials and trusted pointer file. Configure any providers on its Loom deployment under the existing provider policy. (Demonstrate Pipecat recall and capture in a text pipeline #947)

  46. Install the experimental Hermes source plugin in one user profile against its documented 0.23.2 API and configure a namespace-scoped reader or writer token. Automatic turn/transcript synchronization remains off; live write/read qualification is still pending. (Add an experimental native Hermes memory provider #940)

  47. For optional git proxy use, apply additive migration 112_code_repo_git_proxy_identity.sql and roll this build to every pod before setting LOOM_CODE_GIT_PROXY_URL. Configure LOOM_CODE_GIT_PROXY_TOKEN_FILE, LOOM_CODE_GIT_PROXY_CA_FILE and LOOM_CODE_GIT_PROXY_IDENTITY where needed, including web pods that keep mirrors. This proxy mode supports the existing unsandboxed administrative configuration: repository-job sandbox must be off, and self-service code.repo.add additionally requires LOOM_CODE_SELF_SERVICE_GIT_UNSANDBOXED and its platform-admin rule. An entry in LOOM_CODE_REPOS_JSON needs neither self-service permission. A previous build reads and copies repository rows by column name, so the migration supports the roll, but it ignores these settings, warns once at boot and uses direct git. A deployment requiring proxy routing cannot roll back to that build. (Network git of the code index can go through an operator's forward proxy, with an identity per repository #933)

  48. To run the optional team-memory walkthrough, use an authenticated deployment with TLS for remote clients, its operator's signing-key environment and disposable scoped credentials and namespace. The server must support the absent-row precondition and admit the write under its session-protection policy. Stop on a precondition or protection refusal; retain the precondition, compare the returned identity and follow token revocation after the check. This is an application walkthrough, not a release deployment or a captured access-control result. (Document team memory with scoped credentials #956)

Release artifacts

Release assets include the installer, stack bundle, SDK packages, coding-agent plugin archives and agent setup contract. Each carries a <file>.sigstore.json bundle for cosign verify-blob; verify the certificate identity for this release's release.yml tag and the workflow OIDC issuer. Older assets without a bundle cannot pass this file-signature check. Both container images also carry signatures. (#689)

Maintenance checks

The SDK close-cancellation regression holds the first point write unfinished and checks that cancellation is accepted. The separate first-set regression controls its clock so the intended writes share one coalescing window. Cancellation, ordering and completed-write assertions remain; the production SDK and batching window are unchanged. (#932, #934)

The n8n build uses scoped overrides for minimatch 11.0.8, qs 6.16.1 and uuid 11.1.1. A hash-bound assembler backport in stream-json 1.9.1 makes ordinary and reviver __proto__ writes own data properties while retaining the CommonJS API. This is a local mitigation on 1.9.1; it does not claim an upstream fixed version, clearance of other dependency advisories or live n8n qualification. The esbuild 0.28.1 and pnpm 10.34.9 inputs remain. (#855)

The composed tooling change installs compiler indexers, package managers and agent test CLIs from committed integrity locks with vendor lifecycle scripts disabled. A local binder requires matching native optional packages and contained binaries; missing inputs stop installation without a replacement download. Agent platform checks use the locked tree's native executable paths. Published SDK acceptance checks disable lifecycle scripts and verify registry signatures before import. The separate source proposal remains open. These incorporated changes require this release composition's exact-head CI and independent review. (#856)

Hook failure tests require an observed server request within their existing attempt limit and deadlines, with separate fixture state for each attempt. (#888)

Translation reconciliation retains each completed reply's original source partition when planning corrections across source updates. Source, reply, reuse and financial evidence must be authenticated by the caller; planning authorizes no generation or provider call. (#877)

Route-census tests construct the authentication surface in a fresh interpreter and retain the complete route assertions. (#881)

Reviewed cost guides and Portuguese release-process and code-index operations pages are included, with English canonical and incomplete coverage disclosed. (#870)

Translation checks preserve source-attested format names in the code-index table while requiring ordinary descriptions to be translated. (#880)

Translation validation indexes protected ranges and table rows to avoid repeated scans on large reference pages. Request boundaries, labels and source offsets are preserved. (#834)

Original-unit reconciliation permits correction of a faulty completed whole-page translation when the prepared English input is unchanged. Existing request bounds and changed-input replacement restrictions remain. (#839)

Tracker reply parsing gains regression coverage, and CI action dependencies are updated. Additional tests cover budget bounds, memory-history diffs, audit privacy identifiers and port allocation. (#675, #705, #674, #700, #821, #812)

Translation link checks preserve the longest source-attested repository path suffix in a label and compare equivalent localized or English fallback labels through canonical source targets. Source spelling, target, occurrence and untranslated-prose restrictions remain. (#843, #846)

Translation reconciliation accepts optional authenticated raw source origins when retained lines prove current counterparts through supported preparation transforms. Insertions must have unique endpoints; complete original units, separators and history cover the current container. Changed units retain their fault and reply references. Whole-page changes must use current English as the new input and fit the existing UTF-8 request bound without splitting. Missing or ambiguous origins, replaced endpoints and reordered retained lines refuse. Interior source replacements can be reconciled only when retained lines prove the original request boundaries; conservative proof can still refuse replacements whose text prevents unique ordered correspondence. Completed request origins can also be retained when ordinary paragraphs are appended under the final heading and the original request boundaries remain proved. This permits neither replay of an older response against changed input nor generation, publication or expanded coverage claims. (#847, #873, #927, #953)

Lane-cost failures retain per-sample sizes, completion and sanitized stage-error evidence; this is diagnostic support rather than a parser speed or historical-cause claim. Coalescer fixtures control their existing submission-window premise without changing production clocks, budgets, batch limits or ordering assertions. Subprocess cancellation coverage waits for its existing readiness signals while preserving deadlines and process-group cleanup assertions. (#854, #861, #864, #866, #871)

The release guide uses prep/release-X.Y.Z for preparation. Major and minor preparation starts from and targets main; patches start from and target the maintained release/X.Y line. Maintained lines begin at the first final vX.Y.0 commit, candidates use vX.Y.Z-rc.N, and final tags use vX.Y.Z. Existing published history is preserved; naming or retargeting cannot resolve a refused update. This documentation changes no repository setting or existing branch. (#844)

The shared-memory evaluation guide separates cross-client recall and corrections from denied namespace access and revoked or expired credentials. A successful shared read or an unrelated transport failure is not evidence that authorization isolation passed. The guide supplies evaluation criteria, not an executed deployment result. (#894)

The n8n package build uses esbuild 0.28.1 and pnpm 10.34.9 through its manifest and frozen lockfile. This build dependency update adds no Loom wire or stored-data change and does not establish publication, catalog approval or independent runtime acceptance for the community package. (#755)

Erasure classification now distinguishes canonical-owner tables from namespace erasure tables, step-only tables and tables with no memory content. loom.connected_accounts belongs to canonical sign-in owner erasure, including sealed provider identity and credential history; namespace and memory user_id erasure remain separate scopes. This documents the existing owner-erasure implementation and adds no migration, runtime operation or cleanup step. (#917)

task.progress records progress without extending a claimed task's lease. Use task.heartbeat to renew the lease. This corrects the documentation of existing behavior rather than changing task scheduling or deadlines. (#919)

The SDK issue harness compares an optional direct SDK code-index binding with an identical local-tool baseline and grades patches outside the agent. Frozen manifests, usage and partial patches retain unsuccessful attempts; preflight, tool and cleanup refusals keep their execution stage and HTTP status through failed supervision. Controllers stop on refusals before grading or advancing. This is a harness addition, not a published benchmark result, an actual SDK model-tool or CLI comparison, imported coverage evaluation, or a performance claim. (#902)

Translation page validation shares each table parse between label and layout checks while retaining the existing checks and standalone freshness. It adds no cache shared between requests or generated translation changes. (#926)

The connected-account browser check matches the complete connect-button name while disconnect finishes; the action, viewport cases and assertions remain. This changes no application or API behavior. (#930)

The schema reference distinguishes registration from proposal: memory_schema.propose always stores and returns proposed, regardless of the supplied status. Registration retains its active/proposed choice, and the shared input default remains active. This corrects descriptions without changing validation, handlers or wire schemas. (#921)

What changed for clients since v1.2.0

375 operations served (was 361); ops +14 −0; request fields +26 −0; response fields +29 −0.

New operations

  • connection.disconnect — role VIEWER, budget 30000 ms, MCP extended; request fields: kind, org
  • connection.list — role VIEWER, budget 1000 ms, MCP extended; request fields: org
  • knowledge.gaps.draft — role ADMIN, budget 30000 ms, MCP extended; request fields: gap_id, namespace, org
  • knowledge.gaps.list — role ADMIN, budget 300 ms, MCP extended; request fields: limit, namespace, org
  • knowledge.move — role WRITER, budget 8000 ms, MCP extended; request fields: namespace, org, source_uris, to_namespace
  • knowledge.query_history.add — role ADMIN, budget 30000 ms, MCP extended; request fields: cluster, databases, host, name, namespace, org, password, port, secure, user
  • knowledge.query_history.list — role ADMIN, budget 250 ms, MCP extended; request fields: fields, limit, namespace, org
  • knowledge.query_history.remove — role ADMIN, budget 1000 ms, MCP extended; request fields: name, namespace, org
  • knowledge.source.parse_links — role VIEWER, budget 50 ms, MCP extended; request fields: links, namespace, org
  • knowledge.source.push_configure — role ADMIN, budget 60000 ms, MCP extended; request fields: agent, enabled, name, namespace, org, session_id, user_id
  • knowledge.source.push_confirm — role ADMIN, budget 1000 ms, MCP extended; request fields: challenge_fingerprint, generation, name, namespace, org, signing_secret
  • knowledge.source.push_status — role ADMIN, budget 1000 ms, MCP extended; request fields: name, namespace, org, reveal_challenge
  • knowledge.source.update — role WRITER, budget 60000 ms, MCP extended; request fields: add_items, add_links, include, name, namespace, org, remove_items, remove_links
  • memory.definitions.list — role ADMIN, budget 400 ms, MCP extended; request fields: namespace, org

Changed operations

  • assistant.ask: response +mirrored
  • code.callees: request +include_lessons, lesson_namespaces; response +lessons, lessons_left_out
  • code.callers: request +include_lessons, lesson_namespaces; response +lessons, lessons_left_out
  • code.context_pack: request +include_lessons, lesson_namespaces; response +lessons, lessons_left_out
  • code.impact: request +include_lessons, lesson_namespaces; response +lessons, lessons_left_out
  • code.outline: request +include_lessons, lesson_namespaces; response +lessons, lessons_left_out
  • code.repo.add: request +git_proxy_identity
  • code.repo.remote_branches: request +git_proxy_identity
  • code.search: request +include_lessons, lesson_namespaces; response +lessons, lessons_left_out
  • collection.list: request +include_document_counts
  • cost.budget_status: request +user_id; response +namespace_over_cap, user_id
  • knowledge.list: request +namespaces
  • knowledge.source.add: request +credential_mode, include, items, links; response +include, items; required_role ADMIN → WRITER; budget_p99_ms 30000 → 60000
  • knowledge.source.check: response +checked, sampled; required_role ADMIN → WRITER
  • knowledge.source.inspect: response +access, credential_mode, include, items, owner_subject
  • knowledge.source.remove: required_role ADMIN → WRITER
  • knowledge.source.sync_now: required_role ADMIN → WRITER
  • knowledge.synthesize: request +user_id
  • memory.promote: request +approved_sha256; response +ignored_args
  • memory.search: response +term_leg
  • namespace.get: response +budget_alert_inbox, daily_user_cost_cap_usd
  • namespace.register: request +budget_alert_inbox, daily_user_cost_cap_usd
  • search.query: request +live; response +live

Other client changes

  • General search's optional tracker response section can carry mode="search", namespaces, namespace_limit and cited results. Existing duplicate verdicts keep their shape. (General search includes readable tracker issues and pull requests #773)

  • memory.search hits include agent, identifying the writer when row identity is available; otherwise it is empty. (Google ADK memory services use Loom in both SDKs #696)

  • collection.list entries include nullable document_count. It is null unless include_document_counts is requested and counts documents directly in the collection. (Library collections show document counts and refresh states #771)

  • Learning entries include source_resource_id, defaulting to an empty string. learnings.review promotes private definition candidates only to memory. (Draft definitions from query history and search gaps #768)

  • cost.budget_status user mode requires a namespace. spent_24h_usd is nullable, and pct_of_cap is null when user spend is withheld. cap_source adds namespace_user and global_user; nullable namespace_over_cap reports the independent namespace ceiling. Namespace get/list entries expose daily_user_cost_cap_usd and budget_alert_inbox. A positive namespace user cap overrides the global default; zero inherits, and zero at both tiers disables it. (Named users have daily cost caps and inbox warnings #777)

  • knowledge.synthesize accepts user_id, default '-'; the Python convenience method forwards its optional user id. Transport user pins take precedence. Named-user caps cover supported synthesis and query expansion; unattributed work remains namespace-only. (Named users have daily cost caps and inbox warnings #777)

  • Budget warnings arrive asynchronously on a later point check or status poll after the configured share is reached; there is no discovery scan. Repeated observations deduplicate within a UTC day for the same scope, inbox, cap and share. User warning payloads set spent_24h_usd to null for every recipient. Polling and watch streams filter warning user/session scope before pagination; ordinary task events retain inbox visibility. (Named users have daily cost caps and inbox warnings #777)

  • Assistant answer tokens carry their caller's strict Boolean session_protection claim. Descendant answers cannot override it with a transport header, and ordinary delegation does not inherit it. Clients promising protection must stop automatic session-content capture after a mirrored read and preserve that state across child/resumed sessions. (Session-content admission uses durable deployment protection #819)

  • Multi-entry MCP batches replace every response-bearing marked entry, including denied or empty entries, with INVALID_REQUEST and Send this request individually. Notifications remain silent. Single calls and one-entry arrays retain checked reads. (Marked document requests keep accounting behind their privacy boundary #803)

  • Protected history makes enforcing policy rules evaluate potentially matching marked participation as unknown before success/argument filters, unresolved references or negation; it can refuse reads as well as writes. policy.check reports protected history unavailable with null protected leaf values. Shadow policies remain non-enforcing. (Marked document requests keep accounting behind their privacy boundary #803)

  • Marked cost is disclosed only to a verified unrestricted platform admin or authorized namespace/whole-day totals without fine-grained filters. Fine-grained marked cost is withheld; marked-only latency has no ordinary samples and returns zero latency values. This does not disable internal budget checks. (Marked document requests keep accounting behind their privacy boundary #803)

  • Source-push setup operations are experimental, namespace-admin only and non-idempotent, including status. push_status reveals a pending verification token only with reveal_challenge=true; ordinary inspection, audit and idempotency replay omit it. The response's enabled reports the global switch, while status reports source setup state. (Provider notifications request document source syncs #784)

  • Provider notifications use the separate non-RPC POST /hooks/knowledge/{namespace}/{source_id}/{generation}?org=<encoded-org> route with provider proof, rather than an RPC bearer token. Bodies are bounded at 65,536 bytes. A 200 acknowledgement can represent an early synchronization notification and does not prove that work was requested or content ingested. Drive uses channel-token verification rather than signed delivery; other supported providers require their documented HMAC and identity checks. (Provider notifications request document source syncs #784)

  • assistant.ask adds optional response field mirrored, default false. Marked chat output begins with mirrored: true; marked SSE begins with data: {"mirrored":true} before answer text. HTTP/MCP success and error output retains the mark after serialization and protocol unwrapping; MCP structured and JSON text results preserve it, and a marked search summary follows a first JSON marker block. (Assistant answers bind mirror admission and private accounting #835)

  • The server fixes answer_mirrors before spend or model work from active mirrored registrations in the authorized namespace. Active registrations keep the answer marked even when ACL checks exclude every private row. Caller fields, headers, token-mint endpoints and CLI flags cannot supply admission. An initially ordinary answer cannot gain a mirror leg after source activation. Ordinary delegated children omit both answer_mirrors and session_protection; descendants cannot widen either claim with a header. Current document authorization still applies. (Assistant answers bind mirror admission and private accounting #835)

  • Incremental assistant text is provisional. A marked response retains required accounting after headers, joins producer cleanup through repeated cancellation and confirms its required private cost row before terminal completion. Failed required capture or delivery cannot complete successfully. Whole responses and checked source frames retain accounting before publication. (Assistant answers bind mirror admission and private accounting #835)

  • Python and TypeScript assistant streams require [DONE] before returning a completed answer. Incomplete EOF fails; text after [DONE] is ignored while later transport errors still propagate. Stream completion does not qualify durable automatic capture or privacy state across child/resumed sessions; clients promising protection must satisfy the separate session contract. (Assistant answers bind mirror admission and private accounting #835)

Align published version pins and set the coding-agent plugins minimum server version for the new code lessons calls.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
* Draft definitions from query history and search gaps (#768)

* Draft metric definitions from query history and search gaps

Keep generated drafts private until an administrator publishes their content. Register opt-in read-only query-history sources and bound generation by privacy thresholds and namespace caps.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allow an isolated name for mixed-version checks

Use the requested container name for both creation and cleanup. A failed creation leaves any existing container untouched.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Pin query-history setting defaults

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Query-history settings distinguish cadence from deadlines

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Describe definition draft operations and review reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Document retained definition heads during draft recovery

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Show complete definition drafts before publication

Reviewers can inspect the full body and evidence before promoting a definition. Describe private draft destinations and authenticated approval attribution.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition support counts in memory lineage

Private definition drafts retain a supporting-account count without storing identities. Both exact lineage paths use that count while ordinary and assistant-rule candidates keep their existing accounting.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private definition decisions against erasure

Private queue writes retain their proposal time and share a bounded writer lease with namespace deletion. Check erasure under that lease and retain uncertain sends so a delayed decision or model completion cannot restore erased evidence.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private publication and source registry erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Expect retained writer fences in namespace erasure traces

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allocate query-history migration after mirrored erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Require private draft cleanup before older readers serve

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep gap input out of stored draft evidence

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate statement shapes on the composed schema

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition review tests within their type contracts

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Complete definition drafting contracts and Library composition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Curation rollout names the existing absence precondition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Entity filters refuse oversized alias expansion (#800)

* Entity filters refuse oversized alias expansion

Bound raw inputs and distinct canonical aliases before constructing the search filter. Preserve supported filter order and propagate size refusals through optional paths.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* The SDK error inventory includes oversized vector input

The new server refusal uses the existing SDK validation error. Include its code in the exhaustive inventory so the registry assertion continues to cover every server error.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* The alias ceiling states its fixed input contract

The effective entity-filter ceiling is a documented input boundary rather than an operator resource setting. Explain that distinction beside the constant without changing the value or the settings checks.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Analyzer corpus tracks the bounded distinct entity query

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Checked document reads enforce current source permissions (#799)

* Personal-agent tokens name their reader through an explicit signed claim

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Resolve mirrored readers from verified account evidence

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Stage source-bound permission checks for mirror reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep mirror candidates out of shared lexical statistics

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fill mirrored candidate pools after current permission checks

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Hydrate only exact chunks that retain current mirrored access

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Rank complete admitted source copies before projecting public addresses

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Refresh all mirrored search passes before selecting a route

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Collect ordinary search legs separately from fusion

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Use checked source pools in knowledge search

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Read source documents at exact admitted identities

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Check source access on document gets and keep mining ordinary

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Read admitted document chunks including embedding stubs

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Read checked source sections in document outlines

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Refresh source and reader evidence before checked reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Read checked source originals through guarded streams

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Read checked list fields and counts from one admitted copy

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Group list metadata before admission and public pagination

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* List admitted document families before filtering and pagination

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Get and download the newest admitted mirror copy

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Invalidate synthesis proofs after a document identity or permission change

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Assemble synthesis from checked whole document copies

Keep ordinary chunks and document precedence. Pin one mirrored copy per address and discard all mirrored chunks when a checked read fails.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Recheck document access around synthesis calls

Validate every mirrored copy before and after each model call and before returning an answer. Discard generated output after permission loss, drain concurrent calls, and report only documents the synthesis read.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Read admitted source titles without document diagnostics

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Include checked source titles in document inspections

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Retain metadata proofs for live search chunks

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Renew checked search proofs across provider work

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Retain permission deadlines for composed document reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Renew composed document reads before rendering their results

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep checked repository listings out of shared caches

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Renew repository search permissions after both search legs

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Renew document access before unified search assembly

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep checked stack knowledge through final assembly

Prepare mirror access before stack reads and retain the checked selection through
rendering. Rebuild the complete ordinary suffix when its evidence expires.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Prepare checked reads before context-pack providers

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Prepare checked ask requests before provider work

Retain source proofs through answer assembly and rebuild the ordinary result after permission loss. Drain owned provider work before publishing the replacement answer.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* The analyzer corpus covers checked source reads

Drive metadata discovery, admitted chunk and document reads through the existing statement recorder, and regenerate group-membership erasure shapes.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Run checked source regressions under their required CI profiles

Keep signed HTTP cases in authenticated steps under both storage authorities, include the native Postgres cases, and close partial native acquisitions on every exit.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Preserve checked-read contracts across query and provider boundaries

Keep materialized views in place, analyze explicitly formatted SQL without appending another format, and retain the caller stage around checked providers. Align delegation, context-pack and auth-data controls with their actual paths.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep checked-read migrations after query-history storage

Preserve the merged query-history schema, erasure fences and generated references while assigning the unmerged reader-evidence migrations their own numbers.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep the generated checked-read schema and journal controls current

Regenerate the schema from an empty database in the composed migration order. Remove the RENAME regression with its unused parser now that the views change query in place.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* docs: connect the autonomous agent verification workflow (#802)

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* docs: clarify fake-provider extraction output (#804)

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Postgres summary downgrade test publishes its starting vectors (#805)

A new write answers at its commit. Wait for the initial vectors before deleting a source so the downgrade can copy them and settle its document within the delete.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Expose a private cell probe in the Helm chart (#806)

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* General search includes readable tracker issues and pull requests (#773)

* Draft metric definitions from query history and search gaps

Keep generated drafts private until an administrator publishes their content. Register opt-in read-only query-history sources and bound generation by privacy thresholds and namespace caps.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allow an isolated name for mixed-version checks

Use the requested container name for both creation and cleanup. A failed creation leaves any existing container untouched.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Pin query-history setting defaults

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Query-history settings distinguish cadence from deadlines

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Describe definition draft operations and review reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Document retained definition heads during draft recovery

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Show complete definition drafts before publication

Reviewers can inspect the full body and evidence before promoting a definition. Describe private draft destinations and authenticated approval attribution.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition support counts in memory lineage

Private definition drafts retain a supporting-account count without storing identities. Both exact lineage paths use that count while ordinary and assistant-rule candidates keep their existing accounting.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private definition decisions against erasure

Private queue writes retain their proposal time and share a bounded writer lease with namespace deletion. Check erasure under that lease and retain uncertain sends so a delayed decision or model completion cannot restore erased evidence.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private publication and source registry erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Expect retained writer fences in namespace erasure traces

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allocate query-history migration after mirrored erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Require private draft cleanup before older readers serve

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep gap input out of stored draft evidence

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate statement shapes on the composed schema

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition review tests within their type contracts

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Complete definition drafting contracts and Library composition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Search readable issues and pull requests with loom_search

Resolve authorized tracker targets before narrowing capture consent and searching the other planes. Keep duplicate verdicts for scoped defect reports and leave memory and knowledge ranking unchanged for general tracker matches.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Unified search descriptions include readable issues and pull requests

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Refresh persisted tracker links before discovery

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Describe tracker search results and tag scope

Empty duplicate verdicts remain useful evidence, while tag filters apply to memory and knowledge rather than tracker matches.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Curation rollout names the existing absence precondition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Format the unified search argument table

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep tracker search regression fixtures on the planned call path

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Use the configured test target for tracker ASGI requests

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Release preparation uses plugin version 1.2.28

Give the minimum-server warning a version distinct from other pending plugin changes.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Local development documents the fake LLM's test output (#807)

The deterministic fake client can synthesize facts. Match the development guide to that behavior.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Getting started distinguishes fake LLM output from meaningful results

### Changelog category (leave one):

- Documentation (changelog entry is not required)

### Changelog entry (one or two sentences a user of Loom can read without the diff; the release pull request collects these into the release notes):

Not required.

### Why

The fake LLM can synthesize facts and return deterministic answer text for development and tests. The quickstart and first-agent guide distinguish that output from meaningful extraction and grounded answers, which need a configured real LLM. The troubleshooting row explains that an empty extraction can also be valid when there is nothing memorable to retain.

### Scope and risk

The two guides describe existing provider behavior consistently in their introductions, provider table, session and ask explanations, and extraction troubleshooting. The quickstart table also receives formatter padding. Setup commands, conversation logging, expiry and permission guidance remain unchanged. The risk is confusing test output with meaningful results; both guides distinguish them.

### Compatibility, rollout and rollback

- Wire: none.
- Operator steps: none.
- Rollback: reverting restores the inaccurate fake-output descriptions; no stored data or generated files are involved.

### Tests

- Prettier 3.9.9 checks both changed Markdown files.
- The prose checker passes for both files and this description.
- A focused source comparison confirms the six related passages and table padding are the complete diff; headings, first sentences and links remain unchanged.
- `git diff --check` passes.
- Runtime tests were not run because this changes prose only. Full GitHub PR CI passes for this head.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated (an agent wrote most of the diff): Codex; checked against the fake extraction helper, client selection, SDK session flush and answer synthesis paths.

- [x] I can explain every line of this diff and how it interacts with the rest of Loom.

### Checklist

- [x] `make format-check`, `make lint`, `make typecheck` and `make test-unit` pass in GitHub CI; the integration suite also passes. Only the focused checks above ran locally.
- [ ] Generated files are regenerated in this PR, not by hand: the SDK op tables, `docs/reference/configuration.md` and the `docs/reference/` tables (`python scripts/codegen/docs_reference.py`), the reach manifest, `schemas/schema.sql` (`make schema-dump`) and the Codex plugin copies (`python3 scripts/codegen/agent_plugin_sync.py`). Not applicable: no generated source changed.
- [ ] A new migration is additive and numbered after the latest on `main`; no merged migration is edited. Not applicable.
- [ ] A behaviour change to any file a plugin ships (`hooks/`, `lib/`, `bin/`, `skills/`, `.mcp.json`, the Codex `plugin-defaults.json`) bumps both plugin manifests' version, and no other open PR uses that version. Not applicable.
- [ ] The frontend builds (`npm --prefix frontend run build`) and its tests pass, if `frontend/` changed. Not applicable.
- [x] Docs under `docs/` describe the system as it now is, and this description names any operator steps and wire changes.
- [x] Repository text has no dates, people, host names, internal namespaces, ticket or decision numbers; the history goes in this description (`tests/unit/test_public_copy.py`).
- [x] Commits are signed off (`git commit -s`), as CONTRIBUTING.md asks.

* Marked document requests keep accounting behind their privacy boundary (#803)

* Preserve marked request protections for dependency composition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Run marked event controls under both storage authorities

Keep private database cleanup tied to an explicit UUID even when creation loses its reply. Classify the fixed required-publication cleanup ceiling without changing its bound.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Check marked durable folds and scoped SQL reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep marked event migration after checked source reads

Retain the query-history migration and its dependent schema while assigning the unmerged marked-event cutover a distinct journal number.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate marked-reader query shapes after dependency composition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Retain selected source proofs through response publication

Renew the exact selected content after required accounting and awaited sends. Discard abandoned branches and unopened downloads, preserve derived metadata dependencies, and keep cancellation accounting. Probe the physical memory relation when its analysis policy is removed.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Type selected publication proofs explicitly

The strict type checker needs the element type before extending the retained proof list. Runtime behavior is unchanged.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep marked event projections within the privacy boundary

Retain permitted numeric work without copying document identifiers or error prose. Supply expired scope settings only while provisioning the signed analyst view, and keep existing regression controls aligned with their producers.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Format the event storage reference

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep source proofs through metadata and outline publication

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep response completion independent of expired publication proofs

Renew access before headers and every byte-bearing body frame. Empty completion frames must finish after a Content-Length client has consumed the response.

Give each integration test ownership of its offline client signer so in-process snapshots cannot stale the next test issuer. Preserve production key freshness and restore the exact prior snapshot on every exit.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Entity summary reads reject oversized seed sets without partial fallback (#808)

* Bound entity summary seeds and preserve typed refusals

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Optional read docstrings distinguish typed input refusals

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Wrap the composed search failure contract

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Document the fixed summary seed input bound

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Tracker integration restores the test client's keyring view

The in-process server lifespan loads a timestamped keyring snapshot. Restore the original client snapshot at fixture teardown so later tests do not inherit a stale server view. Production freshness checks remain unchanged.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Search cleanup preserves cancellation inside handled caller exceptions

Pass the search body exception through the owned async context rather than reading ambient exception state. A successful search awaited from an exception handler still propagates cancellation after joining its children.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* docs: show metadata-first code retrieval in the cost guide (#818)

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Memory-linked code guide shows an indexed change (#816)

* Memory-linked code guide shows an indexed change

Show how a saved lesson reads before and after a committed expression change and link the existing regression for the state transitions.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Change walkthrough uses trunk and checks capped lesson cards

The link worker maintains trunk commits, while symbol cards can omit a changed lesson. Name trunk and point to the saved memory links when the card leaves it out.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* test: pin audit privacy fixture target IDs (#821)

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Investigation guide uses unified search for document retrieval (#815)

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Governed-page tests wait for a served definition before mutation (#817)

An approval refresh can withhold a parked hit within its bounded wait.
Keep empty no-match pages in the existing preparation deadline so the
case reaches its edit or demotion only after proving the governed hit.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* OIDC tests allocate Loom's port after the identity provider binds (#812)

Selecting both ports before either listener starts can assign the same
unreserved port twice. Keep the identity provider listening while choosing
Loom's port so the allocator excludes it.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Documentation publishes and maintains five translation locales (#764)

### Changelog category (leave one):

- Improvement

### Changelog entry

The documentation site publishes available machine translations in Simplified Chinese, Japanese, Korean, Brazilian Portuguese and Spanish alongside canonical English. Translation maintenance discovers tracked documentation changes and prepares bounded, validated updates after English changes merge.

### Why

Generated translations need published routes and a maintenance path that follows additions, edits, renames and deletions. The site and updater share route mapping, preserve English fallbacks, and expose stale translations without treating pending coverage as current.

### Scope and risk

This change combines localized site rendering, the translation lifecycle workflow and 34 reviewed generated updates. Language links, canonical URLs, alternate-language metadata, sitemaps, Markdown copies and indexing submissions use the available pages. Translation batches preserve verified pending output and repair surviving links before removing obsolete generated pages.

The role-bearing runner executes trusted main code. Hosted validation checks the exact generated branch before merge. The cumulative reservation ledger and batch limits constrain provider calls; a missing allowance defers generation. No server operations, storage schema or tenant authorization changes are included.

### Compatibility, rollout and rollback

- Wire: none. English remains canonical and the default. All five translation locales are selected by default.
- Operator steps: site publication uses the existing Pages workflow. Automatic generation requires restricted runner labels, an initialized budget ledger and an explicitly approved cumulative allowance, as documented in `CONTRIBUTING.md`. No AWS credential is stored in the repository. The allowance defaults to zero.
- Coverage: the existing translated documentation pages are published. Missing pages and retained stale documentation translations remain pending; this change does not include a full translation backfill.
- Rollback: reverting restores the English site and earlier updater. No database migration or running-service change is required.

### Tests

- `make format`, `make format-check`, `make lint` and `make typecheck`: pass.
- Combined translation lifecycle and localized-site tests: 669 passed; the workflow branch-filter correction adds eight passing focused checks.
- README finalization regression: ten passed after reproducing three failures; eleven existing security and lifecycle checks also passed.
- Documentation links, public copy and agent instructions: 451 passed.
- Strict documentation build and rendered checks across all five locales: pass.
- Heavy local browser and full-suite runs are omitted; full suites run in CI.
- Full unit suite, offline tests and required integration checks: required in CI.
- No provider calls were made for integration or validation. The generated updates retain their original machine output and review evidence.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated: Codex integrated the reviewed components and checked their combined diff and validation results.

- [ ] I can explain every line of this diff and how it interacts with the rest of Loom.

### Checklist

- [ ] Local formatting, lint, types and focused tests pass; the full unit suite, offline tests and all other required checks pass in CI.
- [x] Generated files are regenerated rather than edited by hand; the translation bytes match their reviewed source outputs.
- [x] No migration, server wire field, plugin payload or frontend change is included.
- [x] Documentation describes the configured workflow and its remaining coverage and operator requirements.
- [x] Commits are signed off.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Postgres projection tests publish their starting rows (#790)

Related: https://github.com/ClickHouse/loom/pull/764

### Changelog category (leave one):

- CI Fix or Improvement

### Why

Postgres projection tests depend on initial memory rows and vectors being published before they replace or delete authority heads. A new-key `memory.set` can answer at its Postgres commit, before publication. If a later change overtakes the first publication, the tests can time out on an obsolete delivery or assert against rows that never reached ClickHouse.

The vector-retirement, history-retention, reconciliation, summary-retirement, live-claim redaction and cleared-payload tests wait for initial publication before changing those heads. The vector-retirement test also asserts that a live vector exists. The reconciliation, summary-retirement, live-claim redaction and cleared-payload tests retain their divergence counts and text-free vector tombstone assertions.

### Scope and risk

Six integration tests, including the four existing vector-retirement write variants and both live/parked lease cases. The existing publication helper establishes each fixture's initial state. Final publication waits, vector retirement, zero targets, history horizons and redaction assertions remain intact. No timeout, retry, skip, worker setting or production behavior changes.

### Compatibility, rollout and rollback

- Wire: none.
- Operator steps: none.
- Rollback: reverting removes the explicit initial publication preconditions.

### Tests

- `ruff check tests/integration/test_pg_authority.py`: pass.
- `ruff format --check tests/integration/test_pg_authority.py`: pass.
- Python AST parsing of the changed file and `git diff --check`: pass.
- Integration tests were not run locally. Full CI, including both Postgres authority jobs, must pass before merge.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated: Codex traced publication timing and added the publication preconditions and live-vector assertion.

- [ ] I can explain every line of this diff and how it interacts with the rest of Loom.

