Repository navigation
feat(lane-transcripts): open a leg's sealed telemetry into the sink - #17403
Conversation
The puller opens otel.jsonl.gz.sealed, when an artifact carries it beside the transcript, under the payload name otel.jsonl.gz with the transcript's run and job binding, and stores it next to the transcript. The row gains telemetry, telemetry_outcome, telemetry_bytes, telemetry_spans and telemetry_joined_spans, the last counting the spans whose traceId is the row's trace_id. Telemetry that does not open keeps the transcript. Prune removes a row's telemetry with its transcript and clears a missing path. Refs: #17261 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…be read Telemetry that fails to extract from the zip, or decompresses past a 64 MiB cap of its own, is recorded as telemetry_outcome unreadable and the transcript row is still written, so one bad telemetry entry no longer turns the artifact into a fetch failure or aborts the pull. Spans are counted line by line rather than through one split array. Refs: #17261 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Lane run report
Failed GitHub API calls
|
|
claude-review reviewed Code review (incremental, blocking only): PR #17403,
|
|
claude-security-review reviewed Security review (incremental, blocking only): no findingsThis review covers only what changed since
The new wording now says that What I checked:
This review was blocking only, because this lane already reviewed an earlier head. No SUGGESTION findings were dropped. No spend change. critical-findings: 0 |
…on failure A telemetry entry that unzip cannot extract also yields `unreadable` without ever being opened, so the contract row and the openTelemetry JSDoc now say so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Refs: #17261
Summary
The transcript puller opens a leg's sealed telemetry and writes it, with its trace join, to the sink. This is the second half of #17261. #17400 opts the lane sweep into the sealed receiver and uploads
otel.jsonl.gz.sealedbeside the transcript, and the scheduled re-pin #17376 moved the pins to ci-workflows v0.76.3. Before this PR, the puller accepted that second sealed file and left it unopened (pull.mjs,EXTRA_SEALED).Fix
pull.mjs. When an artifact's one extra entry isotel.jsonl.gz.sealed, the pull opens it under the payload nameotel.jsonl.gz, with the transcript's run id and job id binding. It stores the file, gzipped, as<run id>-<attempt>-<job id>.otel.jsonl.gzbeside the transcript (mode 0600). Telemetry that does not open, or does not decompress, is not stored, and the transcript is kept.Row fields. Every pulled row gains these fields:
telemetry: the path, or null;telemetry_outcome:stored,none,sealed_mismatchorunreadable;telemetry_bytes;telemetry_spans;telemetry_joined_spans: the spans whosetraceIdis the row'strace_id.The last field is the trace join. run-claude-cli passes the CLI a
TRACEPARENTunder the lane-outcome trace id (ci-workflows#1136), so a row whose joined count equals its span count shows that every span joined its run attempt.store.mjs.telemetryPathandtelemetrySpansare new.prunemaps telemetry files to their rows, so a 30-day or cap prune of a row's telemetry setstelemetry: null, and a missing telemetry file clears its path.Docs. The lane-transcripts contract and the puller README list the new file and fields. The changes add fields only: the contract lets a writer add fields, and readers treat an unknown field as absent.
The transcript survives bad telemetry. These findings came from the security review. Telemetry that fails to extract from the zip, or decompresses past a 64 MiB cap of its own, is recorded as
unreadable, and the transcript row is still written. This holds the 60 MiB that run-claude-cli's receiver keeps (otlp-receiver.cjs,MAX_TELEMETRY_BYTES). Spans are counted line by line, so a payload of nothing but newlines cannot build one huge array and abort the pull.Telemetry text, like transcript text, reaches only the sink's files. Nothing prints it, and the metrics row holds counts, never content.
Verification
node --test --test-concurrency=4 scripts/lane-transcripts/pull.test.mjs: 60 tests, 60 passed, 0 failed..github/actions/resolve-config/callers.test.mjs, which importsstore.mjs: 290 tests, 290 passed, 0 failed. Five tests are new, and one existing test gained an assertion:unreadable, and its transcript row is kept.sealed_mismatch, and the transcript is kept. A transcript alone isnone.telemetrySpansskips a line that is not JSON.sealed_mismatch.lane_runsDuckDB view (plugins/harness-ops/skills/observability/otel/lane-runs.sql) names its columns, so the new fields are ignored there. The contract says the session-flow reader treats an unknown field as absent.Remaining
maintenance-audit-lane-runs.ymlon main, runpull, and confirm that a row from that run hastelemetry_outcome: storedand atelemetry_joined_spansabove 0. ci(lanes): re-pin to ci-workflows v0.76.0 and opt a lane into the sealed telemetry receiver #17261 closes after that check.pr-pipeline-sandboxsyncsscripts/lane-transcripts/store.mjsbyte for byte (scripts/lanes/check-sync.shthere), so its next resync takes this file at the merged sha.Related
otel.jsonl.gz.sealed.🤖 Generated with Claude Code