Review spec
Dispatch the pre-PR adapter layer against a spec markdown file. Supports two modes — normal (claim-file-driven, lifecycle-integrated) and --bootstrap (standalone, advisory-only, no claim file required). Always runs 2 free Claude subagents (code-reviewer + adversarial); runs Codex CLI + Cursor Cloud Agent if env keys + plan eligibility checks pass; prints copy-paste fenced prompt for optional Composer 2.5 manual paste-back. Prints a progress status table at dispatch and a final aggregated table at completion. Aggregates findings → claim file `review_history` (normal mode) or stdout summary (bootstrap mode). Use when - "review my spec", "check the spec before approving", "run findings on the design doc", "review this draft", "/review-spec --bootstrap <path>".From its SKILL.md
npx -y skills add MarcinSufa/claude-pr-autopilot --skill review-specAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
24.1 KB, as published. Nobody here has run it
/pr-autopilot:review-spec
Pre-PR adapter dispatch on a markdown spec file. Idempotent and re-runnable.
Full algorithm + adapter inventory + cursor-cloud poll contract: docs/superpowers/specs/2026-05-28-pr-autopilot-v0.5-pre-pr-lifecycle-design.md (v2.2 §Pre-PR Adapter Layer + §/review-spec).
v0.5.1 changes (this version): docs/superpowers/specs/2026-05-28-pr-autopilot-v0.5.1-review-spec-improvements.md.
Mode detection (READ FIRST — determines control flow)
If the invocation contains --bootstrap (either --bootstrap=<path> or --bootstrap <path>):
→ go to Bootstrap mode section. Do NOT proceed to "Pre-flight (normal mode)" or "Steps (normal mode)".
Otherwise (no --bootstrap argument):
→ go to Pre-flight (normal mode) section. Do NOT proceed to "Bootstrap mode".
Modes are mutually exclusive. Do not run both. If unclear which mode applies, refuse with:
[pr-autopilot/review-spec] ambiguous invocation; expected either no args (normal mode) or --bootstrap <path>.
Argument parsing protocol
This skill is markdown that Claude interprets. There is no bash getopts. The contract for argument parsing:
- The invocation
/pr-autopilot:review-specwith no arguments → normal mode. - The invocation
/pr-autopilot:review-spec --bootstrap=<path>→ bootstrap mode with<path>as the spec file path. The path is the token immediately following the=sign. - The invocation
/pr-autopilot:review-spec --bootstrap <path>(whitespace-separated) → bootstrap mode with<path>as the spec file path. The path is the first non-flag token after--bootstrap. - If the path token itself begins with
--, Claude MUST treat it as a path (not a flag). E.g.,--bootstrap --weird-name.mdis a path of--weird-name.md. (Rare; documented for completeness.) - If
--bootstrapis followed by nothing (end of invocation), refuse with:[pr-autopilot/review-spec] --bootstrap requires a path argument; got none. - If
--bootstrapis specified TWICE in the same invocation, refuse with:[pr-autopilot/review-spec] duplicate --bootstrap flag; only one path is supported per invocation. - Single-dash
-bootstrapis NOT supported → refuse with[pr-autopilot/review-spec] unknown flag '-bootstrap'; did you mean '--bootstrap'?. - An additional flag
--forcemay appear AFTER the path in bootstrap mode (e.g.,--bootstrap path.md --force). It disables the enforcement guard. See Bootstrap mode B-Pre-flight step 2. - Before running ANY dispatch, Claude MUST verify the path exists via Read tool. If the file does not exist, refuse with:
[pr-autopilot/review-spec] --bootstrap path not found: <path>.
Pre-flight (normal mode)
Required preconditions for normal mode:
- Inside a worktree with
.claude/assignment-claims/<id>.jsonpresent. subStatusisspec_drafting,spec_revising, orspec_review_complete(re-run after manual paste).- Spec file exists at
claimFile.specFileor auto-detect newestspecs/*-<id>.md.
If preconditions fail, refuse with the corresponding error. Do NOT silently fall through to bootstrap mode.
Bootstrap mode (v0.5.1)
Use cases:
- A PR that introduces pr-autopilot to a new repo (bootstrap PR has no assignment yet).
- Reviewing a draft spec before deciding whether to formally assign it.
- Standalone spec review without lifecycle overhead.
B-Pre-flight (bootstrap-specific)
-
The
<path>argument resolves to an existing markdown file (verified via Argument parsing protocol). -
Enforcement guard. Refuse if the current repo is in "real assignment" state. A real assignment state means BOTH conditions hold:
assignments.yamlexists at repo root, AND.claude/assignment-claims/directory exists AND contains at least one.jsonfile.
If both: refuse with:
[pr-autopilot/review-spec] --bootstrap is for repos without active assignments. This repo has assignments.yaml AND active claim file(s). Use /pr-autopilot:assign <id> then /pr-autopilot:review-spec instead. (Override: append --force to this invocation if you really mean to bypass.)Rationale: prevent lazy bypass of the lifecycle. Bootstrap mode is for genuinely-bootstrap state, not a shortcut to skip claim files in established projects.
Exception: if
--forceis appended to the invocation, enforcement is skipped. Use case: testing the skill against a real-assignment repo. Always logged with audit signal[BOOTSTRAP_FORCE](see B-Step 5).
B-Steps (bootstrap-specific)
-
No claim file commits. No subStatus transitions. No git operations on the claim file. The bootstrap mode is read-only with respect to git.
-
Step 1.5 — print initial status table. See
Step 1.5in normal mode "Steps (normal mode)". Same format. Skip the "iter <n>" annotation (bootstrap has no iter tracking). -
Dispatch reviewers identically to normal mode Step 2 (see below). Use the bootstrap
<path>as the<path>placeholder in adapter prompts. -
Composer 2.5 prompt — print to chat per "Step 3" in normal mode. Same format. Path is the bootstrap argument.
-
Aggregate findings. Same logic as normal mode Step 4 (v0.5.2 — 6-column table + Score derivation + outlier footer apply identically). BUT the result is printed to chat ONLY — no claim file write (none exists), no
reviewers[]JSON array persisted. Output structure (same Step 4 final-table format from normal mode):✅ Bootstrap review complete for <path> in <wall-clock-time>s. | Reviewer | Model | Score | Time | Findings (P0/P1/P2) | Verdict | |---|---|---|---|---|---| | feature-dev:code-reviewer | claude-opus-4-7 (feature-dev) | 4/5 | 102s | 0/3/5 | Solid spec; minor polish | | general-purpose adversarial | claude-opus-4-7 (general) | 3/5 | 84s | 3/6/0 | Three blockers | | codex-exec | — | ⏭ skipped | — | — | OPENAI_API_KEY not set | | cursor-cloud-agent | composer-2.5 | <✅ N/5 OR ⏭ skipped (reason)> | <time> | <findings> | <verdict> | | composer-2.5 manual | — | 📋 prompt printed | — | — | Optional paste-back | Aggregate: P0 X, P1 Y, P2 Z, avg <A>/5 over <M> scored channels.Outlier footer (if any single reviewer Score ≤ 2) — same format as normal mode.
If
P0 > 0: also print aggregated P0 list grouped by reviewer (same format as normal mode). IfP0 == 0: print✅ No P0 findings. Spec ready for next step (Marcin approval, then implementation). -
Audit signal. Write a memory to ExoVault via
mcp__exo-vault__write_memory(memoryType:episodic). The memory body MUST be generated by the helper script (NOT constructed in prose by Claude) to guarantee the[BOOTSTRAP_FORCE]token is byte-exact + grep-able for future audits:PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/claude-pr-autopilot/pr-autopilot/0.5.2}" # Without --force: AUDIT_BODY=$(bash "$PLUGIN_ROOT/hooks/bootstrap-force-audit.sh" "<spec-abs-path>") # With --force (B-Pre-flight enforcement guard skipped): AUDIT_BODY=$(bash "$PLUGIN_ROOT/hooks/bootstrap-force-audit.sh" "<spec-abs-path>" force)Then append the reviewer summary + counts to that base body before writing:
<AUDIT_BODY> Reviewers: <list>. Result: <N>P0/<M>P1/<K>P2.Why a script and not prose: per review iter2 finding, prose-only enforcement risks token drift (
[bootstrap-force]vs[BOOTSTRAP_FORCE], etc.) or skip-under-context-pressure. The script hardcodes the exact token and ISO8601 timestamp format. -
Exit clean. No further steps. Re-running with same path is idempotent — same dedup behavior as normal mode (Composer manual paste-back parsed from conversation context, body-hash dedup).
Steps (normal mode)
Step 1 — read claim file + start
Read .claude/assignment-claims/<id>.json. Update subStatus: spec_review_requested. Increment reviewIteration if existing, else set to 1. Commit:
git add .claude/assignment-claims/<id>.json
git commit -m "chore(pr-autopilot): review-spec start for <id> (iter <n>)"
Step 1.5 — print initial status table (v0.5.1, Gap D.1)
Before any dispatch, print to chat:
🔄 **Reviewing spec at `<absolute-path>` (iter <n>)** ...
Dispatching reviewers:
| Reviewer | Type | Status |
|---|---|---|
| feature-dev:code-reviewer | Claude subagent (FREE, always) | ⏳ pending |
| general-purpose adversarial | Claude subagent (FREE, hostile) | ⏳ pending |
| codex-exec | <if env+CLI set: ⏳ pending; else: ⏭ skipped (no OPENAI_API_KEY / codex)> |
| cursor-cloud-agent | <after probe: ⏳ pending (Pro) OR ⏭ skipped (Free/invalid/network)> |
| composer-2.5 manual | FREE paste-back | 📋 prompt will print below |
Then write TodoWrite with a single sentinel item for the whole dispatch (NOT one per reviewer). This respects TodoWrite's hard constraint "Exactly ONE task must be in_progress at any time":
TodoWrite([
{ content: "/review-spec dispatch (iter <n>) — N reviewers in parallel",
activeForm: "Running N reviewers in parallel (iter <n>)",
status: "in_progress" }
])
Where N is the count of channels actually dispatching (e.g., 2 for free-only, 3 with codex, 4 with cursor-cloud-agent on Pro). Per-channel status is shown in the chat status table above; TodoWrite is only the sticky sentinel.
In Step 4, this item flips to completed with activeForm: "Done: N reviewers, X P0 / Y P1 / Z P2".
v0.5.1 limitation: status table updates atomically at Step 4 (since subagent dispatch via Agent tool is foreground/blocking — once a single Agent-call message dispatches multiple subagents in parallel, all complete before control returns). TRUE incremental per-reviewer status updates are v0.6+ (Gap D.2 — see ROADMAP for hooks-based vs MCP-server-based architecture choice). The chat status table above provides the visual map; the single TodoWrite sentinel provides sticky persistence.
Step 2 — dispatch sync adapters in parallel
Graphify dispatch-time filesystem check (v0.5.3+): before dispatching adapters, check whether this repo has a graphify code knowledge graph at graphify-out/graph.json. /review-spec has NO per-PR state file (bootstrap mode has no PR; normal mode hasn't loaded claim state at this point), so the check is purely filesystem-based:
# PR #9 review P1 fix: also gate on `advisory != "off"` so the hint is NEVER injected
# when the user opted out at the config level, even if a graph happens to exist on disk.
# Matches the /step §0.6a contract: advisory=off means "skip entirely" everywhere.
if [ "${config_graphify_advisory:-auto}" != "off" ] \
&& [ -f "graphify-out/graph.json" ] \
&& [ "${config_graphify_promptHint:-true}" = "true" ]; then
_graphifyHintEnabled="true"
else
_graphifyHintEnabled="false"
fi
When _graphifyHintEnabled == "true", the claude-code-reviewer-subagent and claude-self-review prompts (see "Adapter prompts" below) get a graphify hint prepended that tells the subagent to query the graph before grep'ing files.
For cursor-cloud-agent, run the probe FIRST to determine plan eligibility:
# Fallback CLAUDE_PLUGIN_ROOT for manual-copy contexts (per Composer review iter3 P2-2):
# matches the pattern documented in templates/settings-snippet.json
bash "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/claude-pr-autopilot/pr-autopilot/0.5.2}/hooks/cursor-cloud-agent-probe.sh"
PROBE_EXIT=$?
Based on PROBE_EXIT:
- 0 → Cursor Pro available. Will dispatch cursor-cloud-agent below.
- 42 → Cursor Free (plan_required). Skip cursor-cloud-agent. Append to
reviewers[]:{kind:"cursor-cloud-agent", status:"skipped", reason:"plan_required", iteration:<n>}. Echo:ℹ️ Cursor Cloud Agent skipped — requires Cursor Pro. Upgrade: https://cursor.com/settings/billing. - 43 → API key invalid or missing. Skip with louder warning. Append
reason:"invalid_key". Echo:⚠️ Cursor Cloud Agent skipped — API key invalid or missing. Check ~/.claude/settings.json env.CURSOR_API_KEY. - 44 → Network / parse / other error. Skip with retry hint. Append
reason:"probe_error". Echo:ℹ️ Cursor Cloud Agent probe failed (network/parse). Probe re-runs on next /review-spec invocation; no caching. - 45 (v0.5.2) → Cursor account setting blocks Cloud Agent (Privacy Mode Legacy currently). Append
reason:"privacy_mode_legacy". Echo:⚠️ Cursor Cloud Agent skipped — Privacy Mode (Legacy) blocks Cloud Agent. Disable in Cursor Settings → Privacy.
skipReason union (v0.5.2 extension): plan_required | invalid_key | probe_error | no_env | privacy_mode_legacy.
Now dispatch all enabled adapters in parallel via a single Agent-tool-call message containing multiple invocations:
Always (FREE):
- claude-code-reviewer-subagent:
Agent({ subagent_type: 'feature-dev:code-reviewer', description: 'Pre-PR spec review', prompt: <see "Adapter prompts" below, with <path> = claimFile.specFile or bootstrap argument> }) - claude-self-review (hostile, FREE):
Agent({ subagent_type: 'general-purpose', description: 'Adversarial spec review', prompt: <hostile template, same path substitution> })
Opt-in PAID (if env set):
- codex-exec if
OPENAI_API_KEYset ORwhich codexsucceeds:codex exec --json --sandbox read-only --ask-for-approval never \ --output-last-message "/tmp/codex-review-<id>-iter<n>.md" \ "$(cat <prompt-template-with-spec-path>)" - cursor-cloud-agent (ONLY if probe returned 0):
If timeout → recordRUN=$(curl -sX POST https://api.cursor.com/v1/agents \ -H "Authorization: Bearer $CURSOR_API_KEY" \ -H "Content-Type: application/json" \ -d "{\"prompt\":{\"text\":\"<prompt+spec>\"},\"repos\":[{\"url\":\"<repo-url>\"}],\"model\":{\"id\":\"composer-2.5\"},\"autoCreatePR\":false}") # Poll run.id every 5s up to 120s{kind: "cursor-cloud-agent", status: "pending", runId: "<id>", timeoutAt: "<iso>"}inreviewers[]. Re-run/review-specfolds in pending result idempotently.
Note on URL contract: the dispatch always uses production https://api.cursor.com/v1/agents. There is NO env-var override on dispatch. The probe accepts CURSOR_API_URL override ONLY when PR_AUTOPILOT_TEST_MODE=1 is also set (test-mode-only path). This asymmetry is intentional: tests validate probe LOGIC; dispatch is exercised by EVAL 43 ("all keys set" — requires Cursor Pro available) and EVAL 50 (Free graceful skip) as live integration tests, not by mocked unit tests.
Step 3 — Composer 2.5 manual prompt (always — FREE fallback)
Print to chat:
🔵 **OPTIONAL: Composer 2.5 review** — for a 3rd perspective beyond the 2 Claude subagents, copy the block below, open Cursor (Cmd+I), paste, run, paste reply back into Claude.
Then immediately print a single triple-backtick fenced code block. The block content must have the spec path baked in at skill runtime (no <path> placeholder for user to edit):
```
Review the spec at <ABSOLUTE-PATH-RESOLVED-AT-SKILL-RUNTIME> for:
- internal contradictions
- missing edge cases / failure modes
- scope creep vs declared scope_in/scope_out
- alignment with CLAUDE.md + DESIGN.md (read both first)
- test coverage gaps
- security/correctness for code paths described
Return P0/P1/P2 findings with confidence ratings (5/10-10/10) per PUSHBACK.md.
Format: structured markdown table or per-finding section.
```
After Composer responds: user pastes its reply back into this Claude chat, then re-runs /pr-autopilot:review-spec (or --bootstrap with same path). Findings auto-dedup by body hash.
(Advisory only — does NOT block spec_review_complete transition.)
Step 4 — aggregate findings + print final status table
Aggregate findings from all sync adapters that completed. Increment reviewIteration if not already done in Step 1. Append entries to reviewers[] in claim file:
{
"kind": "claude-code-reviewer-subagent" | "claude-self-review" | "codex-exec" | "cursor-cloud-agent" | "composer-2.5-manual",
"status": "complete" | "pending" | "timeout" | "skipped",
"iteration": <n>,
"findings": { "p0": N, "p1": M, "p2": K, "details": "..." },
"completedAt": "<iso>",
"skipReason": "plan_required" | "invalid_key" | "probe_error" | "no_env" | "privacy_mode_legacy" // only if status: skipped
}
Score derivation (v0.5.2 — DERIVED only, no adapter prompt changes):
For each reviewer with status: "complete", derive a 1-5 score from P0/P1/P2 counts using this rubric:
| Conditions | Score |
|---|---|
| 0 P0 + 0 P1 + ≤4 P2 | 5/5 (ship-ready) |
| 0 P0 + 0 P1 + ≥5 P2 | 4/5 (lots of nits) |
| 0 P0 + 1-2 P1 | 4/5 (minor polish) |
| 0 P0 + ≥3 P1 | 3/5 (revise — lots of P1) |
| 1 P0 | 3/5 (1 blocker) |
| 2 P0 | 2/5 (significant issues) |
| ≥3 P0 | 1/5 (block — fundamental) |
Skipped, pending, timeout, manual reviewers (— rows) → no Score; excluded from aggregate mean.
Model column sourcing (per kind):
| Reviewer kind | Source | Default |
|---|---|---|
claude-code-reviewer-subagent | hardcoded by subagent_type | claude-opus-4-7 (feature-dev) |
claude-self-review | hardcoded by subagent_type | claude-opus-4-7 (general) |
codex-exec | parse JSON .usage.model or .model | codex (model unknown) |
cursor-cloud-agent | parse API .run.model.id or .agent.model.id | composer-2.5 (we requested it) |
composer-2.5-manual | N/A | — |
Verdict extraction: first sentence of the reviewer's response. If extraction fails or reviewer omitted a summary, synthesize: 0 P0 → Spec ready, N P1, K P2.; 1 P0 → One blocker — revise.; ≥2 P0 → N blockers — significant revision needed.. Truncate to ≤60 chars (append … if cut).
Print final status table to chat (v0.5.2 — 6 columns):
✅ **Review iter <n> complete in <wall-clock-time>s.**
| Reviewer | Model | Score | Time | Findings (P0/P1/P2) | Verdict |
|---|---|---|---|---|---|
| feature-dev:code-reviewer | claude-opus-4-7 (feature-dev) | 4/5 | 102s | 0/3/5 | Solid spec; minor polish on §5.3 |
| general-purpose adversarial | claude-opus-4-7 (general) | 3/5 | 84s | 3/6/0 | Three blockers; revise before TDD |
| cursor-cloud-agent | composer-2.5 | 5/5 | 67s | 0/0/2 | Looks ready to ship |
| codex-exec | — | ⏭ skipped | — | — | OPENAI_API_KEY not set |
| composer-2.5 manual | — | 📋 pending | — | — | Optional paste-back |
**Aggregate: P0 X, P1 Y, P2 Z, avg <A>/5 over <M> scored channels** → <decision: spec_review_complete OR spec_revising>
Status column absorbed into Score column (v0.5.2): instead of a separate Status column, the Score cell renders ⏭ skipped (reason), 📋 pending, ❌ failed, ⏱ timeout for non-numeric states. Only completed reviewers get a numeric N/5.
Outlier footer (v0.5.2): if ANY single reviewer's derived Score is ≤ 2, append BELOW the aggregate line:
⚠️ Low score outlier: <reviewer-kind> <N/5> (<P0>P0). Aggregate avg hides blocker.
This surfaces the blocker that the mean would otherwise mask (e.g., 4 reviewers at 5/5/5/1 → mean 4.0/5 looks safe but the 1/5 is a real blocker).
Bootstrap mode parity (v0.5.2): identical 6-column table + outlier footer + Score derivation apply in bootstrap mode (B-Step 5). No claim-file write in bootstrap — table printed to chat only.
