flow-review

name	flow-review
description	Phase 3: Review — six tenants assessed by four cognitively isolated agents (reviewer, pre-mortem, adversarial, documentation) launched in parallel. Parent session gathers context, triages findings, and fixes.

FLOW Review — Phase 3: Review

Usage

/flow:flow-review
/flow:flow-review --continue-step

• /flow:flow-review — uses the configured mode from the state file's skills.flow-review config
• /flow:flow-review --continue-step — self-invocation: skip Announce and Update State, dispatch to the next step via Resume Check

Run `phase-enter` as your very first action. If it returns an error, stop immediately and show the error to the user.

${CLAUDE_PLUGIN_ROOT}/bin/flow phase-enter --phase flow-review --steps-total 4

Parse the JSON output. If "status": "error", STOP and show the error.

If "status": "ok", capture the returned fields: project_root, branch, worktree_path, worktree_cwd, relative_cwd, pr_number, pr_url, feature, slack_thread_ts, and plan_file. The autonomy mode is resolved separately in the Mode Resolution section below via resolve-skill-mode.

Use the returned fields for all downstream references. Do not re-read the state file or re-run git commands to gather the same information. Do not cd to the project root — bin/flow commands find paths internally.

Re-anchor cwd

Mono-repo flows started inside a subdirectory (e.g. api/) capture that path as relative_cwd and rely on cwd staying at <worktree>/<relative_cwd> so subsequent bin/flow calls pass the cwd-drift guard. Context loss between skill invocations can reset cwd to the main repo root; the bash block below re-anchors regardless of how the session got here. Substitute the worktree_cwd value from the phase-enter response — a no-op for root-level flows (where it equals worktree_path) and a real re-anchor for mono-repo flows.

cd "<worktree_cwd>"

Six Tenants

The Review phase assesses the work through six tenants. Every finding from every agent must map to one of these tenants. Findings that do not map to a tenant are dropped.

Tenant 1 — Architecture. Does the code follow the project's conventions, rules, and planned approach? Deviations from CLAUDE.md, .claude/rules/, and the implementation plan are findings.

Tenant 2 — Simplicity. Is there unnecessary complexity? Duplicated logic, missed abstractions, over-engineering, conditionals that could be flattened, names that could be clearer.

Tenant 3 — Maintainability. Can a newcomer understand this code without context from the conversation that produced it? Implicit assumptions, undocumented patterns, names that only make sense with tribal knowledge.

Tenant 4 — Correctness. Does the code actually work? Logic errors, edge cases, off-by-one errors, null handling gaps, error propagation, race conditions, and security vulnerabilities (injection, auth bypass, data exposure).

Tenant 5 — Test coverage. Every production line must be exercised by a named test. Any uncovered line is a Real finding. Meaningful assertions, edge cases covered, error paths exercised. Gaps are proven by adversarial tests that fail.

Tenant 6 — Documentation. Do the docs match the code after these changes? CLAUDE.md, .claude/rules/, README, doc comments, and inline comments that no longer reflect the code's actual behavior.

Concurrency

This flow is one of potentially many running simultaneously — on this machine (multiple worktrees) and across machines (multiple engineers). Your state file (.flow-states/<branch>/state.json) is yours alone. Never read or write another branch's state. All local artifacts (logs, plan files, temp files) are scoped by branch name. GitHub state (PRs, issues, labels) is shared across all engineers — operations that create or modify shared state must be idempotent.

Mode Resolution

Resolve commit and continue on every entry — fresh invocation and --continue-step self-invocation alike — from the state file's skills.flow-review config via resolve-skill-mode. The state file is the single source of truth for skill autonomy; there are no --auto/--manual flags.

On a --continue-step self-invocation, recover the worktree directory before resolving the branch. The resume path skips phase-enter (which normally cds into the worktree), and the branch resolution just below is cwd-dependent — so a session whose cwd reset to the main-repo root would otherwise resolve the integration branch instead of the feature branch. bin/flow resume-anchor reads the session-keyed phase-anchor marker and returns the recovered worktree_cwd:

${CLAUDE_PLUGIN_ROOT}/bin/flow resume-anchor

Parse the JSON output and branch on status:

• "ok" — cd into the returned worktree_cwd, then resolve the branch below from the recovered directory.
• "no_marker" — no marker to recover; proceed with the cwd-based branch detection below as-is.
• "error" — the marker was corrupt; do NOT cd to any returned path. Treat it exactly like no_marker and proceed with the cwd-based detection below.

Resolve the current branch first: run git worktree list --porcelain, note the project root (the path on the first worktree line), find the worktree entry whose path matches the current working directory, and take the branch refs/heads/<name> line from that entry (strip the refs/heads/ prefix). Call this <branch>. Then run the resolver:

${CLAUDE_PLUGIN_ROOT}/bin/flow resolve-skill-mode --skill flow-review --branch <branch>

Parse the JSON output. commit and continue are each "auto" or "manual":

• commit=auto — auto-fix and auto-commit all findings.
• commit=manual — require explicit approval of changes and routing decisions.
• continue=auto — auto-advance to Learn when Review completes.
• continue=manual — prompt before advancing to Learn.

Self-Invocation Check

If --continue-step was passed, this is a self-invocation from a previous step. Skip the Announce banner and the phase-enter call (do not enter the phase again). Run ## Mode Resolution above (it runs on every entry), then proceed directly to the Resume Check section.

Announce

At the very start, output the following banner in your response (not via Bash) inside a fenced code block:

```text
──────────────────────────────────────────────────
  FLOW v2.7.0 — Phase 3: Review — STARTING
──────────────────────────────────────────────────
```

Logging

After every Bash command completes, log it to .flow-states/<branch>/log using bin/flow log.

Run the command first, then log the result. Pipeline the log call with the next command where possible (run both in parallel in one response).

${CLAUDE_PLUGIN_ROOT}/bin/flow log <branch> "[Phase 3] Step X — desc (exit EC)"

Get <branch> from the state file.

Resume Check

Read review_step from the state file (default 0 if absent).

• If 1 — Step 1 is done. Skip to Step 2.
• If 2 — Steps 1-2 are done. Skip to Step 3.
• If 3 — Steps 1-3 are done. Skip to Step 4.
• If 4 — All steps are done. Skip to Done.

---

Step 1 — Gather

Collect all artifacts needed by the agents. No analysis — just artifact collection.

Read the plan file. Read files.plan from the state file to get the plan file path. Use the Read tool to read the plan file at <project_root>/<files.plan path> — the .flow-states/ tree lives at the project root, not inside the worktree, so the <project_root>/ prefix is required (a raw relative read resolves under the worktree and the validate-worktree-paths hook blocks it).

Read project conventions. Use the Read tool to read the project CLAUDE.md at <worktree_path>/CLAUDE.md. Use the Glob tool to find all .claude/rules/*.md files at <worktree_path>/.claude/rules/*.md, then read each file.

Resolve the integration branch. Before constructing the diff ranges, run bin/flow base-branch to retrieve the base branch the flow coordinates against (the integration branch captured at flow-start). Capture its stdout — call the value <base_branch> — and substitute it into the diff commands below. A repo whose default branch is staging produces <base_branch> = staging; a standard repo produces <base_branch> = main.

${CLAUDE_PLUGIN_ROOT}/bin/flow base-branch

Capture both diffs to files. bin/flow capture-diff runs both git diff origin/<base_branch>...HEAD (full) and the -w variant (substantive — whitespace-only changes filtered out) and writes the results to canonical paths under .flow-states/<branch>/. The agents read those files via the Read tool instead of receiving the diff bytes inline in their prompts, keeping the parent skill's prompt budget bounded as PR size grows. Substitute <branch> with the flow's branch name and <base_branch> with the value captured above.

${CLAUDE_PLUGIN_ROOT}/bin/flow capture-diff --branch <branch> --base <base_branch>

Parse the JSON output. The full field is the path to the full diff file — call it <full_diff_file> and pass it to the reviewer agent (context-rich). The substantive field is the path to the substantive diff file — call it <substantive_diff_file> and pass it to the pre-mortem, adversarial, and documentation agents (context-sparse). On PRs where formatters (cargo fmt, prettier, black) reformat many files, the substantive diff excludes formatting noise and preserves the agents' turn budget for behavioral analysis.

On capture-diff error. When the JSON status is "error", surface the message. A missing-revision message — the base ref origin/<base_branch> is not present in this worktree (git reports "unknown revision" or "ambiguous argument") — means the integration ref was never fetched here. Fetch it once and retry capture-diff:

git fetch origin <base_branch>

Then re-run the capture-diff command above exactly once. If the retry still returns status == "error", HALT with a structured error and report the message — do not launch the agents with a missing diff.

Compute affected doc paths.

The documentation agent's turn budget is bounded; pointing it at the full <worktree_path>/docs/ tree on moderately-sized PRs causes autocompact-thrash before findings are produced. The skill computes a narrowed list of doc paths likely affected by the diff via filename heuristics, and the agent investigates ONLY those paths.

Capture the changed-file list:

git diff --name-only origin/<base_branch>...HEAD

From the changed-file list, derive <doc_paths> (a list of paths under <worktree_path> to embed inline in the documentation agent prompt):

• For each skills/<name>/SKILL.md in the diff → add <worktree_path>/docs/skills/<name>.md
• For each phase skill (skills/flow-<phase>/SKILL.md) in the diff → also add <worktree_path>/docs/phases/phase-<N>-<phase>.md (look up <N> from flow-phases.json — phase order is the array index plus one)
• For each .claude/rules/*.md in the diff → include the rule path itself so the agent can check cross-references in sibling rule files
• Always include <worktree_path>/CLAUDE.md
• If the diff touches src/state.rs, src/phase_*.rs, or flow-phases.json → add <worktree_path>/docs/reference/flow-state-schema.md
• If the diff touches agents/*.md → add <worktree_path>/docs/reference/agents.md if it exists

Capture this list as <doc_paths> for use in Step 2. The narrowed list is exhaustive for documentation drift in this PR — the documentation agent does not need to walk the full docs tree.