### Checklist

- [x] Focused formatting, lint and syntax checks pass.
- [ ] Full CI and both Postgres authority jobs pass.
- [x] All six tests retain their existing timeouts and final assertions.
- [x] No generated file, schema, migration, wire field, plugin or frontend change is included.
- [x] Commits are signed off.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Desktop browser launch survives installer terminal hangup (#822)

Ignore HUP before forking the asynchronous Linux opener and restore the parent handler afterwards. The opener cannot establish its own signal handling before it is scheduled. Exercise inherited handling with a self-HUP before the existing exact-once URL assertion.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Named users have daily cost caps and inbox warnings (#777)

* Draft metric definitions from query history and search gaps

Keep generated drafts private until an administrator publishes their content. Register opt-in read-only query-history sources and bound generation by privacy thresholds and namespace caps.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allow an isolated name for mixed-version checks

Use the requested container name for both creation and cleanup. A failed creation leaves any existing container untouched.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Pin query-history setting defaults

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Query-history settings distinguish cadence from deadlines

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Describe definition draft operations and review reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Document retained definition heads during draft recovery

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Show complete definition drafts before publication

Reviewers can inspect the full body and evidence before promoting a definition. Describe private draft destinations and authenticated approval attribution.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition support counts in memory lineage

Private definition drafts retain a supporting-account count without storing identities. Both exact lineage paths use that count while ordinary and assistant-rule candidates keep their existing accounting.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private definition decisions against erasure

Private queue writes retain their proposal time and share a bounded writer lease with namespace deletion. Check erasure under that lease and retain uncertain sends so a delayed decision or model completion cannot restore erased evidence.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private publication and source registry erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Expect retained writer fences in namespace erasure traces

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allocate query-history migration after mirrored erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Require private draft cleanup before older readers serve

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep gap input out of stored draft evidence

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate statement shapes on the composed schema

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition review tests within their type contracts

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Complete definition drafting contracts and Library composition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Search readable issues and pull requests with loom_search

Resolve authorized tracker targets before narrowing capture consent and searching the other planes. Keep duplicate verdicts for scoped defect reports and leave memory and knowledge ranking unchanged for general tracker matches.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Unified search descriptions include readable issues and pull requests

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Refresh persisted tracker links before discovery

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Describe tracker search results and tag scope

Empty duplicate verdicts remain useful evidence, while tag filters apply to memory and knowledge rather than tracker matches.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Add named-user cost caps and inbox budget warnings

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Include the user budget migration in the schema journal inventory

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* User budgets cover synthesis and scoped warning streams

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Name budget source entry points and invariants

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Define user-budget activation and rollback boundaries

Older replicas neither preserve user policy nor protect private warnings. Activate after the fleet upgrade and settle and remove warning events before a maintenance rollback, then restore saved policy on upgrade.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allocate user budgets after query-history registrations

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Reuse the shared writer fence for budget warnings

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Number the budget upgrade rule after mirror erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Curation rollout names the existing absence precondition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Format the unified search argument table

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep tracker search regression fixtures on the planned call path

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Use the configured test target for tracker ASGI requests

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate schema for user cost budgets

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Align user budget contracts with checked synthesis

Carry warning identity columns through task event reads and register the warning pipeline. Preserve named-user budget assertions while updating checked-search fixtures and their explicit keyword contracts, and regenerate the SDK surface and statement catalog.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate schema and statement catalog for user budgets

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Test marked namespace spend through the shared budget snapshot

The native marked-spend check uses the same rolling spend read as budget
status and admission, including when the namespace has no cost cap.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Bound leased sends by the received server clock

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Withhold exact user spend from budget status and warnings

Named-user cap verdicts and warning thresholds count all ledger spend.
Status reveals exact user spend only to a platform admin without namespace
role restrictions; user warnings always redact the amount for recipients.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Translation requests stay bounded and preserve technical literals (#797)

Related: https://github.com/ClickHouse/loom/pull/792
Related: https://github.com/ClickHouse/loom/pull/764

### Changelog category (leave one):

- Bug Fix (user-visible misbehavior in a release)

### Changelog entry

Machine translation bounds documentation requests, preserves source-attested names and technical values, and verifies the assembled page before writing it.

### Why

Splitting long pages only at second-level headings leaves some sections larger than one translation reply can hold. A truncated reply stops generation. The request planner now caps each source slice at 16 KiB while retaining paragraph, list-item, code, HTML and table-row boundaries.

The cost estimate uses the same slices that translation sends. The assembled page is checked again for content invariants and table layout before it is written. An indivisible oversized unit is refused before a provider call. Indented code remains whole, including tables that site extensions render differently; an oversized ambiguous block is refused.

Correct names and filenames can otherwise be rejected as untranslated English, while translated return-value strings can pass. Source context distinguishes protected spellings from ordinary descriptions. It checks preserved values in their original structural positions and keeps adjacent explanatory prose subject to translation checks.

### Scope and risk

Shared Markdown tokens move into the request-splitting module so the verifier and planner use one lexer. Maintained Markdown token ranges identify list items and table rows; the assembled table check parses both the site dialect and CommonMark tables. Code source ranges mask literal tags before HTML analysis. Site cell text is measured after serialization and raw HTML restoration. An indentation scan also covers the site extension’s two-space ordered-list continuations, which differ from CommonMark’s marker-width indentation. Their container rules differ. This avoids a partial table grammar that missed quoted tables and raw HTML contexts. Parser versions and runtime requirements are pinned; provider packages retain their versions. The architecture page describes the planning and validation boundaries. Source bytes, page-wide anchors and order survive reconstruction. The workflow installs its isolated, hash-locked runtime before checking freshness. Provider, model, output-token limit, retry, concurrency, budget, runner and permission settings retain their values.

Source-context checks use the same English page and source identity throughout retry, assembly and stored-page validation. Table context survives slices that begin after the header. Protection applies to attested name spans, repository file labels and declared technical values; capitalization, quotation marks and table position alone do not exempt English prose. The verifier rejects removed or changed protected spellings. A documented model default is tied to its source page, setting and default-value cell. Tracker contract type families are tied to their source page, ordered field group and Family cell, with tests against the implemented schema. Their literal spellings are preserved while ordinary headings, qualifiers and descriptions still require translation. Stored-page lifecycle validation uses the same source identity and both table parsers before staging, importing or activating output. The existing all-document inventory, README/subset safeguards, proof checks and HTML-tag counter remain in place. These checks do not certify semantic fidelity or current-source publication eligibility.

A source-byte bound reduces oversized requests; it does not guarantee that every model response fits. Existing truncated-reply refusal remains in place. A Markdown unit larger than the bound needs an English-source edit before translation can proceed.

### Compatibility, rollout and rollback

- Wire: none.
- Operator steps: none. This source change does not alter running generation jobs.
- Rollback: reverting restores the previous request planner and label checks. Generated output and budget history remain unchanged.

### Tests

- Nine regression cases fail before the change, covering oversized sections, multibyte input, protected Markdown units, provider admission, usage accounting and assembled table validation.
- Focused regressions cover two-space ordered-list continuations in plain text, quotes and admonitions, tab-indented lists, optional outer pipes, row loss, delimiters, single-column tables, nested containers, raw HTML and HTML elements spanning headings. Header and body cells cannot lose all text in CommonMark-only tables, including when empty inline markup remains in either dialect. Literal HTML in indented and nested fenced code stays protected. Translated cell text and valid pipe reflow are accepted in plain text, quotes, lists and admonitions. Existing section/assembly checks also pass.
- Focused HTML token controls cover the shared lexer. A parsed workflow comparison checks that runtime installation precedes the first CLI call; generation remains conditional on stale work and all other workflow settings retain their values.
- Three affected English inputs reconstruct exactly, with every source request at most 16 KiB. No provider was called for validation.
- Focused source-context cases cover exact names and file labels, ordinary residual descriptions, escaped pipes, literal-value changes, headerless table slices, retry feedback, assembled-page refusal and stored-page checks. Model-default regressions reject wrong settings, columns, paths, mutations and moves, and preserve full table context for headerless slices. Existing cached responses are examined without provider calls; remaining genuine differences stay reported.
- The related main composition passes 84 focused cases across lifecycle, README/subset/proof controls, stored metadata, bounded staging, tracker types and CI input guards. Stored-page regressions first reproduce eight source-context/table-layout failures; the corrected lifecycle rejects missing rows, extra cells and changed protected names. Existing cached tracker output passes the corrected type-family check without another provider call; the earlier untranslated phrase still rejects.
- Full suites must pass in CI on this final head. The tokenizer and all-document lifecycle changes are merged.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated: Codex traced the oversized requests, implemented Markdown slicing and source-context checks, and checked focused regressions.

- [ ] I can explain every line of this diff and how it interacts with the rest of Loom.

### Checklist

- [x] Focused formatting, lint and regression checks pass.
- [ ] Full CI passes on the final main-based head.
- [ ] Fresh technical and content reviews have no findings.
- [x] Provider settings, financial guards and active generation inputs are preserved.
- [x] Commits are signed off.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Documentation navigation keeps the reader's language (#794)

Related: https://github.com/ClickHouse/loom/pull/764

### Changelog category (leave one):

- Documentation

### Why

Readers can choose an available language from every documentation page. When that page has no translation, the language link names an existing translated landing page. Native READMEs link to their available documentation.

Translated pages keep available sidebar and footer destinations in the reader’s language. Missing translations carry an explicit English label. Language controls, search labels and translation-status notices use the page’s language. Videos use their translated fallback text for accessible labels.

### Scope and risk

The site projects navigation per page without changing shared navigation objects. Equivalent-page metadata remains separate from landing-page choices. The README generator maintains its existing owned regions and reuses the configured site base and landing priority. Quickstart and other available owned pages provide native website links when translated indexes are absent. Shared interface copy participates in site indexing and README maintenance inputs. Video labels are derived during rendering while preserving media URLs, controls and other attributes. Retained obsolete landing pages keep their README links until the selected locale retires them; retirement repairs that locale's generated navigation in the same batch.

The change covers English and the five configured translation languages. Translation coverage and search-tokenizer quality require separate validation.

### Compatibility, rollout and rollback

- Wire: none.
- Operator steps: none.
- Rollback: reverting restores the prior navigation and generated README regions.

### Tests

- Nine focused checks for navigation ownership, language choices, native stale notices and generated README destinations pass.
- The supplied strict Material fixture covers 13 pages across all six languages, with native search labels, active navigation, English fallbacks, Markdown notices, metadata and sitemap checks.
- Thirteen focused video-label and page-language checks pass, including English preservation, escaping and multiple-video cases.
- Nine focused lifecycle cases pass, including landing removal, preferred-landing fallback, subset ownership, selected-page progress, deferred README activation, stale/conflicted output and proof preservation. Two new cases reproduce the failure before the fix and verify the site still serves an unselected retained landing.
- Eight focused fallback and retirement cases pass, including quickstart-only and other-page landings, retained locale ownership, selected-page progress, the last native landing and unchanged site/indexing behavior. The two missing-link cases fail before the fallback change.
- Nine focused README and retirement checks pass. Quoted filenames and Unicode paths fail before URL encoding and produce complete, encoded links after it.
- Full suites run in CI and must pass on the public head before merge. Deployed routes remain a separate publication check.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated: Codex implemented navigation projection, native interface copy and accessible video labels. Independent technical and interface-copy review is required on the public head.

- [ ] I can explain every line of this diff and how it interacts with the rest of Loom.

### Checklist

- [x] Focused checks exercise the changed user paths.
- [ ] Full CI passes on the public head.
- [ ] Fresh technical and interface-copy reviews have no findings.
- [x] Generated README regions come from the existing generator.
- [x] Commits are signed off.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Session-content admission uses durable deployment protection (#819)

* Guard automatic session operations with a durable deployment latch

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Generate the deployment session latch schema

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Session admission validates complete Keeper storage metadata

ClickHouse engine_full includes each primary key. Match the actual storage clause while refusing wrong paths and keys, and classify automatic document-search fallbacks. Regenerate schema references and pin the new setting and error contracts.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Session admission preserves request context and attests positive writes

Scope query tags through proof cleanup and reject positive writes on a backend that was not validated. Reuse strict Boolean claim validation and isolate the MCP refusal phase so both entry points remain within their size limits.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Admission heartbeat and lease checks follow real callbacks

Assert repair precedes the holder read. Control only the lease test clocks and renewal wakes so healthy renewals cross the original expiry without assuming a prompt wall-clock schedule; keep the half-TTL cancellation policy unchanged.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate analyzer shapes for the composed latch schema

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

---------

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Retain completed translation units across English edits (#824)

Related: https://github.com/ClickHouse/loom/pull/797

### Changelog category (leave one):

- Improvement

### Changelog entry (one or two sentences a user of Loom can read without the diff; the release pull request collects these into the release notes):

Add a pure translation reconciliation API that retains authenticated completed requests across English edits and plans affected or unfinished ranges within the request bound.

### Why

Planning every request again from current English changes the cuts of completed work. A partial run also needs to distinguish completed replies from original units that have no completion. The reconciler takes a complete authenticated original-unit manifest, preserves whole completed payloads and exact separator ownership, and returns current scoped dispositions or explicit conflicts.

Headerless table gaps use the existing bounded planner with their complete source context. Changed-input corrections carry bounded ASCII feedback in the canonical SYSTEM, without replaying an old-input assistant response. Empty feedback preserves the existing SYSTEM and initial message.

### Scope and risk

This adds a pure planning API, contextual request splitting, initial-feedback construction and architecture documentation. The caller remains responsible for authenticating original plans, completion statuses, selected receipts and refusals. A metadata digest or missing cache file cannot establish that a completion is absent. The API does not resolve replies, admit provider calls or publish translations.

Repeated or unsafe mappings return conflicts. Replacing a completed whole-page unit and assigning unproven cuts among several unmatched original units remain unsupported. Every resolved or generated unit needs current-context validation, followed by whole-page invariants and semantic/completeness review. The main risk is incorrect source-range or separator ownership; focused regressions check exact reconstruction and preserved completed payloads.

### Compatibility, rollout and rollback

- Wire: no operation or field changes. The translation Python API gains reconciliation helpers, optional source context on `chunks.split`, and optional `initial_feedback` on `translate_text`.
- Operator steps: none. Consumers opt into the pure API; this PR adds no CLI, generation loop, dependency or execution controls. Default model, provider settings and spending code are unchanged.
- Rollback: reverting removes these APIs and their documentation. There are no migrations or stored-data changes to undo.

### Tests

The focused Python run passed 66 cases: reconciliation, canonical initial feedback and three existing default splitter regressions. It covers partial completion, mixed empty/LF separators, headerless table tails, source insertions/deletions, ambiguous or unsafe ranges, refusal, fault coalescing, request bounds and rejection before provider calls.

```sh
PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. python -m pytest --noconftest -o addopts='' -q \
  tests/unit/test_i18n_reconcile.py \
  tests/unit/test_i18n_initial_feedback.py \
  tests/unit/test_i18n_request_bounds.py::test_one_large_section_splits_at_complete_table_rows \
  tests/unit/test_i18n_request_bounds.py::test_a_short_multibyte_page_still_obeys_the_byte_bound \
  tests/unit/test_i18n_request_bounds.py::test_slices_keep_code_html_and_anchor_heading_pairs_together
```

Targeted Ruff and whitespace checks pass. The writing checker validates the architecture text and this description. Broad local suites were not run; full PR CI and independent review remain required before merge. No provider calls were made.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated: Codex wrote the implementation and tests, checked source-range and request-payload assertions, and ran the focused checks above. Maintainer review remains required.

- [ ] I can explain every line of this diff and how it interacts with the rest of Loom.

### Checklist

- [ ] `make format-check`, `make lint`, `make typecheck` and `make test-unit` pass; the integration suite passes if server behaviour changed.
- [ ] Generated files are regenerated in this PR, not by hand: the SDK op tables, `docs/reference/configuration.md` and the `docs/reference/` tables (`python scripts/codegen/docs_reference.py`), the reach manifest, `schemas/schema.sql` (`make schema-dump`) and the Codex plugin copies (`python3 scripts/codegen/agent_plugin_sync.py`). Not applicable: no generated surfaces changed.
- [ ] A new migration is additive and numbered after the latest on `main`; no merged migration is edited. Not applicable: no migrations changed.
- [ ] A behaviour change to any file a plugin ships (`hooks/`, `lib/`, `bin/`, `skills/`, `.mcp.json`, the Codex `plugin-defaults.json`) bumps both plugin manifests' version, and no other open PR uses that version. Not applicable: no plugin files changed.
- [ ] The frontend builds (`npm --prefix frontend run build`) and its tests pass, if `frontend/` changed. Not applicable: no frontend files changed.
- [x] Docs under `docs/` describe the system as it now is, and this description names any operator steps and wire changes.
- [x] Repository text has no dates, people, host names, internal namespaces, ticket or decision numbers; the history goes in this description (`tests/unit/test_public_copy.py`).
- [x] Commits are signed off (`git commit -s`), as CONTRIBUTING.md asks.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Docker smoke checks code-linked memories over MCP

### Changelog category (leave one):

- CI Fix or Improvement (changelog entry is not required)

### Changelog entry (one or two sentences a user of Loom can read without the diff; the release pull request collects these into the release notes):

Not required.

### Why

The Docker smoke verifies compiler indexes and restart persistence. Code-linked memory integration tests exercise handlers in process, so they do not check a packaged server returning a lesson's changed state over MCP after its source is reindexed. Extend the existing smoke with that round trip.

### Scope and risk

The existing Python fixture saves a lesson over MCP, waits for the link worker to mark it fresh, commits an expression change, and waits for the ingest and link workers to report the new commit and changed link. The lifecycle also waits for the scheduled Python compiler worker to publish that commit. Restart verification reads the changed link again, then deletes that smoke memory and checks its absence.

The installed environment template enables scheduled ingest. Before the first boot, the smoke explicitly sets `LOOM_CODE_INGEST_ENABLED=false` in its disposable environment so its initial seed and compiler checks finish before scheduled ingest starts. It then sets the same switch to `true` and recreates its existing server service. Scheduled compiler flags retain their existing installed values; this change does not disable them during the initial seed. Worker intervals, memory settling, product defaults, compiler checks and the 45-minute CI job timeout are unchanged. Each lifecycle invocation has one shared 1,020-second bound at the default worker intervals. Waiting for the normal Python compiler cadence can add about nine to ten minutes to CI; both architecture jobs must establish that the complete smoke fits its existing timeout. The status endpoint can return a cached repository head after the link has changed. The smoke polls until that endpoint also reports the exact edited commit, within the same lifecycle deadline. The other language fixtures retain their existing checks. A failed smoke still tears down its own Compose stack and volumes.

The installed smoke configuration uses single-user mode, a local embedder and the fake LLM. These assertions check transport, indexing, link states and persistence; they do not measure model quality or authenticated-client behavior.

### Compatibility, rollout and rollback

- Wire: none.
- Operator steps: none. The ingest switch is set only in the disposable smoke environment.
- Rollback: reverting removes the added smoke checks; no deployed data or schema changes are involved.

### Tests

- Ruff 0.16.9 lint and format checks pass for both changed scripts.
- The prose checker and `git diff --check` pass.
- A focused extracted environment-updater check verifies the explicit false-to-true phase configuration and source ordering. Fifteen earlier tiny extracted-helper checks cover MCP response errors, saved-value and namespace checks, fresh and changed states, exact commit checks and source compilation. They do not start a server or run a native smoke.
- Nine focused checks run the extracted status poller with inert MCP responses. The previous poller passes five and fails four; the corrected poller passes all nine. The checks cover transient cached heads, a permanently stale head reaching the existing deadline, and rejection of incorrect link commits, symbol paths and included lesson states.
- The previous native amd64 run returned the fresh link, then failed when the status endpoint returned an older cached head immediately after the scheduled Python compiler applied the edit. This corrects the smoke observation; it does not increase timeouts or change the status cache.
- No local server, database, Docker, browser or provider test ran. Full current-head PR CI passes, including both existing Docker Quickstart architecture jobs. The packaged checks pass the fresh link, committed source edit, exact changed commit, restart readback and exact-key deletion assertions.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated (an agent wrote most of the diff): Codex; checked against the existing smoke, worker configuration, MCP contract and code-link integration test.

- [x] I can explain every line of this diff and how it interacts with the rest of Loom.

### Checklist

- [x] `make format-check`, `make lint`, `make typecheck` and `make test-unit` pass in GitHub CI, including the integration and native image checks. Only the focused checks above ran locally.
- [ ] Generated files are regenerated in this PR, not by hand: the SDK op tables, `docs/reference/configuration.md` and the `docs/reference/` tables (`python scripts/codegen/docs_reference.py`), the reach manifest, `schemas/schema.sql` (`make schema-dump`) and the Codex plugin copies (`python3 scripts/codegen/agent_plugin_sync.py`). Not applicable: no generated source changed.
- [ ] A new migration is additive and numbered after the latest on `main`; no merged migration is edited. Not applicable.
- [ ] A behaviour change to any file a plugin ships (`hooks/`, `lib/`, `bin/`, `skills/`, `.mcp.json`, the Codex `plugin-defaults.json`) bumps both plugin manifests' version, and no other open PR uses that version. Not applicable.
- [ ] The frontend builds (`npm --prefix frontend run build`) and its tests pass, if `frontend/` changed. Not applicable.
- [x] Docs under `docs/` describe the system as it now is, and this description names any operator steps and wire changes. No documentation source changed.
- [x] Repository text has no dates, people, host names, internal namespaces, ticket or decision numbers; the history goes in this description (`tests/unit/test_public_copy.py`).
- [x] Commits are signed off (`git commit -s`), as CONTRIBUTING.md asks.

* Provider notifications request document source syncs (#784)

* Draft metric definitions from query history and search gaps

Keep generated drafts private until an administrator publishes their content. Register opt-in read-only query-history sources and bound generation by privacy thresholds and namespace caps.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allow an isolated name for mixed-version checks

Use the requested container name for both creation and cleanup. A failed creation leaves any existing container untouched.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Pin query-history setting defaults

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Query-history settings distinguish cadence from deadlines

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Describe definition draft operations and review reads

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Document retained definition heads during draft recovery

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Show complete definition drafts before publication

Reviewers can inspect the full body and evidence before promoting a definition. Describe private draft destinations and authenticated approval attribution.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition support counts in memory lineage

Private definition drafts retain a supporting-account count without storing identities. Both exact lineage paths use that count while ordinary and assistant-rule candidates keep their existing accounting.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private definition decisions against erasure

Private queue writes retain their proposal time and share a bounded writer lease with namespace deletion. Check erasure under that lease and retain uncertain sends so a delayed decision or model completion cannot restore erased evidence.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Fence private publication and source registry erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Expect retained writer fences in namespace erasure traces

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Allocate query-history migration after mirrored erasure

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Require private draft cleanup before older readers serve

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep gap input out of stored draft evidence

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Regenerate statement shapes on the composed schema

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Keep definition review tests within their type contracts

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Complete definition drafting contracts and Library composition

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Search readable issues and pull requests with loom_search

Resolve authorized tracker targets before narrowing capture consent and searching the other planes. Keep duplicate verdicts for scoped defect reports and leave memory and knowledge ranking unchanged for general tracker matches.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Unified search descriptions include readable issues and pull requests

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Refresh persisted tracker links before discovery

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>

* Describe tracker search results and tag scope

Empty duplicate verdicts remain useful evidence, while tag filters apply to memory and knowledge rather tha…
Keep the release version pins while incorporating the merged source, generated clients and documentation. Every other file retains main's contents.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep token-cache time separate from model-call policy freshness. Exercise token rotation under an explicit customer and controlled process clock so stale policy cannot mask the expected 401 behavior.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Bind the marked work bag to the serialization check so later event tests cannot inherit it.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
(cherry picked from commit c315568)
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep compiler and agent CLI dependencies reproducible across native and container installs. Use a patched locked workflow package manager and check release package signatures before importing code with install scripts disabled.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Disable package lifecycle scripts and require matching native packages from the committed graph before binding launchers. Retain baseline CPU variants and offline shim rebuilding, and test missing-native failures with network and process leak controls.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
### Changelog category (leave one):

- Not for changelog (changelog entry is not required)

### Why

Reconcile the existing release preparation with the merged source through the preparation branch's required pull-request workflow. The resulting tree retains main's source, generated clients and documentation together with the release version pins and two test-isolation repairs.

### Scope and risk

This forward merge includes main through #839. Relative to that main snapshot, 22 version paths set the release pins: the server, SDKs, frontend, root packages and skills use 1.3.0; the chart uses 0.3.0 with appVersion 1.3.0; the paired coding-agent plugins use 1.2.30 and require server 1.3.0. Two existing test files also isolate their clock and marked request context. The clock fixture replaces only the gateway module dependency, and the mirror-failure test restores its prior request-work context through the existing binding helper. Production freshness and required-tail refusals remain unchanged. Every other file and mode matches that main snapshot.

The existing preparation branch and its history are retained. The release pull request into main follows after this protected update lands. A release candidate, its soak and final publication remain separate steps.

### Compatibility, rollout and rollback

- Wire and operator changes inherited from the included features are listed in the release notes below. The version reconciliation adds no wire operation, schema migration, feature activation or default setting.
- Keep mirrored-source creation, query history, gap drafting and independent source push disabled unless their documented activation requirements are met. Native analyst provisioning and ordinary-event compatibility require their supported runtime proof.
- Final composition, generated notes, plugin versions and every required check must be reconciled again before the release pull request merges. Version collisions must be resolved against the versions actually merged on main.
- Rolling upgrades and rollback retain the data, accounting, erasure and protection requirements in the release notes. No tag or publication is performed by this update.

### Tests

- `prepare.py check 1.3.0`, `prepare.py check-chart 1.3.0` and `scripts.codegen.agent_plugin_sync --check` pass.
- The included version checkpoint passed 108 focused release-version and release-preparation checks and all eight Node plugin server-version checks. Full local suites were not run.
- The combined focused regression passes all 44 checks, including mirror serialization followed by consumer-event emission, the bearer-cache module and both model-policy denial controls. The two source repairs also retain their diagnosed failing reproductions.
- Fresh generated notes use this complete source and history. The generated wire appendix is retained verbatim in the curated notes.
- Full unit, integration, coverage, schema, SDK, frontend, plugin and security checks run in remote CI and must pass on this head and its tested merge composition.
- Internal independent whole-source and public-content review runs alongside CI.

### AI assistance in the code (leave one; writing this description with AI is fine):

- AI-generated: Codex reconciled the merge, checked source provenance, incorporated the published test-isolation repairs and prepared the version pins and release notes.

- [x] I can explain every authored change and its interaction with the included source.

### Checklist

- [x] Commits are signed off.
- [x] Version and chart checks pass; both plugins and their minimum server version agree.
- [x] Non-version content matches the included main snapshot except the two published, focused-tested isolation repairs.
- [ ] Fresh independent whole-source and complete public-body review reports no findings or nits.
- [ ] Every applicable remote CI check passes for the exact tested composition.
- [ ] The protected branch's actual merge requirements are satisfied.
- [ ] Concurrent plugin versions and current main are reconciled before the final release merge.

## Release notes for v1.3.0

### New features

- Framework integrations provide retrievers, memory stores, toolsets, conversation sessions and native storage through the Python and TypeScript SDKs. Python extras include `langchain`, `llamaindex`, `openai-agents`, `crewai`, `autogen`, `pydantic-ai`, `ag2`, `agent-framework`, `claude-memory`, `google-adk`, `haystack`, `pipecat` and `agno`. TypeScript integrations use optional peers and package subpaths. Install only the dependencies your application uses. (#607, #609, #650, #653, #655, #658, #661, #663, #666, #696, #729, #736, #737, #750, #765, #766, #779)
- Native file-tree memory tools create, view, edit, delete and rename documents shared across both SDKs. Python uses the `claude-memory` extra; TypeScript uses the matching integration subpath. A store groups files within a namespace and is not an authorization boundary: private user trees require separate namespaces. (#736)
- Native TypeScript memory storage persists threads, messages and resource working memory. Construct one `LoomMemoryStorage` per canonical user resource. Writes clear default expiry; retention, erasure, refusal and deletion policies still apply. `memory.set` accepts `if_match_sha256="absent"` for conditional first creation under a concrete identity. An older server refuses that condition rather than writing unconditionally. (#779)
- The Python and TypeScript `LoomSession` implementations persist conversation history. Processes with the same language, agent, user and session resume the same history. (#737)
- Separate workflow and plugin packages expose AI memory, remember, recall and knowledge-search tools for applications configured with a Loom endpoint and scoped token. Their publication and catalog acceptance are separate from the SDK release. (#754, #763)
- `LOOM_LLM_PROVIDER=local` runs model jobs against a chat-completions server you operate. `LOOM_ASSISTANT_PROVIDER=local` does the same for the assistant and requires `LOOM_ASSISTANT_MODEL`. The installer can discover a local model server; `LOOM_LOCAL_MODELS=ollama` selects that setup without a prompt. (#634)
- `knowledge.gaps.list` and the Library show questions that repeated searches could not answer. Recording is off by default through `LOOM_KNOWLEDGE_GAPS_ENABLED` and requires capture permission. (#598)
- `knowledge.source.parse_links` identifies pasted links before registration. Source registration accepts item lists and file-kind filters, while `knowledge.source.update` adds or removes source items without recreating the source. The CLI supports `--link`. (#562)
- `knowledge.list` accepts several `namespaces`. The Library's **All documents** view lists documents from every readable namespace and collection; **Everywhere** expands a collection search to every readable place. (#615, #686)
- `knowledge.move` relocates documents with their chunks, vectors, originals and tables. The Library offers **Move to…**, drag destinations and **Add to collection…**, with a confirmation explaining relocation and sharing. (#672, #734, #772)
- Uploading a folder preserves its nested structure as collections. The Library shows collection trees, document counts, folder chips and sharing information. Empty writable collections accept dropped files and folders. (#525, #613, #680, #771, #772)
- Ask AI opens beside the page on a wide screen, over it on a narrow one and as a sheet on a phone. Search can answer a question in place and continue the same conversation in chat. A deployment without an assistant URL retains Loom's own reads without chat controls. (#580, #596)
- Code answers include memories linked to their top files, with stale-code status and the commit at which the code changed. `include_lessons: false` or `code_result_lessons` disables these lessons. (#638, #659)
- `memory.definitions.list` reports governed definitions and lapse reasons. `memory.search` gives an exact term or alias a governed-definition leg and reports its outcome. (#679, #725)
- Operator routes support the first cell's attach sequence. A later cell registers only after the first serves; `loom cell remove` deletes an unattached cell with no placements. (#681)
- Additional extension manifests, experimental CLI hooks and team marketplace configuration make Loom available to more coding clients. See the integration documentation for each client's installation and tested status. (#594, #642, #694, #730)
- Admins can draft definitions from eligible search gaps or an opt-in query-history source, then review and publish them from the learnings queue. Query-history credentials are sealed; the model sees normalized SQL, which still contains identifiers. Distinct source accounts do not establish distinct people or anonymity. Drafts and supporting evidence remain private to namespace admins until publication. (#768)
- `loom_search` and `search.query` include readable tracker issues and pull requests for ordinary questions, with namespace citations. Scoped defect reports keep the duplicate-check verdict, including a successful empty result. (#773)
- The Helm chart exposes an optional private cell identity listener through a ClusterIP Service using `cellProbe.port`. The default zero keeps the listener disabled. (#806)
- Named users can have separate rolling-24-hour cost caps within a namespace, with threshold warnings sent to an existing namespace-owned task inbox. The user cap does not raise the namespace ceiling; concurrent calls can exceed these soft caps. (#777)
- Experimental source push setup accepts provider notifications as hints to request normal document-source syncs. Push is off by default; polling remains the recovery path. Notifications cannot write document content or widen access. (#784)

- `loom tracker-feed` maintains GitHub issues, pull requests and comments in a team-owned ClickHouse tracker record. Contract views apply observed repository visibility and hide content when the observation expires. Setup requires a dedicated record, separate writer and reader credentials, allowlisted repositories and an operator-scheduled feed. (#786)

### Improvements and fixes

- New installs verify a memory write/read round trip, offer to connect detected coding clients and open the dashboard where supported. `LOOM_CONNECT` controls client setup and `LOOM_NO_OPEN=1` suppresses the browser. Port conflicts stop installation before pulling or starting containers; pull errors point to proxy settings. (#574, #583)
- SDK memory adapters honor the calibrated abstention floor. Results above the full-value threshold use stored clips up to 1,200 characters, and `loom_recall` keeps valid JSON by omitting whole lower-ranked results when its size bound is reached. Explicit `defer=False` extracts inline. (#643, #647, #664)
- Framework context providers reject padded user ids; the corresponding optional dependency stays within its tested minor series. Exact memory deletion requires a server response containing the writer field; older responses return no deletion id. (#696, #746)
- On single-user installs, HTTP requests may omit their identity where supported. Writes use agent `http`, and searches read every agent; an explicit agent `"-"` keeps its meaning. Python `Agent()` and `SyncAgent()` support omitted identity. (#576, #595)
- Memory writes with a user compare against that user's rows and namespace pins. A restatement under a different key cannot merge into another user's row. (#701)
- Namespace erasure includes promoted shared rows. Redaction scrubs shared copies, and promotion or demotion invalidates cached answers and advances the write mark across replicas. (#572, #684, #697)
- Definition keys identify their own rows: conflict resolution cannot merge different keys. Shared definitions with a sensitive or deleted source stay out of search. (#656, #685)
- Linked-code memories whose definitions moved can follow those definitions across larger changes, within the documented bounds. Changed, unresolved or removed code lowers a memory by three search places by default rather than excluding it. (#612, #635)
- Document source sync limits new documents through `LOOM_KNOWLEDGE_SOURCE_MAX_DOCUMENTS` while still updating or retiring held documents. Cleanup has its own live per-pass bound, `LOOM_KNOWLEDGE_SOURCE_RETIRE_DOCUMENTS_PER_PASS`, default 40. Metadata edits survive sync until the source file changes. (#662, #683, #791)
- Namespaces matching `acl-` followed by 24 hexadecimal digits are reserved for permission-aware source mirrors and excluded from ordinary requests, org-wide reads and exports. Mirror creation remains disabled; its additive schema and erasure coverage are included. (#665, #670, #678, #778)
- Dashboard actions remain visible with their required role. Namespace and audit labels explain what they show. Source-list errors include a request id and Retry; run details copy complete ids. Standard control keys remain, while application-specific shortcuts and their hints are removed. (#551, #589, #602, #649, #698, #699)
- Coding-agent hooks launch Node directly, match temporary recall filenames literally and avoid running Python merely to locate pytest. Plugin 1.2.30 warns when its Loom server is older than 1.3.0. (#566, #573, #638, #733)
- MCP descriptions use callable tool names and explain how to reach unlisted operations through `loom_op`. The readonly toolset also restricts implicit `loom_op` routing. Context instructions request an initial task pack when hooks have not supplied it. (#570, #645)
- The CLI's commented-settings merge preserves unrelated settings and prints a manual entry when it cannot merge. Re-run the relevant client installation command to replace an obsolete client configuration. (#740)
- Indexing comment-heavy template strings and normalizing SDK URL/key-prefix slashes take linear time. Finished dependency-install tunnels release their slots after the configured half-close timeout. (#564, #565, #579)
- Under Postgres memory authority, `memory.update` resolves a committed key before its ClickHouse projection appears. (#781)
- The analyst probe checks its signed key canary before trusting a newer catalog version. The signed key canary remains mandatory. (#793)
- Translation checks distinguish literal comparisons from complete HTML markup. Shell launchers and skill files retain their committed LF bytes during checkout. (#767, #792)
- Checked document reads require current source permissions, reader identity and the exact content version before serving mirrored content. Public mirrored-source creation remains disabled until the fleet, request and session protections are complete. (#799)
- Checked document requests keep private diagnostics and derived activity behind their privacy boundary and finish required accounting before publishing content. Loss of the selected source/version proof refuses the complete response frame; after headers, the unchecked body is aborted. It does not rerender a replacement. Public mirrored-source creation remains disabled. (#803)
- Deployment session protection survives process switches and tenant erasure. Unverifiable protection state refuses session-content operations before their content is processed. (#819)
- Plugin 1.2.30 explains repository tags for code-related facts while keeping unrelated facts outside that repository filter. (#787)
- Available machine translations accompany canonical English in Simplified Chinese, Japanese, Korean, Brazilian Portuguese and Spanish. Documentation navigation retains the selected language. Translation maintenance preserves completed units across English edits, bounds refinement requests and checks technical literals before publishing reviewed content. (#764, #794, #797, #824, #829, #830, #833, #838)
- Guides explain persistent coding-agent memory, code-linked memories, metadata-first retrieval and the difference between fake-model test output and meaningful answers. The code-linked memory guide points to the packaged MCP checks. (#726, #727, #807, #815, #816, #818, #831)
- The desktop installer opens the dashboard when its terminal closes before the browser helper starts. (#822)