Update the single TodoWrite sentinel item added in Step 1.5 to completed:
TodoWrite([
{ content: "/review-spec dispatch (iter <n>) — N reviewers in parallel",
activeForm: "Done: N reviewers, X P0 / Y P1 / Z P2 (in <wallclock>s)",
status: "completed" }
])
The per-reviewer outcomes (skipped, failed, completed counts) are visible in the chat status table above. Skipped channels are reflected in the aggregate P0/P1/P2 counts (counted as 0 findings) and the reviewers[] JSON array on the claim file. The TodoWrite sentinel is intentionally summary-level — Claude's chat sidebar shows it as one completed item with the aggregate counts in its label.
Step 5 — Composer manual paste handling
If conversation context contains a pasted Composer reply since the last /review-spec invocation, parse it, dedup by body hash (do not duplicate entries), append as {kind: "composer-2.5-manual", status: "complete", iteration: <n>, ...}. Advisory only — re-aggregation may flip state if new P0 surfaces.
Step 6 — decision based on aggregated P0 count
-
0 P0 →
subStatus: spec_review_complete. Echo (after the final status table from Step 4):📋 Spec ready for next step. Marcin: when ready, run /pr-autopilot:approve-spec. (Optional: paste Composer 2.5 findings + re-run /review-spec for additional perspective.) -
≥1 P0 →
subStatus: spec_revising. Echo aggregated P0 list grouped by reviewer (after the final status table):⚠️ Spec revision needed. <N> P0 findings: [P0 from <reviewer-kind>, confidence 8/10] <title> — <body> ... Agent: address findings, re-run /pr-autopilot:review-spec.
Step 7 — commit claim file update
git add .claude/assignment-claims/<id>.json
git commit -m "chore(pr-autopilot): review-spec iter <n> for <id> — <N> P0, <M> P1"
Adapter prompts (PUSHBACK.md rubric)
Graphify hint prefix (v0.5.3+): when _graphifyHintEnabled == "true" (set in Step 2), prepend the following block to BOTH claude-code-reviewer-subagent and claude-self-review prompts:
**Code knowledge graph available:** This repo has a graphify-built knowledge graph at
`graphify-out/graph.json`. BEFORE grep'ing for symbols or reading source files, query
the graph: `graphify explain "<symbol>"` returns the node + connections + community
in ~1-3k tokens vs ~30-100k for a multi-file grep. Use `graphify path "A" "B"` for
dependency-trace. If `graphify` errors with "command not found" (CLI not installed
locally), fall back to grep + Read without retrying.
When _graphifyHintEnabled == "false", the hint block is omitted entirely (no instruction injection).
claude-code-reviewer-subagent prompt template:
Review the spec at <path> for engineering rigor:
- internal contradictions
- missing edge cases / failure modes
- scope creep vs scope_in/scope_out
- alignment with CLAUDE.md + DESIGN.md (read both first)
- test coverage gaps
- security/correctness for any code paths described
Return P0/P1/P2 findings with confidence ratings (5/10-10/10) per PUSHBACK.md rubric.
Format: structured markdown table or per-finding section.
claude-self-review (hostile) prompt template:
You are a HOSTILE reviewer reading the spec at <path>. Find WEAKNESSES:
- what is underspecified / hand-wavy
- what assumptions might be wrong
- what edge cases / error paths / race conditions are missed
- what scope creep is hidden in scope_in
- what tests would actually catch the failures the author hand-waved
Be ruthless. Do NOT be polite. Tighten the spec by finding weaknesses,
not by validating the author's choices.
Return P0/P1 ONLY (no P2 noise) with confidence ratings.
Idempotency
- Normal mode: re-running
/review-specafterspec_review_completeis safe — it re-dispatches reviewers and may flip back tospec_revisingif new P0 appear. - Bootstrap mode: re-running with same path is safe — prints fresh status tables + aggregated findings to chat. No state persistence between runs (no claim file).
- Pending Cursor Cloud Agent runs get re-polled on re-run.
- Manual Composer paste dedup'd by body hash across iterations.
- TodoWrite items from a previous run are overwritten by Step 1.5's fresh TodoWrite call (idempotent re-render).