Skip to content

Latest commit

 

History

History
62 lines (60 loc) · 11.4 KB

README.md

File metadata and controls

62 lines (60 loc) · 11.4 KB

Architecture decision records

Standalone, individually addressable copies of the ADRs recorded in the phase implementation notes (now archived under ../archive/phases/). Extraction was mechanical: each record is verbatim from its source file, plus a provenance header. ADRs are historical — they record why a decision was made at the time; where reality has moved on, the linked issues/PRs in each record's text are the trail.

ADR Title Source
ADR-1 Postgres is the ref authority 08-spike-implementation-notes.md
ADR-2 S3 upload strictly before the ref flip 08-spike-implementation-notes.md
ADR-3 Namespace mapping via separate ref stores 08-spike-implementation-notes.md
ADR-4 Frozen contracts first, one integration seam 09-phase1-implementation-notes.md
ADR-5 Multi-org schema, single-org serve mode 09-phase1-implementation-notes.md
ADR-6 Soft delete via a unique constraint spanning deleted rows 09-phase1-implementation-notes.md
ADR-7 Promote the spike, don't rewrite it 10-phase2-implementation-notes.md
ADR-8 Production auth on the git path; git-only serve mode dropped 10-phase2-implementation-notes.md
ADR-9 verify-sig delegates to git's own verification machinery 10-phase2-implementation-notes.md
ADR-10 Signing-key registry = meta table + CLI (dashboard stand-in) 10-phase2-implementation-notes.md
ADR-11 Ephemeral namespace policy quirk 10-phase2-implementation-notes.md
ADR-12 URL-scheme factories with lifecycle interfaces 11-local-backends-implementation-notes.md
ADR-13 Local mode is intentionally single-node 11-local-backends-implementation-notes.md
ADR-14 Read plumbing lives in internal/gitread, still os/exec git 12-phase3-implementation-notes.md
ADR-15 Cursor formats — keyset everywhere, tip-anchored for commits 12-phase3-implementation-notes.md
ADR-16 refs.created_at column for branches[].created_at 12-phase3-implementation-notes.md
ADR-17 File streaming = http.ServeContent over a request-scoped temp file 12-phase3-implementation-notes.md
ADR-18 Archive is a custom walk into archive/tar + gzip, not git archive 12-phase3-implementation-notes.md
ADR-19 Hydrate serializes per pack with an flock; temp-download + atomic rename 12-phase3-implementation-notes.md
ADR-20 requireAuth takes variadic scopes; all must be present 12-phase3-implementation-notes.md
ADR-21 Repos are keyed by repo_id everywhere; URL path elements resolve via meta 13-phase4-implementation-notes.md
ADR-22 CAS mismatch = 412, lost race / merge conflict = 409 13-phase4-implementation-notes.md
ADR-23 Write pipeline = per-request scratch odb over the cache, then publish 13-phase4-implementation-notes.md
ADR-24 NDJSON ingestion is incremental with hard per-frame caps 13-phase4-implementation-notes.md
ADR-25 Ref policies enforced API-side through the hook's engine 13-phase4-implementation-notes.md
ADR-26 Result envelope statuses; 499 is not emittable 13-phase4-implementation-notes.md
ADR-27 Jobs are a queue seam — River on Postgres, in-process goroutines in local mode 14-phase5-implementation-notes.md
ADR-28 Push events are emitted by a CASRef decorator, not by handlers 14-phase5-implementation-notes.md
ADR-29 Webhook endpoint management is a self-defined CRUD surface 14-phase5-implementation-notes.md
ADR-30 Upstream sync = mirror fetch + loopback push through our own front door 14-phase5-implementation-notes.md
ADR-31 Forking is synchronous — a one-shot deterministic clone inside POST /repos 14-phase5-implementation-notes.md
ADR-32 Upstream credentials and webhook secrets are stored plaintext 14-phase5-implementation-notes.md
ADR-33 Git-push eventing crosses processes — durable River inserts on pg, synchronous deliveries in local mode 14-phase5-implementation-notes.md
ADR-34 LFS lives on the storage host — presigned transfers, proxy fallback 15-phase6-implementation-notes.md
ADR-35 LFS objects are keyed by base RepoID — namespaces share one object space 15-phase6-implementation-notes.md
ADR-36 Named-repo aliases re-dispatch to the v1 handlers — one implementation, two path styles 15-phase6-implementation-notes.md
ADR-37 Rate limiting is a pre-auth, per-node, per-client-IP token bucket 15-phase6-implementation-notes.md
ADR-38 Runtime image is alpine + external git; the binary is static 16-phase7-implementation-notes.md
ADR-39 Replicas are an opt-in profile of the dev compose file, symmetric by construction 16-phase7-implementation-notes.md
ADR-40 River migrations run under a session-level advisory lock 16-phase7-implementation-notes.md
ADR-41 Registered LFS objects live at a canonical key legacy presigned URLs cannot address issue #51, PR #61
ADR-42 LFS reclamation is mark-and-sweep over reachable pointer files issue #73, PR #117
ADR-43 Absence reconciliation covers registered-but-missing bytes issue #101
ADR-44 Outbound push forwarding rides a separate push-claim lane issue #90
ADR-45 Sync egress policy is unified behind one dial-time guard issue #139
ADR-46 s3wal — refs and metadata ride one conditional-CAS object per repo group issue #133
ADR-47 Cache repos get split commit-graphs, written by full hydrates issue #307
ADR-48 Metrics ride an opt-in, separate listener with fixed labels issue #325
ADR-49 Read-side git children run under one per-request budget issue #313
ADR-50 Local scheduling state is row-owned issue #266
ADR-51 S3WAL metadata v2 is one versioned document, converted by a fenced command issue #335