Derive adversarial test setup.

The adversarial agent writes a single probe test file inside the project's test tree so the language test runner can discover and execute it. The exact path is owned by the project — declared via the project's bin/test --adversarial-path invocation — the same way each project owns the four bin/{format,lint,build,test} toolchain decisions. Worktree removal at Phase 4 Complete disposes of the probe as a side effect of removing the worktree directory, so no separate cleanup hook is needed.

Run this from the current working directory (the agent's worktree_cwd captured at flow-start — the worktree root for project-root flows, or the service subdirectory .worktrees/<branch>/<service>/ for mono-repo flows) and capture stdout:

bin/test --adversarial-path

Strip trailing whitespace (newline, spaces, tabs) from the captured stdout before using the value — bin/test is project- owned bash that may print the path via echo (which appends a newline) or include trailing whitespace from line-continuation quoting. The contract is a single-line path; the skill normalizes defensively.

If `bin/test` exits 2, surface the stderr message and halt — the project must configure `bin/test --adversarial-path` (uncomment the runner block and set the matching path comment in `bin/test`) before Review can run. Do NOT proceed to Step 2.

This is an infrastructure halt, not a decision point. There is a single option only — fix the unconfigured stub: uncomment the runner block in bin/test and set the matching path comment, then re-run Review. Do NOT enumerate alternatives ("(1) skip the adversarial agent, (2) abort the workflow, (3) configure bin/test") — every non-fix path silently weakens the quality gate.

Per .claude/rules/anti-patterns.md "Never Offer to Skip Workflow Steps" and .claude/rules/fix-infrastructure-bugs.md "Fix Infrastructure Bugs Immediately": when an infrastructure halt fires inside a workflow, the response is to fix the underlying problem and resume — never to bypass the step that detected it.

The returned path may be absolute or relative. A relative path is resolved against the cwd you ran bin/test --adversarial-path from (the worktree_cwd), so it lands inside the worktree. An absolute path must already point inside the worktree — the validate-worktree-paths hook will block the adversarial agent's Write tool call otherwise; surface that rejection as a finding if it happens. The recommended convention is a path relative to the cwd, matching the assets/bin-stubs/test.sh examples (tests/test_adversarial_flow.rs, test/adversarial_flow_test.rb, etc.).

Resolve the absolute plugin bin/flow path.

The adversarial agent runs <test_command> from inside its own sub-agent session, where the plugin-root prefix is not expanded and the literal token would not resolve. Resolve the absolute path here in the parent and substitute it into the agent prompt instead:

${CLAUDE_PLUGIN_ROOT}/bin/flow plugin-bin-flow

Capture the trimmed stdout as <flow_cli> — an absolute path ending /bin/flow.

If `plugin-bin-flow` exits non-zero (the plugin root is unset, empty, or non-absolute), surface the stderr message and halt. Do NOT embed the non-path error string into the adversarial agent prompt, and do NOT fall back to the unexpanded plugin-root token. Resolve the plugin root and re-run Review.

Capture these values for Step 2 (use the trimmed path verbatim, including extension — the project's bin/test owns both):

• <temp_test_file> = (trimmed output of bin/test --adversarial-path)
• <flow_cli> = (trimmed output of bin/flow plugin-bin-flow)
• <test_command> = the bash invocation shown below, with <temp_test_file> and <flow_cli> substituted by the values captured above

The adversarial agent invokes <test_command> with a 10-minute Bash tool timeout (timeout: 600000) — CI runs can take 3–4 minutes and the default 2-minute timeout would background the process, defeating the gate (per .claude/rules/ci-is-a-gate.md):

<flow_cli> ci --test --file <temp_test_file>

The adversarial agent always launches when bin/test --adversarial-path returns a configured path. If the project's bin/test does not support a --file flag (or cannot compile a single file in isolation), the agent will surface that as a finding rather than silently skipping.

Audit tombstone staleness.

${CLAUDE_PLUGIN_ROOT}/bin/flow tombstone-audit

Parse the JSON output. If the stale array is non-empty, note the stale tombstones for removal in Step 4. Each entry has pr, merged_at, and file fields identifying which test function to remove and from which file. If the command fails (exit non-zero) or the JSON contains a status field with value "threshold_error" or "error", note no stale tombstones — the audit is best-effort and skipped on API failure.

Record step completion:

${CLAUDE_PLUGIN_ROOT}/bin/flow set-timestamp --set review_step=1

To continue to Step 2, invoke flow:flow-review --continue-step using the Skill tool as your final action. Do not output anything else after this invocation.

---

Step 2 — Launch agents

You MUST launch ALL applicable agents listed below in a single response. Never skip an agent because another agent already returned findings. Each agent surfaces independent risk categories that other agents miss — skipping one defeats cognitive isolation. Do not proceed past this step until every applicable agent has been launched and returned.

The four agent launches go in ONE response — the four Agent tool calls themselves, and nothing else. Issue NO other tool call (no Bash, no Read, no Grep, no Skill, no fifth Agent call) between the first agent's launch and the fourth agent's return. Each launch is recorded automatically by FLOW's PreToolUse:Agent hook into phases.flow-review.agents_returned — there is nothing for you to record, and no per-agent state mutation runs in this window. Classification runs ONLY after all four agents have returned, never interleaved between launches. Interleaving a tool call between launches forces the agents into sequential launch-wait-classify runs instead of one concurrent batch — quadrupling Review's wall-clock cost — and reading one agent's findings before launching the next re-introduces the cross-agent bias that cognitive isolation exists to break (.claude/rules/cognitive-isolation.md).

Launch all applicable agents in a single response using multiple Agent tool calls. All agents are independent — they share no state and can run concurrently. Each agent is cognitively isolated from the conversation that produced the code, eliminating self-reporting bias.

Reviewer agent — context-rich (receives diff, plan, CLAUDE.md, rules):

Use the Agent tool with:

• subagent_type: "flow:reviewer"
• description: "Context-isolated code review"

Provide all artifacts in the prompt with labeled sections. The diff is passed as a file path the agent reads via the Read tool rather than as inline bytes:

DIFF_FILE: <full_diff_file>

PLAN: (full plan file content)

CLAUDE.MD: (full CLAUDE.md content)

RULES: (each .claude/rules/ file, prefixed with its filename)

Prefix the prompt with:

"You are reviewing code you did not write. The path to the full diff (DIFF_FILE) is provided below; Read it via the Read tool before analyzing. The plan, the project CLAUDE.md, and all project rules are provided inline. Review the diff for architecture adherence, simplicity, correctness, and security."

Pre-mortem agent — context-sparse (receives only the substantive diff):

Use the Agent tool with:

• subagent_type: "flow:pre-mortem"
• description: "Pre-mortem incident analysis"

Provide the substantive diff as a file path the agent reads via the Read tool. Embed SUBSTANTIVE_DIFF_FILE: <substantive_diff_file> in the prompt and prefix with:

"This PR was merged and caused a production incident. The path to the substantive diff (whitespace-only changes filtered) is provided below (SUBSTANTIVE_DIFF_FILE). Read the file via the Read tool before analyzing. Investigate the codebase and write the incident report. Security failure modes are explicitly in scope."

Adversarial agent — context-sparse (receives substantive diff, temp file path, test command, CLAUDE.md path, branch name). Always launch.

Use the <temp_test_file> and <test_command> derived in Step 1. The path is supplied verbatim — the project's bin/test --adversarial-path already chose it (including the file extension), so the agent does not pick a path or extension itself.

Use the Agent tool with:

• subagent_type: "flow:adversarial"
• description: "Adversarial test generation"

Provide the substantive diff as a file path the agent reads via the Read tool. Embed SUBSTANTIVE_DIFF_FILE: <substantive_diff_file> in the prompt, along with:

• The temp test file path (<temp_test_file>, including extension)
• The test command (<test_command>)
• The path to the project CLAUDE.md
• The branch name

The agent must Read the substantive diff file before analyzing.

Documentation agent — context-sparse (receives substantive diff, narrowed doc-paths list, doc roots):

Use the Agent tool with:

• subagent_type: "flow:documentation"
• description: "Documentation and maintainability review"

Provide the substantive diff as a file path the agent reads via the Read tool. Embed SUBSTANTIVE_DIFF_FILE: <substantive_diff_file> in the prompt, along with:

• The narrowed list of doc paths (<doc_paths> from Step 1) — embed inline, one path per line, under a DOC_PATHS: header
• The path to the project CLAUDE.md
• The path to the .claude/rules/ directory (for cross-reference checks only — the agent must investigate ONLY the listed doc paths for documentation drift, not the full <worktree_path>/docs/ tree)

Prefix the prompt with:

"You are a new team member reading this PR for the first time. The path to the substantive diff (whitespace-only changes filtered) is provided below (SUBSTANTIVE_DIFF_FILE). Read the diff file via the Read tool before analyzing, along with the NARROWED LIST of doc paths likely affected by this PR (under the DOC_PATHS header). Read each listed doc path and check it against the diff for drift. Do NOT walk the full <worktree>/docs/ tree — the listed paths are exhaustive for documentation drift in this PR. Investigate the codebase for comprehension barriers as usual."

Wait for all agents to return.

Classify each agent's response. Apply the classes in priority order — read-overflow first, truncation second, external failure third, normal completion otherwise. Priority order matters for two reasons. A read-overflow return has zero findings and no END-OF-FINDINGS marker, so it ALSO satisfies Class 1's marker-absence trigger; evaluating read-overflow first prevents an overflow from being misclassified as truncation, whose diff-partition-only remedy never bounds the read that overflowed. And a truncated agent's prose may coincidentally contain external-failure substrings, but the correct response to truncation is re-invocation against a narrowed partition, not recording the agent as skipped.

Class 0 — Read overflow. For each high-investigation agent (reviewer, documentation), check whether the returned output has zero structured **Finding blocks AND no END-OF-FINDINGS marker AND contains a context-overflow marker (prompt is too long, context length, context window, too long), matched ASCII-case-insensitively on the agent's full response. This is the agent overflowing its context on a single oversized read — most often a whole-file read of a large CLAUDE.md — before producing any finding. A read-overflow return looks identical to truncation by marker absence, so it MUST be evaluated before Class 1: Class 1's diff-partition-only remedy does not bound the read that overflowed, so classifying an overflow as truncation re-invokes into the same overflow indefinitely.

The zero-findings precondition is load-bearing, exactly as in Class 2: an agent that produced one or more **Finding blocks AND whose prose mentions "too long" (e.g., a maintainability finding about an overlong function) is NOT read-overflowed — it is a normal completion whose findings happen to use the term. Class 0 fires only when the response is structurally empty (no findings) and an overflow marker is the explanation.

When read-overflow is detected on an agent, re-invoke it once under a bounded-read protocol.

Slice the diff per family. The documentation agent already consults CLAUDE.md via Grep + ranged Read (it never whole-reads CLAUDE.md), so the CLAUDE.md read is already bounded. To bound the diff read too, slice the substantive diff per file family. From the changed-file list (Step 1), derive the directory families present in the diff (src/, tests/, agents/, skills/, .claude/, docs/) and run capture-diff once with a --family per present family (trailing slash; substitute <branch>/<base_branch> as in Step 1):

${CLAUDE_PLUGIN_ROOT}/bin/flow capture-diff --branch <branch> --base <base_branch> --family src/ --family tests/

Parse the JSON family_slices array — each entry's path is a bounded per-family substantive diff file under .flow-states/<branch>/.

Re-invoke per family slice. Re-invoke the overflowed agent once per non-empty family slice, with SUBSTANTIVE_DIFF_FILE set to that family's slice path instead of the whole substantive diff. Keep the agent's other inputs (narrowed doc-paths list, CLAUDE.md path, rules dir) unchanged. Every path named in the re-invocation prompt MUST stay inside <worktree_path>/ or this flow's own .flow-states/<branch>/ subtree, per the HARD-GATE in the Class 1 recovery below.

Combine findings. Combine findings from every family re-invocation as if they had come from a single run. Each finding still maps to one of the six tenants for triage in Step 3.

If a bounded re-invocation STILL returns an overflow marker with zero findings, the per-family slice was itself too large — a single non-partitionable family (e.g. a one-file src/ diff) cannot be sliced any smaller. Before declaring the tenant unavailable, apply the second recovery axis, split-by-finding-type (see .claude/rules/cognitive-isolation.md "Partition strategies"). Re-invoke the documentation agent twice — split by finding-type AND investigation depth, so each pass runs only half the agent's investigation and stays under budget. BOTH passes receive SUBSTANTIVE_DIFF_FILE (the whole substantive diff, or the largest non-empty family slice): the diff is the bounded comparison anchor both halves need, so neither pass can be starved of it. Both passes consult CLAUDE.md and .claude/rules/ only via Grep + ranged Read (the documentation agent's bounded read invariant), never whole-file — the prose corpus is not forbidden, only read in bounded form:

• Maintainability pass. Produces only Maintainability (Tenant 3) findings from the diff plus grep-anchored source investigation. It may Grep CLAUDE.md and .claude/rules/ (ranged) to confirm whether a pattern is documented — the discriminator between a documented pattern and a comprehension barrier — but it SKIPS the systematic per-DOC_PATHS: drift comparison the drift pass owns.
• Drift pass. Produces only Documentation (Tenant 6) findings by checking each DOC_PATHS: doc against the diff, consulting CLAUDE.md and .claude/rules/ via Grep + ranged Read. It SKIPS the codebase-comprehension (source-file) investigation the maintainability pass owns.

Both re-invocations MUST honor the path-scoping HARD-GATE in the Class 1 recovery below — every path named in either prompt stays inside <worktree_path>/ or this flow's own .flow-states/<branch>/ subtree. Combine findings from both passes as if they had come from a single run; each finding still maps to one of the six tenants for triage in Step 3. A pass that returns the END-OF-FINDINGS marker with zero findings is a legitimate empty result (both passes receive the diff and the bounded prose corpus, so neither is starved) — fold it into the combined set as "no findings for that tenant," not as a failure.

