Skip to content

Feat/task memory evidence pipeline - #63

Merged
AnzoBenjamin merged 3 commits into
mainfrom
feat/task-memory-evidence-pipeline
Aug 24, 2026
Merged

Feat/task memory evidence pipeline#63
AnzoBenjamin merged 3 commits into
mainfrom
feat/task-memory-evidence-pipeline

Conversation

@AnzoBenjamin

@AnzoBenjamin AnzoBenjamin commented Aug 24, 2026

Copy link
Copy Markdown
Owner

This change is Reviewable

SDK file mutations reported failures with console.error, so hosts that own a structured logger never saw them and the diagnostic bypassed redaction. An optional logger now threads from run.ts through changeFile/changeFiles, replaceRange and writeAuditFindings, and the emitted payload carries only path, change type, byte length and error code.

replace_range now splices on the raw file bytes so line terminators outside the edited range survive an edit, pins a byte-exact commit expectation alongside the LF-normalized freshness hash so a CRLF<->LF-only external rewrite cannot be silently reverted, and returns its declared { file, errorMessage } shape for oversize, unparsed and out-of-project input instead of throwing a ZodError.

write_audit_findings tightens its input schema at the boundary that feeds structuralReceipt: single-line and control-character hygiene on coverage entries and finding paths, trimming plus uniqueness across every coverage list, identifier bounds on sessionSlug/shardId/snapshotId, non-empty coverage on snapshot-bound calls, and the rules stated on the fields themselves since every rejection collapses to one generic message. changeFile also reports the attempted action from the state it observed in-lock and cancels its operation registration on every pre-commit failure path.

Adds sdk/src/__tests__/run-mutation-dispatch.test.ts covering the dispatch preconditions and the compact-receipt mutation side channel, and expands the change-file, replace-range and write-audit-findings suites over the new behaviour.
Buffer concurrent tool evidence per step and flush once to avoid revision clobbering and repeated normalize/checksum; programmatic calls commit immediately. Capture request goal for root agent and compile with focus paths so relevance outranks recency under budget. Keep hash-verified reads across workspace revisions, guard pathless evidence by revision, and cap evidence per kind to prevent review churn evicting reads. Fix read anchor and mutation hash canonicalization so recorded evidence reconciles fresh next session; persist aborted runs and unify compact-receipt mutation selection via findFileMutationResult with shared changedActions between journal and FilesystemMutationEvent.
@vercel

vercel Bot commented Aug 24, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
openbuff Ready Ready Preview, v0 Aug 24, 2026 3:44pm

@AnzoBenjamin
AnzoBenjamin merged commit b9bdcbe into main Aug 24, 2026
16 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant