Enterprise session audit

Product UI: open a completed session → Review audit panel (clone/mount receipt, specialist ledger, Download audit JSON). See UI guide — session blade.

Review audit panel with specialist ledger

Problem

Regulated customers need more than a findings list: proof of what was reviewed, by which specialist, with what tools/context, confidence, and verdict chain — recoverable after worker restarts.

Expert panel (roles)

Expert	Concern	Design decision
Compliance / GRC	Tamper-evident provenance	Specialist responses stored as SHA-256 + redacted excerpt; context receipt (clone/mount, verified SHA)
Security architect	No secret leakage	Scrub tokens in notes/excerpts; never log clone tokens
Staff eng (platform)	Multi-worker safety	Postgres `FOR UPDATE SKIP LOCKED` + job lease reclaim for crash recovery
SRE	Resume after pod kill	Checkpoint units + re-enqueue incomplete sessions on worker boot
Product / UX	Human report	Markdown session report + per-subagent appendix + Download .md
QA	Observable runs	`specialistRuns[]` with paths, files, tools, findingsSummary, avgConfidence

Data model

SessionAudit
  context: ContextReceipt          # code plane binding
  specialistRuns[]: SpecialistRun  # one per subagent step
  tools: ToolTraceSummary          # graph/sandbox calls
  judge: JudgeNoiseSummary
  timings?: SessionTimings         # wall clocks for bottleneck analysis
  zeroFindings?: …
  heal?: …

SpecialistRun (per subagent step)
  role, unitId, unitLabel, model, runner
  startedAt, endedAt, durationMs, status
  pathsReviewed[], filesReviewed[]
  toolCallCount, usedGraph
  findingCount, findingsSummary[{title,severity,confidence,path,line}]
  avgConfidence
  responseSha256, responseExcerpt (redacted)
  stepIndex

SessionTimings
  sessionStartedAt, sessionEndedAt, totalDurationMs
  stages[]: { stage, startedAt, endedAt, durationMs, message }
  units[]:  { unitId, label, durationMs, roles, specialistMaxMs, status }
  summary:  longestStage/Unit/Specialist, byStageMs, specialistRunsSumMs, toolsSumMs

Stored on review_sessions.metadata.audit (+ live session.audit).
Also denormalized as metadata.timings for analytics without unpacking the full audit.
Report: metadata.report.markdown includes Timing / bottlenecks + Subagent / specialist audit trail.

Resume / self-heal after worker crash

Checkpoint after planning/units (session_checkpoints + file mirror).
On worker start: reclaimStale → running jobs with expired lease → pending.
resumeIncompleteSessions enqueues resumable sessions (not blocked by zombie running rows).
Claim path also adopts expired leases (STEW_JOB_LEASE_MS, default 120s).
Orchestrator resume: true skips completed unit ids from checkpoint.

Controls

Env	Meaning
`STEW_JOB_LEASE_MS`	Stale lock age before reclaim (default 120000)
`STEW_SESSION_REPORT_LLM`	`0` off; default on for stewardship narrative
`STEW_INLINE_WORKER`	`0` = external workers only

Non-goals (v1)

Cryptographic notarization / WORM storage (future: S3 Object Lock)
Full raw LLM transcript retention by default (cost/PII) — hash + excerpt only

Problem​

Expert panel (roles)​

Data model​

Resume / self-heal after worker crash​

Controls​

Non-goals (v1)​