enrich-and-review

Runs the enrichment pipeline over plan chunk directories.

SKILL.md

GitHub

Enrich and Review — Enrichment Pipeline for Plan Stubs

Run the enrichment-review pipeline on a chunk directory: dispatch Sonnet enricher agents, then Opus reviewers, sequentially. Every phase below is a manual EM procedure; worked detail: wiki.

Never a stage of an engineering ask's one submission or its relay fallback: it serves backlog runs and hands-on re-plans.

<!-- Single-sourced at snippets/project-rag-preamble.md; synced by `bin/verify-snippet-sync project-rag-preamble --check/--fix`. Edit the source, never this block. --> <!-- BEGIN project-rag-preamble (synced from snippets/project-rag-preamble.md) -->

Code lookup: project-rag first. ToolSearch("select:mcp__project-rag__project_staleness_check,mcp__project-rag__project_symbol,mcp__project-rag__project_symbol_callers,mcp__project-rag__project_symbol_references,mcp__project-rag__project_symbol_brief,mcp__project-rag__project_referencers,mcp__project-rag__project_semantic_search,mcp__project-rag__project_rag_instructions") project_staleness_check; callers project_symbol_callers/_references; impact project_referencers; else project_rag_instructions. Bodies, members, UE engine types: load project_symbol_source, project_engine_* by name; a project_symbol stub is not "corpus can't answer". If a project-rag call errors or its tools are absent (daemon down), fall back to Grep/Read and carry on. Friction: memo project-rag-em / gh issue create -R dbc-oduffy/project-rag.

<!-- END project-rag-preamble -->

Arguments

$ARGUMENTS: a directory path → chunk directory; specific stub IDs ("2A 2B 2C") → enrich only those; "all" → everything at "Pending enrichment"; --reviewers "name1,name2" → explicit reviewer override, dispatched in listed order (first = domain pass, second = architectural/generalist), replacing Phase 5 auto-detection (PM-directed dual review).

Phase 0: Plan Review Gate

HALT before enriching anything unless the source plan's header carries a **Review:** line matching "Reviewed by [name] on [date]", "Skipped per PM direction", or the byte-exact marker string pinned at coordinator/skills/staff-session/SKILL.md Step 8, item 2 (read it from there, never hand-type a second copy). No match → "This plan has not been through review. Route it through /review first, or confirm PM override to skip."

<!-- engine-gap: field=plan.review_marker_present producer=unknown memo=engine-gap-markers-name-a-memo-that-was-never-filed.md -->

Phase 1: Discover Stubs

Read the tracker README/chunk index; classify each stub at "Pending enrichment" as survey-type (external assets, unfamiliar codebases), plan-type (known codebase, file paths + steps), or manual (non-delegatable). Report the split.

Phase 2: Independence Verification

Before parallel dispatch, build a file→stub map from each stub's Files-Affected/Scope; stubs sharing a file enrich sequentially. Report the parallel/sequential split.

<!-- engine-gap: field=tracker.stub_file_overlap producer=unknown memo=engine-gap-markers-name-a-memo-that-was-never-filed.md -->

Phase 2.5: Write-Ahead Status Update

Before dispatching any enrichers: tracker.advance_status (coordinator_core/ops/tracker/advance_status.py) with tracker_path=<tracker README>, stub_ids=<about-to-be-enriched>, to_status="Enrichment in progress" — flips status atomically, does NOT self-commit. Commit the tracker README change (scoped) immediately after, before any enricher launches — the commit makes the WAL record durable, not the op call.

Negative-spec: this op's write target (a caller-supplied tracker README path) carries no ratified DR carve-out; check the op's "Provisional classification" docstring note first.

Phase 3: Dispatch Enrichers

Do not ask whether to dispatch — invoking this skill IS the request for the dispatch this step names; it dissolves no gate this skill's own body names.

Optional repo map via misc-session-and-guards.py rag-freshness-gate (pass its path in the prompt). Run enabled plugins' root-level enricher-pre-pass.md fragments in your own context and name their artifacts in enricher prompts; fold enricher-survey.md fragments matching project_type into the survey-type prompt (fall back to the enricher's generic protocol if none match).

Independent stubs → parallel Sonnet enrichers (Task, subagent_type: "enricher", model: "sonnet", run_in_background: true, single message). Dependent stubs → sequential. Manual stubs → report to PM/Coordinator.

Phase 4: Resolve Coordinator Flags

Read each enriched stub for NEEDS_COORDINATOR: flags, decide from project context/design docs/PM direction, write the resolution back replacing the flag. Escalate if uncertain.

Phase 4.5: Pre-Review Status Update

Same op as Phase 2.5 with stub_ids=<enriched stubs>, to_status="Under review"; commit immediately after.

Phase 5: Dispatch Reviewers

Do not ask whether to dispatch — invoking this skill IS the request for the dispatch this step names; it dissolves no gate this skill's own body names.

Selection: explicit --reviewers override, or auto-detect (classify enriched-stub work type, merge this plugin's routing.md with every enabled plugin's root-level routing.md fragment, match). Persist findings to the provisioned sidecar, no EM pre-scaffold — each reviewer's brief carries its state/subagent-share/<session>/<provision_key>.md path (provision_report-injected); the reviewer writes ReviewOutput there and returns DONE: <sidecar-path> | verdict: <OK|WARN|BLOCKED> | findings: <N> (detail: wiki).

Sequential with a fix-application gate: dispatch Reviewer 1 (scope = all enriched stubs) → STOP, Reviewer 1 applies and verifies its ledger against the returned sidecar path (every finding, spot-checked diff) → only then dispatch Reviewer 2 on the corrected stubs (fresh sidecar path) → STOP, Reviewer 2 applies and verifies the same way. A single reviewer skips the second pass but keeps the fix-application step. Conflicting feedback: apply unless it conflicts with stated requirements or PM direction, document overrides with rationale in the stub, escalate genuine uncertainty.

The enriched-artifact duty is contract, not brief text. The contract is injected at dispatch for every REVIEWER_ROSTER member. A resolved reviewer whose contract_blocks: row in coordinator/subagent-sandbox-policy.yaml does not list enriched-artifact-review-contract, or that has no row, gets the path coordinator/snippets/enriched-artifact-review-contract.md in its dispatch, never a paraphrase: no brief restates the duty's text (A-DUTY-ONLY-A-BRIEF-CARRIES-DISAPPEARS-WITH-THE-EM).

Phase 6: Update Tracker

Same op with stub_ids=<reviewed stubs>, to_status="Enriched and reviewed", commit immediately after. Note manual-flagged stubs and any needing PM decision.

Completion

Report: stubs enriched, stubs reviewed (and by whom), outstanding flags/PM decisions, which stubs are ready for executor dispatch, and — when this ran unattended — the accepted losses at state/audits/2026-09-07-em-carried-obligations-census/accepted-losses.md, named, not summarised: four duties on this path have no carrier once no EM is in the loop, and the report is where an operator learns the run did not perform them.