fix(pipeline): wave resume broken for composition pipelines — aggregate artifacts unregistered + new workspace path

[nextlevelshit]

Two layered bugs make wave resume unusable for any composition pipeline that uses aggregate or iterate

Empirical baseline

ops-pr-respond-20260427-190848-9617 on PR #1407 ran fetch-pr → parallel-review (6 audits) → merge-findings (aggregate) → triage (failed: schema compile error from Perl-only regex (?!, fixed separately). At failure, triage had already produced a valid triaged-findings.json (19 actionable / 3 deferred / 1 rejected) and merge-findings had written merged-findings.json.

wave resume <run-id> then wave resume <run-id> --force both failed:

Error: pipeline execution failed: ❌ Phase 'triage' failed: cannot resume from 'triage':
  prior step 'parallel-review' has no workspace artifacts

Error: pipeline execution failed: step "triage" failed: failed to inject artifacts:
  required artifact 'merged-findings' from step 'merge-findings' not found

Bug 1 — aggregate / iterate steps don't register output artifacts

internal/pipeline/composition.go executeAggregate (line 450-492) writes step.Aggregate.Into to disk and stashes the bytes in tmplCtx.SetStepOutput, but never calls store.RegisterArtifact. Compare internal/pipeline/executor.go:4517-4523 for prompt/command steps.

DB inspection of failed run:

sqlite> SELECT step_id, name FROM artifact WHERE run_id='ops-pr-respond-20260427-190848-9617';
fetch-pr|pr-context
triage|triaged-findings

Only 2 artifacts. Should be 3 (also merge-findings:merged-findings). Same gap likely exists for iterate (no collected-output artifact registered) and branch outputs.

Downstream consequences:

• inject_artifacts in any step depending on an aggregate output can't find it after restart.
• WebUI OUT pills are blank on aggregate/iterate steps (separate symptom of same bug).
• Resume preflight refuses because the dependency artifact is unregistered.

Bug 2 — resume creates a new workspace path

The failed run wrote everything under .agents/workspaces/ops-pr-respond-20260427-190848-9617/. wave resume keeps the run-id in DB but creates a fresh workspace dir .agents/workspaces/ops-pr-respond-20260427-201303-8495/ (timestamp at resume time). The first step run inside the new workspace (triage) then can't see prior step outputs that live under the original path.

Quote from resume error message:

Inspect workspace artifacts:
  ls .agents/workspaces/ops-pr-respond-20260427-201303-8495/

The path printed in the error doesn't match the persisted run-id. Old workspace 190848-9617 still on disk with triaged-findings.json, pr-context.json, etc.

Proposed fixes

Bug 1:

• In composition.go executeAggregate, after os.WriteFile, call c.store.RegisterArtifact(runID, step.ID, name, outputPath, "json", size). Requires plumbing runID into CompositionExecutor (currently only store is held; runID flows in via Execute somewhere — needs a constructor add or per-Execute call).
• Same fix on executeIterate for the implicit collected-output array (collectIterateOutputs).
• Backfill: also register on executeBranch outputs if any.

Bug 2:

• Resume should reuse the original workspace path from pipeline_run.workspace_path or from the most recent checkpoint's workspace_path.
• The checkpoint table already has workspace_path per step — resume can read it and feed the parent run executor a workspace override.
• Alternative: symlink <old-ws>/<step> into the new workspace dir if a per-resume timestamp dir is required for state isolation.

Acceptance

• executeAggregate registers an artifact in DB for every successful aggregate step.
• executeIterate registers a collected-output artifact (or named per step.ID + "/collected-output").
• WebUI shows OUT pill for aggregate/iterate steps populated.
• wave resume <run-id> on a composition pipeline that failed after aggregate/iterate succeeds at finding upstream artifacts — without --force and without re-running prior steps.
• Regression test: ops-pr-respond on a representative PR, kill the run after merge-findings, wave resume — pipeline picks up at triage reading the registered merged-findings artifact.

Related

• feat(pipeline): ops-pr-respond — full PR review-to-resolution composition #1401 (ops-pr-respond) — original feature, called out sub-pipeline OUT-pill DB registration as a known gap.
• audit(webui): composition runs render flat — overview + detail views need tree-aware redesign #1412 (webui composition audit) — same bug class manifests as missing/incorrect OUT pills.
• fix(pipeline): impl-finding mount workspace silently drops commits — git push has no remote #1413 (impl-finding worktree) — current validation blocked by these bugs.

Source

Failed run ops-pr-respond-20260427-190848-9617 on PR #1407 (2026-04-27).

[nextlevelshit]

Status update (2026-04-28)

Both layered bugs were addressed:

PR	Fix
#1441	Aggregate / iterate steps now register their Into / collected-output artifacts in the run-scoped artifact table. wave resume reads step workspaces from the original run's tree instead of fresh empty dir at resume timestamp. WithWorkspaceRunID executor option + workspaceRunIDFor accessor.
#1443	Sub-pipeline inject can resolve persona-step outputs
#1491	Command-step output_artifacts registered in ArtifactPaths (separate but related — filter-scope and file-each-followup were silently shipping 0-byte deliverables to downstream personas)
#1447	Skip stall watchdog for composition steps

Resume preflight should now find both aggregate outputs (merged-findings, etc.) and command-step outputs (scoped-findings, etc.) in the artifact table.

End-to-end re-validation of wave resume on a composition pipeline pending — will run at end of session.

Will close after a successful resume of a composition run.

[nextlevelshit]

Final closeout — all four sub-bugs fixed and validated end-to-end

wave resume for composition pipelines is now functional. Live evidence from this session resuming ops-pr-respond-20260428-191622-c773 past triage and into resolve-and-verify:

Sub-bug	Fix	Evidence
Aggregate / iterate outputs not registered as artifacts	#1441	merge-findings + collected-output now in DB + ArtifactPaths
Workspace heuristic rejected composition + worktree priors	#1508	regression test TestValidateGenericStepSequence_CompositionAndWorktree
Bare-name artifact paths lost on resume → sub-pipeline injection failed	direct main e67a733e	resolve-each loop body found pr-context, fan-out spawned 14+ children
Reaper killed live sub-pipeline children at 5min (no PID, no heartbeat, age fallback)	direct main 26ca1506	apply-fix children completed at 6m6s, 9m21s, 11m48s, 12m39s — would have been reaped pre-fix

Closing. The end-of-loop general_error: context canceled we hit at +30m was a separate concern (errgroup cascade after one adapter hit default_timeout_minutes: 30), addressed by bumping to 90m in commit 92f206c2. Out of scope for this issue — it's a composition-fan-out timeout-budget tuning question, not a resume-correctness bug.