- Assistant answers mark mirrored-source access before output and keep mirror admission fixed for the answer's lifetime. Shared HTTP and MCP success/error responses preserve the privacy flag. Python and TypeScript assistant stream clients reject incomplete answers. Public mirrored-source creation remains disabled. (#835)
- Recall Lab reuses score rows while a draft query changes and keeps immediate clearing. (#786)

### Compatibility and deployment

- A Postgres-backed process reads cell state before serving. A process without `LOOM_CELL` refuses an enforced or partial cell; a failed initial Postgres read fails boot, except while migrations are pending, when the process remains unready. (#586)
- An enforced cell checks tenancy on writes, including the tenancy row itself. `LOOM_PG_DSN` must connect directly to Postgres. Older pods on an already prepared later cell must be rolled in one step once migrating boot installs the triggers. `loom doctor` detects missing, altered or disabled triggers and lost session markers. (#599, #676)
- CLI writes check the deployment and require the grants in upgrade step 15. Without `LOOM_CELL`, they refuse a cell past recording. Commands configured for a cell refuse another cell's stores; revocations that cannot restore access retain their documented outage behavior. New org names are bounded at 2048 UTF-8 bytes. (#668, #671, #673)
- `memory.promote` for a definition requires `approved_sha256` and promotes the namespace-wide row by the token's actor. Sensitive values, empty terms, conflicting terms or aliases and the shared-definition cap refuse promotion. (#667)
- Native HTTPS uses `LOOM_TLS_CERT_FILE` and `LOOM_TLS_KEY_FILE`, with certificate reload in each web process. The chart's `tls.secretName` configures the certificate mount and HTTPS probes. The image launcher refuses `UVICORN_*` variables; remove those variables before upgrading and use the documented Loom settings. (#756)
- Each metric retains up to `otel_metrics_cardinality_limit` series per process, default 20,000; additional attribute sets accumulate in an overflow series. (#648)
- Query-history and gap drafting introduce private definition candidates. Before downgrading to an older reader, use a maintenance window: disable both drafting features, quiesce drafting/review/registration/worker traffic, drain queued/model work, settle uncertain sends and retained leases, remove all versions of private candidate kinds and all registrations, and verify zero on every readable storage replica. Keep private producers disabled throughout the downgrade. Published definitions retain their existing lifecycle. (#768)
- `memory.search` and `memory.search_trace` accept at most 64 raw `entity_filter` entries, including repeated names. Canonical alias expansion is bounded at 256 names. Entity-summary retrieval separately bounds the case-insensitive distinct union of the canonical filter, query handles and known vocabulary tokens at 256 names. Overflow of either 256-name bound returns 422 `VECTOR_INPUT_TOO_LARGE` through retries, `memory.ask` widening and composed search, without a partial ranking. Narrow the query or filter before retrying. (#800, #808)
- `POST /v1/me/personal_tokens.create` issues the explicit personal-agent token kind only from the person's own primary sign-in session. The optional signed `reads_as_creator` claim defaults to false and ordinary delegation does not copy it. No operation-transport RPC is added by this route. (#799)
- Source membership and listing evidence must be present and current for checked reads. `LOOM_KNOWLEDGE_SOURCE_ACL_STALE_SECONDS` defaults to 3600; no backfill infers domain membership from an email address. Before rolling back after mirrored writes, stop those writers, settle outstanding work and follow the documented mirror erasure and telemetry sweep. Keep mirrored creation disabled on an incomplete fleet. (#799)
- Ordinary event inserts and signed, namespace-scoped analysis remain available across serving builds. `agent_events` retains ordinary columns and values; a permanent `ordinary_events_only` CHECK rejects marked events and reserved mirror namespaces. Marked requests and mirror workers write the separate `agent_events_private` store, which has no analyst grant or analyst-visible physical metadata. (#825)
- The signed `analysis_agent_events` catalog view unions ordinary and private events and masks protected fields before caller predicates and grouping. Ordinary `agent_events` reads retain their signed namespace scope; the analyst cannot read the private table. Trusted cost, budget and history-policy reads cover both stores, and user erasure scrubs both through `agent_events.scrub`. Private events use the ordinary ledger's retention and tiering policy. (#803, #825)
- Analyst administration uses `loom ch-setup --apply-analyst-sql` with explicitly configured existing access-management credentials. It applies only the analyst boundary under the live proof and lease, without rebuilding reader credentials. `--print-analyst-sql` refuses with a nonzero result and prints no SQL or secrets. General `--print-sql` remains available for database and reader preparation; print and apply modes are mutually exclusive. (#825)
- Session-content operations require exactly one `X-Loom-Session-Protection: v1` header once deployment protection is set. Missing, repeated or unknown declarations grant no promise. The gate includes dual-use `memory.set`, `memory.bulk_set`, `memory.search` and code/context operations used by automatic clients. Missing protection or unverifiable deployment state returns nonretryable 409 `SESSION_PROTECTION_REQUIRED`. HTTP checks before reading the body; MCP checks before content validation, identity defaults, replay or implicit session runs. (#819)
- The optional `LOOM_KNOWLEDGE_MIRRORED_SOURCES_ENABLED` setting now exists, defaults false and requires restart. Setting it true persists a deployment-wide latch before startup admits clients or workers; the heartbeat repairs that protection. Setting it false does not clear the latch. With the switch false, an undeclared client still needs a fresh service-principal Keeper absence proof; a declared client does not exercise that default path. This setting does not make public mirrored-source creation available. (#819)
- Named-user budget status withholds exact spend and its percentage except for a verified platform admin without namespace-role restrictions. Admission and routing still count all spend, including marked requests. User warnings always withhold exact spend, including webhooks and operator-triggered delivery. Namespace and org totals retain their existing numeric form. (#777)
- Before a downgrade after budget warnings were produced, stop request/background budget-check producers and status polling, drain scheduled deliveries, settle uncertain writes and writer leases, then remove only `budget.warning` events and verify completion and zero warnings on every storage replica. Save the global and namespace policies first: older namespace writes reset the new cap/inbox fields. Keep the feature disabled on older builds and reapply saved policy only after the full fleet supports it again. (#777)

- Tracker feeds support a single self-hosted server or shared storage with `ReplacingMergeTree` or `SharedReplacingMergeTree`; self-hosted replicated tables are refused. Feed connections retain the deployment's storage-authority, organization and cell checks. The dedicated writer must have verified limited direct and effective grants and remain separate from the serving reader. (#786)
- Assistant mirror admission uses a separate strict server-issued `answer_mirrors` Boolean. Missing admission is false, and mirror discovery also requires the caller's separate `session_protection` promise. An older verifier accepting the signed claim does not enforce it. Keep mirrored-source creation disabled until every serving build and consumer enforces the documented privacy protocol. No migration or backfill is added by this protocol change. (#835)

### Operator steps

1. Follow [the upgrade checklist](https://clickhouse.github.io/loom/operations/upgrades/) and apply the additive migrations through the normal migrating boot. This release's cell and shared-memory changes have additional steps below.
2. After every older replica is gone, follow upgrade step 14 to identify promoted rows left by earlier namespace erasures. Use the ledger's census cutoff and namespace when applying the documented tombstone or purge mutation, including hidden versions. Only a namespace-list erasure authorizes a purge. Erasures beyond the ledger's retention cannot be identified by that query. (#572)
3. Run `loom secrets reseal` with this build so document-source secrets and any existing source-push registry secrets are covered. No synthetic push setup is needed for an empty registry. (#562, #784)
4. Before migrating reserved mirror namespaces, resolve any migration refusal with the preceding build: erase the affected namespace contents, remove its code repository and delete the namespace, then retry. (#665)
5. Apply upgrade step 15's CLI grants, including ClickHouse `SELECT` on `loom.storage_authority`, the Postgres cell-state grants, connection limit and bucket settings. Complete step 16 before first-cell attach. (#668, #673)
6. Upgrade every router replica before attaching cells or upgrading cell processes. The cell processes read router state on each heartbeat; older routers do not have the required per-replica cache. Validate readiness and session tenancy markers after rollout. (#681, #693)
7. Keep the document-source limit at or below 50,000 while any older replica serves. After rollout, the live setting may be raised. (#662)
8. After all older replicas are gone, have an admin promote each existing shared definition again so its approved hash and token-stamped actor are recorded. The new live cap `memory_shared_definitions_max` defaults to 500. (#667)
9. For local assistant models, configure the model and endpoint, mount the assistant secret and add explicit network-policy egress to the model server's port. (#634)
10. Before enabling TLS, mount a certificate/key pair and change the upstream proxy to HTTPS in the same deployment window. Deployments using launcher-level `UVICORN_*` variables must remove them even when TLS remains off. (#756)
11. A desktop-launched coding client must find Node on its startup `PATH`; configure the client's environment when Node is installed only through a version manager. Upgrade Loom to 1.3.0 before relying on plugin 1.2.30's new server calls. (#573, #638)
12. Install an SDK extra or optional peer for each framework integration used. Separate plugin/community packages require their own installation and publication prerequisites; the release does not imply catalog approval. (#696, #750, #754, #763, #765, #766)
13. Before analyst provisioning, use a ClickHouse release that supplies `system.constraints`, as the pinned database release does. Missing CHECK metadata refuses the cutover. The operator must verify this on the actual database before the candidate starts. (#825)
14. Apply the additive query-history registration migration and upgrade every server and worker before registering a source or creating private candidates. `knowledge_query_history_enabled` defaults off; gap drafting also requires `knowledge_gaps_enabled`. Review the bounded read/pass/source/group/draft settings before enabling either feature. (#768)
15. For the private cell probe, configure Postgres and authentication, choose `cellProbe.port` and finish both web and worker rollouts before preparation. Use the Service origin for the probe URL. Configure this port only through the chart value, not `env`, `extraEnv`, `settingsFile` or `secrets.extra`. Set zero to disable it. (#806)
16. Apply the additive source-ACL schema before serving checked reads. Keep public mirrored creation disabled until the documented fleet, request, session and enablement protections are complete. (#799)
17. Apply migrations 103–106 in order for checked reads and marked accounting. Migration 106 pauses and waits for the three local refresh schedulers, changes their definitions and resumes them; complete the cutover on every refresh-owning replica. Existing ordinary aggregate types are preserved and prior input is not replayed. Keep mirrored creation disabled. (#799, #803)
18. Apply migration 107 before configuring named-user budgets. While any serving build is older or unknown, keep `LOOM_DEFAULT_DAILY_USER_COST_CAP_USD=0`, every namespace's `daily_user_cost_cap_usd=0` and `budget_alert_inbox=''`. The last value disables warning configuration; it does not require emptying an inbox and does not drain existing sends. The global user cap and `LOOM_COST_BUDGET_ALERT_SHARE` are live; the share defaults to 0.8. Do not rely on user caps during overlap. (#777)
19. Apply migration 108 and preserve its deployment protection state. Keep the mirrored-source switch and both private-drafting switches off on every server and worker through the mixed fleet. Verify ordinary undeclared client calls through the supported route, including the required Keeper proof. Missing metadata, backend identity changes or uncertain proof writes refuse admission; adding a promise header is not a repair. Only the documented complete mirror cleanup can clear protection, last; tenant erasure and a local switch change cannot. (#819)
20. Apply migration 109. Keep `LOOM_KNOWLEDGE_SOURCE_PUSH_ENABLED=false` unless opting into source push; this switch is separate from mirrored-source protection. Later setup needs a fleet supporting the registry's erasure, an explicit HTTPS `LOOM_PUBLIC_BASE_URL` origin and provider-specific confirmation. Source credential/item changes retire setup. Before rollback after use, disable push and complete the documented fenced registry cleanup, including historical versions, while preserving unrelated namespace data. (#784)
21. Documentation publication uses the existing site workflow. Automatic translation generation requires the documented restricted runner, initialized budget ledger and explicit cumulative allowance, which defaults to zero. This release does not configure that allowance or authorize generation. (#764)

22. Apply migration 110 after migration 109. It creates private event storage, adds the ordinary-table CHECK and updates union readers without moving, erasing or replaying existing rows or aggregate history. Complete its pause/wait/update/resume of both refresh definitions on every refresh-owning replica. Before grants, analyst provisioning verifies exact CHECK/table identities on every writer, freezes and drains the finite set of prepared INSERTs, then requires a zero contamination census. Existing contaminated rows remain untouched for explicit operator resolution. ALTER completion or client disconnection alone does not settle a prepared write. (#825)
23. `LOOM_CH_ANALYSIS_OBSERVER_URLS` is a restart-only list of additional trusted writer HTTP origins, default `[]`. The configured service origin is included and the existing service credentials are reused. Every writer UUID needs exactly one attested route. Complete native policy-catalog visibility, principal/role identity and non-replaying POST semantics are required; missing/duplicate writers, hidden metadata, redirects, HTTPS downgrade, identity drift or incomplete replies refuse provisioning. Intermediaries must preserve those semantics. Use the guarded analyst apply command for administrative setup. (#825)
24. Before an older build returns after mirrored enablement, settle outstanding writes and complete the documented mirror cleanup, including `agent_events_private`, marked receipts/audit figures and the cost rollup, before clearing deployment protection. Disabling a feature or downgrading schema alone does not establish privacy. Keep mirrored creation disabled until the separate fleet, client, restore and final erasure requirements are complete. (#803, #819, #825)

25. To opt into tracker feeds, create the dedicated record with separate writer/reader credentials, set the tracker organization, initialize the record and register its read-only contract views, then configure and schedule each allowlisted repository. Attest the supplied token's read-only permissions: a successful API read does not certify its grants. Renew installation tokens externally and configure each replacement before expiry. (#786)
26. Tracker visibility observations expire after 15 minutes and are not an instant permission mirror. Turns are bounded to 120 seconds, 100 items per page and 2–100 pages per stream. Incremental polling does not detect deletions. Stop every writer before erasing raw rows, comments and checkpoints or retiring a repository. Namespace deletion removes Loom's derived state and grants, not the external record or its backups. Rollback stops the scheduled writer; observation expiry hides serving rows without erasing the record. (#786)

### Release artifacts

Release assets include the installer, stack bundle, SDK packages, coding-agent plugin archives and agent setup contract. Each carries a `<file>.sigstore.json` bundle for `cosign verify-blob`; verify the certificate identity for this release's `release.yml` tag and the workflow OIDC issuer. Older assets without a bundle cannot pass this file-signature check. Both container images also carry signatures. (#689)

### Maintenance checks

Translation validation indexes protected ranges and table rows to avoid repeated scans on large reference pages. Request boundaries, labels and source offsets are preserved. (#834)

Original-unit reconciliation permits correction of a faulty completed whole-page translation when the prepared English input is unchanged. Existing request bounds and changed-input replacement restrictions remain. (#839)

Tracker reply parsing gains regression coverage, and CI action dependencies are updated. Additional tests cover budget bounds, memory-history diffs, audit privacy identifiers and port allocation. (#675, #705, #674, #700, #821, #812)

### What changed for clients since v1.2.0

373 operations served (was 361); ops +12 −0; request fields +22 −0; response fields +25 −0.

#### New operations
- `knowledge.gaps.draft` — role ADMIN, budget 30000 ms, MCP extended; request fields: gap_id, namespace, org
- `knowledge.gaps.list` — role ADMIN, budget 300 ms, MCP extended; request fields: limit, namespace, org
- `knowledge.move` — role WRITER, budget 8000 ms, MCP extended; request fields: namespace, org, source_uris, to_namespace
- `knowledge.query_history.add` — role ADMIN, budget 30000 ms, MCP extended; request fields: cluster, databases, host, name, namespace, org, password, port, secure, user
- `knowledge.query_history.list` — role ADMIN, budget 250 ms, MCP extended; request fields: fields, limit, namespace, org
- `knowledge.query_history.remove` — role ADMIN, budget 1000 ms, MCP extended; request fields: name, namespace, org
- `knowledge.source.parse_links` — role VIEWER, budget 50 ms, MCP extended; request fields: links, namespace, org
- `knowledge.source.push_configure` — role ADMIN, budget 60000 ms, MCP extended; request fields: agent, enabled, name, namespace, org, session_id, user_id
- `knowledge.source.push_confirm` — role ADMIN, budget 1000 ms, MCP extended; request fields: challenge_fingerprint, generation, name, namespace, org, signing_secret
- `knowledge.source.push_status` — role ADMIN, budget 1000 ms, MCP extended; request fields: name, namespace, org, reveal_challenge
- `knowledge.source.update` — role ADMIN, budget 60000 ms, MCP extended; request fields: add_items, add_links, include, name, namespace, org, remove_items, remove_links
- `memory.definitions.list` — role ADMIN, budget 400 ms, MCP extended; request fields: namespace, org

#### Changed operations
- `assistant.ask`: response +`mirrored`
- `code.callees`: request +`include_lessons`, `lesson_namespaces`; response +`lessons`, `lessons_left_out`
- `code.callers`: request +`include_lessons`, `lesson_namespaces`; response +`lessons`, `lessons_left_out`
- `code.context_pack`: request +`include_lessons`, `lesson_namespaces`; response +`lessons`, `lessons_left_out`
- `code.impact`: request +`include_lessons`, `lesson_namespaces`; response +`lessons`, `lessons_left_out`
- `code.outline`: request +`include_lessons`, `lesson_namespaces`; response +`lessons`, `lessons_left_out`
- `code.search`: request +`include_lessons`, `lesson_namespaces`; response +`lessons`, `lessons_left_out`
- `collection.list`: request +`include_document_counts`
- `cost.budget_status`: request +`user_id`; response +`namespace_over_cap`, `user_id`
- `knowledge.list`: request +`namespaces`
- `knowledge.source.add`: request +`include`, `items`, `links`; response +`include`, `items`; budget_p99_ms 30000 → 60000
- `knowledge.source.check`: response +`checked`, `sampled`
- `knowledge.source.inspect`: response +`include`, `items`
- `knowledge.synthesize`: request +`user_id`
- `memory.promote`: request +`approved_sha256`; response +`ignored_args`
- `memory.search`: response +`term_leg`
- `namespace.get`: response +`budget_alert_inbox`, `daily_user_cost_cap_usd`
- `namespace.register`: request +`budget_alert_inbox`, `daily_user_cost_cap_usd`


### Other client changes

- General search's optional `tracker` response section can carry `mode="search"`, `namespaces`, `namespace_limit` and cited `results`. Existing duplicate verdicts keep their shape. (#773)
- `memory.search` hits include `agent`, identifying the writer when row identity is available; otherwise it is empty. (#696)
- `collection.list` entries include nullable `document_count`. It is null unless `include_document_counts` is requested and counts documents directly in the collection. (#771)
- Learning entries include `source_resource_id`, defaulting to an empty string. `learnings.review` promotes private definition candidates only to `memory`. (#768)
- `cost.budget_status` user mode requires a namespace. `spent_24h_usd` is nullable, and `pct_of_cap` is null when user spend is withheld. `cap_source` adds `namespace_user` and `global_user`; nullable `namespace_over_cap` reports the independent namespace ceiling. Namespace get/list entries expose `daily_user_cost_cap_usd` and `budget_alert_inbox`. A positive namespace user cap overrides the global default; zero inherits, and zero at both tiers disables it. (#777)
- `knowledge.synthesize` accepts `user_id`, default `'-'`; the Python convenience method forwards its optional user id. Transport user pins take precedence. Named-user caps cover supported synthesis and query expansion; unattributed work remains namespace-only. (#777)
- Budget warnings arrive asynchronously on a later point check or status poll after the configured share is reached; there is no discovery scan. Repeated observations deduplicate within a UTC day for the same scope, inbox, cap and share. User warning payloads set `spent_24h_usd` to null for every recipient. Polling and watch streams filter warning user/session scope before pagination; ordinary task events retain inbox visibility. (#777)
- Assistant answer tokens carry their caller's strict Boolean `session_protection` claim. Descendant answers cannot override it with a transport header, and ordinary delegation does not inherit it. Clients promising protection must stop automatic session-content capture after a mirrored read and preserve that state across child/resumed sessions. (#819)
- Multi-entry MCP batches replace every response-bearing marked entry, including denied or empty entries, with `INVALID_REQUEST` and `Send this request individually`. Notifications remain silent. Single calls and one-entry arrays retain checked reads. (#803)
- Protected history makes enforcing policy rules evaluate potentially matching marked participation as unknown before success/argument filters, unresolved references or negation; it can refuse reads as well as writes. `policy.check` reports `protected history unavailable` with null protected leaf values. Shadow policies remain non-enforcing. (#803)
- Marked cost is disclosed only to a verified unrestricted platform admin or authorized namespace/whole-day totals without fine-grained filters. Fine-grained marked cost is withheld; marked-only latency has no ordinary samples and returns zero latency values. This does not disable internal budget checks. (#803)
- Source-push setup operations are experimental, namespace-admin only and non-idempotent, including status. `push_status` reveals a pending verification token only with `reveal_challenge=true`; ordinary inspection, audit and idempotency replay omit it. The response's `enabled` reports the global switch, while `status` reports source setup state. (#784)
- Provider notifications use the separate non-RPC `POST /hooks/knowledge/{namespace}/{source_id}/{generation}?org=<encoded-org>` route with provider proof, rather than an RPC bearer token. Bodies are bounded at 65,536 bytes. A 200 acknowledgement can represent an early synchronization notification and does not prove that work was requested or content ingested. Drive uses channel-token verification rather than signed delivery; other supported providers require their documented HMAC and identity checks. (#784)

- `assistant.ask` adds optional response field `mirrored`, default false. Marked chat output begins with `mirrored: true`; marked SSE begins with `data: {"mirrored":true}` before answer text. HTTP/MCP success and error output retains the mark after serialization and protocol unwrapping; MCP structured and JSON text results preserve it, and a marked search summary follows a first JSON marker block. (#835)
- The server fixes `answer_mirrors` before spend or model work from active mirrored registrations in the authorized namespace. Active registrations keep the answer marked even when ACL checks exclude every private row. Caller fields, headers, token-mint endpoints and CLI flags cannot supply admission. An initially ordinary answer cannot gain a mirror leg after source activation. Ordinary delegated children omit both `answer_mirrors` and `session_protection`; descendants cannot widen either claim with a header. Current document authorization still applies. (#835)
- Incremental assistant text is provisional. A marked response retains required accounting after headers, joins producer cleanup through repeated cancellation and confirms its required private cost row before terminal completion. Failed required capture or delivery cannot complete successfully. Whole responses and checked source frames retain accounting before publication. (#835)
- Python and TypeScript assistant streams require `[DONE]` before returning a completed answer. Incomplete EOF fails; text after `[DONE]` is ignored while later transport errors still propagate. Stream completion does not qualify durable automatic capture or privacy state across child/resumed sessions; clients promising protection must satisfy the separate session contract. (#835)

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Retain the protected preparation history and the source branch ancestry
for the final release without changing their shared tree.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep current main source and dependencies while aligning release versions.
Both coding-agent plugins use 1.2.36 and require server 1.3.0.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Preserve the merged coordination transport and its tests without changing
the release version pins.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Preserve the merged documentation with the unified release versions and
the coding-agent plugins at version 1.2.36.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep the merged SDK and documentation changes with the unified release
versions and the request-witness fixture correction.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Preserve the preparation ancestry and the current main source while retaining the release versions and plugin minimum.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
…ing-state

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
…rows

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
…ntity

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
@zlareb1
zlareb1 marked this pull request as ready for review November 18, 2026 22:07
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Validate published dependency metadata against the tag and install each Python extra and npm peer in a clean hosted environment.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Run synthetic metadata checks through the original test environment so packaging is available. Describe the canonical parent status and time predicate in the chunk regression module.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
….3.0

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Include the published team-memory walkthrough in the release composition and keep the qualified release changes intact.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep the release composition aligned with the published documentation without changing its qualified code.

Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
@zlareb1
zlareb1 merged commit 08eeb3d into main Nov 19, 2026
63 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant