From cc6bfbf42dbcf46456ce04c4c3ff45b2a4ab8d73 Mon Sep 17 00:00:00 2001 From: Sutu Sebastian Date: Mon, 16 Nov 2026 14:42:57 +0200 Subject: [PATCH 1/2] changeset: consolidate the 0.7.0 release notes MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The 130 pending changesets describe internals as often as they describe what a 0.6.1 user notices, so shipping them verbatim makes the 0.7.0 release body a changelog instead of notes. Fold them into 14 themed entries — the mechanism 0.4.0 used (b21eaff "changeset: consolidate the 0.4.0 release notes") — so `changeset version` writes one curated "what's new" section plus its themed items, and CHANGELOG.md, the GitHub Release body, and the docs site's /changelog stay one text. Every item traces to the changeset that produced it: each item line ends in a hidden `` comment naming its sources, and this message carries the same theme -> sources mapping. Four items cover sibling changesets that describe one behaviour (the author fingerprint, the revert memory, the published review.json key names, and corpus grading), and a change that both adds a feature and breaks a shape appears twice — once in its theme, once under breaking — so no source is lost and no breaking step goes unnamed. Start point: .scratch/factory/roadmap-next/RELEASE-NOTES-draft.md, written against 104 of these changesets. Refreshed against every pending changeset on main 7d488f975: 24 arrived after the draft, and 2 more (backtest-reverts, run-print-stage) predate it but were missing from it. Headline (00-what-s-new.md, 5 sentences): Verdict 0.7.0 is a breadth release. It adds a local code index, a read-only `serve` surface with a browser UI, and a backtest harness that mines past bug fixes from your own history, replays them against your agent, and scores the result. Review output gains an auditable provenance receipt, diff-coverage attestation, advisory risk scoring, and an optional validation pass that floors each finding on the evidence behind it, while the CI gate now checks coverage by default. Posting grows a secret guard, honest per-host capability reporting, and a summary-comment option, GitLab and Forgejo join GitHub as full publish targets, and lens packs, declared recipes, and reported verification make the review pipeline itself data. Thirteen pre-1.0 changes alter behaviour — each names the migration step below — and the known limits close the notes. Theme -> sources (minor unless noted, all under .changeset/): 00-what-s-new.md <- the headline, no items 01-new-commands-and-flags.md <- backtest-mine backtest-replay backtest-reverts backtest-score bounded-foreground-waits fix-all-batch handoff-verb incident-memory ingest-sarif panel-verification post-summary-comment reaction-scoped-fix recipes-format report-html risk-queue run-agent run-print-stage skill-management thread-resolve 02-local-code-index.md <- blast-radius-index-hint fetch-index-impact index-blast-radius index-callers index-corpus-bugs index-graph index-impact-budget index-scanners index-transitive-tests 03-review-quality-and-provenance.md <- author-routing author-routing-hint author-routing-record author-routing-skill author-routing-trend context-sources coverage-scoring-basis diff-coverage-attestation fetch-reverts ignore-overrides ignore-suppression learnings-propose null-ranges-malformed post-content-digest post-content-recovery post-live-authority post-verdict-reconciliation prompt-key-shapes prompt-rigor-key reactions-trend resolution-rate rigor-validation risk-score rule-exposure rule-pack-import withheld-capabilities 04-safety-and-trust.md <- curl-token-argv learnings-evidence line-break-guards mention-neutralization post-secret-guard punctuated-mentions run-argument-credential setup-verdict-hardening skill-install-preflight store-lock-and-listing strikethrough-mentions verify-indirection verify-out-owner-only verify-sandbox verify-sandbox-macos 05-agents-and-mcp.md <- agent-definition-display mcp-config-hosts mcp-mutation-class run-repair-loop self-review-loop 06-hosts-and-ci.md <- doctor-provider-adopt-gitlab forgejo-post github-action gitlab-incremental gitlab-publish ready-calibrate-gitlab 07-serve-and-ui.md <- cache-content-budget cache-small-content inventory-budget-exact serve-live-updates serve-s1 serve-s2-discovery serve-s3-cost-precision serve-ui vanished-threads 08-lens-packs.md <- lens-dispatch-deltas lens-packs-s1 pack-add-backup-warning pack-layers pack-staging-stale pack-test packs-lock 09-measurement.md <- bench-adjudication bench-corpus-followups bench-keyword-guard bench-real-pilot benchmark-clean-controls benchmark-controls-metrics 10-breaking-changes.md <- bench-real-pilot bundle-name-guard gate-coverage-default gate-live-all-hosts gate-tier-validation honest-capabilities index-import-shadow managed-parent-symlinks pack-source-allowlist provenance-receipt provider-from-remote typed-truncation verify-env-allowlist verify-indirection verify-out-owner-only verify-static-gate-bypasses write-bundle-symlinks 11-fixes.md <- backtest-partial-clone fetch-refusal-next-steps next-usage-basis-plain pr-wait-inconsistent-reads read-verb-re-enter record-cost-name-warnings release-asset-sigpipe review-residuals upgrade-hints-quoting windows-watchdog-no-hang windows-watchdog-resume-retry 12-docs.md <- defaults-at-a-glance docs-site-restructure troubleshooting 13-known-limits.md <- verify-indirection verify-sandbox-macos verify-static-gate-bypasses 04-safety-and-trust.md, 11-fixes.md, 12-docs.md, and 13-known-limits.md ship as patch; the rest as minor. One minor changeset is enough to cut 0.7.0. Breaking (13 items in 10-breaking-changes.md), each with its _Migrate:_ step: the `index blast-radius` truncation receipt, `gate` checking coverage by default, `gate --live` reading every host, unknown gate tiers, the default-deny verify child environment, the verify command shapes that used to run (`verify-networked` and the new `verify-indirection` arm), the owner-only `.verify.out`, guarded bundle paths, provider inference from origin, the `capabilities` split, the `pack add` source allowlist, the benchmark `provenance.base_sha`/`head_sha` requirement, and the rebuilt-code-index requirement. Known limits (13-known-limits.md) are the three from STA-436's 2026-11-16 status comment: the macOS sandbox covers the network half only, the string gate is a tripwire rather than a boundary (`sh "$V"` still passes it), and a backslash before a CRLF line ending is not joined as a continuation. Counts: 130 changesets read, 14 consolidated entries, 126 item lines (123 content items plus the 3 known-limit lines), 13 breaking, 5-sentence headline. Verification: bun fmt --check .changeset/*.md -> All 14 files are formatted. go test ./internal/archtest/ -count=1 -> ok (changeset frontmatter guard) bash evals/one.sh doc-coverage -> PASS (135 cases) --- .changeset/00-what-s-new.md | 15 ++++++ .changeset/01-new-commands-and-flags.md | 25 ++++++++++ .changeset/02-local-code-index.md | 15 ++++++ .../03-review-quality-and-provenance.md | 27 +++++++++++ .changeset/04-safety-and-trust.md | 21 ++++++++ .changeset/05-agents-and-mcp.md | 11 +++++ .changeset/06-hosts-and-ci.md | 12 +++++ .changeset/07-serve-and-ui.md | 15 ++++++ .changeset/08-lens-packs.md | 13 +++++ .changeset/09-measurement.md | 11 +++++ .changeset/10-breaking-changes.md | 21 ++++++++ .changeset/11-fixes.md | 17 +++++++ .changeset/12-docs.md | 9 ++++ .changeset/13-known-limits.md | 9 ++++ .changeset/agent-definition-display.md | 5 -- .changeset/author-routing-hint.md | 7 --- .changeset/author-routing-record.md | 7 --- .changeset/author-routing-skill.md | 7 --- .changeset/author-routing-trend.md | 8 ---- .changeset/author-routing.md | 9 ---- .changeset/backtest-mine.md | 16 ------- .changeset/backtest-partial-clone.md | 7 --- .changeset/backtest-replay.md | 24 ---------- .changeset/backtest-reverts.md | 38 --------------- .changeset/backtest-score.md | 24 ---------- .changeset/bench-adjudication.md | 18 ------- .changeset/bench-corpus-followups.md | 8 ---- .changeset/bench-keyword-guard.md | 12 ----- .changeset/bench-real-pilot.md | 17 ------- .changeset/benchmark-clean-controls.md | 7 --- .changeset/benchmark-controls-metrics.md | 15 ------ .changeset/blast-radius-index-hint.md | 14 ------ .changeset/bounded-foreground-waits.md | 9 ---- .changeset/bundle-name-guard.md | 27 ----------- .changeset/cache-content-budget.md | 5 -- .changeset/cache-small-content.md | 5 -- .changeset/context-sources.md | 5 -- .changeset/coverage-scoring-basis.md | 14 ------ .changeset/curl-token-argv.md | 11 ----- .changeset/defaults-at-a-glance.md | 8 ---- .changeset/diff-coverage-attestation.md | 12 ----- .changeset/docs-site-restructure.md | 5 -- .changeset/doctor-provider-adopt-gitlab.md | 18 ------- .changeset/fetch-index-impact.md | 28 ----------- .changeset/fetch-refusal-next-steps.md | 14 ------ .changeset/fetch-reverts.md | 33 ------------- .changeset/fix-all-batch.md | 5 -- .changeset/forgejo-post.md | 15 ------ .changeset/gate-coverage-default.md | 30 ------------ .changeset/gate-live-all-hosts.md | 16 ------- .changeset/gate-tier-validation.md | 8 ---- .changeset/github-action.md | 7 --- .changeset/gitlab-incremental.md | 16 ------- .changeset/gitlab-publish.md | 9 ---- .changeset/handoff-verb.md | 27 ----------- .changeset/honest-capabilities.md | 21 -------- .changeset/ignore-overrides.md | 5 -- .changeset/ignore-suppression.md | 5 -- .changeset/incident-memory.md | 38 --------------- .changeset/index-blast-radius.md | 19 -------- .changeset/index-callers.md | 21 -------- .changeset/index-corpus-bugs.md | 16 ------- .changeset/index-graph.md | 19 -------- .changeset/index-impact-budget.md | 5 -- .changeset/index-import-shadow.md | 5 -- .changeset/index-scanners.md | 35 -------------- .changeset/index-transitive-tests.md | 20 -------- .changeset/ingest-sarif.md | 16 ------- .changeset/inventory-budget-exact.md | 5 -- .changeset/learnings-evidence.md | 7 --- .changeset/learnings-propose.md | 8 ---- .changeset/lens-dispatch-deltas.md | 21 -------- .changeset/lens-packs-s1.md | 17 ------- .changeset/line-break-guards.md | 11 ----- .changeset/managed-parent-symlinks.md | 5 -- .changeset/mcp-config-hosts.md | 18 ------- .changeset/mcp-mutation-class.md | 7 --- .changeset/mention-neutralization.md | 23 --------- .changeset/next-usage-basis-plain.md | 8 ---- .changeset/null-ranges-malformed.md | 13 ----- .changeset/pack-add-backup-warning.md | 7 --- .changeset/pack-layers.md | 27 ----------- .changeset/pack-source-allowlist.md | 20 -------- .changeset/pack-staging-stale.md | 8 ---- .changeset/pack-test.md | 5 -- .changeset/packs-lock.md | 35 -------------- .changeset/panel-verification.md | 9 ---- .changeset/post-content-digest.md | 5 -- .changeset/post-content-recovery.md | 5 -- .changeset/post-live-authority.md | 5 -- .changeset/post-secret-guard.md | 15 ------ .changeset/post-summary-comment.md | 16 ------- .changeset/post-verdict-reconciliation.md | 5 -- .changeset/pr-wait-inconsistent-reads.md | 10 ---- .changeset/prompt-key-shapes.md | 13 ----- .changeset/prompt-rigor-key.md | 17 ------- .changeset/provenance-receipt.md | 5 -- .changeset/provider-from-remote.md | 5 -- .changeset/punctuated-mentions.md | 13 ----- .changeset/reaction-scoped-fix.md | 32 ------------- .changeset/reactions-trend.md | 5 -- .changeset/read-verb-re-enter.md | 10 ---- .changeset/ready-calibrate-gitlab.md | 15 ------ .changeset/recipes-format.md | 19 -------- .changeset/record-cost-name-warnings.md | 10 ---- .changeset/release-asset-sigpipe.md | 13 ----- .changeset/report-html.md | 7 --- .changeset/resolution-rate.md | 6 --- .changeset/review-residuals.md | 16 ------- .changeset/rigor-validation.md | 33 ------------- .changeset/risk-queue.md | 21 -------- .changeset/risk-score.md | 25 ---------- .changeset/rule-exposure.md | 5 -- .changeset/rule-pack-import.md | 5 -- .changeset/run-agent.md | 27 ----------- .changeset/run-argument-credential.md | 16 ------- .changeset/run-print-stage.md | 16 ------- .changeset/run-repair-loop.md | 5 -- .changeset/self-review-loop.md | 5 -- .changeset/serve-live-updates.md | 5 -- .changeset/serve-s1.md | 9 ---- .changeset/serve-s2-discovery.md | 7 --- .changeset/serve-s3-cost-precision.md | 8 ---- .changeset/serve-ui.md | 7 --- .changeset/setup-verdict-hardening.md | 8 ---- .changeset/skill-install-preflight.md | 14 ------ .changeset/skill-management.md | 29 ----------- .changeset/store-lock-and-listing.md | 19 -------- .changeset/strikethrough-mentions.md | 14 ------ .changeset/thread-resolve.md | 17 ------- .changeset/troubleshooting.md | 8 ---- .changeset/typed-truncation.md | 11 ----- .changeset/upgrade-hints-quoting.md | 9 ---- .changeset/vanished-threads.md | 5 -- .changeset/verify-env-allowlist.md | 20 -------- .changeset/verify-indirection.md | 48 ------------------- .changeset/verify-out-owner-only.md | 10 ---- .changeset/verify-sandbox-macos.md | 17 ------- .changeset/verify-sandbox.md | 21 -------- .changeset/verify-static-gate-bypasses.md | 18 ------- .changeset/windows-watchdog-no-hang.md | 8 ---- .changeset/windows-watchdog-resume-retry.md | 8 ---- .changeset/withheld-capabilities.md | 18 ------- .changeset/write-bundle-symlinks.md | 5 -- 144 files changed, 221 insertions(+), 1797 deletions(-) create mode 100644 .changeset/00-what-s-new.md create mode 100644 .changeset/01-new-commands-and-flags.md create mode 100644 .changeset/02-local-code-index.md create mode 100644 .changeset/03-review-quality-and-provenance.md create mode 100644 .changeset/04-safety-and-trust.md create mode 100644 .changeset/05-agents-and-mcp.md create mode 100644 .changeset/06-hosts-and-ci.md create mode 100644 .changeset/07-serve-and-ui.md create mode 100644 .changeset/08-lens-packs.md create mode 100644 .changeset/09-measurement.md create mode 100644 .changeset/10-breaking-changes.md create mode 100644 .changeset/11-fixes.md create mode 100644 .changeset/12-docs.md create mode 100644 .changeset/13-known-limits.md delete mode 100644 .changeset/agent-definition-display.md delete mode 100644 .changeset/author-routing-hint.md delete mode 100644 .changeset/author-routing-record.md delete mode 100644 .changeset/author-routing-skill.md delete mode 100644 .changeset/author-routing-trend.md delete mode 100644 .changeset/author-routing.md delete mode 100644 .changeset/backtest-mine.md delete mode 100644 .changeset/backtest-partial-clone.md delete mode 100644 .changeset/backtest-replay.md delete mode 100644 .changeset/backtest-reverts.md delete mode 100644 .changeset/backtest-score.md delete mode 100644 .changeset/bench-adjudication.md delete mode 100644 .changeset/bench-corpus-followups.md delete mode 100644 .changeset/bench-keyword-guard.md delete mode 100644 .changeset/bench-real-pilot.md delete mode 100644 .changeset/benchmark-clean-controls.md delete mode 100644 .changeset/benchmark-controls-metrics.md delete mode 100644 .changeset/blast-radius-index-hint.md delete mode 100644 .changeset/bounded-foreground-waits.md delete mode 100644 .changeset/bundle-name-guard.md delete mode 100644 .changeset/cache-content-budget.md delete mode 100644 .changeset/cache-small-content.md delete mode 100644 .changeset/context-sources.md delete mode 100644 .changeset/coverage-scoring-basis.md delete mode 100644 .changeset/curl-token-argv.md delete mode 100644 .changeset/defaults-at-a-glance.md delete mode 100644 .changeset/diff-coverage-attestation.md delete mode 100644 .changeset/docs-site-restructure.md delete mode 100644 .changeset/doctor-provider-adopt-gitlab.md delete mode 100644 .changeset/fetch-index-impact.md delete mode 100644 .changeset/fetch-refusal-next-steps.md delete mode 100644 .changeset/fetch-reverts.md delete mode 100644 .changeset/fix-all-batch.md delete mode 100644 .changeset/forgejo-post.md delete mode 100644 .changeset/gate-coverage-default.md delete mode 100644 .changeset/gate-live-all-hosts.md delete mode 100644 .changeset/gate-tier-validation.md delete mode 100644 .changeset/github-action.md delete mode 100644 .changeset/gitlab-incremental.md delete mode 100644 .changeset/gitlab-publish.md delete mode 100644 .changeset/handoff-verb.md delete mode 100644 .changeset/honest-capabilities.md delete mode 100644 .changeset/ignore-overrides.md delete mode 100644 .changeset/ignore-suppression.md delete mode 100644 .changeset/incident-memory.md delete mode 100644 .changeset/index-blast-radius.md delete mode 100644 .changeset/index-callers.md delete mode 100644 .changeset/index-corpus-bugs.md delete mode 100644 .changeset/index-graph.md delete mode 100644 .changeset/index-impact-budget.md delete mode 100644 .changeset/index-import-shadow.md delete mode 100644 .changeset/index-scanners.md delete mode 100644 .changeset/index-transitive-tests.md delete mode 100644 .changeset/ingest-sarif.md delete mode 100644 .changeset/inventory-budget-exact.md delete mode 100644 .changeset/learnings-evidence.md delete mode 100644 .changeset/learnings-propose.md delete mode 100644 .changeset/lens-dispatch-deltas.md delete mode 100644 .changeset/lens-packs-s1.md delete mode 100644 .changeset/line-break-guards.md delete mode 100644 .changeset/managed-parent-symlinks.md delete mode 100644 .changeset/mcp-config-hosts.md delete mode 100644 .changeset/mcp-mutation-class.md delete mode 100644 .changeset/mention-neutralization.md delete mode 100644 .changeset/next-usage-basis-plain.md delete mode 100644 .changeset/null-ranges-malformed.md delete mode 100644 .changeset/pack-add-backup-warning.md delete mode 100644 .changeset/pack-layers.md delete mode 100644 .changeset/pack-source-allowlist.md delete mode 100644 .changeset/pack-staging-stale.md delete mode 100644 .changeset/pack-test.md delete mode 100644 .changeset/packs-lock.md delete mode 100644 .changeset/panel-verification.md delete mode 100644 .changeset/post-content-digest.md delete mode 100644 .changeset/post-content-recovery.md delete mode 100644 .changeset/post-live-authority.md delete mode 100644 .changeset/post-secret-guard.md delete mode 100644 .changeset/post-summary-comment.md delete mode 100644 .changeset/post-verdict-reconciliation.md delete mode 100644 .changeset/pr-wait-inconsistent-reads.md delete mode 100644 .changeset/prompt-key-shapes.md delete mode 100644 .changeset/prompt-rigor-key.md delete mode 100644 .changeset/provenance-receipt.md delete mode 100644 .changeset/provider-from-remote.md delete mode 100644 .changeset/punctuated-mentions.md delete mode 100644 .changeset/reaction-scoped-fix.md delete mode 100644 .changeset/reactions-trend.md delete mode 100644 .changeset/read-verb-re-enter.md delete mode 100644 .changeset/ready-calibrate-gitlab.md delete mode 100644 .changeset/recipes-format.md delete mode 100644 .changeset/record-cost-name-warnings.md delete mode 100644 .changeset/release-asset-sigpipe.md delete mode 100644 .changeset/report-html.md delete mode 100644 .changeset/resolution-rate.md delete mode 100644 .changeset/review-residuals.md delete mode 100644 .changeset/rigor-validation.md delete mode 100644 .changeset/risk-queue.md delete mode 100644 .changeset/risk-score.md delete mode 100644 .changeset/rule-exposure.md delete mode 100644 .changeset/rule-pack-import.md delete mode 100644 .changeset/run-agent.md delete mode 100644 .changeset/run-argument-credential.md delete mode 100644 .changeset/run-print-stage.md delete mode 100644 .changeset/run-repair-loop.md delete mode 100644 .changeset/self-review-loop.md delete mode 100644 .changeset/serve-live-updates.md delete mode 100644 .changeset/serve-s1.md delete mode 100644 .changeset/serve-s2-discovery.md delete mode 100644 .changeset/serve-s3-cost-precision.md delete mode 100644 .changeset/serve-ui.md delete mode 100644 .changeset/setup-verdict-hardening.md delete mode 100644 .changeset/skill-install-preflight.md delete mode 100644 .changeset/skill-management.md delete mode 100644 .changeset/store-lock-and-listing.md delete mode 100644 .changeset/strikethrough-mentions.md delete mode 100644 .changeset/thread-resolve.md delete mode 100644 .changeset/troubleshooting.md delete mode 100644 .changeset/typed-truncation.md delete mode 100644 .changeset/upgrade-hints-quoting.md delete mode 100644 .changeset/vanished-threads.md delete mode 100644 .changeset/verify-env-allowlist.md delete mode 100644 .changeset/verify-indirection.md delete mode 100644 .changeset/verify-out-owner-only.md delete mode 100644 .changeset/verify-sandbox-macos.md delete mode 100644 .changeset/verify-sandbox.md delete mode 100644 .changeset/verify-static-gate-bypasses.md delete mode 100644 .changeset/windows-watchdog-no-hang.md delete mode 100644 .changeset/windows-watchdog-resume-retry.md delete mode 100644 .changeset/withheld-capabilities.md delete mode 100644 .changeset/write-bundle-symlinks.md diff --git a/.changeset/00-what-s-new.md b/.changeset/00-what-s-new.md new file mode 100644 index 000000000..da5abf972 --- /dev/null +++ b/.changeset/00-what-s-new.md @@ -0,0 +1,15 @@ +--- +"@stainless-code/verdict": minor +--- + +**What's new in 0.7.0** + +Verdict 0.7.0 is a breadth release. It adds a local code index, a read-only `serve` surface with a browser +UI, and a backtest harness that mines past bug fixes from your own history, replays them against your +agent, and scores the result. Review output gains an auditable provenance receipt, diff-coverage +attestation, advisory risk scoring, and an optional validation pass that floors each finding on the +evidence behind it, while the CI gate now checks coverage by default. Posting grows a secret guard, honest +per-host capability reporting, and a summary-comment option, GitLab and Forgejo join GitHub as full publish +targets, and lens packs, declared recipes, and reported verification make the review pipeline itself data. +Thirteen pre-1.0 changes alter behaviour — each names the migration step below — and the known limits close +the notes. diff --git a/.changeset/01-new-commands-and-flags.md b/.changeset/01-new-commands-and-flags.md new file mode 100644 index 000000000..973d3bb43 --- /dev/null +++ b/.changeset/01-new-commands-and-flags.md @@ -0,0 +1,25 @@ +--- +"@stainless-code/verdict": minor +--- + +**New commands and flags** + +- **`pr-wait`** — wait for a PR to merge, close, change checks, or change review state, with a bounded timeout and typed timeout data. [Bounded foreground waits](docs/manual.md#bounded-foreground-waits) +- **`ingest`** — import SARIF 2.1.0 results into a bundle's findings store, sharing one dedupe and suppression path with model reviews. [Static-analysis ingestion](docs/manual.md#local-review-no-pr) +- **`threads resolve`** — close the review threads whose findings your store marks addressed, one receipt per run, with `--dry-run` and typed refusals for threads verdict did not author. [Guarded submission](docs/manual.md#guarded-submission-and-operator-surfaces) +- **`risk`** — rank every local bundle into an advisory triage queue by band and 0–100 score, without a host call or an index build. [Triage risk queue](docs/manual.md#triage-risk-queue) +- **`report --html`** — export the six dashboard views as a static site, published through staging and backup siblings with rollback and symlink refusals. [Static dashboard export](docs/manual.md#static-dashboard-export) +- **`run --agent`** — run your own agent CLI against an existing bundle, assert exactly one fresh `review.json`, then hand off to the unchanged record/post validation. [Your agent, launched](docs/manual.md#your-agent-launched-run---agent) +- **`panel`** — plan independent verifier dispatches locally; keep-biased results enrich findings and summaries without voting or dropping anything. [Panel verification](docs/manual.md#panel-verification) +- **Recipes** — `.verdict/recipes.json` declares a lens set, synthesizer hint, and budget as strict versioned data, read at the PR base and hashed into the bundle. [Recipes](docs/manual.md#recipes-declared-review-pipelines) +- **`skill show`, `skill status`, `skill uninstall`** — inspect one payload file byte-identically, gate on drift with an exit code, and remove only the files a receipt owns. [Install routes](docs/manual.md#install-routes-and-where-commands-run) +- **`post --summary comment`** — maintain one separate PR-level summary comment, updated in place across all three publish hosts; the default `off` keeps output byte-identical. [Guarded submission](docs/manual.md#guarded-submission-and-operator-surfaces) +- **`fix --all`** — the batch form of the bounded fix loop: every open inline finding at or above `--min-severity`, one pass per sig under the same ceilings, per-sig outcomes in `data.batch`, and a `fix.patch` written beside the ledger. [Fix loops](docs/manual.md#fix-loops-bounded-verify-gated) +- **`fix --selected`** — address exactly the sigs a human 👍 selected on the root verdict comment, read through one shared reaction reader; a selection file is written only when something was reacted to. [Fix loops](docs/manual.md#fix-loops-bounded-verify-gated) +- **`handoff`** — one markdown projection of the open inline findings `fix` would take, selecting through the same selector, with `--json`, `--format md`, and repeatable `--paths`; read-only, with no ledger and no host call. [Handoff](docs/manual.md#handoff-the-fix-brief-for-someone-else) +- **`run --print --stage `** — render one recipe stage at a time, so CI can dispatch the four stages under different models. [Recipes](docs/manual.md#recipes-declared-review-pipelines) +- **`backtest reverts`** — record the reverts git already knows about into a sorted, byte-stable `.verdict/reverts.jsonl`, with `.verdict/reverts.md` as the fenced reading surface. [Reverts](docs/manual.md#reverts) +- **`backtest mine`** — mine SZZ candidates from a local clone with zero model calls and zero network, writing a timestamp-free `mine.json`. [Backtest](docs/manual.md#backtest-szz-candidates) +- **`backtest replay`** — replay mined candidates against your own agent in a throwaway worktree per candidate, bounded by review count, projected budget, and timeout. [Replay](docs/manual.md#replay) +- **`backtest score`** — grade a replay's findings against the candidate bug lines and report a catch rate, off-target findings per PR, and cost per review as a pure read. [Score](docs/manual.md#score) +- **`learn --incident`** — attach one revert or incident to a recorded bundle in a sorted, idempotent `.verdict/incidents.jsonl` ledger. [Knowledge base](docs/manual.md#knowledge-base-scoped-local-rules) diff --git a/.changeset/02-local-code-index.md b/.changeset/02-local-code-index.md new file mode 100644 index 000000000..acc1b1c7c --- /dev/null +++ b/.changeset/02-local-code-index.md @@ -0,0 +1,15 @@ +--- +"@stainless-code/verdict": minor +--- + +**Local code index** + +- **`verdict index`** — build a local Go index with the standard library, stored by tree SHA so an empty commit reuses a tree and a rename parses nothing. [Code index](docs/manual.md#code-index) +- **Graph queries** — `index tests-for`, `index co-changed`, and `index owners` answer candidate tests, co-change partners, and the CODEOWNERS rule at a tree. [Code index](docs/manual.md#code-index) +- **Three more languages** — the index now reads TypeScript, JavaScript, and Python through two dependency-free lexical scanners, with no tree-sitter and no cgo. [Code index](docs/manual.md#code-index) +- **`index blast-radius`** — one change's reach: callers to depth 2 plus candidate tests, co-change partners, and owners for the changed paths. [Code index](docs/manual.md#code-index) +- **Transitive tests** — `tests-for` walks the reverse import graph up to `--depth` hops and labels tests it reached through imports. [Code index](docs/manual.md#code-index) +- **`fetch --index`** — hand an index you already built to `fetch` or `fetch-local` and the bundle gains an `impact` artifact and a fenced `## impact` section. [Impact hand-off](docs/manual.md#impact-hand-off-fetch---index) +- **Missing-index hints** — `index blast-radius` and `fetch` now name the revision and tree an absent index is about, with the exact shell-quoted build step that answers for it. [Code index](docs/manual.md#code-index) +- **Scanner fixes** — four corpus bugs are fixed: a transitive test no longer disappears, relative imports resolve, a Python attribute call is no longer misattributed, and a Go selector outside its package is counted unresolved. [Code index](docs/manual.md#code-index) +- **One impact budget** — the impact section's byte budget is shared across symbols, callers, and tests, each list carrying its own omitted count. [Code index](docs/manual.md#code-index) diff --git a/.changeset/03-review-quality-and-provenance.md b/.changeset/03-review-quality-and-provenance.md new file mode 100644 index 000000000..a69c6b69a --- /dev/null +++ b/.changeset/03-review-quality-and-provenance.md @@ -0,0 +1,27 @@ +--- +"@stainless-code/verdict": minor +--- + +**Review quality and provenance** + +- **Explicit context** — `fetch --context` and `fetch-local --context` attach named context files, and findings cite them. [Explicit context](docs/manual.md#explicit-context) +- **Coverage basis** — every coverage read now names what its percentage counts: full hunk spans, both sides and context, not just changed lines. [Diff coverage](docs/manual.md#diff-coverage-agent-attestation) +- **Coverage attestation** — `review.json` gains an optional `coverage[]` checked against the bundle's own hunks, with `--require-coverage` to refuse a partial claim and one body line on a partial attestation. [Diff coverage](docs/manual.md#diff-coverage-agent-attestation) +- **`ranges: null` is malformed** — a coverage claim whose `ranges` is present but not a string now counts as unattested instead of as the whole file. [Diff coverage](docs/manual.md#diff-coverage-agent-attestation) +- **Ignore overrides** — repeatable `--ignore ` and `--no-default-excludes` on fetch, both recorded in the exclusion receipt. [Exclusions](docs/manual.md#exclusions-and-policy-silence) +- **Suppression receipts** — excluded diff bytes and withheld findings are receipted, with `post --min-severity` and `report --include-withheld` to read them. [Exclusions](docs/manual.md#exclusions-and-policy-silence) +- **Proposed learnings** — propose inactive one-rule packs from resolved threads, accept them through the pack importer, or reject them permanently with a reason. [Proposed learnings](docs/manual.md#proposed-learnings) +- **Rule import** — `learn --pack --id ` imports a structured rule object and keeps its pack provenance and hash. [Import a rule object](docs/manual.md#import-a-rule-object) +- **Rule exposure** — `learn --status` gains `recalled_in`, `cited_in`, `exposure`, and `acceptance` per rule, reporting unmeasured recall as null rather than zero. [Knowledge base](docs/manual.md#knowledge-base-scoped-local-rules) +- **Risk scoring** — an optional `risk` claim in `review.json` plus `report --risk` and `report --max-risk`, which reports the larger of your claim and a deterministic floor. [Triage risk queue](docs/manual.md#triage-risk-queue) +- **Reaction proposals** — opt-in reaction proposals from posted comments, plus per-rule acceptance and false-hit rates. [Reaction proposals](docs/manual.md#reaction-proposals-and-rule-rates) +- **Resolution metrics** — live resolution rollups, clean-review metrics, and calibration receipts, with clean reports getting their own headline. [Resolution and clean reviews](docs/manual.md#resolution-and-clean-reviews) +- **Withheld capabilities** — `post.json` records, per capability, what a host could not carry and why, so a degraded publish stays auditable. [Guarded submission](docs/manual.md#guarded-submission-and-operator-surfaces) +- **Payload receipts** — a post payload receipt now hashes the rendered review content and the reviewed head, not diff positions or range metadata. [Guarded submission](docs/manual.md#guarded-submission-and-operator-surfaces) +- **Partial publication** — post only the comments and summary parts missing from the fetched remote content, and rearm at a new head with `rearmed` and `previous_head` recorded. [Guarded submission](docs/manual.md#guarded-submission-and-operator-surfaces) +- **Live authority** — post refreshes live reviews and comments first and refuses without writing when that read fails, so a retry no longer needs a re-fetch. [Guarded submission](docs/manual.md#guarded-submission-and-operator-surfaces) +- **Author fingerprint** — `fetch` writes an advisory `author.json` (`{agent, model, confidence, evidence[]}`) from five harness signals, suggests a reviewer from another model family, and `record.json` carries the author/reviewer pair that `trend` tallies. [Bundle namespace](docs/manual.md#bundle-namespace-core-files-vs-x-) +- **Revert memory** — a changed path that matches the tracked reverts ledger gives `fetch` a bounded, fenced `reverts.txt`, and the review prompt treats a revert as evidence the change broke something once. [Reverts](docs/manual.md#reverts) +- **Validation pass** — a trusted `prompts/validate.md` pass writes `validation.json` per sig, `rigor_policy` floors each severity tier, and `post --require-rigor` narrows it per invocation: advisory by default, a receipted withhold in gate mode. [Rigor floors](docs/manual.md#rigor-floors-and-the-validation-pass) +- **Published key names** — `prompts/review.md` tags the evidence rung as `rigor` and the in-bounds flag as `in-bounds`, and an archtest holds every published key to `review.schema.json`. [Rigor floors](docs/manual.md#rigor-floors-and-the-validation-pass) +- **Verdict reconciliation** — binding verdicts reconcile separately from summary and inline content, so a retried approval does not duplicate a GitLab summary note. [Submit per host](docs/manual.md#submit-post-per-host) diff --git a/.changeset/04-safety-and-trust.md b/.changeset/04-safety-and-trust.md new file mode 100644 index 000000000..bdea65550 --- /dev/null +++ b/.changeset/04-safety-and-trust.md @@ -0,0 +1,21 @@ +--- +"@stainless-code/verdict": patch +--- + +**Safety and trust** + +- **Tokens off the argv** — Forgejo and Bitbucket reads hand the credential to curl on stdin or a 0600 file, so it is gone from `ps` and `/proc//cmdline`. [Providers](docs/manual.md#providers-non-github-hosts) +- **Secret guard** — `post` refuses a secret-shaped value before the first host call, naming the shape and never the value, with `--allow-secret-values` as the explicit exception. [Guarded submission](docs/manual.md#guarded-submission-and-operator-surfaces) +- **Mentions neutralized** — `record` escapes `@`-mentions once, where the body is built, so recorded, posted, dry-run, and digested bytes cannot diverge. [Mentions](docs/manual.md#mentions-in-the-published-text) +- **Punctuated mentions** — `(@alice)`, `**@alice**`, and `_@alice_` are escaped while real addresses stay byte-identical. [Mentions](docs/manual.md#mentions-in-the-published-text) +- **Strikethrough mentions** — `~~@alice~~` is escaped too, while a lone tilde in prose stays as written. [Mentions](docs/manual.md#mentions-in-the-published-text) +- **Learning evidence** — cited paths and comment identifiers must stay inside quoted learning evidence, and unsafe ones refuse before a learning is proposed or accepted. [Learnings](docs/manual.md#learnings-durable-repo-memory) +- **Unicode line breaks** — U+2028 and U+2029 now refuse a serve URL, refuse a learning path, and flatten in a rendered `pack test` line. [Serve](docs/manual.md#serve-local-http-surface) +- **Store locking** — every `.verdict/kb.jsonl` write takes one advisory lock; a second writer refuses `store-busy` instead of dropping a row. [Findings store](docs/manual.md#findings-store--triage-authority) +- **Enforced sandbox** — `verify --sandbox` runs the verify command under bubblewrap on Linux with no network, a private PID namespace, empty `~/.ssh`/`~/.aws`/`~/.docker`, and no silent fallback. [Enforced sandbox](docs/manual.md#enforced-sandbox-verify---sandbox) +- **macOS sandbox** — `verify --sandbox` now denies every `bind()` and `connect()` on macOS through the built-in `sandbox-exec`, loopback and unix sockets included, so filesystem isolation stays Linux-only. [Enforced sandbox](docs/manual.md#enforced-sandbox-verify---sandbox) +- **Variable command words refuse** — `E=rm; $E -rf ./target` and the braced, quoted, `export`ed, loop, concatenated, and own-line spellings now refuse `verify-indirection`, including after `env`/`timeout`/`xargs` wrappers and inside `sh -c`. [Known residuals](docs/manual.md#known-residuals-string-scan-ceiling) +- **`.verify.out` is owner-only** — the verify stream is written `0600` on Unix, like `.run.out`, because a command that prints a credential leaves the value in it. [Enforced sandbox](docs/manual.md#enforced-sandbox-verify---sandbox) +- **Launch-record redaction** — a secret-shaped `run --agent` operand no longer reaches the recorded `output_tail` or the 0644 child log. [What `run` asserts](docs/manual.md#what-run-asserts) +- **setup-verdict hardening** — the action passes its inputs through environment variables and verifies a cached archive before reuse, failing closed on a tampered one. [The runner is your CI](docs/manual.md#the-runner-is-your-ci) +- **Install preflight** — `skill install` probes exactly the directories it will write before the first byte, so a read-only target leaves nothing half-installed. [Install routes](docs/manual.md#install-routes-and-where-commands-run) diff --git a/.changeset/05-agents-and-mcp.md b/.changeset/05-agents-and-mcp.md new file mode 100644 index 000000000..ef046be1b --- /dev/null +++ b/.changeset/05-agents-and-mcp.md @@ -0,0 +1,11 @@ +--- +"@stainless-code/verdict": minor +--- + +**Agents and MCP** + +- **Bounded repair loop** — `run --rounds N` re-prompts an agent with the shape checker's fix list, stops on the first valid round, and refuses `repair-exhausted`. [Your agent, launched](docs/manual.md#your-agent-launched-run---agent) +- **Self-review loop** — a bounded local self-review recipe for Claude, Codex, and OpenCode with round receipts and coverage bound to the verified diff identity. [Agent self-review loop](docs/manual.md#agent-self-review-loop) +- **MCP config for six more hosts** — `mcp config` writes snippets for six more hosts and carries `--read-only` or `--operator` into the argv it prints. [Host config](docs/manual.md#host-config) +- **Declared mutation classes** — every MCP tool states its local mutation class, and a call hidden by `--read-only` returns the typed `read-only-tool` refusal. [MCP server](docs/manual.md#mcp-server-stdio-experimental-until-10) +- **Agent preflight** — a missing agent binary is caught before prompt assembly, with install guidance, and printed or recorded pass-through arguments are limited to safe model and cost options. [Agents, arguments, and `--print`](docs/manual.md#agents-arguments-and---print) diff --git a/.changeset/06-hosts-and-ci.md b/.changeset/06-hosts-and-ci.md new file mode 100644 index 000000000..db30d130f --- /dev/null +++ b/.changeset/06-hosts-and-ci.md @@ -0,0 +1,12 @@ +--- +"@stainless-code/verdict": minor +--- + +**Hosts and CI** + +- **Forgejo publish** — `post` submits one atomic Forgejo review with the body, verdict, reviewed commit, and every inline finding, and withholds what the host cannot anchor rather than guessing. [Submit per host](docs/manual.md#submit-post-per-host) +- **GitLab publish** — each finding becomes one positioned discussion, the review body a summary note, and an approval verdict a real MR approval. [Submit per host](docs/manual.md#submit-post-per-host) +- **GitLab incremental** — `fetch --provider gitlab --base-sha` writes the same `incremental.diff` contract the GitHub compare path writes, and hosts without a compare adapter say so in a note. [Providers](docs/manual.md#providers-non-github-hosts) +- **GitLab readiness and calibration** — `ready` advises on merge requests and plain `calibrate` joins live GitLab threads through `glab`, on the same blocker and outcome vocabulary as GitHub. [Calibration](docs/manual.md#calibration-predictions-vs-ground-truth) +- **`doctor --provider`** — preflight one host at a time, with a per-host probe table and credentials reported by presence, never by value. [Doctor preflight](docs/manual.md#doctor-preflight) +- **GitHub Action** — the review and merge gate ships as a first-party composite action that installs a pinned verified release and uses your runner, your agent, and your keys. [The runner is your CI](docs/manual.md#the-runner-is-your-ci) diff --git a/.changeset/07-serve-and-ui.md b/.changeset/07-serve-and-ui.md new file mode 100644 index 000000000..d3e4e94cb --- /dev/null +++ b/.changeset/07-serve-and-ui.md @@ -0,0 +1,15 @@ +--- +"@stainless-code/verdict": minor +--- + +**Serve and the browser UI** + +- **`verdict serve`** — a read-only HTTP surface over the current repository, loopback by default, with a session-cookie exchange, strict CSP, Host and Origin gates, and per-response source traceability. [Serve](docs/manual.md#serve-local-http-surface) +- **Root discovery** — `serve --root` finds nested repositories under a bounded traversal backed by a local 24-hour registry, and reports cache, truncation, and skipped patterns. [Serve](docs/manual.md#serve-local-http-surface) +- **Cost and precision buckets** — `/v1/cost` adds repository, ISO-week, and model buckets and `/v1/precision` adds category and tier counts, keeping unmeasured prices null and marking corrupt aggregates partial. [Serve](docs/manual.md#serve-local-http-surface) +- **Browser UI** — `serve` opens an embedded UI with inbox, review, threads, cost, precision, and gate views, dark mode, live refresh, and static `data.json` reading. [Serve](docs/manual.md#serve-local-http-surface) +- **Live invalidation** — bundles invalidate as they change, with `--interval`, `--sweep`, and `--no-watch` to control it. [Serve](docs/manual.md#serve-local-http-surface) +- **Vanished thread feeds** — a thread feed removed mid-read now reads as absent instead of corrupt. [Serve](docs/manual.md#serve-local-http-surface) +- **Cache content budget** — serve cache hashing is bounded to 16 MiB per bundle lookup and bypasses the cache when the budget is exceeded. [Serve](docs/manual.md#serve-local-http-surface) +- **Small-file rewrites** — artifacts up to 64 KiB are hashed, so a rewrite that preserves size and mtime is still detected. [Serve](docs/manual.md#serve-local-http-surface) +- **Inventory accounting** — serve inventory accounting stays inside its byte budget even for oversized artifacts. [Serve](docs/manual.md#serve-local-http-surface) diff --git a/.changeset/08-lens-packs.md b/.changeset/08-lens-packs.md new file mode 100644 index 000000000..dd699a430 --- /dev/null +++ b/.changeset/08-lens-packs.md @@ -0,0 +1,13 @@ +--- +"@stainless-code/verdict": minor +--- + +**Lens packs** + +- **Built-in lenses as data** — `perf`, `api-contract`, and `test` now ship as a pack resolved into every bundle as `lenses.txt`, with the four hashed prompt files byte-identical. [Lens packs](docs/manual.md#lens-packs) +- **No per-lens cap** — a lens returns every finding its evidence supports: the built-in pack declares only `max_files: 12`, and the dispatch template leads with the lens's domain. [Lens packs](docs/manual.md#lens-packs) +- **`pack add` and a lock** — vendor a pack from a directory or git source, validated before it loads and pinned in a strict `.verdict/verdict.lock`. [Lens packs](docs/manual.md#lens-packs) +- **Layered packs** — `pack list` reports every loaded pack with its layer, source, tree hash, ids, and drift, and a repo pack is read from the trusted base so a PR cannot edit the pack that reviews it. [Lens packs](docs/manual.md#lens-packs) +- **`pack test`** — run a pack's eval cases offline. [Lens packs](docs/manual.md#lens-packs) +- **Leftover backup sibling** — a failed sweep of a previous tree is reported as success plus a warning naming the leftover, instead of refusing an add whose new tree is already durable. [Lens packs](docs/manual.md#lens-packs) +- **Stale staging** — a crashed re-add's leftover staging or backup sibling refuses `pack-staging-stale` with the path to delete. [Lens packs](docs/manual.md#lens-packs) diff --git a/.changeset/09-measurement.md b/.changeset/09-measurement.md new file mode 100644 index 000000000..6fa2b8ab9 --- /dev/null +++ b/.changeset/09-measurement.md @@ -0,0 +1,11 @@ +--- +"@stainless-code/verdict": minor +--- + +**Measurement** + +- **Clean controls** — the benchmark corpus accepts clean controls with empty ground truth, ships two seeded clean cases, and keeps seeded fixtures labelled as not a model-quality claim. [Benchmark corpus](docs/benchmark.md) +- **Reproducible metrics** — typed severity weights, ordered k-of-n recall, control false-positive rates, weighted F1, signal-to-noise, and explicit denominators, with provenance and argv recorded. [Benchmark corpus](docs/benchmark.md) +- **Adjudicated rates** — a versioned per-run adjudication record reclassifies an unmatched finding as `real-unlisted`, `matches-ground-truth`, or `false-positive`, and `bench/runner -render` recomputes strict and adjudicated precision and recall without re-running a model. [Benchmark corpus](docs/benchmark.md) +- **First real cases** — four real cases from `urfave/cli` (three bugs and one clean control), `-max-usd` stops scheduling at measured spend and marks the run uncitable, and refused samples are counted in the cost rollup. [Benchmark corpus](docs/benchmark.md) +- **Corpus grading guard** — a keyword the case's own diff spells refuses, a case left with no phrase asks for another keyword, and a real bundle whose `full.diff` is missing, unreadable, or prose is refused instead of graded leniently. [Benchmark corpus](docs/benchmark.md) diff --git a/.changeset/10-breaking-changes.md b/.changeset/10-breaking-changes.md new file mode 100644 index 000000000..c697edc8a --- /dev/null +++ b/.changeset/10-breaking-changes.md @@ -0,0 +1,21 @@ +--- +"@stainless-code/verdict": minor +--- + +**Breaking changes (pre-1.0)** + +Package is on `0.x`, so these ship as a minor bump; each one names the step that migrates it. + +- **`index blast-radius` truncation shape** — the integer `truncated` count and `depth_truncated` are replaced by the shared boolean receipt (`truncated`, `truncation_reason`, `returned`, `total`, `next_offset`), and a complete result omits the receipt instead of carrying `truncated: false`. _Migrate:_ read `truncation_reason` and `returned`; treat an absent receipt as complete. [Code index](docs/manual.md#code-index) +- **`gate` checks coverage by default** — the new `coverage` dimension fails when the gated HEAD has no receipt, a receipt for another HEAD, no `coverage[]` attestation, or under 100% attested. _Migrate:_ `--coverage warn`, or the SDK `Coverage` field, keeps the old passing behaviour while reviews attest nothing. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) +- **`gate --live` reads every host** — a non-GitHub bundle used to pass the live dimension without reading threads; it now reads the bundle's own host and fails closed when it cannot. _Migrate:_ `--live auto` warns and gates on the local dimensions only. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) +- **Unknown gate tiers refuse** — an unrecognised tier is `bad-flags` on the MCP `report` tool and in `Report.Check`/`Gate.MaxSeverity`, not only on the CLI. _Migrate:_ send a tier from the published set, or omit the field. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) +- **`verify` child environment is default-deny** — only baseline locale, terminal, shell, XDG, and toolchain names pass, and a secret-shaped name is dropped even when it matches an allowlist prefix. _Migrate:_ list what you need in `.verdict/verify-env` or `--env-allowlist`; dropped names appear in `warnings[]`. [Child environment](docs/manual.md#child-environment-what-inherits-what-is-explicit) +- **`verify` refuses command shapes it used to run** — `wget --remote-name`/`-i`, `curl -K`, `busybox wget`, `pip3 download `, `telnet`, `git archive --remote`, and the rest refuse `verify-networked` in attached and clustered spellings, and a command word built from a variable (`E=rm; $E -rf ./target`, including the braced, quoted, `export`ed, loop, and wrapper spellings) refuses `verify-indirection`. _Migrate:_ write the tool word plainly and rewrite the command to a plain fetch your recipe can run; every refusal keeps its `try:` hint. [Known residuals](docs/manual.md#known-residuals-string-scan-ceiling) +- **`.verify.out` is owner-only** — the verify stream is written `0600` on Unix where it was `0644`, because a command that prints a credential leaves the value in it. _Migrate:_ read it as the user that ran `verify`, or copy the bytes inside that same step; on Windows the file keeps the directory's inherited ACL. [Enforced sandbox](docs/manual.md#enforced-sandbox-verify---sandbox) +- **Bundle paths are guarded** — a bundle name that is not one path element (`a/b`, `..`, `CON`, `x:`) refuses `bad-flags`, and a symlinked bundle root or component refuses before any write, including a link that points inside the repository. _Migrate:_ one plain name per bundle, and a real directory where a link used to work. [Bundle namespace](docs/manual.md#bundle-namespace-core-files-vs-x-) +- **Provider is inferred from origin** — with no `--provider` flag, an unknown or conflicting origin now refuses `provider-unknown` instead of defaulting to GitHub. _Migrate:_ pass `--provider` explicitly, or keep a plain GitHub origin. [Providers](docs/manual.md#providers-non-github-hosts) +- **`capabilities` splits, `truncated` widens** — the binding verdict becomes `approve` and `request_changes` with `review_verdicts` kept as the compatibility sum, `unsupported` names each absent capability, and `data.truncated` widens from `boolean` to `boolean | row[]`. _Migrate:_ branch on the two new flags, and narrow `truncated` before you read it as a boolean. [Published schemas](docs/manual.md#published-schemas--drift-rule) +- **`pack add` source allowlist** — only `https://host/path` with no userinfo, `ssh://user@host/path`, and `user@host:path` are accepted; anything else refuses before a git call, and a lock row carrying one refuses at `pack list`, `fetch`, `post`, and `record`. _Migrate:_ rewrite the source in one of the three forms. [Lens packs](docs/manual.md#lens-packs) +- **Benchmark cases name their commits** — a real corpus case now requires `provenance.base_sha` and `provenance.head_sha` (40-hex), the two commits the reviewed diff spans, and a bundle stamps the same pair. _Migrate:_ add both commits to every real `case.json`; a case without them fails validation. [Benchmark corpus](docs/benchmark.md) +- **Rebuild indexes** — indexes written by the previous scanner over-report callers, so rebuild them. _Migrate:_ `verdict index build --at HEAD`. [Code index](docs/manual.md#code-index) diff --git a/.changeset/11-fixes.md b/.changeset/11-fixes.md new file mode 100644 index 000000000..ae670180d --- /dev/null +++ b/.changeset/11-fixes.md @@ -0,0 +1,17 @@ +--- +"@stainless-code/verdict": patch +--- + +**Fixes** + +- **Read-verb refusals** — `report`, `next`, and `show` refusals carry a shell-quoted `re-enter:` line naming the invocation you typed. [Machine surface](docs/manual.md#machine-surface) +- **`next --coverage` with no positionals** — leads with one `error:` line naming what is missing, and a plain `Basis:` line is accepted without a warning. [Machine surface](docs/manual.md#machine-surface) +- **Unnamed cost models** — `record` warns once per `cost.json` `models[]` entry with no `name`, and still accepts and prices the record. [Cost control](docs/manual.md#cost-control-budget-estimate-consent) +- **Over-cap and missing-index refusals** — `diff-too-large` and `missing-index` now carry a `next:` line naming the exclusions that fit, and the exact build command. [Over-cap diffs](docs/manual.md#over-cap-diffs-chunked-fetch--manual-recipe) +- **Five small repairs** — an agent that fails just before its budget reads `agent-failed`, MCP refusals carry their warnings, `index-capped` names the bytes it measured, `index status` points at the query, and a crashed agent's partial review can no longer block on a FIFO. [Machine surface](docs/manual.md#machine-surface) +- **Windows watchdog** — a command that failed to start and survived the cleanup kill now returns at once, naming both errors, instead of hanging forever. [Bounded foreground waits](docs/manual.md#bounded-foreground-waits) +- **Quoted upgrade hints** — `verdict upgrade` quotes its project-local `cd` hint so a path with a space or quote pastes resolvably, and the uninstall sibling hint reads as a sentence. [Machine surface](docs/manual.md#machine-surface) +- **Release asset check** — the release script's completeness check no longer reports a shipped asset as missing when the asset list is large. [Release train](CONTRIBUTING.md) +- **`pr-wait` inconsistent reads** — a head that is not yet the same on both APIs is retried instead of failing the bounded wait, and `inconsistent_observations` keeps the retry visible. [Bounded foreground waits](docs/manual.md#bounded-foreground-waits) +- **Windows watchdog resume** — a command that failed before starting under load retries the thread lookup instead of refusing `resume suspended verify process`. [Bounded foreground waits](docs/manual.md#bounded-foreground-waits) +- **`backtest mine` in a partial clone** — a `--filter=blob:none` clone refuses `partial-clone` with the step that fetches the missing objects, instead of stalling. [Backtest](docs/manual.md#backtest-szz-candidates) diff --git a/.changeset/12-docs.md b/.changeset/12-docs.md new file mode 100644 index 000000000..7ce03aa90 --- /dev/null +++ b/.changeset/12-docs.md @@ -0,0 +1,9 @@ +--- +"@stainless-code/verdict": patch +--- + +**Docs** + +- **Refusal code reference** — `docs/troubleshooting.md` lists every refusal code with its cause and fix, and a drift test fails the build when one is missing. [Refusal codes](docs/troubleshooting.md) +- **Docs site structure** — grouped sections with short sidebar labels, a Markdown twin for every page, and a glossary route from `verdict --help`. [docs/](docs/README.md) +- **Defaults at a glance** — a generated table names every default a bare run takes and the verb it shapes, so it cannot drift from the shipped binary. [Defaults](docs/manual.md#defaults-at-a-glance) diff --git a/.changeset/13-known-limits.md b/.changeset/13-known-limits.md new file mode 100644 index 000000000..08b66f7b4 --- /dev/null +++ b/.changeset/13-known-limits.md @@ -0,0 +1,9 @@ +--- +"@stainless-code/verdict": patch +--- + +**Known limits** + +- **The macOS sandbox covers the network half only** — `verify --sandbox` denies `bind()` and `connect()` on macOS but has no filesystem namespace, so credential paths, `/proc`, `/run`, and a private `/tmp` are hidden only on Linux under `bubblewrap`; filesystem isolation on macOS needs a Linux VM or a container running that recipe. [Enforced sandbox](docs/manual.md#enforced-sandbox-verify---sandbox) +- **The string gate is a tripwire, not a boundary** — `sh "$V"`, a script named by a variable, still passes the pre-exec scan, which is why `verify --sandbox` is the boundary and a spelling outside the gate's reach is a gate gap to arm rather than a vulnerability. [Known residuals](docs/manual.md#known-residuals-string-scan-ceiling) +- **A backslash before a CRLF line ending is not joined as a continuation** — the gate refuses that shape rather than passing it, so a CRLF continuation fails safe. [Known residuals](docs/manual.md#known-residuals-string-scan-ceiling) diff --git a/.changeset/agent-definition-display.md b/.changeset/agent-definition-display.md deleted file mode 100644 index b4ec8a62e..000000000 --- a/.changeset/agent-definition-display.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Check missing agent binaries before prompt assembly, include installation guidance, and restrict printed and recorded pass-through arguments to safe model and cost options. diff --git a/.changeset/author-routing-hint.md b/.changeset/author-routing-hint.md deleted file mode 100644 index b2bc7f790..000000000 --- a/.changeset/author-routing-hint.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -When `author.json` names an agent, `fetch`'s next step says so and suggests a -reviewer from another model family. Advisory only: nothing blocks, and a bundle -with no fingerprint carries no suggestion. diff --git a/.changeset/author-routing-record.md b/.changeset/author-routing-record.md deleted file mode 100644 index 67638bb09..000000000 --- a/.changeset/author-routing-record.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`record.json` carries the author/reviewer pair: `author` (`{agent, model}` from -`author.json`, null when the bundle has none) and `review_models`, every model -name `cost.json` prices. An absent fingerprint stores null, never a human. diff --git a/.changeset/author-routing-skill.md b/.changeset/author-routing-skill.md deleted file mode 100644 index db3c2a5a8..000000000 --- a/.changeset/author-routing-skill.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -The skill tells the orchestrator to prefer a reviewer and verifier from another -model family when `author.json` names an agent, and to record the model it -actually ran. The fingerprint never pins a model. diff --git a/.changeset/author-routing-trend.md b/.changeset/author-routing-trend.md deleted file mode 100644 index 6950d9035..000000000 --- a/.changeset/author-routing-trend.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`trend` groups outcomes by the author/reviewer pair: per-bundle rows carry -`author` and `review_models` from `record.json`, and `by_pair` tallies each -`/` combination, with `unknown` for what a bundle never -recorded. diff --git a/.changeset/author-routing.md b/.changeset/author-routing.md deleted file mode 100644 index d449eddb1..000000000 --- a/.changeset/author-routing.md +++ /dev/null @@ -1,9 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`fetch` writes `.verdict/state//author.json` -(`schemas/author.schema.json`) from five harness signals in the PR title, body, -head ref, and commits: `{agent, model, confidence, evidence[]}`. It is advisory -data that never blocks or changes the review, and no signal writes no file — -absent means unknown, never human. diff --git a/.changeset/backtest-mine.md b/.changeset/backtest-mine.md deleted file mode 100644 index ba3eceb68..000000000 --- a/.changeset/backtest-mine.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Mine SZZ candidates from a local clone. `backtest mine --since -<90d|12w|18mo|2y|ref> [--until ] [--run ]` blames each later -`fix:`/`Revert "..."` commit's deleted lines to the merged or squash PR that -introduced them and writes -`.verdict/backtest//{run.json,mine.json}` with zero model calls and zero -network. `--run` ids are at most 42 bytes, the room a replay bundle name has -for one, so a mined run is always replayable. Candidates carry the PR head, base, reconstruction, merge/head dates, -and the union of anchored `path:line:kind` bugs; exclusions are counted by -reason and never refused. `mine.json` is timestamp-free, so a rerun over the -same window is byte-identical. The artifact is not citable: SZZ blame is a -hypothesis. Replay and score read it. The verb is CLI-only because replay -spends on model runs. diff --git a/.changeset/backtest-partial-clone.md b/.changeset/backtest-partial-clone.md deleted file mode 100644 index b8ad46c90..000000000 --- a/.changeset/backtest-partial-clone.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`backtest mine` in a `--filter=blob:none` clone used to stall with no candidate -for as long as the network kept answering; it now refuses `partial-clone` and -says how to get the missing objects back. diff --git a/.changeset/backtest-replay.md b/.changeset/backtest-replay.md deleted file mode 100644 index b0cd2fe6c..000000000 --- a/.changeset/backtest-replay.md +++ /dev/null @@ -1,24 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Replay mined candidates against your own agent. `backtest replay --run ---agent [--max-reviews ] [--budget-usd ] [--timeout ] -[--rounds ] [-- ]` walks the run's `mine.json` newest merge -first, builds a fresh local bundle at each candidate's own `base..head`, -launches the agent in a detached `git worktree` of that head (removed again -after each review, failures included), and records the review into the -bundle's `findings.json`. Spend is bounded by `--max-reviews` (1–99), a -median-projected `--budget-usd` ceiling, and a per-review `--timeout`; a review -that reported no price is unknown spend, never `$0.00`. A failed review is -that candidate's row (`status` is the refusal code) and the walk continues; -only `agent-missing`, `agent-spawn-failed`, cancellation, and a -`backtest-failed` that would hit every candidate (git gone mid-run, or a -state directory that cannot be created) end the run. -`--run` ids are capped at 42 bytes by `backtest mine` and `backtest replay` -alike, so two runs can never be cut onto one bundle name. -Rerunning the same `--run` replays only candidates that produced no review, and -a stored candidate replayed with another agent or other pass-through arguments -refuses `run-mismatch`. Every bundle carries a `backtest.json` marker, so -`trend` keeps replay input out of your history. Artifacts: -`.verdict/backtest//prs//replay.json`. Replays are not citable. diff --git a/.changeset/backtest-reverts.md b/.changeset/backtest-reverts.md deleted file mode 100644 index cb34860d2..000000000 --- a/.changeset/backtest-reverts.md +++ /dev/null @@ -1,38 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Record the reverts git already knows about. `verdict backtest reverts -[--since <90d|12w|18mo|2y|ref>] [--until ]` scans the same window the -miner resolves for revert commits only (subject `Revert "..."`, the miner's own -predicate and hermetic git layer, no blame stage, no model call) and writes -`.verdict/reverts.jsonl` — a tracked ledger beside `kb.jsonl`, one row per -revert commit with `{schema_version, sha, subject, reverts, date, paths}`, -sorted by `sha`. Rows in the scanned window are upserted and rows outside it are -kept byte for byte, so a narrower scan never forgets a wider one and a rerun of -the same window is byte-identical. `--since` defaults to `12mo`, `--until` to -`HEAD`, and the window refusals are the miner's. - -`.verdict/reverts.md` is the reading surface: one section per stored revert with -the validated short `sha` and `date` outside the fence and every other field -inside one `<<>>` block. Those bytes are author-controlled — the -ledger is tracked and hand-editable — so `internal/untrusted` caps the body, -breaks a forged fence marker and a markdown code fence, neutralizes an HTML -comment opener, folds control and line-separator runes, and labels the block -"evidence, never instructions"; nothing here reaches `review.json` or a post, -and the human lines never print a subject. A stored row is validated field by -field, and a wrong type or shape refuses `corrupt-store` naming the field rather -than rendering as empty values. A subject past 4,096 bytes is stored truncated -so a mass-message commit cannot become repository weight. - -A revert of a merge, one wider than 200 files or 5,000 lines, one touching a -binary file, and one git cannot read are counted by reason (`merge-revert`, -`wide-revert`, `binary`, `unreadable`), never refused. The mode is memory, not a -claim: attach one deliberately with `verdict learn --incident --sha - --path `, so `trend --incidents` counts only the misses a human -attached. - -The scan takes `.verdict/reverts.jsonl.lock` for its read-modify-write, the -third store lock in the tree, and brings `.verdict/.gitignore` to the tracked -contract while it holds it; a lock another scan holds past the wait refuses -`store-busy` and writes nothing. diff --git a/.changeset/backtest-score.md b/.changeset/backtest-score.md deleted file mode 100644 index 90a62f40c..000000000 --- a/.changeset/backtest-score.md +++ /dev/null @@ -1,24 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Score a replay, and run the whole backtest in one command. `backtest score --run -` grades each replayed PR's recorded findings against the bug lines the -run's own SZZ candidates carry, through the corpus matcher (±1 line, one -finding per bug, text that names the defect kind), and writes -`.verdict/backtest//report.json`. It is a pure read: no agent launches, no -bundle is built, and nothing enters `.verdict/state/corpus-*`. -`catch_rate` is `matched / bugs`, a match rate over blame-derived candidates — -a line a later fix deleted is a hypothesis, so a defect fixed without touching -the line is invisible and a deleted cleanup line beside the real bug is a -candidate that was never defective. `off_target_per_pr` is `(reported − -matched) / prs`; an unmatched finding is not proven false, so the report never -calls it one. -`cost_per_review_usd` is null when nothing was priced, and -`spend.unpriced_reviews` marks a partial total as a lower bound. A per-signal -table splits the rate by `fix` and `revert`, `unclassified` counts bugs no -finding can name, and every zero denominator renders `null`. `backtest mine` -now records each bug's seed `signal` in `mine.json`. Bare `backtest --since - --agent ` composes mine, replay, and score on one run id. The -score report is also the verb's `--json` data, published as -`schemas/backtest.schema.json`. Scores are not citable. diff --git a/.changeset/bench-adjudication.md b/.changeset/bench-adjudication.md deleted file mode 100644 index 0cf00bf43..000000000 --- a/.changeset/bench-adjudication.md +++ /dev/null @@ -1,18 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Publish adjudicated precision and recall beside the strict figures. A -versioned per-run adjudication record reclassifies an unmatched finding as -`real-unlisted`, `matches-ground-truth` (the defect a ground-truth bug names in -words the case's keywords miss, refused when it names a bug the run already -matched or sits outside that bug's window), or `false-positive`. Every entry -names its finding by case, sample, and sig, quotes the code in the rationale, -and names who made the call; ground truth is never edited. `bench/runner --render -adjudication [-out ]` recomputes both figures -without re-running a model, and the report and its Markdown carry the -adjudicated rates, the counts per verdict, the record's path and sha256, and a -per-case keyword-miss list. Strict figures stay untouched, and with no record -the adjudicated block is absent rather than copied from them. A malformed or -contradicted record refuses typed (`bad-adjudication`, `unknown-sig`, -`rationale-required`, `run-mismatch`). diff --git a/.changeset/bench-corpus-followups.md b/.changeset/bench-corpus-followups.md deleted file mode 100644 index adf1de9b2..000000000 --- a/.changeset/bench-corpus-followups.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -A corpus bundle whose `full.diff` is prose rather than a diff no longer grades -leniently — it refuses, and `backtest score` names `backtest replay` when the -broken bundle is a replay bundle. The cohort skip warning now names only a -sibling that belonged to the cohort. diff --git a/.changeset/bench-keyword-guard.md b/.changeset/bench-keyword-guard.md deleted file mode 100644 index d098804ee..000000000 --- a/.changeset/bench-keyword-guard.md +++ /dev/null @@ -1,12 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Refuse a corpus keyword the case's own diff spells, including as one part of a -split selector, a spaced selector, or a prefix of a longer identifier. The -built-in `kind` and category phrases the diff spells are instead dropped from -that case's grading, so race and bounds cases stay curatable; a case left with -no phrase is refused and asks for a keyword the diff does not spell. A real -bundle whose `full.diff` is missing or unreadable is refused instead of graded -with the unfiltered built-ins. Grading also normalizes the finding's text, so a -phrase that opens a sentence ("Empty positional: ...") still names the defect. diff --git a/.changeset/bench-real-pilot.md b/.changeset/bench-real-pilot.md deleted file mode 100644 index be6dd6029..000000000 --- a/.changeset/bench-real-pilot.md +++ /dev/null @@ -1,17 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add the first four real benchmark cases (urfave/cli: three bugs and one clean -control) and harden the benchmark runner for paid runs. - -Breaking: real corpus cases now require `provenance.base_sha` and -`provenance.head_sha` (40-hex), the commits the reviewed diff spans. A bundle -stamps the same pair, so a case names exactly what was reviewed. Older real -cases without them fail validation. - -`bench/runner -max-usd ` stops scheduling once measured spend reaches `n`, -lists every unrun sample in `skipped`, and marks the run uncitable. The cost -rollup counts refused samples (`cost.refused_samples`, `cost.refused_usd`) -instead of hiding them. A case's ground-truth keywords must not appear in its -diff, and every ground-truth line, `also_lines` included, must be an added line. diff --git a/.changeset/benchmark-clean-controls.md b/.changeset/benchmark-clean-controls.md deleted file mode 100644 index 99aaeb6cc..000000000 --- a/.changeset/benchmark-clean-controls.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Accept clean benchmark controls with empty ground truth and seeded or real -provenance. Include two seeded clean cases and exercise clean review recording -and calibration through the CLI. Seeded fixtures are not a model-quality claim. diff --git a/.changeset/benchmark-controls-metrics.md b/.changeset/benchmark-controls-metrics.md deleted file mode 100644 index 1c7f368cc..000000000 --- a/.changeset/benchmark-controls-metrics.md +++ /dev/null @@ -1,15 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Make corpus benchmark claims reproducible with typed severity weights, -ordered k-of-n recall, control false-positive rates, weighted F1, signal to -noise, and explicit denominators and observation counts. Preserve unknown -cost and latency as null. The benchmark runner records provenance and command -argv and writes a static Markdown report alongside JSON. Seeded runs are -labelled "not a model-quality claim". - -Metric names distinguish false discovery rate from control-PR false-positive -rate. `control_fp_rate` counts controls with any finding; raw counts use -`control_findings_per_case`. Weighted F1 is named -`severity_weighted_pooled_f1` and its pooled formula is documented. diff --git a/.changeset/blast-radius-index-hint.md b/.changeset/blast-radius-index-hint.md deleted file mode 100644 index 7735eea76..000000000 --- a/.changeset/blast-radius-index-hint.md +++ /dev/null @@ -1,14 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Make `index blast-radius` say which revision and checkout its missing index is -about, and how to answer for it — on both surfaces. The warning names the ref -the operator typed beside the tree it resolved to; the human line now ends -` — run ` (it used to name no command at all, because `next:` reaches only -`--json`), and `next:` carries the same step for JSON. The step is shell-quoted -(`--at 'feature/$USER'`, like `--from` already was), and a hand-off that named a -bare tree SHA gets words instead of a command — `verdict index build at a commit -whose tree is ` — because `build` resolves a commit and refuses every tree -SHA with `bad-at`. The unindexed counts stay `null`, and refusal codes, exit -codes, and the envelope keep their shape. diff --git a/.changeset/bounded-foreground-waits.md b/.changeset/bounded-foreground-waits.md deleted file mode 100644 index 30d467238..000000000 --- a/.changeset/bounded-foreground-waits.md +++ /dev/null @@ -1,9 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add bounded foreground GitHub PR waits through `pr-wait`, MCP `pr_wait`, and -Go and TypeScript clients. Conditions cover merge, closure, checks and review -changes. Timeouts return typed data. `ready --wait-for build --timeout D` waits -for CI completion and returns not-ready on timeout. GitLab and Forgejo waits -refuse explicitly; no daemon or saved cursor is introduced. diff --git a/.changeset/bundle-name-guard.md b/.changeset/bundle-name-guard.md deleted file mode 100644 index 5d2d34e32..000000000 --- a/.changeset/bundle-name-guard.md +++ /dev/null @@ -1,27 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Refuse a bundle name that is not one path element. Every verb that names a -bundle — adopt, calibrate, config explain/preview, fetch, fix, gate, index -blast-radius, next, post, ready, record, report, run, show, threads, triage, -verify, and the MCP tools that take a `pr` — resolves the name through one -shared guard before it joins `.verdict/state`, so `../other`, `a/b`, `a\b`, -`.`, `..`, and the empty name refuse `bad-flags` with one line instead of -reading outside the bundle home. The same guard refuses, on every OS, the -spellings Windows re-reads as another path or as a device: any `:`, a trailing -`.` or space, and a DOS device basename (`CON`, `PRN`, `AUX`, `NUL`, `COM1`- -`COM9`, `LPT1`-`LPT9`). `index blast-radius` reported the same input under its -own message and now renders the shared one. A PR number, a `local-` slug, and a -`corpus-` name still resolve as before. - -The MCP `next` tool answers a findings store it cannot read with the typed -`corrupt-store` refusal instead of `{"item": null}`, so a caller no longer -reads a broken store as "nothing to act on". - -`fetch` and `fetch-local` refuse a bundle directory that is a symlink, with the -`bad-config` refusal, before the first write: a link's target is an unrelated -location even when it resolves inside the repository. `trend` reads each listed -bundle directory it finds instead of resolving the listed name through the -bare-name fallback, so a `42` beside a `local-42` is no longer counted twice, and -`42` keeps its own findings. diff --git a/.changeset/cache-content-budget.md b/.changeset/cache-content-budget.md deleted file mode 100644 index 56b97026e..000000000 --- a/.changeset/cache-content-budget.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Bound serve cache content hashing to 16 MiB per bundle lookup, shared across artifacts, chunks, and override fragments. Bypass the cache when the content budget is exceeded. diff --git a/.changeset/cache-small-content.md b/.changeset/cache-small-content.md deleted file mode 100644 index 9e32405c1..000000000 --- a/.changeset/cache-small-content.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Hash bundle artifacts up to 64 KiB when validating the serve cache, detecting rewrites that preserve size and mtime. diff --git a/.changeset/context-sources.md b/.changeset/context-sources.md deleted file mode 100644 index 1d78cedbe..000000000 --- a/.changeset/context-sources.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Attach explicit context files with `fetch --context` and `fetch-local --context`, and cite context sources on findings. diff --git a/.changeset/coverage-scoring-basis.md b/.changeset/coverage-scoring-basis.md deleted file mode 100644 index da599aa50..000000000 --- a/.changeset/coverage-scoring-basis.md +++ /dev/null @@ -1,14 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -The coverage read names what its percentage counts. `next --coverage` and the -`coverage-incomplete` refusal said only "% of changed lines attested", while the -evaluator scores every hunk span on both sides, context included, so a reviewer -who attested every changed line but not its hunk context read a lower number -than they had covered. Both read points now state the basis beside the number, -for example `coverage: 54% of full hunk spans, both sides and context (12/22 -lines, 1 files)`. The record warning stderr line uses the same basis, and the -constant lives once in `internal/core/coverage` so the surfaces cannot drift. -The `--json` keys are unchanged; the `warnings` string now names the same -basis. The machine-built body line is unchanged. diff --git a/.changeset/curl-token-argv.md b/.changeset/curl-token-argv.md deleted file mode 100644 index 4e56dd16c..000000000 --- a/.changeset/curl-token-argv.md +++ /dev/null @@ -1,11 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Host tokens no longer appear in curl's command line. Forgejo and Bitbucket -reads hand the credential to curl as a config document on stdin, and a Forgejo -write stages it in a 0600 file under a private temp dir because its body owns -stdin; the file is removed after the call. The token, password, and -`Authorization` header are gone from the argv a local `ps` or -`/proc//cmdline` can read. A credential carrying a newline refuses before -any process spawns. diff --git a/.changeset/defaults-at-a-glance.md b/.changeset/defaults-at-a-glance.md deleted file mode 100644 index b739fedca..000000000 --- a/.changeset/defaults-at-a-glance.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`docs/manual.md` gains a "Defaults at a glance" table: one row per default a -bare run takes, naming the verb each one shapes. The table is generated from -the typed defaults registry and the flag parsers, so it cannot drift from the -shipped binary. diff --git a/.changeset/diff-coverage-attestation.md b/.changeset/diff-coverage-attestation.md deleted file mode 100644 index b430f6d58..000000000 --- a/.changeset/diff-coverage-attestation.md +++ /dev/null @@ -1,12 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Attest diff coverage. `review.json` gains an optional `coverage[]` of -`{file, ranges?}`, checked against the bundle's own `hunks.tsv`: `post` and -`record` warn once and add one body line on a partial attestation, and the new -`--require-coverage` flag refuses `coverage-incomplete` without bypassing -`--force`. `next --coverage` reports the unattested ranges before recording, and -`post.json`/`record.json` carry the measured summary. Generated and vendored -paths are exempt; chunked bundles keep `chunks[]` as their only attestation. -Legacy reviews render byte-identically. diff --git a/.changeset/docs-site-restructure.md b/.changeset/docs-site-restructure.md deleted file mode 100644 index afa6ff646..000000000 --- a/.changeset/docs-site-restructure.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Restructured the docs site into grouped sections with short sidebar labels, added a Markdown twin for every page, and pointed `verdict --help` at the new glossary route (`/verdict/docs/glossary`). diff --git a/.changeset/doctor-provider-adopt-gitlab.md b/.changeset/doctor-provider-adopt-gitlab.md deleted file mode 100644 index 323f387ca..000000000 --- a/.changeset/doctor-provider-adopt-gitlab.md +++ /dev/null @@ -1,18 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`doctor --provider github|gitlab|gitcafe|forgejo|bitbucket|azure` preflights one -host instead of always assuming GitHub: the same enum and `bad-provider` -refusal as `fetch --provider`, defaulting to `github` so a bare `doctor` keeps -its rows byte for byte. Each host has its own probe table — `gh` and its scope -rows for GitHub, `glab` presence, `glab auth status`, and a `glab api user` -reachability read for GitLab, the `cafe` CLI for gitcafe, and `curl`/`az` plus -the credential variables their fetch paths read (`FORGEJO_TOKEN`, -`BITBUCKET_TOKEN`, `AZURE_DEVOPS_EXT_PAT`, …) for the rest, reported by presence -and never by value. A non-GitHub preflight shows no `gh` row. - -`adopt` now hands local findings to any bundle verdict can post to, so a GitLab -bundle adopts exactly like a GitHub one; its re-anchor reads the stored hunks -and shells no host CLI. Bundles on a host with no submit path still refuse -`provider-unimplemented`, with the host named in the message. diff --git a/.changeset/fetch-index-impact.md b/.changeset/fetch-index-impact.md deleted file mode 100644 index 1826e8c64..000000000 --- a/.changeset/fetch-index-impact.md +++ /dev/null @@ -1,28 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Ship the index hand-off. `fetch` and `fetch-local` take -`--index `: the operator names an index they already built — a -revision in the reviewed repository, or a directory holding `.verdict/index` — -and the bundle gains an `impact` artifact and a `## impact` section on -`toc.txt`, inside the existing nonce fence. Nothing is built, refreshed, or -discovered: without the flag no section is written and an earlier run's -`impact.json` is removed. - -`impact.json` (named in `data.artifacts`) holds the answered tree and commit, -`stale`, `coverage`, the diff's changed `paths`, the row `limit`, and one -measured object per list — `symbols`, `callers` (with `unresolved`), and -`tests` (with `depth_truncated`) — each carrying `count`, `truncated`, and -`rows`; a list the index could not answer reads `null`, never `0`. `sources` -carries `index:`, naming the tree that answered. The `toc.txt` section is -capped at 8 KiB and ends with `truncated: line(s) omitted to fit the 8 KiB -impact cap` when rows do not fit, and every path and symbol it renders is -escaped through the shared fence rule, so a hostile name can neither close the -fence nor start a line of its own. - -A directory with no index store refuses `bad-flags`, an unresolvable ref -`bad-at`, a tree the store never built `missing-index`, and a truncated tree -file `index-corrupt`; a named index answering for another tree, and a bundle -head this clone cannot see, warn visibly and proceed. The four hashed prompt -files are untouched, so `prompts_hash` does not move. diff --git a/.changeset/fetch-refusal-next-steps.md b/.changeset/fetch-refusal-next-steps.md deleted file mode 100644 index 0ceffe46c..000000000 --- a/.changeset/fetch-refusal-next-steps.md +++ /dev/null @@ -1,14 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Make two fetch refusals name their next step. An over-cap diff refuses -`diff-too-large` with a `next:` line naming the exclusions that already fit a -review — `--ignore ` (repeatable on `fetch` and `fetch-local`), a -committed `.verdict/ignore` line, or paths after `--` on `fetch-local` — because -the cap is measured after them; the cap itself is unchanged. A tree the index -store never built refuses `missing-index` with the revision the operator typed -beside the tree it resolved to (`no index for HEAD (tree )`) and the exact -build command that answers for it (`next: verdict index build --at HEAD`, or -`--from ` for a directory hand-off). Refusal codes, exit codes, and the -`--json` envelope keep their shape; only the message text is longer. diff --git a/.changeset/fetch-reverts.md b/.changeset/fetch-reverts.md deleted file mode 100644 index 10138e50f..000000000 --- a/.changeset/fetch-reverts.md +++ /dev/null @@ -1,33 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Give a review the revert memory that belongs to its diff. When a changed path -matches a row of the tracked `.verdict/reverts.jsonl` ledger (`verdict backtest -reverts`), `fetch` and `fetch-local` write -`.verdict/state//reverts.txt`: one `<<>>` block per matching -revert inside `internal/untrusted`'s fence, with the machine facts and the -author's paths first and the author's subject last, so a cap can only ever cut -subject text. A match is the same path, or a row path that is a directory -prefix of it; rows are read through one validating reader -(`internal/reverts`) and sorted by `sha`, so the file is a function of the -ledger and not of its line order. - -The note is bounded at 20 reverts or 64 kB, and the truncation line names the -cap it hit and points at `.verdict/reverts.md` for the rest. It joins the -bundle contract: `data.artifacts["reverts.txt"]`, the `--show-prompts` dump, and -`provenance.json`'s `bundle` inputs all carry it, so the receipt records its -bytes without copying an author's text. - -Memory is context, never orders. `prompts/review.md` gains one clause beside -`kb.txt` telling the reviewer to read the file when `fetch` reports it and to -treat a revert as evidence the change broke something once — the prompt-set hash -moves once, to -`71e81dd0a06320e6252cd6ddc49e0d9591a61684a254568a653a7b338bcb1857`, and the -record and fetch goldens follow it. - -Nothing is a gate. No intersection removes the file, an absent ledger is -silence, and a ledger this build cannot read warns once naming the store and the -field defect and writes nothing: `fetch` only reads the ledger, so a scan -holding its lock never blocks a fetch, and no revert byte reaches `review.json` -or a post. diff --git a/.changeset/fix-all-batch.md b/.changeset/fix-all-batch.md deleted file mode 100644 index bd1cdcdfb..000000000 --- a/.changeset/fix-all-batch.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `fix --all`, the batch form of the bounded fix loop: it selects every open inline finding at or above `--min-severity` (default `important`) through the same selector `report` and `next` rank with, takes one pass per sig under the unchanged 3-passes-per-sig and one-file ceilings, and names each sig's outcome (`ready`, `ceiling`, `scope-escape`, `no-anchor`) in `data.batch`. `--paths ` (repeatable) scopes a batch to matching anchor paths. A batch refuses `no-verify-cmd` unless the proving command is supplied by `--verify-cmd` or the bundle's `patch.json` `verifyCmd`. Every pass now also writes `fix.patch` beside the ledger — the worktree's unified diff against `HEAD`, never applied and never pushed — and refuses `fix-patch-secret` rather than write a credential shape into the bundle. diff --git a/.changeset/forgejo-post.md b/.changeset/forgejo-post.md deleted file mode 100644 index 009fa49b0..000000000 --- a/.changeset/forgejo-post.md +++ /dev/null @@ -1,15 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`post` publishes to Forgejo pull requests over the REST API: one atomic review -carrying the body, the verdict, the reviewed commit, and every inline finding, -with `APPROVED` for an approval. A review comment anchors on a file line -number (`new_position` on the new side, `old_position` on the old), so a -finding with no numeric line is withheld, a LEFT anchor on a file the change -deletes is withheld, and a multi-line range lands on its end line; each loss is -counted in the envelope `warnings`. A 200 that answers `PENDING` refuses -`post-failed` instead of recording a receipt for a draft nobody sees, and a -pre-existing draft authored by the token user refuses with a submit-or-discard -hint. `data.capabilities` now reports Forgejo's post, inline anchors, and both -binding verdicts. diff --git a/.changeset/gate-coverage-default.md b/.changeset/gate-coverage-default.md deleted file mode 100644 index aa7d702cd..000000000 --- a/.changeset/gate-coverage-default.md +++ /dev/null @@ -1,30 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Gate on diff coverage by default. `verdict gate` gains a `coverage` dimension -that reads the receipt `post.json`/`record.json` already store for the gated -HEAD and fails when that receipt is absent, names another HEAD, carries no -`coverage[]` attestation (a review that never attested coverage reads -`unknown`, not `pass`), or attests under 100% — the reviewed ranges are not -re-evaluated, so the gate reports the number the store holds. `--coverage -require|warn` sets the posture and defaults to `require`; `--coverage warn` -reports the same shortfall as one `warn:` line and never fails the gate: it is -how to keep the old behavior and the migration while a review attests nothing. An all-exempt -diff passes once a receipt exists, a chunked bundle is graded on the chunk -attestation its receipt stores, and `--coverage` alone is still a usage error. -`post` and `record` keep the warning posture (`--require-coverage` opts into -the refusal), the SDK's `GateRequest` gains the same `Coverage` field with the -same default, and the published envelope schema adds `coverage` to the gate -dimension ids. - -A chunked bundle is graded too, because `chunks[]` is its only attestation and -nothing stored that fact. `post.json` and `record.json` gain -`chunk_coverage: {attested, total}`, written beside the existing measured -summary at the point the merge already validates `chunks[]`: the manifest is -the denominator, and only a claim the manifest can name counts toward the -numerator, so the receipt can never attest more than the review wrote. Every -chunk attested passes; a shortfall fails naming the count (`1 of 2 chunk(s) -unattested`); a receipt written by an older binary carries no such key and -still reads `unknown` under `require`. Non-chunked receipts are unchanged byte -for byte. diff --git a/.changeset/gate-live-all-hosts.md b/.changeset/gate-live-all-hosts.md deleted file mode 100644 index 3805a4eb9..000000000 --- a/.changeset/gate-live-all-hosts.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`gate --live` reads threads on every host and fails closed when it cannot. The -live-thread dimension skipped the read for any bundle whose provider was not -GitHub — it warned `live recheck skipped: threads unreadable here` -and passed the gate without looking, so `--live strict` on a GitLab, GitCafe, -Forgejo, Bitbucket, or Azure bundle was a pass that proved nothing. The probe -now reads the bundle's own host through the same per-host adapter the `threads` -refresh uses (GitHub keeps its resolution-only `gh api graphql` read, unchanged -byte for byte), counts unresolved threads the same fail-closed way, and treats a -read that stopped before the host's last page as unread rather than clean: -`--live auto` warns and gates on the local dimensions, `--live strict` fails -closed with the existing strict refusal. Only local bundles skip the lookup now, -because they carry no host threads at all. diff --git a/.changeset/gate-tier-validation.md b/.changeset/gate-tier-validation.md deleted file mode 100644 index 233e36018..000000000 --- a/.changeset/gate-tier-validation.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -An unknown gate tier now refuses `bad-flags` everywhere, not only on the CLI. The MCP `report` tool's `flags.check` -and the Go SDK's `Report.Check` and `Gate.MaxSeverity` used to pass any string through, so `check: "bogus"` gated on -every finding and could flip an informational-only bundle from passing to failing. The CLI keeps its existing -`--check` / `--max-severity` usage error. diff --git a/.changeset/github-action.md b/.changeset/github-action.md deleted file mode 100644 index d1bb1e991..000000000 --- a/.changeset/github-action.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Package the review and merge gate as a first-party composite GitHub Action. -The runner is your CI — pass your own agent and keys, install a pinned verified -verdict release, and reuse the pipeline through the shipped example workflow. diff --git a/.changeset/gitlab-incremental.md b/.changeset/gitlab-incremental.md deleted file mode 100644 index e0776e277..000000000 --- a/.changeset/gitlab-incremental.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Incremental re-review on GitLab. `fetch --provider gitlab --base-sha ` -(or `--since`) calls `glab api -projects//repository/compare?from=&to=`, joins the response's -per-file `diff` entries into `incremental.diff`, and names the base in -`incremental_base` — the same artifact, note, and `.compare.err` contract the -GitHub compare path already writes. A failed or unreadable compare still -finishes the fetch and writes `incremental.diff.note` instead of refusing. -Hosts without a compare adapter (GitCafe, Forgejo, Bitbucket, Azure) now accept -`--base-sha`/`--since` and write the note naming the unsupported provider -rather than refusing the fetch. The delta feeds the ordinary record/advance -path, so a second review round reopens delta-touched findings and records at -the fetched head. diff --git a/.changeset/gitlab-publish.md b/.changeset/gitlab-publish.md deleted file mode 100644 index ac875f389..000000000 --- a/.changeset/gitlab-publish.md +++ /dev/null @@ -1,9 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`post` publishes to GitLab merge requests: each finding becomes one positioned -discussion, the review body becomes a summary note, and an approval verdict -becomes a real MR approval. `capabilities` reports GitLab's approval and -request-changes support separately, and a request-changes verdict on GitLab -warns that the host has no such event instead of silently pretending. diff --git a/.changeset/handoff-verb.md b/.changeset/handoff-verb.md deleted file mode 100644 index d40722e12..000000000 --- a/.changeset/handoff-verb.md +++ /dev/null @@ -1,27 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verdict handoff [--min-severity ] [--paths ]... -[--format md] [--json]` prints one markdown projection of the open inline -findings `fix` would take, for a human or a different agent to pick up where -`fix` stops. It selects through the same selector `fix --all` uses, so the two -verbs cannot pick different work, and each row carries the full sig, the -`path:line` locator, the severity and rigor rung, the fix text and code -suggestion, the cited KB rule ids, the verify command the bundle records, and -the `brief` block `fix` prints — the identical bytes, because both verbs render -it from one function. - -`--json` carries the same rows and the same `brief_text`; the markdown and the -envelope also carry one caution line, before any quoted prose, that finding text -quotes PR content an attacker controls. `--min-severity` defaults to -`important` (`off` for no floor) and repeatable portable `--paths` globs narrow -the projection. Nothing open at the floor refuses `no-open-findings`, naming the -floor and the globs. `--format md` — the default — also beats a -`REVIEW_OUTPUT_FORMAT=json` exported for the whole job, because flags outrank -env; passing both `--json` and `--format md` is a usage error rather than a -silent winner. - -The verb is read-only: no ledger, no patch, no store write, no host call. The -MCP `handoff` tool serves the same `data` and is declared non-mutating; the Go -facade gains `Client.Handoff` and the TypeScript client a `handoff` method. diff --git a/.changeset/honest-capabilities.md b/.changeset/honest-capabilities.md deleted file mode 100644 index 5e2d2cfe4..000000000 --- a/.changeset/honest-capabilities.md +++ /dev/null @@ -1,21 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Make `data.capabilities` honest. The binding verdict splits into `approve` and -`request_changes`, because hosts differ: GitHub carries both, while GitLab -approves a merge request but has no request-changes review event — a bound -request-changes there still posts the discussions and summary note, and says in -a warning that the verdict did not bind. `review_verdicts` stays in place as -the compatibility sum `approve && request_changes`, so a consumer written before -the split reads the same truth; new readers branch on the two flags. -`thread_resolution` reads true on GitHub and GitLab, where `threads resolve` -closes a verdict-authored thread, and false on Forgejo, which has no documented -resolve endpoint, and on the fetch-only hosts. The new `unsupported` object -carries one machine-readable -reason code per absent capability (`fetch-only`, `no-request-changes-event`, -`no-approve-verdict`, `no-inline-anchors`, `no-resolve-mutation`, -`unknown-host`). The published envelope schema types `capabilities` instead of -leaving it open; only `provider` and `fetch` stay required, so an envelope -written before the split still validates while every new fetch emits the full -object. diff --git a/.changeset/ignore-overrides.md b/.changeset/ignore-overrides.md deleted file mode 100644 index 9f8c23eeb..000000000 --- a/.changeset/ignore-overrides.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add repeatable fetch ignore overrides and a switch to disable built-in exclusions, recording both in exclusion receipts. diff --git a/.changeset/ignore-suppression.md b/.changeset/ignore-suppression.md deleted file mode 100644 index 85a111f4c..000000000 --- a/.changeset/ignore-suppression.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Record receipts for excluded diff bytes and suppressed findings. Add post severity floors and show quiet findings in reports and trend counts. diff --git a/.changeset/incident-memory.md b/.changeset/incident-memory.md deleted file mode 100644 index b52e2e1ef..000000000 --- a/.changeset/incident-memory.md +++ /dev/null @@ -1,38 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verdict learn --incident --sha --path ` attaches one -revert or incident to a recorded bundle, writing -`.verdict/incidents.jsonl` at the repo root: one row per `(bundle, sha)` pair -with `{schema_version, bundle, sha, paths, note, url, detected_at}`, kept -sorted by bundle then sha, where `bundle` is the resolved bundle directory's -name (an alias records the same pair as the full name). `--path` repeats (sorted-deduped), `--note` caps at -2,000 bytes, `--url` takes `http(s)` only, `--at` defaults to now, and the -bundle must already carry `findings.json`. The attach is idempotent — the -second run reports `existing`, moves no byte, and answers with the stored row — -and `--dry-run` reports what the write would do, writing nothing. - -`verdict trend --incidents` joins that ledger to the bundles' findings stores -and reports `incidents`, `measured`, `unmeasured`, `hit`, `missed`, `miss_rate` -(`null` when nothing measured), and per-bundle rows: an incident is a hit when -one non-`false-positive` finding covers one of its paths, exact or as a -directory prefix. A bundle whose store will not read is unmeasured rather than -a miss; a corrupt ledger refuses `corrupt-store`; no ledger reads a zero tally. -The unflagged output stays byte-identical. - -The flag joins the CLI, MCP (`learn` gains `incident`/`sha`/`path`/`note`/ -`url`/`at`, `trend` gains `incidents`), Go SDK (`pkg/verdict`), the TypeScript -client (`clients/ts`), and the published JSON Schemas -(`learnIncident`, `trendIncidents`, `trendIncidentBundle`). - -Attaches are serialized by an OS advisory lock on -`.verdict/incidents.jsonl.lock`: a second attach waits up to five seconds and -refuses `store-busy`, while a crashed holder releases the lock at once. The -lock file persists and is ignored — the tracked ignore contract -(`.verdict/.gitignore`, written by `init` and kept current by the attach) gains -the `incidents.jsonl.lock` line. - -An attach that would push the ledger past the 8 MiB the reader opens refuses -`ledger-too-large` and writes nothing, instead of leaving a file no later read -(attach or `trend --incidents`) can open. diff --git a/.changeset/index-blast-radius.md b/.changeset/index-blast-radius.md deleted file mode 100644 index 0482b2e03..000000000 --- a/.changeset/index-blast-radius.md +++ /dev/null @@ -1,19 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `index blast-radius --pr |--paths ` — one change's reach. The -changed declarations come from the fetched bundle's `full.diff` (no host call) -or from named files; their callers are walked to depth 2, once per target and -cycle-safe; and candidate tests, co-change partners, and CODEOWNERS owners ride -along for the same paths. A diff section with no line hunks — a rename or a -mode change — reads as every declaration in its new path and warns that it did. -A `++` or `--` line inside a hunk is a content line, so its changed line and the -hunk's anchors stay put. A depth-2 caller names the changed declaration it is -reached from, not the depth-1 caller in between. A `--pr` that holds a path -separator or is `.` or `..` refuses as `bad-flags`, since `--pr` names a bundle. -Every list publishes its count and truncation count, `co_changed` publishes its -distinct partner count beside the row count, and an absent index reads the -symbol, caller, and test counts as `null`, never 0. The read-only MCP `index` -tool serves the new query in both its `--paths` and `--pr` forms, and the -generated envelope schema types its output for the triage risk score. diff --git a/.changeset/index-callers.md b/.changeset/index-callers.md deleted file mode 100644 index 01757e5be..000000000 --- a/.changeset/index-callers.md +++ /dev/null @@ -1,21 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `verdict index`: a local Go code index built with the standard library -(`go/parser`, `go/ast`) and stored under `.verdict/index/` by tree SHA. -`index build [--from ] [--at ]` writes one tree file plus -content-addressed shards, so an empty commit reuses the tree, a pure rename -parses nothing, and an edit parses one blob. `index status` reports the stored -trees and shards, `index prune` keeps the newest four trees and the shards they -reference, and `index callers [--at ] [--limit ]` returns -call sites labelled `exact` within a package or `lexical` across an import -edge; method values, cross-package value selectors that may be fields, dynamic dispatch, reflection, and generic instantiation -are counted as unresolved, never guessed. A package qualifier that matches no -package directory or import path refuses `bad-flags` rather than reading as no -callers. Caps -refuse `index-capped` with the measured number, a truncated shard refuses -`index-corrupt`, and a query against a tree that is not HEAD's warns and sets -`stale: true`. The MCP `index` tool is read-only and serves `callers` and -`status`; the SDK forwards every subcommand. `.verdict/index/` joins the -tracked ignore contract. diff --git a/.changeset/index-corpus-bugs.md b/.changeset/index-corpus-bugs.md deleted file mode 100644 index 5d831c9c4..000000000 --- a/.changeset/index-corpus-bugs.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -The index accuracy corpus's four scanner bugs are fixed, so the published -scorecard and the mechanism labels now match the manual. `index tests-for` -names a test whose own package imports the declaring package at -`transitive-import, depth 1` instead of dropping it, and `imports-module` -resolves a relative specifier (`import { tidy } from "../lib"`) against the -importing file's directory — including in a tree that also carries a `go.mod`, -where a scanned package's import path is now the directory rather than the Go -module path. `index callers` no longer attributes a Python attribute call -(`log.save(...)`) to a same-named module-level function, and a Go selector -written on a value outside the declaring package (`s.Save`, `a.Touch`) is -counted in `counts.unresolved` instead of disappearing, so an empty answer no -longer claims nothing was left unattributed. diff --git a/.changeset/index-graph.md b/.changeset/index-graph.md deleted file mode 100644 index 16b7a8e8d..000000000 --- a/.changeset/index-graph.md +++ /dev/null @@ -1,19 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Extend `verdict index` with three graph queries. `index tests-for -|--symbol [--at ] [--limit ]` lists candidate tests -labelled `same-package`, `references`, or `imports-module` — tests worth -reading, never coverage claims; only `_test.go` files appear. `index co-changed - [--commits ] [--since ] [--min-count ] [--limit ]` -reads one bounded `git log --no-merges` window (merges excluded, renames as two -paths), reports partner paths with co-change counts and four-decimal Jaccard -ratios, echoes the window it read, warns on a shallow clone, and caches the -window under `.verdict/index/cochange/` (two newest, timestamp-free bytes). -`index owners ` answers the CODEOWNERS rule at the tree — GitHub's file -order, last matching rule wins, the documented glob subset only — with -`owners: null` when nothing matches instead of an owner named nobody. The same -`target` object (path, owners, source) rides on `tests-for` and `co-changed`. -The read-only MCP `index` tool serves the three new queries and publishes their -query enum; the SDK forwards them. diff --git a/.changeset/index-impact-budget.md b/.changeset/index-impact-budget.md deleted file mode 100644 index c50e0465d..000000000 --- a/.changeset/index-impact-budget.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Share the impact section byte budget across symbols, callers, and tests, with omitted counts for each list. diff --git a/.changeset/index-import-shadow.md b/.changeset/index-import-shadow.md deleted file mode 100644 index ce2b71261..000000000 --- a/.changeset/index-import-shadow.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Suppress module-function callers through import qualifiers shadowed by local bindings in Python and TypeScript. Rebuild indexes written with the previous scanner. diff --git a/.changeset/index-scanners.md b/.changeset/index-scanners.md deleted file mode 100644 index baf144d8b..000000000 --- a/.changeset/index-scanners.md +++ /dev/null @@ -1,35 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Scan TypeScript, JavaScript, and Python in `verdict index`. A build now reads -`.ts`, `.tsx`, `.mts`, `.cts`, `.js`, `.jsx`, `.mjs`, `.cjs`, `.py`, and -`.pyi` alongside Go, so `index callers`, `index tests-for`, and `index status` -answer for those files too. Go stays parsed with `go/parser`; the other three -are read by two dependency-free lexical scanners — no tree-sitter, no cgo, no -third-party module — that tokenize the file and record declarations, imports, -and call sites. Every scanned file publishes its `language`, a build publishes -`data.languages` with one row per scanned language and how many of the indexed -files it holds, and a caller row names the language it came from. - -A scanned language resolves one way only: `exact` never appears, and a name is -attributed as `lexical` inside the directory that declares it, so a caller list -for a scanned symbol is a floor rather than a proof. Languages never cross: a -caller row, and a test candidate — however it was found — must share the -declaration's language family, so Go, Python, and the JavaScript family each -answer only for themselves. - -Test candidates follow the language's own naming — `.test.` and `.spec.` names -and `__tests__` directories for the JavaScript family, `test_*.py`, -`*_test.py`, `conftest.py`, and `tests`/`test` directories for Python — and a -test in another directory is a candidate only when it imports the queried file -by a specifier that resolves to it, `from lib import mod` included. A blob two -languages both claim is parsed per file and cached under neither, so a shared -file never borrows another language's parse and every build counts those parses -as parses rather than reuse. - -A file over the 1 MiB cap stays in the tree with `oversize: true` and counts as -skipped rather than indexed; an extension the index does not scan never enters -the tree, so it is not counted at all. A store written by an earlier build -(store version 2 now) reads as unbuilt in `index status` and is rebuilt once by -the next `index build`. diff --git a/.changeset/index-transitive-tests.md b/.changeset/index-transitive-tests.md deleted file mode 100644 index d8e32a65b..000000000 --- a/.changeset/index-transitive-tests.md +++ /dev/null @@ -1,20 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Reach candidate tests through the import closure. `index tests-for -|--symbol [--at ] [--depth ] [--limit ]` walks the -reverse import graph from the declaring package — breadth first, cycle safe, -`--depth` hops (default 3, maximum 8; 1 limits the walk to direct imports) — and -labels tests reached through their package's imports as `transitive-import`, -alongside `same-package`, `references`, and `imports-module`. Each row carries -the import hops its own imports took, or its package's when its own imports reach -nothing, so a sibling test that imports the target directly cannot hide it — and -a direct candidate at depth 1 never reads as a transitive one. `depth_truncated` counts -the candidates whose own chain is one hop past the cap. The walk uses each -language's own import data (Go import paths, TypeScript -and JavaScript specifiers, Python modules and the submodule a `from … import …` -names) and never crosses a language family. `index blast-radius` reuses it, so -its candidate-test list gains transitive rows and a `tests.depth_truncated` -count of the candidates its fixed depth cap hid; the read-only MCP `index` tool -takes `depth` too. diff --git a/.changeset/ingest-sarif.md b/.changeset/ingest-sarif.md deleted file mode 100644 index 415a9a2e6..000000000 --- a/.changeset/ingest-sarif.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verdict ingest --sarif ` ingests SARIF 2.1.0 results -into the bundle's `findings.json` store with `source: "tool"`, sharing one -store, one dedupe path (`verdict/sig`, `sig`, and `(tool, fingerprint)`), and one -suppression/withheld path with model reviews. Severities map through -`skills/verdict/schemas/sarif-map.json`, out-of-diff results stay -`anchored: false`, excluded paths land withheld (`source: "exclusion"`), and tool -findings keep `rigor: "said"` (`rigor_source: "default"`). - -`report` and `trend` publish `by_source` (`model` and `tool`) breakdowns, -`ingest` joins the CLI, MCP default tool registry (`mutates: localWrites`), Go -SDK (`pkg/verdict`), TypeScript client (`clients/ts`), and published JSON -Schemas. diff --git a/.changeset/inventory-budget-exact.md b/.changeset/inventory-budget-exact.md deleted file mode 100644 index de4457d11..000000000 --- a/.changeset/inventory-budget-exact.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Keep serve inventory accounting within its byte budget even for oversized artifacts. diff --git a/.changeset/learnings-evidence.md b/.changeset/learnings-evidence.md deleted file mode 100644 index 84f1b04f7..000000000 --- a/.changeset/learnings-evidence.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Keep cited paths and comment identifiers inside quoted learning evidence. -Refuse unsafe paths and identifiers before proposing or accepting a learning, -including proposals generated before this validation was added. diff --git a/.changeset/learnings-propose.md b/.changeset/learnings-propose.md deleted file mode 100644 index d58d9f934..000000000 --- a/.changeset/learnings-propose.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Propose inactive one-rule packs from normalized resolved review threads, accept -them through the existing pack importer, or reject them permanently with a human -reason. All three lifecycle modes support a write-free dry-run. Accepted rules -enter scoped review recall; quoted thread evidence never activates itself. diff --git a/.changeset/lens-dispatch-deltas.md b/.changeset/lens-dispatch-deltas.md deleted file mode 100644 index f53cfafef..000000000 --- a/.changeset/lens-dispatch-deltas.md +++ /dev/null @@ -1,21 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Drop the per-lens findings cap and lead the dispatch with its domain. A lens -now returns every finding its evidence supports: the built-in pack declares -only `max_files: 12` (so `lenses.txt` headers read `max_files=12`, no -`max_findings=`), and all three bodies ask for "at most ~12 files" without a -return count. Aggregation already decides what survives, so the cap only -discarded evidence the lens had paid for. The shared dispatch template in -`prompts/reviewer-subagent.md` gains one leading line, `Domain: `, matching the framing `prompts/review.md` and -`prompts/incremental-review.md` already ask for. `max_findings` stays a valid -pack and recipe declaration (1..20); no shipped pack sets one. - -`prompts/reviewer-subagent.md` is not in the hashed prompt set, so -`prompts_hash` does not move. The pack hash moves from -`06820abbd0a6…` to `120173839a93…` and the lens hashes from -`api-contract:2aa23d13`, `perf:67c1d05a`, `test:456be8d2` to `62c9ccf8`, -`6225fd98`, `efd2eb4e`; the `fetch` goldens and the two eval literals that pin -them follow. diff --git a/.changeset/lens-packs-s1.md b/.changeset/lens-packs-s1.md deleted file mode 100644 index e4ae438ca..000000000 --- a/.changeset/lens-packs-s1.md +++ /dev/null @@ -1,17 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Ship the built-in lenses as a pack. The three lenses `prompts/review.md` step 4 -names today — `perf`, `api-contract`, and `test` — now live as data in -`skills/verdict/packs/builtin/`: a strict JSON `pack.json` (published as -`schemas/pack.schema.json`) plus one markdown body per lens. `fetch` and -`fetch-local` resolve the embedded pack into every bundle as `lenses.txt` -(each body delimited and labelled with its id and eight-hex hash), list it in -`data.artifacts`, and append `lens_pack=@` and -`lenses=[:,…]` to each `plan.txt` group line. A lens whose globs -match none of a group's files is not offered for that group. The loader is -capped (64 KiB/file, 1 MiB/pack, 256 files, 64 lenses), refuses `bad-pack`, -`pack-too-large`, and `pack-unsafe-path` (absolute paths, `..`, symlinks), and -hashes the canonical tree. The four hashed prompt files are byte-identical, so -`prompts_hash` does not move. diff --git a/.changeset/line-break-guards.md b/.changeset/line-break-guards.md deleted file mode 100644 index dc0ad8b7a..000000000 --- a/.changeset/line-break-guards.md +++ /dev/null @@ -1,11 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Refuse Unicode line breaks where only C0 controls were refused before. A URL -path, a learning's cited path, and a rendered `pack test` case line each tested -`unicode.IsControl`, which excludes U+2028 and U+2029 — runes a renderer treats -as a newline. Those characters now refuse the `serve` URL before routing -(`404 no such route`), refuse a learning path as `bad-scope` with the message -naming line breaks, and flatten to `�` in the `pack test` line. Parent -components, glob syntax, and every earlier refusal are unchanged. diff --git a/.changeset/managed-parent-symlinks.md b/.changeset/managed-parent-symlinks.md deleted file mode 100644 index e423d2a87..000000000 --- a/.changeset/managed-parent-symlinks.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Refuse symlinks at every managed bundle component before writing, including links to targets inside the repository. Symlinked repository roots remain supported. diff --git a/.changeset/mcp-config-hosts.md b/.changeset/mcp-config-hosts.md deleted file mode 100644 index 98586e9d1..000000000 --- a/.changeset/mcp-config-hosts.md +++ /dev/null @@ -1,18 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verdict mcp config` now covers six more hosts and the server's posture flags. -`--read-only` or `--operator` (mutually exclusive, like `verdict mcp`) puts the -matching argv after `mcp` in the -printed, `--json`, and `--write` snippets. A `verdict` entry whose argv is one -of verdict's own shapes (`["mcp"]` plus any combination of the two flags) is -updated in place on `--write` with the usual backup; any other argv still -refuses typed (`mcp-config-foreign`). - -New JSON `mcpServers` hosts merge with `--write`: `gemini` -(`~/.gemini/settings.json`), `qwen` (`~/.qwen/settings.json`), and `windsurf` -(`~/.codeium/windsurf/mcp_config.json`). `opencode` (`opencode.json`), -`copilot` (`.vscode/mcp.json`), and `goose` (`~/.config/goose/config.yaml`) -are print-only: the snippet renders each host's own shape and `--write` -refuses typed naming the host. diff --git a/.changeset/mcp-mutation-class.md b/.changeset/mcp-mutation-class.md deleted file mode 100644 index 89cc4a07b..000000000 --- a/.changeset/mcp-mutation-class.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Declare every MCP tool's local mutation class explicitly and derive read-only -exposure from it. Calls hidden by `--read-only` now return the typed -`read-only-tool` tool refusal with a remediation hint. diff --git a/.changeset/mention-neutralization.md b/.changeset/mention-neutralization.md deleted file mode 100644 index 048bb017d..000000000 --- a/.changeset/mention-neutralization.md +++ /dev/null @@ -1,23 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Neutralize `@`-mentions in the review body and in each finding's inline -comment. Both are written by an agent and published verbatim, so an `@handle` -in that prose pinged that person the moment the review was submitted. `record` -now escapes every mention once, where the text is constructed into the record — -the same shared merge `post` runs — so the recorded body, each finding's posted -comment, the `post --dry-run` output, and the payload digest carry the same -bytes instead of diverging at submit time. A finding's signature is computed -over the escaped bytes, so merging the same bundle twice keeps one signature. - -The escape is U+200B ZERO WIDTH SPACE between the `@` and the handle: invisible -in rendered Markdown, still plain text, and no host resolves a mention across -it. Fenced code blocks, inline code spans, HTML comments (finding markers -included), and the basis/footer blocks are left byte-identical, and an address -such as `user@example.com` is not a mention because the byte before the `@` can -sit in an address's local part while the handle must be `[A-Za-z0-9-]+`. A -finding's -`suggestion` is code and stays byte-identical too — a zero-width space inside a -patch would be an edit nobody can see. The rewrite is idempotent and has no -flag: neutralization is not opt-in. diff --git a/.changeset/next-usage-basis-plain.md b/.changeset/next-usage-basis-plain.md deleted file mode 100644 index 609a1af02..000000000 --- a/.changeset/next-usage-basis-plain.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verdict next --coverage` with no positionals now leads with one `error:` line -that names what is missing, like the other verbs. A plain `Basis:` line directly -under the quoted `Confidence:` line is accepted without a `basis-replaced` -warning. diff --git a/.changeset/null-ranges-malformed.md b/.changeset/null-ranges-malformed.md deleted file mode 100644 index 9c5776093..000000000 --- a/.changeset/null-ranges-malformed.md +++ /dev/null @@ -1,13 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -A `ranges: null` coverage claim is malformed, not whole-file. The attestation -reader treated a present-but-null `ranges` as "no ranges given", which is the -same shape as an omitted key, so it counted the claim as the file's whole hunk -set: `record --dry-run --require-coverage` reported the diff 100% attested and -the gate passed on a claim that carried no ranges at all. A `ranges` key that is -present and is not a string — `null` included — now warns once and counts as -unattested, exactly like a reversed or unparsable range. A claim still widens -only by saying `"all"` or by omitting the key, so a malformed claim can never -attest more than it says. diff --git a/.changeset/pack-add-backup-warning.md b/.changeset/pack-add-backup-warning.md deleted file mode 100644 index d5eca05ae..000000000 --- a/.changeset/pack-add-backup-warning.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verdict pack add` reports a failed sweep of the previous tree as success plus a warning -that names the leftover dot-prefixed sibling, instead of refusing an add whose new tree -and lock are already durable. The loader already skips that sibling. diff --git a/.changeset/pack-layers.md b/.changeset/pack-layers.md deleted file mode 100644 index 1fada600d..000000000 --- a/.changeset/pack-layers.md +++ /dev/null @@ -1,27 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Layer lens packs and list them. `verdict pack list [--json]` reports every -loaded pack (built-in, org from the personal settings `packs` array, repo from -`.verdict/packs//`) with its layer, source, canonical tree hash, lens -and rule ids, and drift (null until `verdict.lock`). Repo packs are read from -the trusted-base controls snapshot, so a PR cannot edit the pack that reviews -it; `post`/`record` re-hash the snapshot and refuse `controls-stale`. Duplicate -ids in one layer refuse `bad-pack`; a cross-layer override needs -`"replace": [""]` (downward only) and otherwise refuses -`pack-conflict`. A layered set tags each lens in a `plan.txt` group line with -the full name of the pack that supplied it (`lenses=[@:,…]`), -so two packs sharing a lens id stay distinguishable on the line. A settings -`packs` entry inside `.verdict/packs/` refuses -`bad-pack`, because that tree is the repo layer read at the PR base. Packs may -ship KB rules (`rules/.md`, severity mapped through the manifest) that -join the fenced `kb.txt` context, never the trusted prompt. Manifest keys must -be exact lowercase, non-regular pack files and a symlinked pack root refuse -`pack-unsafe-path`, and body reads are bounded. A regular file directly under -`.verdict/packs/` is not a pack: `pack list` warns once and `fetch` ignores -it. A symlink or special file under `.verdict/packs/`, or a `.verdict/packs` -that is not a real directory, refuses naming the joined path — `pack list` -as `pack-unsafe-path`, `fetch` as `controls-unavailable` — so neither read -calls one ignorable. The read-only `pack` MCP tool, `Client.Pack`/`ListPacks`, -and the TypeScript `pack` method serve the same list. diff --git a/.changeset/pack-source-allowlist.md b/.changeset/pack-source-allowlist.md deleted file mode 100644 index 587f6ce8c..000000000 --- a/.changeset/pack-source-allowlist.md +++ /dev/null @@ -1,20 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verdict pack add` accepts only the three published git source forms — `https://host/path` -with no userinfo, `ssh://user@host/path`, and `user@host:path` — and refuses every other -remote-shaped string before any git call or write. The grammar is one allowlist shared by -`pack add` and `.verdict/verdict.lock`, so the one-slash scheme, `%`-escaped passwords, -token-only and two-`@` https URLs, `http://`, `git+https://`, IPv6 literals with -userinfo, and a query or fragment (`?private_token=…`, `#…`) all refuse as -`bad-pack-source`, and a hand-written lock row carrying one refuses -as `bad-lock-source` at `pack list`, `fetch`, `post`, and `record`. No source is echoed -raw: every refusal, human line, and `pack list` row renders through one display that -prints a parsed source and `` for anything else. git's own failure text is -never rendered: a failed clone, rev-parse, or checkout refuses as `pack-git-failed` with -fixed wording for the phase and its exit code, because that text can carry the source URL, -its query, or the ref. -The published lock pattern follows the same grammar: it bounds each DNS label at the 63 -bytes the loader enforces and refuses a `://` inside an scp path, so a schema check and -`pack list` agree on every source shape. diff --git a/.changeset/pack-staging-stale.md b/.changeset/pack-staging-stale.md deleted file mode 100644 index 16d2217d9..000000000 --- a/.changeset/pack-staging-stale.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verdict pack add` refuses a crashed re-add's leftover staging or backup sibling as -`pack-staging-stale` when it cannot sweep it, naming the path to delete instead of -suggesting a `pack.json` repair. The loader still ignores those dotted siblings, so -`pack list` and `fetch` are unaffected. diff --git a/.changeset/pack-test.md b/.changeset/pack-test.md deleted file mode 100644 index 777619ce2..000000000 --- a/.changeset/pack-test.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `verdict pack test` to run offline lens-pack eval cases. diff --git a/.changeset/packs-lock.md b/.changeset/packs-lock.md deleted file mode 100644 index 783a2d356..000000000 --- a/.changeset/packs-lock.md +++ /dev/null @@ -1,35 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `verdict pack add [@ref]`, which validates a lens pack with -the existing loader, vendors its tree under `.verdict/packs//` -(including the `evals/` and `examples/` payloads), and pins it in -`.verdict/verdict.lock`. The lock's tree hash covers those payload trees too, -so two packs that differ only in an eval case record different hashes. The lock is strict JSON with one row -per repo pack, sorted by name: `name`, `source` (`dir:` or -`git:@`), `version`, the canonical tree `hash`, and an integrity -`sig` over those fields. A `dir:` source is a slash path with no control -character, backslash, or leading `-`; neither spelling carries a userinfo -password. Re-adding a name replaces its tree and entry and -prints the old and new hash; a source inside `.verdict/packs/` refuses -`bad-pack`. Git sources fetch hermetically (no system config, scrubbed -`HOME`/`XDG_*`, no prompts, no hooks, file protocol disabled, shallow clone -without tags or submodules) and record the resolved HEAD commit; the git calls -share a five-minute budget and the clone is checked against the pack caps -before it loads. A URL -userinfo password or token refuses `bad-pack` before the clone and never -reaches a printed line, a refusal message, or `verdict.lock`; the bare -`user@host` forms keep working, and a hand-written lock row carrying one -refuses too. - -`fetch`, `fetch-local`, `post`, and `record` now refuse `pack-lock-missing` -when a repo pack has no lock entry and `pack-lock-drift` when the tree hash, -pack directory, or `sig` disagrees, from one shared check with no bypass -flag. `.verdict/verdict.lock` joins the controls inventory, so it is read at -the PR base and an edited snapshot copy refuses `controls-stale`. `pack list` -fills its `drift` column with null for a matching repo pack and the refusal -code otherwise; the human rows show it, and a lock row whose pack directory -is gone lists with `drift: "pack-lock-drift"`. The built-in pack stays pinned -by the binary and org packs stay in per-operator settings, so neither is -locked. diff --git a/.changeset/panel-verification.md b/.changeset/panel-verification.md deleted file mode 100644 index 4573d8df0..000000000 --- a/.changeset/panel-verification.md +++ /dev/null @@ -1,9 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `verdict panel`, a local planner for independent verifier dispatches. -Keep-biased results enrich findings and record/post/report summaries without -voting, dropping findings, or exposing verifier evidence in posted bodies. -New artifacts live under `verification/`; the pinned `panel` roster is unchanged. -The verifier prompt joins the provenance hash; existing receipts retain their hash. diff --git a/.changeset/post-content-digest.md b/.changeset/post-content-digest.md deleted file mode 100644 index 11dadb5ca..000000000 --- a/.changeset/post-content-digest.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Compute post payload receipts from rendered review content and the reviewed head, excluding diff positions, event selection, and range metadata. diff --git a/.changeset/post-content-recovery.md b/.changeset/post-content-recovery.md deleted file mode 100644 index 1c3bfec88..000000000 --- a/.changeset/post-content-recovery.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Post only comments and summary parts missing from fetched remote content. Recover partial publication after re-fetching, recognize identical findings at moved lines, and record new-head rearming with `rearmed` and `previous_head` in post output and receipts. Local receipts no longer block publication or override approval gates. diff --git a/.changeset/post-live-authority.md b/.changeset/post-live-authority.md deleted file mode 100644 index 8fdec9068..000000000 --- a/.changeset/post-live-authority.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Refresh live remote reviews and comments before publication, refusing without writes if the read fails. Retrying a partial publication no longer requires re-fetching. Match identical comment bodies one-to-one, pairing exact anchors before moved lines so distinct missing findings are preserved. diff --git a/.changeset/post-secret-guard.md b/.changeset/post-secret-guard.md deleted file mode 100644 index cac8a7129..000000000 --- a/.changeset/post-secret-guard.md +++ /dev/null @@ -1,15 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Refuse to publish secret-shaped values. `post` scans the summary body and each -finding's body and suggestion before the first host call, and a credential -value — a PEM private key, an AWS key id, a GitHub, Slack, Google, Stripe, -Anthropic, or OpenAI token — refuses `post-secret-shaped`, naming the site and -the matched shape but never the value: a posted review cannot be unsent. -`--dry-run` warns in the same words and still exits 0. `--allow-secret-values` -publishes the review anyway, and `post.json` records the shapes it let through -in `secret_shapes_allowed[]`. The patterns move to `internal/envsecret`, so -`fetch`'s sensitive-content scan and this guard share one table; fetch's -bundles and goldens are byte-identical, and so is the receipt of a review that -carries no such value. diff --git a/.changeset/post-summary-comment.md b/.changeset/post-summary-comment.md deleted file mode 100644 index db968377c..000000000 --- a/.changeset/post-summary-comment.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add opt-in PR summary comment support via `post --summary comment|off` (default -`off`). When `off`, `post` output, host calls, and receipts stay byte-identical. -When `comment`, `post` maintains one separate PR-level summary comment carrying -a `` marker (distinct from review and -finding markers, never editing the PR description), ignoring comments by other -authors, updating the existing comment in place across GitHub (`PATCH`), -GitLab (`PUT`), and Forgejo (`PATCH`), linking to the submitted review and -posted inline comments where hosts supply URLs, and reporting duplicate matches -in the receipt without deleting older comments. `--dry-run` reports the planned -summary action (`create`, `update`, or `none`) without writing. A failed summary -write never loses the accepted review's receipt: `summary.action` is `failed` -with the reason, and the next post updates the summary. diff --git a/.changeset/post-verdict-reconciliation.md b/.changeset/post-verdict-reconciliation.md deleted file mode 100644 index 6b3b2021a..000000000 --- a/.changeset/post-verdict-reconciliation.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Reconcile binding verdicts separately from summary and inline content. Retry failed approvals without duplicate GitLab summary notes, ignore Forgejo drafts, map Forgejo anchors before matching, and reject malformed nested live notes. Preserve verified patch provenance and clarify partial-publication retry advice. diff --git a/.changeset/pr-wait-inconsistent-reads.md b/.changeset/pr-wait-inconsistent-reads.md deleted file mode 100644 index 5e9173dfe..000000000 --- a/.changeset/pr-wait-inconsistent-reads.md +++ /dev/null @@ -1,10 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Retry an inconsistent `pr-wait` observation instead of failing the whole -bounded wait: a GraphQL head that is not yet the REST head (or a torn view -document) counts as an inconsistent observation, the wait sleeps one interval -and reads again, and only `--once`, a hard `gh` failure, or a state still -inconsistent at the timeout refuses `pr-read-failed`. The pr-wait data now -carries `inconsistent_observations` so a retried read stays visible. diff --git a/.changeset/prompt-key-shapes.md b/.changeset/prompt-key-shapes.md deleted file mode 100644 index 85f4d0116..000000000 --- a/.changeset/prompt-key-shapes.md +++ /dev/null @@ -1,13 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Repair `in_bounds` to `in-bounds` on a review comment, beside the existing -`inBounds`/`rung`/`file` aliases. - -The benchmark's tool-free prompt keeps the `review.json` coverage entry shape -(`{"file": …, "ranges": …}`) it used to replace away with the rest of the submit -step. A bench test pins four key shapes — the comment keys `rigor` and -`in-bounds`, and the coverage entry's `file` and `ranges` — and fails when the -derived prompt drops one, or spells one from the transform's own copy instead of -the shipped prompt. diff --git a/.changeset/prompt-rigor-key.md b/.changeset/prompt-rigor-key.md deleted file mode 100644 index 5261ba0cb..000000000 --- a/.changeset/prompt-rigor-key.md +++ /dev/null @@ -1,17 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Publish the review.json key names the prompt only described. `prompts/review.md` -now tags a finding's evidence rung in the comment key `rigor`, publishing its -values as `said|pointed|walked|ran|reproduced`, writes the in-bounds flag as -`in-bounds`, and shows a coverage entry as `{"file": …, "ranges": …}`; the -`SKILL.md` router names `rigor` too. A pilot review emitted `rung` and -`in_bounds` — the aliases `record` repairs as warnings, which strict CI refuses -— because no prompt published the keys the schema defines. - -An archtest holds every key the prompt publishes to review.schema.json and its -`rigor` value list to the schema enum, so the two cannot drift apart again. The -prompt set moves `prompts_hash`, and the record and fetch goldens follow it: -`prompts_hash`, the `run_id` that folds it, and the two edited prompt rows of -each provenance input list. diff --git a/.changeset/provenance-receipt.md b/.changeset/provenance-receipt.md deleted file mode 100644 index 1edb5f130..000000000 --- a/.changeset/provenance-receipt.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Record a deterministic `.verdict/state//provenance.json` receipt during `fetch` / `fetch-local` (listing resolved prompt/panel/bundle/KB/learnings/override/control/config/exclusions metadata and `defaults.kb_policy.max_chars` / `REVIEW_KB_MAX_CHARS` truncation without copying secrets or rule text), expose `--provenance` on `report` and `post` (`data.receipt.inputs[]`, `data.truncated[]`, and full human table), and append a single `> Inputs:` summary line to the footer when `provenance.json` is present. Published schema: `data.receipt` (`$defs/provenanceReceipt`) is declared alongside the existing keys, and `data.truncated` widens from `boolean` to `boolean | provenanceTruncatedRow[]` on `reportData` and `postData`, so `ReportData.truncated` and `PostData.truncated` type as a union in the TypeScript client (no compat shim, no `schema_version` bump). diff --git a/.changeset/provider-from-remote.md b/.changeset/provider-from-remote.md deleted file mode 100644 index a2ecd3873..000000000 --- a/.changeset/provider-from-remote.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Infer the provider from origin when no provider flag is given, refuse unknown hosts, and record the selection source in PR.json. diff --git a/.changeset/punctuated-mentions.md b/.changeset/punctuated-mentions.md deleted file mode 100644 index 6f376da1e..000000000 --- a/.changeset/punctuated-mentions.md +++ /dev/null @@ -1,13 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Mentions behind punctuation or emphasis are escaped too. The neutralizer opened -a mention only at a line start or after a space or tab, so `(@alice)`, -`**@alice**`, and `_@alice_` were published as written and still notified the -person they named — the exact notification the escape exists to prevent. Any -byte that cannot sit in the local part of an email address now opens a mention, -so those shapes are escaped while `user@example.com` is left byte-identical. -Fenced code blocks, inline code spans, HTML comments including the finding -markers, and a finding's `suggestion` stay untouched, and the rewrite remains -idempotent. diff --git a/.changeset/reaction-scoped-fix.md b/.changeset/reaction-scoped-fix.md deleted file mode 100644 index f45938a48..000000000 --- a/.changeset/reaction-scoped-fix.md +++ /dev/null @@ -1,32 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Reaction-scoped fix rounds: a GitHub `fetch` now writes -`.verdict/state//selected.json` naming the finding sigs whose root -verdict review comment carries a human 👍, and the new `fix --selected` batch -addresses exactly those sigs. - -The reaction read moves into one shared reader (`threadscore`) that `calibrate ---reactions` and `fetch` both use, so the pagination and the bot/self exclusions -are the same code: a bot's 👍 and a comment's own poster's 👍 never count, and a -reply is never a root, so only a comment verdict posted can select its sig. - -`selected.json` is written only when something was reacted to — an empty -selection would claim a scope nobody asked for — a re-fetch after the reactions -go away deletes a stale one, and a failed reaction read keeps the last good -selection and warns instead of narrowing the work. GitLab and Forgejo write no -file, because verdict reads no reactions there. - -`fix --selected` takes one bounded pass per selected sig with the same ceilings, -ledger, and `data.batch` outcomes as `--all`, prints the selection it read -(`data.selection` carries the path), and defaults its floor to `off` — the -reaction is the endorsement a severity floor would otherwise re-judge. It -refuses rather than printing an empty batch: `missing-selection` when no file -exists, `no-selected-findings` when the named sigs are no longer open inline, -and `corrupt-store` when the file is unreadable. It is also the one batch that -runs on a PR bundle, whose instructions land the fix through a re-fetch and -`post`. - -Schemas: new `schemas/selected.schema.json` (docs/selected bundle document), and -the additive optional `data.selection` in the fix data profile. diff --git a/.changeset/reactions-trend.md b/.changeset/reactions-trend.md deleted file mode 100644 index 4f04e5ba8..000000000 --- a/.changeset/reactions-trend.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add opt-in posted GitHub reaction proposals and per-rule acceptance and false-hit rates. diff --git a/.changeset/read-verb-re-enter.md b/.changeset/read-verb-re-enter.md deleted file mode 100644 index 5d86569ee..000000000 --- a/.changeset/read-verb-re-enter.md +++ /dev/null @@ -1,10 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Give `report`, `next`, and `show` refusals a `re-enter:` line. A missing bundle -printed a `try:` hint with no command to re-run, so the way forward was a retype; -the refusal now names the invocation the operator typed, shell-quoted and with -the flags they passed (`re-enter: verdict report acme/app 5101 --json`). This -covers `next`'s store-error path, which renders its own envelope. Refusal codes, -exit codes, and the envelope keep their shape; only stderr gains a line. diff --git a/.changeset/ready-calibrate-gitlab.md b/.changeset/ready-calibrate-gitlab.md deleted file mode 100644 index d46ba6611..000000000 --- a/.changeset/ready-calibrate-gitlab.md +++ /dev/null @@ -1,15 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`ready` advises on GitLab merge requests and plain `calibrate` joins live -GitLab threads, both through `glab` behind the same per-provider tables the -GitHub reads already sit behind (GitHub output byte-identical). `ready` reads -the merge request (`draft`, `has_conflicts`, `detailed_merge_status`, -`description`, head pipeline) plus the pipeline and discussion lists, so the -same draft, conflict, `ci-failing`, `threads-open`, and `description-thin` -blockers report on either host; `calibrate` matches the posted discussion -bodies (finding marker included) against the live discussions and reads -changed files through the repository compare, so the same fixed, -resolved-unfixed, open, and unknown outcomes join on either host. Bundles on -the remaining hosts still refuse `provider-unimplemented` naming the host. diff --git a/.changeset/recipes-format.md b/.changeset/recipes-format.md deleted file mode 100644 index de60f82e5..000000000 --- a/.changeset/recipes-format.md +++ /dev/null @@ -1,19 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Declare review pipelines as data: `.verdict/recipes.json` is a strict, versioned -control document whose named entries state a lens set (with per-lens model hints -and dispatch caps), a synthesizer hint, and a budget, so a pipeline can be -diffed and shared instead of living in prompt prose. The format is published as -`schemas/recipes.schema.json` and the file joins the base-owned controls -inventory, so the document a PR is reviewed by is read at the PR base and a head -edit never steers it. The loader resolves a name — repository entry first, then -the embedded `default` — into the four fixed stages (`lenses`, `synthesize`, -`dedupe`, `rank`) and a canonical `sha256` over the selected entry; a bundle with -no recipe control resolves the built-in and refuses nothing. Two refusal codes -join the published set: `bad-recipe` (one diagnostic naming the JSON path and -the key) and `unknown-recipe` (the message lists the available names). Positions -are declarations, not selections: verdict never dispatches the stages and never -selects a model. The `run --recipe` surface that renders the plan lands on top -of this format. diff --git a/.changeset/record-cost-name-warnings.md b/.changeset/record-cost-name-warnings.md deleted file mode 100644 index c6ccad9cc..000000000 --- a/.changeset/record-cost-name-warnings.md +++ /dev/null @@ -1,10 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`record` (and `record --dry-run`) now warns once per `cost.json` `models[]` -entry with no `name`, for example `cost.json: models[0] has no name — the name -is what attribution reads; add "name"`. The record stays accepted and priced — -warnings only, never a refusal or rewrite — and `record --json` carries the -same text in `warnings[]`. Only a missing or non-string name warns; an empty -string passes, since the schema sets no minimum length. diff --git a/.changeset/release-asset-sigpipe.md b/.changeset/release-asset-sigpipe.md deleted file mode 100644 index 2fab711ed..000000000 --- a/.changeset/release-asset-sigpipe.md +++ /dev/null @@ -1,13 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`scripts/release.sh` no longer reports a shipped release asset as missing. The -completeness check piped the asset list into `grep -q`, which exits at the first -match; the writer then took SIGPIPE, `pipefail` failed the pipeline, and the -asset landed in the missing list — so a complete release could refuse its own -body ("release v0.6.1 is not complete (missing: …)"). The race lost only under -load, which is why it showed up in a parallel eval run. The membership test is -now a bash whole-line match with no producer that can break, and the -`release-body` eval pins it with an asset list over 64 KiB whose first entry -must still be found. diff --git a/.changeset/report-html.md b/.changeset/report-html.md deleted file mode 100644 index 62ab5719b..000000000 --- a/.changeset/report-html.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Export the six dashboard views with `report --html [--root ]`. -Refresh through sibling staging and backup directories, roll back publication -failures, and refuse stale recovery artifacts or unsafe destinations. diff --git a/.changeset/resolution-rate.md b/.changeset/resolution-rate.md deleted file mode 100644 index a07417f85..000000000 --- a/.changeset/resolution-rate.md +++ /dev/null @@ -1,6 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Publish live resolution metrics and save calibration receipts. Add offline -resolution rollups, clean-review metrics, and clean report headlines. diff --git a/.changeset/review-residuals.md b/.changeset/review-residuals.md deleted file mode 100644 index 476a0de72..000000000 --- a/.changeset/review-residuals.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -A batch of small fixes from recent review rounds: - -- An agent that exits non-zero just before its time budget runs out is now reported `agent-failed`, not - `agent-timeout`. The watchdog records its own budget kill instead of comparing the clock afterwards. -- MCP refusal results carry the refusal's `warnings`, both as `warn:` lines in the text and as a `warnings` field in - the structured result. -- Human-mode output prints a refusal's warnings as `warn:` lines before `error:`, so `pack list` no longer hides a - stray-file warning when it refuses. -- The `index-capped` refusal names the bytes it measured. Its hint no longer suggests raising a cap that cannot be - configured. -- `index status` points at the query, not at `index build`, once HEAD is indexed. -- Reading a crashed agent's partial review can no longer block on a FIFO. diff --git a/.changeset/rigor-validation.md b/.changeset/rigor-validation.md deleted file mode 100644 index 1964e61bb..000000000 --- a/.changeset/rigor-validation.md +++ /dev/null @@ -1,33 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Gate a review on evidence rather than on its author's confidence. A trusted -`prompts/validate.md` pass judges each candidate finding — hedge, -reproducibility, anchor, evidence rung, severity — and writes -`.review//validation.json`: one `CONFIRMED`, `REJECTED`, or -`DOWNGRADED` result per sig, with a reason of 10-500 characters. The pass runs -in the invoker, so `verdict` still spawns no model; `post` and `record` read -the file, hash its exact bytes into `validation_hash`, and record -`validation_counts`, `validation_mode`, and `rigor_floors` beside it. - -`rigor_policy` in the trusted-base `.verdict/config.json` floors each severity -tier (`mode: warn|gate`, `critical`/`important`/`informational` at -`said|pointed|walked|ran|reproduced`), and `post --require-rigor -=` narrows it per invocation, implying gate mode for the named -tiers. Advisory is the default: a rejection warns with the validator's reason -and the finding still posts. In gate mode a rejected finding is withheld -through the D2 receipt (`withheld.source: "validation"`, status still `open`) -rather than deleted, a finding below its floor refuses `rigor-below-floor`, and -`--force` publishes it while recording the reason and the offending sigs in -`post.json.forced[]`. A critical that still clears the important floor is -downgraded to `important` with a `validation` history event instead of -refusing, and no verdict ever upgrades a severity. The validator's prose is -stored and never quoted as a finding's evidence. - -`validation.json` is validated strictly: a stale head refuses -`validation-stale`, and an unknown key, unpublished verdict, short reason, a -downgrade that skips a rank, a duplicate sig, or a sig that resolves to no -single finding refuses `bad-validation`; a malformed base policy refuses -`bad-rigor-policy`. `prompts/validate.md` joins the hashed prompt set, so -`prompts_hash` moves once for this change. diff --git a/.changeset/risk-queue.md b/.changeset/risk-queue.md deleted file mode 100644 index d872089f6..000000000 --- a/.changeset/risk-queue.md +++ /dev/null @@ -1,21 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Rank the local bundles into an advisory triage queue. `risk -[--status open|all] [--limit ]` reads every bundle under `.verdict/state` -and prints one row per bundle — its recomputed band, the 0-100 score -(`report --risk`'s reach 35 / churn 25 / category 25 / rung 15, computed from -the bundle's own `impact.json`, `full.diff`, and open findings), and whether it -needs a human — sorted by band, then score, then bundle name. `--status open` -(the default) keeps host-open and local bundles and names the merged and closed -ones it skipped; `--status all` keeps everything. `--limit` (default 50, max 200) bounds the rows and carries the flattened receipt while the tallies keep -describing the whole queue. A bundle with no `findings.json` is queued, never -dropped, and a missing or unusable `impact.json` reads reach unmeasured at zero -points with a counted warning, so an unindexed bundle never looks like a clean -zero. Pure read: no host call, no index build, nothing written. The `data` -object is described by `$defs/riskData` in the generated envelope schema -(additive: `riskData`, `riskBands`, `riskRow`, plus the receipt keys every -capped list publishes), the TypeScript client gains `client.risk()`, the `risk` -MCP tool serves the same data read-only, and the Go facade gains `Client.Risk`. -Gating one bundle stays `report --max-risk`. diff --git a/.changeset/risk-score.md b/.changeset/risk-score.md deleted file mode 100644 index f4d4de194..000000000 --- a/.changeset/risk-score.md +++ /dev/null @@ -1,25 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add advisory change-risk scoring (`review.json.risk`, `report --risk`, -`report --max-risk`, and `trend` risk-band rollups). `review.json` accepts an -optional `risk: {level, rationale, assessed_head}` claim validated during -`record` and persisted to `record.json` / `post.json` (with stale-head detection -via `risk-stale-head`). `report --risk` reports `claimed`, deterministic store -`floor` inputs, and `reported = max(claimed, floor)` (keeping unassessed -changes `unknown` and incomplete evidence at `>= medium`); `report --max-risk` -adds an offline fail-closed gate dimension. `trend` adds per-band outcome counts -under `by_risk` and `per_bundle[].risk`. Exposed across CLI, MCP `report`, -Go SDK `Client.Report`, JSON schemas, and TypeScript client types. - -`report --risk` also publishes `data.risk.score`, an advisory 0–100 triage -score built only from the bundle's own artifacts: `impact.json` for reach, -`full.diff` for churn, the worst open-finding category, and the weakest -open-finding rung. Each component echoes the points it earned and the weight it -carried, so every value is recomputable without a config read, and `measured` -counts the components whose inputs were read: a bundle with no `impact.json` -reads `reach.measured: false`, `reach.points: 0`, and `score.measured: 3` -instead of a silent zero. Human mode prints one `risk score:` line after -`risk floor inputs:`. The score rides `--risk` alone, so `--max-risk` keeps -gating on `reported` and a band and its score can never disagree. diff --git a/.changeset/rule-exposure.md b/.changeset/rule-exposure.md deleted file mode 100644 index 51d57191c..000000000 --- a/.changeset/rule-exposure.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add per-rule `recalled_in`, `cited_in`, `exposure`, and `acceptance` columns to `learn --status` (text and `--json`), reusing the bundle overlap join and reporting missing recall manifests as `unmeasured`/`null`. diff --git a/.changeset/rule-pack-import.md b/.changeset/rule-pack-import.md deleted file mode 100644 index 1ffee7bc7..000000000 --- a/.changeset/rule-pack-import.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Import structured rule objects with `learn --pack --id `, preserving pack provenance and hash in the tracked KB controls. diff --git a/.changeset/run-agent.md b/.changeset/run-agent.md deleted file mode 100644 index 1a3568bf2..000000000 --- a/.changeset/run-agent.md +++ /dev/null @@ -1,27 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Launch your own agent. `verdict run --agent ` -runs the operator's agent CLI against an existing bundle — one process, one -bundle, one shot — then asserts exactly one fresh `review.json` and hands off to -the unchanged `record`/`post` validation. It never retries, re-prompts, repairs, -or reads the model's prose, and it never posts. - -Agents are data: the shipped table names `claude`, `codex`, `cursor`, and -`opencode` with their binaries and non-interactive argument shapes, so a new -agent is a new row. Everything after a bare `--` passes to the agent unchanged, -`--print` renders the exact command (POSIX-quoted) and exits zero without -spawning, and `--timeout` sets the shared watchdog budget. - -Every failure is typed with a hint: `agent-missing`, `agent-spawn-failed`, -`agent-timeout`, `agent-failed` (whose exit code becomes the process exit code so -CI sees the real failure), `no-review-artifact`, `stale-review-artifact`, -`unknown-agent`, and `agent-record-failed`. The launch is recorded in -`.verdict/state//launcher.json` — agent, binary, argv shape with secrets -redacted, timings, exit code — with the child's full output in `.run.out`. - -The verb is CLI-only: spawning a local agent is not an MCP tool, a facade method, -or a TypeScript client method. Verdict invokes your agent only when you ask: no -daemon, queue, scheduler, webhook, model selection, or interpretation of model -output, and `run` itself makes no network calls. diff --git a/.changeset/run-argument-credential.md b/.changeset/run-argument-credential.md deleted file mode 100644 index fd15d3893..000000000 --- a/.changeset/run-argument-credential.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Keep an argument credential out of the launch record. `verdict run --agent` -redacted a secret-shaped flag's operand in `launcher.json`'s `argv`, but the same -credential still reached the bundle twice: the recorded `output_tail` was scrubbed -only against secret-named environment values, so a child that echoed its own argv -— or printed it in a stray debug line — left the operand in plaintext, and the -raw child log `.run.out` was written `0644`. - -The output scrub now hides the operand the argv redactor already identifies, from -one walk that feeds both the record's `argv` and its tail, so the two cannot -disagree about what a credential is. `.run.out` stays the child's raw stream for -debugging and is written `0600`. The environment scrub, the tail bound, `--print`, -and `launcher.json`'s shape are unchanged. diff --git a/.changeset/run-print-stage.md b/.changeset/run-print-stage.md deleted file mode 100644 index 7ce5f7230..000000000 --- a/.changeset/run-print-stage.md +++ /dev/null @@ -1,16 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Render one recipe stage at a time with `run --print --stage --recipe -`. `n` is 1-based over the recipe's four stages (`1 lenses`, `2 -synthesize`, `3 dedupe`, `4 rank`); the rendered argv is the agent's own -flags, the stage's declared model hint as that agent's model argument, your -`--` arguments, and a directive naming the stage, its hints and caps, and the -`review.json` it reads and writes. Every stage reads the bundle and writes -`review.json`, so stages chain through the artifact `record` and `post` already -consume and your CI can dispatch them under different models. Verdict still -launches nothing and sequences nothing: `--stage` requires `--print` and -`--recipe` (`bad-flags` otherwise, as does an index outside 1–4 or a -pass-through `--model`/`-m` against a declared hint), `--json` carries -`data.stage` (`{"index", "kind"}`), and plain `--print` is unchanged. diff --git a/.changeset/run-repair-loop.md b/.changeset/run-repair-loop.md deleted file mode 100644 index 1a81c9bd7..000000000 --- a/.changeset/run-repair-loop.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`run` repairs an invalid `review.json` with `--rounds N` (1–5, default 1): round 1 sends today's directive unchanged, later rounds re-prompt the agent with the prior review bytes and the shape checker's fix list fenced under the run's nonce, and the loop stops on the first valid round. `--timeout` is per round, `--json` carries the loop in `data.repair`, and exhaustion refuses `repair-exhausted`. diff --git a/.changeset/self-review-loop.md b/.changeset/self-review-loop.md deleted file mode 100644 index 42f00073b..000000000 --- a/.changeset/self-review-loop.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Add a bounded local self-review recipe for Claude, Codex, and OpenCode, with round receipts, offline evals, and benchmark instructions. Require complete recorded coverage, bound to the verified diff identity on dirty local bundles. Refuse stale dirty-tree coverage after tracked files change since fetch. diff --git a/.changeset/serve-live-updates.md b/.changeset/serve-live-updates.md deleted file mode 100644 index f420637b1..000000000 --- a/.changeset/serve-live-updates.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add live bundle invalidations to `verdict serve`, with `--interval`, `--sweep`, and `--no-watch` controls. diff --git a/.changeset/serve-s1.md b/.changeset/serve-s1.md deleted file mode 100644 index 1f5495764..000000000 --- a/.changeset/serve-s1.md +++ /dev/null @@ -1,9 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `verdict serve`, a read-only HTTP surface over the current repository -(loopback by default; `--host` is the documented exposure opt-in): -`/v1/health`, `/v1/roots`, `/v1/bundles`, the bundle detail, and `show`, with a -per-port session-cookie exchange for the browser, strict CSP, Host and Origin -gates, and `data.sources[]` traceability on every composed response. diff --git a/.changeset/serve-s2-discovery.md b/.changeset/serve-s2-discovery.md deleted file mode 100644 index 2add83bce..000000000 --- a/.changeset/serve-s2-discovery.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`serve --root` discovers nested repositories with bounded traversal and a local -24-hour registry. The roots and inbox responses report cache, truncation, and -skipped-pattern state. diff --git a/.changeset/serve-s3-cost-precision.md b/.changeset/serve-s3-cost-precision.md deleted file mode 100644 index d85ba3d78..000000000 --- a/.changeset/serve-s3-cost-precision.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `/v1/cost` repository, ISO-week, and model buckets plus `/v1/precision` -category and tier counts from the stored thread feed. Unmeasured model prices, -fixed outcomes stay `null`, no rates are published, and corrupt aggregates are -marked partial. diff --git a/.changeset/serve-ui.md b/.changeset/serve-ui.md deleted file mode 100644 index 883a3f765..000000000 --- a/.changeset/serve-ui.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verdict serve` now opens an embedded browser UI with inbox, review detail, -threads, cost, precision, and gate views. The UI supports dark mode, refreshes -after live invalidations, and can read static `data.json` exports. diff --git a/.changeset/setup-verdict-hardening.md b/.changeset/setup-verdict-hardening.md deleted file mode 100644 index d89523885..000000000 --- a/.changeset/setup-verdict-hardening.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Pass setup-verdict inputs through environment variables and cache release archives. On reuse, refresh the signed checksums and build-commit -metadata, verify the archive, then extract a fresh binary without downloading the -archive again. A tampered archive fails closed and remains cached until the entry -is removed. Cache reuse requires release metadata access. diff --git a/.changeset/skill-install-preflight.md b/.changeset/skill-install-preflight.md deleted file mode 100644 index 229bbfa22..000000000 --- a/.changeset/skill-install-preflight.md +++ /dev/null @@ -1,14 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verdict skill install` now checks its targets before the first write. On POSIX -exactly the directories an install will write are probed before the first byte, -so a read-only skills dir anywhere in a `--detect` set refuses with -`skill-install-failed`, names that target, and leaves the earlier targets -absent instead of already installed. A target that changes between that probe -and the write still fails, at write time. Windows reports that failure at write -time only, because its writability probe is a stat. A receipt path that cannot -be replaced refuses with `skill-foreign-conflict` the same way, before the -payload lands. A single-target install refuses on the same probe, and a -repeated `--agent` dedupes to one install per target instead of refusing. diff --git a/.changeset/skill-management.md b/.changeset/skill-management.md deleted file mode 100644 index 84b4d81d9..000000000 --- a/.changeset/skill-management.md +++ /dev/null @@ -1,29 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verdict skill` grows its inspection and removal half. `skill show ` -prints one embedded payload file byte-identically (with `--json`, one envelope -carrying `path` and the same bytes in `content`); a traversal, absolute, or -unknown path refuses `skill-path-not-inventory` and names `skill files`. -`skill status [--dir|--agent...|--detect] [--diff] [--json]` is the read-only -drift report: it writes nothing, reuses the install plan to report -create/update/keep/blocked per target with the receipt versus embedded version, -and exits 0 in sync and 1 when drifted, so CI can gate without -`install --dry-run`. `--diff` adds one unified diff per changed file, capped at -64 KiB overall with an explicit truncation marker and `truncated` in JSON. -`--agent` now repeats on `install`, `status`, and `uninstall`: targets are -deduplicated and handled in argv order, while `--dir` stays single. -`skill uninstall [--dir|--agent...] [--force] [--dry-run] [--json]` deletes -only the inventory paths the `.verdict-managed` receipt owns, then the receipt, -and leaves every foreign file in place; a target that is a symlink -(`skill-target-file`) or an inventory path that escapes the target through a -symlinked directory (`skill-foreign-conflict`) refuses before any deletion; a -target without a receipt refuses -`skill-dir-foreign` unless `--force` (which still removes only inventory -paths), an absent install refuses `skill-not-installed`, and `--dry-run` -deletes nothing. `skill install` likewise refuses a symlinked target -(`skill-target-file`) and a payload directory that escapes the target through -a symlink (`skill-foreign-conflict`) before it writes anything. -`verdict uninstall`'s skill-copy hint now names -`verdict skill uninstall` instead of `rm -rf`. diff --git a/.changeset/store-lock-and-listing.md b/.changeset/store-lock-and-listing.md deleted file mode 100644 index 275002aff..000000000 --- a/.changeset/store-lock-and-listing.md +++ /dev/null @@ -1,19 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Every write to `.verdict/kb.jsonl` now serializes on the same OS advisory lock -the incident ledger uses: `learn --rule` and `learn --pack` append under it, -`learn --retire` and `learn --restore` rewrite under it above their read, and -the credit rewrite a `record`/`triage` flip performs holds it across its -read-modify-write. A second writer waits five seconds for the lock: `learn` -then refuses `store-busy` instead of dropping a row, and a flip drops the -staged credit with the usual retry warning instead of losing an append. The -persistent `.verdict/kb.jsonl.lock` joins the tracked `.verdict/.gitignore` -contract, which the write brings up to date under that lock so parallel writers -cannot race one another's rewrite of the shared file. `trend`, adopt's `--from auto` twin scan, and the serve listings now -stat a bundle home before listing it: a `.verdict/state` that is not a -directory, or will not list, refuses `corrupt-store` naming it instead of -reading as an empty store — an absent home still reads as empty, a bundle home -that exists but cannot list never reads as "no bundles", and serve keeps its -own `Stat`/`ReadDir` seams to reach the same decision. diff --git a/.changeset/strikethrough-mentions.md b/.changeset/strikethrough-mentions.md deleted file mode 100644 index bf2d2f5d2..000000000 --- a/.changeset/strikethrough-mentions.md +++ /dev/null @@ -1,14 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Mentions wrapped in strikethrough are escaped too. The neutralizer walked back -over a run of `*` or `_` to tell emphasis markup from the tail of an address, -but `~` is also a valid local-part byte and was left out of that walk, so -`~~@alice~~` was published as written and GitHub read it as a strikethrough -mention — the exact notification the escape exists to prevent. A run of `~~` now -counts as markup the way `**` does, so that shape is escaped while -`foo~@example.com`, `user@example.com`, `(dev@example.com)`, and a lone tilde in -prose such as `approximately ~5 minutes` stay byte-identical. The rewrite -remains idempotent, and fenced code blocks, inline code spans, HTML comments, -and a finding's `suggestion` are still untouched. diff --git a/.changeset/thread-resolve.md b/.changeset/thread-resolve.md deleted file mode 100644 index e51ef52ec..000000000 --- a/.changeset/thread-resolve.md +++ /dev/null @@ -1,17 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Add `threads resolve [--thread ]... [--dry-run]`, the -receipt-backed thread mutation: without `--thread` it closes every open -self-authored thread whose linked finding the local store marks addressed; -with `--thread` it closes only the named ids, and a human-authored thread -refuses. A thread is resolvable only when its root carries verdict's marker -and its root author equals the authenticated token user (read once per run; -an unreadable login refuses before any write); a forged marker skips as -`not-self-authored` by default and refuses `not-verdict-authored` when named. -`--dry-run` lists the plan and makes zero writes; a real run makes one -host write per thread and records `threads-resolve.json`. GitHub resolves -through `resolveReviewThread`, GitLab through the discussion resolved flag, and -`thread_resolution` now reads true on both; Forgejo refuses typed because its -API has no documented resolve endpoint. diff --git a/.changeset/troubleshooting.md b/.changeset/troubleshooting.md deleted file mode 100644 index 77f93c3ab..000000000 --- a/.changeset/troubleshooting.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`docs/troubleshooting.md` lists every refusal code the binary can emit, each -with its cause and its fix. A drift test fails the build when a code ships -without an entry (or an entry names a retired code), and the manual's -refusal section links the page. diff --git a/.changeset/typed-truncation.md b/.changeset/typed-truncation.md deleted file mode 100644 index 859d318d6..000000000 --- a/.changeset/typed-truncation.md +++ /dev/null @@ -1,11 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -Publish a shared truncation receipt across CLI `--json` `data` envelopes (`fetch`, `report`, `next`, `show`, `trend`, `threads`, and `serve`), bundle artifacts, and MCP text/structured results, and make `next`/`show` report truncated misses honestly instead of refusing or claiming nothing exists. - -Breaking (pre-1.0, within schema_version "1" by operator decision): index -blast-radius lists replace the integer `truncated` count and `depth_truncated` -with the shared boolean receipt (`truncated`, `truncation_reason`, `returned`, -`total`, `next_offset`); complete results omit the receipt instead of carrying -`truncated: false`. diff --git a/.changeset/upgrade-hints-quoting.md b/.changeset/upgrade-hints-quoting.md deleted file mode 100644 index 8a6d2d3a3..000000000 --- a/.changeset/upgrade-hints-quoting.md +++ /dev/null @@ -1,9 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verdict upgrade` quotes the project-local `run: cd ` hint with the -POSIX single-quote rule, so a project path with a space or `'` pastes -resolvably; `verdict uninstall`'s sibling skill-copy refusal now reads as a -sentence (`remove it with the tool that installed the skill copy at `) -instead of rendering prose in the `remove it with:` command slot. diff --git a/.changeset/vanished-threads.md b/.changeset/vanished-threads.md deleted file mode 100644 index baf8a040d..000000000 --- a/.changeset/vanished-threads.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Treat a thread feed removed during a serve read as absent instead of corrupt. diff --git a/.changeset/verify-env-allowlist.md b/.changeset/verify-env-allowlist.md deleted file mode 100644 index c24366f5f..000000000 --- a/.changeset/verify-env-allowlist.md +++ /dev/null @@ -1,20 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verify` now filters the child process environment with a default-deny -allowlist instead of forwarding the full parent `os.Environ()`. Only baseline -locale, terminal, shell, XDG, and toolchain variables (`PATH`, `HOME`, `LANG`, -`LC_*`, `TERM`, `TMPDIR`, `TZ`, `CI`, `USER`, `LOGNAME`, `SHELL`, color flags, -`XDG_*_HOME`, Go/C/Rust/JS/Python toolchain cache and home variables, plus -Windows OS runtime variables on `GOOS=windows`) pass by default, and any -variable matching `envsecret.IsSecret` (such as `FOO_TOKEN`, `GITHUB_TOKEN`, -`SSH_AUTH_SOCK`, or `LC_TOKEN`) is dropped even when it matches an allowlist -prefix. An explicit opt-in keeps the named variable even when it matches -`envsecret.IsSecret`. Operators can opt specific variable names in via -`--env-allowlist ` (also wired as `envAllowlist` on the `verify` -MCP tool and `EnvAllowlist` on `verdict.VerifyRequest`) or `.verdict/verify-env` -(one name per line, `#` comments and blank lines ignored, read from the -base-snapshotted `controls` inventory rather than the untrusted PR worktree). -Dropped variable names — never values — are reported in deterministic sorted -order in `warnings[]`. diff --git a/.changeset/verify-indirection.md b/.changeset/verify-indirection.md deleted file mode 100644 index b2e0d8ccf..000000000 --- a/.changeset/verify-indirection.md +++ /dev/null @@ -1,48 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -The verify pre-exec gate refuses a command word it cannot read: `E=rm; $E -rf -./target` and `E=curl; $E -o ./f http://host/x` passed every arm because the -tool word sat in the variable, and the resolved command then ran. `verify` now -refuses the shape as `verify-indirection`, quoting the spelling it saw and -naming the remedy (write the tool word plainly; `verify --sandbox` refuses it -too, because the gate runs before the sandbox wrapper). The braced (`${E}`), quoted (`"$E"`), `export`ed, -assignment-prefix, loop (`for E in curl; do $E ...`), concatenated (`$A$B`), -and own-line (`E=rm\n$E -rf ./target`) spellings close with it, and so do the -wrappers that run a later word (`env -u HOME $E`, `timeout -s KILL 5 $E`, -`timeout 5s $E`, `xargs -I {} $E`, `command`, `nice`, `time`, `nohup`, -`stdbuf`, `exec`, `then`, `do`). Variables in argument position still execute -(`echo $USER`, `PORT=9 curl http://127.0.0.1:9/$PORT/f`, `$BIN/tool --version`), -so the false-positive boundary is unchanged for a command written plainly. - -The gate does not resolve variables — the value can come from the command's own -environment, or from a later or conditional assignment — so it refuses rather -than guessing, and a variable holding a whole command line (`sh "$V"`) stays a -documented residual of the tripwire. A wrapper's own option grammar decides -which later word bash runs, so after `builtin`/`command`/`env`/`exec`/`nice`/ -`nohup`/`stdbuf`/`time`/`timeout`/`xargs` any variable in that command refuses: -that is the deliberate cost of not modelling each wrapper's flags, so -`timeout 300 make test` stays clean while `timeout 300 make test $TARGET` -refuses. A shell carrying an inline command (`sh -c "$E pwned"`, `bash -ec`, `sh -cu`) -has that command string judged as its own command position, so -`sh -c "echo $E"` stays clean; `eval` and `sudo` refuse earlier as -`verify-destructive`. -The three arms that read a command word — glob fragments, indirection and -command position — now read the segmented text, which splits at newlines, so a -later line is a command position of its own: `echo ok\nr? -rf x` and -`echo ok\n$(echo rm) -rf x` refuse instead of joining the first segment. (The -whole-string greps already saw every line.) A `\`-newline continuation is one -command and joins before segmentation, so `echo \` + newline + `"$USER"` stays -clean. Leading redirections (`&1`, `>> f`) are skipped to reach -the command word; a quoted word stays one word, so `LABEL="hello world" $E ok` -and `>"out file" $E` refuse rather than putting `world` in command position; a -substitution body (`$($E pwned)`, backticks, `cat <($E pwned)`) is judged as its -own command position — quote-aware, so `$($E "(")` closes where bash closes it — -while `$((…))` arithmetic and single-quoted literals (`awk '{print $($NF)}'`) -are left alone. - -`SECURITY.md` now says the same thing -the README and the manual say: the refusals are a tripwire, `verify --sandbox` -is the boundary, and a spelling outside the gate's reach is a gate gap to arm, -not a vulnerability. diff --git a/.changeset/verify-out-owner-only.md b/.changeset/verify-out-owner-only.md deleted file mode 100644 index fb9bf03aa..000000000 --- a/.changeset/verify-out-owner-only.md +++ /dev/null @@ -1,10 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verify` writes `.verify.out` at `0600` on Unix instead of `0644`: the stream is -the command's raw output, so a command that prints a credential leaves the value -in it, and the file is now the operator's to read and nobody else's — the same -permission `run` gives `.run.out`. The recorded tail in `patch.json` and the -deep-validation receipts stay redacted as before. Windows has no mode bits to -set, so there the file keeps the directory's inherited ACL. diff --git a/.changeset/verify-sandbox-macos.md b/.changeset/verify-sandbox-macos.md deleted file mode 100644 index ee3e8aaa7..000000000 --- a/.changeset/verify-sandbox-macos.md +++ /dev/null @@ -1,17 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -`verify --sandbox` now enforces the sandbox on macOS as well as Linux. -macOS runs the command under the built-in `sandbox-exec` with -`(version 1)(allow default)(deny network*)`: every `bind()` and `connect()` -is denied, loopback and unix sockets included, so a command that -listens or reaches a listener fails. macOS has no filesystem namespace, so -the mount shape stays Linux-only — credential paths, `/proc`, `/run`, and a -private `/tmp` are still hidden by `bubblewrap` on Linux, and on macOS -filesystem isolation needs a Linux VM or container running that recipe. The same probe runs before anything -executes, and a host with no mechanism (Windows included), a missing runner, -or a runner that cannot build its sandbox still refuses `sandbox-unavailable` -and runs nothing. The manual and `docs/troubleshooting.md` now call the -static gate a tripwire rather than a boundary and name what isolates on each -platform. diff --git a/.changeset/verify-sandbox.md b/.changeset/verify-sandbox.md deleted file mode 100644 index 935962d3b..000000000 --- a/.changeset/verify-sandbox.md +++ /dev/null @@ -1,21 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`verify --sandbox` runs the verify command inside an enforced bubblewrap -sandbox on Linux: no network, a private PID namespace with its own `/proc` -(so no host PID or `/proc//environ` is readable), an empty `/run` (so -`docker.sock`, the session bus, and agent sockets have no unix-socket route -out), a fresh `/dev`, a private `/tmp`, `--new-session`, and -`--die-with-parent`. `~/.aws`, `~/.ssh`, `~/.config/gh`, `~/.docker`, -`~/.kube`, `~/.config/gcloud`, `~/.azure`, and `~/.gnupg` are mounted empty; -`~/.netrc`, `~/.git-credentials`, `~/.npmrc`, and `~/.pypirc` are bound to -`/dev/null`; the reviewed root stays writable. It probes the same flags with -a no-op before anything executes and refuses `sandbox-unavailable` — running -nothing and leaving `patch.json` and `.verify.out` untouched — on every other -OS, when `bwrap` is missing, or when namespaces are restricted; it never -silently falls back to an unsandboxed run. The opt-in reaches the MCP tool -(`sandbox`), the Go facade (`Sandbox`), and the TypeScript client -(`--sandbox` argv); the patch record, deep-group receipt, and `--json` data -carry `sandboxed` when it was enforced. Runs without the flag are -byte-identical to before. diff --git a/.changeset/verify-static-gate-bypasses.md b/.changeset/verify-static-gate-bypasses.md deleted file mode 100644 index c630f7808..000000000 --- a/.changeset/verify-static-gate-bypasses.md +++ /dev/null @@ -1,18 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -The verify pre-exec gate closes twelve static bypasses. `wget --remote-name`, -`wget -i`, `curl -K`/`--config`, `busybox wget`, `pip3 download `, -`telnet`, curl's `telnet://` scheme, and `git archive --remote` now refuse as -`verify-networked`; a secret-shaped variable from the shared `envsecret` -classification in a networked command (`curl .../$AWS_KEY`, -`curl -H "X: $AWS_KEY"`) refuses as `verify-credential-read`; and -`git --exec-path` plus `tar -I` refuse as `verify-destructive`. Both short -options match the attached and clustered spellings a shell accepts too -(`tar -Itouch-fz`, `tar -cIf`, `wget -i/tmp/urls`, `wget -qi /tmp/urls`), not -only the separated form. Every refusal keeps the quoted fragment and `try:` -rewrite hint. The documented boundary is -unchanged: a plain `curl` GET, `/tmp` scratch writes, and non-secret variables -in URLs still execute, and the adversarial corpus rows that pinned each gap -move from `xfail` to `strict` with benign neighbours beside them. diff --git a/.changeset/windows-watchdog-no-hang.md b/.changeset/windows-watchdog-no-hang.md deleted file mode 100644 index b50f3580e..000000000 --- a/.changeset/windows-watchdog-no-hang.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -On Windows, a command that failed to start — never attached to its job object, -or left with no thread to resume — and then survived verdict's cleanup kill -left the run waiting on it forever. That spawn failure now returns at once, -naming both the spawn error and the kill error, instead of hanging. diff --git a/.changeset/windows-watchdog-resume-retry.md b/.changeset/windows-watchdog-resume-retry.md deleted file mode 100644 index 4ced83f32..000000000 --- a/.changeset/windows-watchdog-resume-retry.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -On Windows, a watchdog-run command could fail before starting with `resume -suspended verify process: no thread resumed` under load. The thread lookup is -now retried a few times before the spawn fails, and a child whose thread never -appears is still killed with its job. diff --git a/.changeset/withheld-capabilities.md b/.changeset/withheld-capabilities.md deleted file mode 100644 index f0e499d00..000000000 --- a/.changeset/withheld-capabilities.md +++ /dev/null @@ -1,18 +0,0 @@ ---- -"@stainless-code/verdict": minor ---- - -`post.json` records what a host withheld, so a degraded publish stays auditable -after the envelope is gone. The new optional `withheld[]` lists one entry per -capability the host could not carry, with the capability name and the same -machine-readable reason code `fetch` publishes in `data.capabilities.unsupported` -(`no-request-changes-event`, `no-approve-verdict`, `no-inline-anchors`, …) plus -the `requested_event` the run asked for. A bound `REQUEST_CHANGES` on GitLab -therefore leaves `review.state=commented` — the state the host accepted — _and_ -an entry saying `request_changes` was withheld, so it is no longer -indistinguishable from a plain comment; the key is absent when nothing was -withheld, and `schema_version` stays `"1"` (additive: no field renamed or -removed). The `inline_anchors` capability now has a gate behind its claim: a -host advertising `inline_anchors: false` is sent no anchored comment, and the -merge keeps every finding in the body (keep-biased) with a warning naming the -host, instead of writing a position the host would refuse. diff --git a/.changeset/write-bundle-symlinks.md b/.changeset/write-bundle-symlinks.md deleted file mode 100644 index 87bd89a19..000000000 --- a/.changeset/write-bundle-symlinks.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -"@stainless-code/verdict": patch ---- - -Refuse symlinked bundle directories before `post` or `record` writes. From 2cb5e0ad45a99068362baf7f81516ef3ea6359ee Mon Sep 17 00:00:00 2001 From: Sutu Sebastian Date: Mon, 16 Nov 2026 14:43:53 +0200 Subject: [PATCH 2/2] changeset: use US spelling in the 0.7.0 notes MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The notes are user-facing copy, so a British spelling has no place in them: behaviour -> behavior (00-what-s-new.md, 10-breaking-changes.md) unrecognised -> unrecognized (10-breaking-changes.md) labelled -> labeled (09-measurement.md) Swept all 14 files for the rest of the class — colour, honour, favour, cancelled/-lled, licence, defence, grey, centre, programme, catalogue, modelling, judgement, artefact, towards, whilst, and every -ise/-isation verb — and found no other hit. Copy only: no item, mapping, or bump level moves. --- .changeset/00-what-s-new.md | 2 +- .changeset/09-measurement.md | 2 +- .changeset/10-breaking-changes.md | 4 ++-- 3 files changed, 4 insertions(+), 4 deletions(-) diff --git a/.changeset/00-what-s-new.md b/.changeset/00-what-s-new.md index da5abf972..a2005ced6 100644 --- a/.changeset/00-what-s-new.md +++ b/.changeset/00-what-s-new.md @@ -11,5 +11,5 @@ attestation, advisory risk scoring, and an optional validation pass that floors evidence behind it, while the CI gate now checks coverage by default. Posting grows a secret guard, honest per-host capability reporting, and a summary-comment option, GitLab and Forgejo join GitHub as full publish targets, and lens packs, declared recipes, and reported verification make the review pipeline itself data. -Thirteen pre-1.0 changes alter behaviour — each names the migration step below — and the known limits close +Thirteen pre-1.0 changes alter behavior — each names the migration step below — and the known limits close the notes. diff --git a/.changeset/09-measurement.md b/.changeset/09-measurement.md index 6fa2b8ab9..2ab0a1fc3 100644 --- a/.changeset/09-measurement.md +++ b/.changeset/09-measurement.md @@ -4,7 +4,7 @@ **Measurement** -- **Clean controls** — the benchmark corpus accepts clean controls with empty ground truth, ships two seeded clean cases, and keeps seeded fixtures labelled as not a model-quality claim. [Benchmark corpus](docs/benchmark.md) +- **Clean controls** — the benchmark corpus accepts clean controls with empty ground truth, ships two seeded clean cases, and keeps seeded fixtures labeled as not a model-quality claim. [Benchmark corpus](docs/benchmark.md) - **Reproducible metrics** — typed severity weights, ordered k-of-n recall, control false-positive rates, weighted F1, signal-to-noise, and explicit denominators, with provenance and argv recorded. [Benchmark corpus](docs/benchmark.md) - **Adjudicated rates** — a versioned per-run adjudication record reclassifies an unmatched finding as `real-unlisted`, `matches-ground-truth`, or `false-positive`, and `bench/runner -render` recomputes strict and adjudicated precision and recall without re-running a model. [Benchmark corpus](docs/benchmark.md) - **First real cases** — four real cases from `urfave/cli` (three bugs and one clean control), `-max-usd` stops scheduling at measured spend and marks the run uncitable, and refused samples are counted in the cost rollup. [Benchmark corpus](docs/benchmark.md) diff --git a/.changeset/10-breaking-changes.md b/.changeset/10-breaking-changes.md index c697edc8a..b51050570 100644 --- a/.changeset/10-breaking-changes.md +++ b/.changeset/10-breaking-changes.md @@ -7,9 +7,9 @@ Package is on `0.x`, so these ship as a minor bump; each one names the step that migrates it. - **`index blast-radius` truncation shape** — the integer `truncated` count and `depth_truncated` are replaced by the shared boolean receipt (`truncated`, `truncation_reason`, `returned`, `total`, `next_offset`), and a complete result omits the receipt instead of carrying `truncated: false`. _Migrate:_ read `truncation_reason` and `returned`; treat an absent receipt as complete. [Code index](docs/manual.md#code-index) -- **`gate` checks coverage by default** — the new `coverage` dimension fails when the gated HEAD has no receipt, a receipt for another HEAD, no `coverage[]` attestation, or under 100% attested. _Migrate:_ `--coverage warn`, or the SDK `Coverage` field, keeps the old passing behaviour while reviews attest nothing. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) +- **`gate` checks coverage by default** — the new `coverage` dimension fails when the gated HEAD has no receipt, a receipt for another HEAD, no `coverage[]` attestation, or under 100% attested. _Migrate:_ `--coverage warn`, or the SDK `Coverage` field, keeps the old passing behavior while reviews attest nothing. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) - **`gate --live` reads every host** — a non-GitHub bundle used to pass the live dimension without reading threads; it now reads the bundle's own host and fails closed when it cannot. _Migrate:_ `--live auto` warns and gates on the local dimensions only. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) -- **Unknown gate tiers refuse** — an unrecognised tier is `bad-flags` on the MCP `report` tool and in `Report.Check`/`Gate.MaxSeverity`, not only on the CLI. _Migrate:_ send a tier from the published set, or omit the field. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) +- **Unknown gate tiers refuse** — an unrecognized tier is `bad-flags` on the MCP `report` tool and in `Report.Check`/`Gate.MaxSeverity`, not only on the CLI. _Migrate:_ send a tier from the published set, or omit the field. [Gate](docs/manual.md#gate-one-composed-exit-code-for-ci) - **`verify` child environment is default-deny** — only baseline locale, terminal, shell, XDG, and toolchain names pass, and a secret-shaped name is dropped even when it matches an allowlist prefix. _Migrate:_ list what you need in `.verdict/verify-env` or `--env-allowlist`; dropped names appear in `warnings[]`. [Child environment](docs/manual.md#child-environment-what-inherits-what-is-explicit) - **`verify` refuses command shapes it used to run** — `wget --remote-name`/`-i`, `curl -K`, `busybox wget`, `pip3 download `, `telnet`, `git archive --remote`, and the rest refuse `verify-networked` in attached and clustered spellings, and a command word built from a variable (`E=rm; $E -rf ./target`, including the braced, quoted, `export`ed, loop, and wrapper spellings) refuses `verify-indirection`. _Migrate:_ write the tool word plainly and rewrite the command to a plain fetch your recipe can run; every refusal keeps its `try:` hint. [Known residuals](docs/manual.md#known-residuals-string-scan-ceiling) - **`.verify.out` is owner-only** — the verify stream is written `0600` on Unix where it was `0644`, because a command that prints a credential leaves the value in it. _Migrate:_ read it as the user that ran `verify`, or copy the bytes inside that same step; on Windows the file keeps the directory's inherited ACL. [Enforced sandbox](docs/manual.md#enforced-sandbox-verify---sandbox)