Only after BOTH axes are exhausted — per-family slicing AND both split-by-finding-type passes STILL returning an overflow marker (the diff slice handed to a pass is itself the oversized read) — note the tenant unavailable in the Step 3 triage summary and proceed. Do NOT fabricate the agent's findings (see the HARD-GATE at the end of this step) and do NOT split infinitely. The agent's launch is already recorded in agents_returned by FLOW's PreToolUse:Agent hook, so the phase-finalize required-agents gate is satisfied — only its findings are missing.

Class 1 — Truncation. For each high-investigation agent (reviewer, documentation), check whether the returned output contains the literal END-OF-FINDINGS completion marker as the final structural element. Marker absence alone means the agent was truncated by maxTurns exhaustion — regardless of whether any partial **Finding block was produced. An agent that exhausts its turn budget DURING investigation (before producing any finding) is the case the recovery path most needs to catch: requiring a partial finding to trigger Class 1 would silently classify early-truncation as "found nothing." See .claude/rules/cognitive-isolation.md "Context Budget + Truncation Recovery" — "Absence of the marker means the agent was truncated, not that it found nothing."

When truncation is detected on an agent:

1. Identify the partition strategy. For documentation (a DOC_PATHS:-driven agent), partition the diff by file family (src/, tests/, agents/, skills/, .claude/, docs/). For reviewer, partition by tenant family (architecture + simplicity vs. correctness + security).
2. Re-invoke the truncated agent once per non-empty partition, with the scope narrowed to that partition only. Keep the agent's other inputs (plan, CLAUDE.md, rules, narrowed doc-paths list) unchanged.
3. Combine findings from every invocation as if they had come from a single run. Each finding still maps to one of the six tenants for triage in Step 3.

If a re-invocation itself returns without the completion marker, that is double-truncation — the partition was still too large. Note the agent as truncated in the triage summary (Step 3) rather than splitting infinitely. The user decides whether to accept partial coverage or rerun Review on a smaller subset of the diff.

Retry prompts MUST NOT instruct the sub-agent to Read file paths outside `/`. The `validate-pretool` Agent-path prompt-body scanner blocks any Agent call whose `prompt` field embeds an out-of-worktree path; the autonomous-flow-strict response shape in `validate-worktree-paths` ensures Reads on paths the hook already blocks return structured JSON errors to the sub-agent instead of user prompts. If the truncated agent's investigation required out-of-worktree files, drop the requirement entirely instead of redirecting the agent toward a different out-of-worktree path. See `.claude/rules/cognitive-isolation.md` "Context Budget + Truncation Recovery" and `src/hooks/agent_prompt_scan.rs`.

Class 2 — External failure. When the agent has produced zero structured **Finding blocks AND no END-OF-FINDINGS marker AND the response contains a canonical external-failure marker (rate_limit, 429, usage_limit, API Error, rate limit exceeded), the agent hit an upstream API or quota failure rather than running out of turns mid-investigation. The agent has produced no findings the parent can use, but a future flow-review re-invocation could succeed.

The zero-findings precondition is load-bearing: an agent that produced one or more **Finding blocks AND mentions rate_limit in its prose (e.g., a security finding about rate-limiting code) is NOT externally-failed — it is a normal completion with findings that happen to discuss the term. Class 2 only fires when the response is structurally empty (no findings at all) and the failure marker is the explanation.

Substring match is ASCII-case-insensitive on the agent's full response. False positives in this class would discard legitimate findings (the agent's **Finding blocks never reach Step 3 triage), so the zero-findings precondition exists specifically to prevent the substring-in-prose case.

Re-invoke the externally-failed agent once with its original prompt — the upstream limit may have cleared. If the re-invocation returns the END-OF-FINDINGS marker (or, for a non-investigation agent, exits cleanly), treat it as Class 3. If it returns zero findings with an external-failure marker a second time, the agent's findings are genuinely unavailable for this Review pass: note it in the triage summary (Step 3) and proceed. The agent's launch is already recorded in agents_returned by FLOW's PreToolUse:Agent hook, so the phase-finalize required-agents gate is satisfied — only its findings are missing. Do NOT fabricate the agent's findings (see the HARD-GATE below).

Class 3 — Normal completion. The response contains the END-OF-FINDINGS marker (high-investigation agents) or has exited cleanly (other agents). Findings flow to Step 3 triage unchanged.

After all four agents have returned. The four launches are already recorded in phases.flow-review.agents_returned by FLOW's PreToolUse:Agent hook — there is no per-agent recording step and no skip path. The only post-return work is the classification above and the Step 3 triage that follows.

When an agent's findings are unavailable — a Class 2 external failure that persisted through one re-invocation, or a Class 1 truncation that double-truncated — you MUST NOT synthesize that agent's findings inline. Note the agent as unavailable in the triage summary and move on. Fabricating an agent's analysis from session memory defeats cognitive isolation per `.claude/rules/cognitive-isolation.md` "Never Supplement Agent Work From the Parent Session" and produces an audit trail that falsely shows "agent reviewed X" when "parent reviewed X" is what actually happened.

The probe file lives inside the worktree's test tree, so worktree removal at Phase 4 Complete (or /flow:flow-abort) disposes of it automatically as a side effect of git worktree remove. The basename glob is also pre-listed in .git/info/exclude (test_adversarial_flow.*, *_adversarial_flow_test.rb) so the throwaway probe never appears in a user's git status output alongside intentional changes.

Record step completion:

${CLAUDE_PLUGIN_ROOT}/bin/flow set-timestamp --set review_step=2

To continue to Step 3, invoke flow:flow-review --continue-step using the Skill tool as your final action. Do not output anything else after this invocation.

---

Step 3 — Triage

Triage findings from each agent in order: reviewer, pre-mortem, adversarial, documentation. For each finding, classify it as Real (fix in Step 4) or False positive (dismiss with rationale).

There is no filing path. All real findings are fixed during Code Review — see .claude/rules/review-scope.md.

Supersession check

Run the supersession check before classification. The supersession test catches code that the current PR has made permanently redundant — code that would leave the PR's behavior unchanged if deleted. Routing such code to Step 4 for deletion regardless of file location keeps dead-on-merge code from surviving into main.

Run the supersession test from .claude/rules/supersession.md. For every finding, ask: "Would deleting the code this finding describes leave the PR's behavior unchanged?"

If yes, the finding is in-scope for deletion regardless of which file the code lives in — route it to Step 4 for deletion.

If no, proceed with the Real / False positive classification below.

If uncertain whether the code is superseded, treat as "no" and proceed with the classification.

Classification

Real — a credible issue supported by evidence. Includes structural issues like duplicate code, missing abstractions, and naming problems. Route to Step 4 for fixing. All real findings are fixed in this PR — regardless of which file they live in.

False positive — speculative, not supported by the code, or already covered by tests. Discard with rationale. After classifying each false positive, record it:

${CLAUDE_PLUGIN_ROOT}/bin/flow add-finding --finding "<description>" --reason "<reason>" --outcome "dismissed" --phase "flow-review"

Truncation check

Step 2's recovery path re-invokes any agent that returned without the END-OF-FINDINGS completion marker (per .claude/rules/cognitive-isolation.md "Context Budget + Truncation Recovery"), so by Step 3 every high-investigation agent's output either ends with the marker (natural completion) or has been re-invoked across partitions until it does — with the combined findings already merged into a single set for that agent.

The remaining truncation-detection responsibility in Step 3 is for the double-truncation case: an agent whose Step 2 re-invocation itself returned without the marker. Note that agent in the triage summary so the user knows coverage was partial. Pre-mortem and adversarial agents do not declare the marker (their investigation surface is naturally bounded by the diff itself); judge their completeness by whether they produced at least one structured **Finding block or an explicit "No findings" report.

Triage summary

Show each finding with its source agent, tenant, triage decision, and rationale inside a fenced code block:

```text
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  FLOW — Review — Step 3: Triage — SUMMARY
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  Reviewer
  --------
  - [T1 Architecture] [REAL] <finding description>
  - [T2 Simplicity] [FALSE POSITIVE] <reason>

  Pre-Mortem
  ----------
  - [T4 Correctness] [REAL] <finding description>

  Adversarial
  -----------
  - [T5 Test coverage] [REAL] <finding description>

  Documentation
  -------------
  - [T6 Documentation] [REAL] <finding description>

  Truncated agents: none

  Real findings to fix : N
  False positives      : N

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
```

If all agents report no findings, show the triage summary with zero findings, then skip the commit and proceed directly to Done.

Record step completion:

${CLAUDE_PLUGIN_ROOT}/bin/flow set-timestamp --set review_step=3

To continue to Step 4, invoke flow:flow-review --continue-step using the Skill tool as your final action. Do not output anything else after this invocation.

---

Step 4 — Fix

Fix all real findings from Step 3.

If no real findings exist, skip this step and proceed to Done.

Fix each finding

For each real finding, fix the issue in code. After fixing each finding, record it:

${CLAUDE_PLUGIN_ROOT}/bin/flow add-finding --finding "<description>" --reason "<reason>" --outcome "fixed" --phase "flow-review"

After fixing all findings, run CI once. Use a 10-minute Bash tool timeout (timeout: 600000) — CI runs can take 3–4 minutes and the default 2-minute timeout would background the process, defeating the gate (per .claude/rules/ci-is-a-gate.md).

${CLAUDE_PLUGIN_ROOT}/bin/flow ci

`bin/flow ci` must be green before committing. If CI fails, identify the breaking fix and iterate until green.

Remove stale tombstones

After fixing all findings and running CI above, remove any stale tombstones identified in Step 1. For each stale entry:

• Open the file from the audit output
• Find the test function guarding the stale PR
• Remove the entire test function (including its doc comment and #[test] attribute)
• If the removal leaves an empty section comment (e.g. // --- Tombstone tests --- with no tests below it), remove the section comment too

Stale tombstone removal is a mechanical operation — no judgment call needed. The tombstone-audit command already verified that the PR was merged before the oldest open PR was created, meaning no active branch could resurrect the deleted code.

Back navigation

If a finding is too significant to fix in Review:

If commit=auto, fix it directly without asking.

If commit=manual, use AskUserQuestion:

• Go back to Code — implementation issue
• Go back to Plan — plan was missing something

Go back to Code: update Phase 3 to pending, Phase 2 to in_progress, then invoke flow:flow-code.

Go back to Plan: update Phase 3 to pending, Phase 2 to in_progress, then invoke flow:flow-plan.

Commit

Set the continuation context and flag before committing. The self-invocation carries no mode flag — the resumed run re-resolves commit/continue from the state file via ## Mode Resolution:

${CLAUDE_PLUGIN_ROOT}/bin/flow set-timestamp --set "_continue_context=Set review_step=4, then self-invoke flow:flow-review --continue-step."

${CLAUDE_PLUGIN_ROOT}/bin/flow set-timestamp --set _continue_pending=commit

Invoke /flow:flow-commit.

Record step completion:

${CLAUDE_PLUGIN_ROOT}/bin/flow set-timestamp --set review_step=4

To continue to Done, invoke flow:flow-review --continue-step using the Skill tool as your final action. Do not output anything else after this invocation.

---

Done — Update state and complete phase

Finalize the phase (complete + Slack notification in one call):

${CLAUDE_PLUGIN_ROOT}/bin/flow phase-finalize --phase flow-review --branch <branch> --thread-ts <slack_thread_ts>

Omit --thread-ts if slack_thread_ts was not returned by phase-enter.

Parse the JSON output.

Handle the required_agent_not_returned error reason. When the response shape is {"status":"error","reason":"required_agent_not_returned","missing":[...],"message":"..."}, one or more required agents are absent from agents_returned. Because FLOW's PreToolUse:Agent hook records every launch, a missing agent means that agent was never launched in this Review pass. The required-agents gate ran before any state mutation, so the phase has not been advanced.

Recovery is to launch the missing agents. Re-run Step 2's launch for each agent named in missing[] (reusing the <full_diff_file> and <substantive_diff_file> paths captured in Step 1 and each agent's Step 2 prompt template). The launch is recorded by the hook, so no separate recording call is needed. Classify each return (Class 1/2/3) and route any findings to Step 3 triage. Then re-run phase-finalize.

Do NOT advance to the COMPLETE banner until every agent named in missing[] has been launched AND a subsequent phase-finalize call returns {"status":"ok",...}.

When the response is {"status":"error", ...} for any OTHER reason, report the error and stop.

Use the formatted_time field in the COMPLETE banner below. Do not print the timing calculation.

Output in your response (not via Bash) inside a fenced code block:

```text
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  ✓ FLOW v2.7.0 — Phase 3: Review — COMPLETE (<formatted_time>)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
```

STOP. Parse `continue_action` from the `phase-finalize` output above to determine how to advance.

1. Use continue_action from the phase-finalize output — phase-finalize computes it from the state file's skills.flow-review.continue config. If continue_action is "invoke" → continue=auto. If continue_action is "ask" → continue=manual.

2. If continue=auto → invoke flow:flow-complete directly using the Skill tool. Do NOT run bin/flow status. Do NOT use AskUserQuestion. This is the FINAL action in this response — nothing else follows.

3. If continue=manual → you MUST do all of the following before proceeding: a. Run bin/flow status via Bash and print its stdout in your response inside a fenced code block:

   ${CLAUDE_PLUGIN_ROOT}/bin/flow status
   

   b. Use AskUserQuestion: "Phase 3: Review is complete. Ready to begin Phase 4: Complete?" Options: "Yes, start Phase 4 now", "Not yet", "I have a correction or learning to capture" c. If "I have a correction or learning to capture": ask what to capture, invoke /flow:flow-note, then re-ask with only "Yes, start Phase 4 now" and "Not yet" d. If Yes → invoke flow:flow-complete using the Skill tool e. If Not yet → print the paused banner below f. Do NOT invoke flow:flow-complete until the user responds

Do NOT skip this check. Do NOT auto-advance when the mode is manual.

If Not yet, output in your response (not via Bash) inside a fenced code block:

```text
══════════════════════════════════════════════════
  ◆ FLOW — Paused
  Run /flow:flow-complete when ready.
══════════════════════════════════════════════════
```

---

Hard Rules

• Always run bin/flow ci after any fix made during Review
• Never transition to Learn unless bin/flow ci is green
• Fix every real finding from agent triage — do not leave findings unaddressed
• Follow the project CLAUDE.md conventions when fixing
• All analysis comes from cognitively isolated agents — the parent session never reviews the diff itself
• Parent session gathers, launches, triages, and fixes — it does not analyze
• Every finding must map to one of the six tenants — findings that do not map are dropped
• One commit for all Review fixes (Step 4), not one commit per finding
• After each step completes, advance to the next step via self-invocation — never pause or wait for user input between steps (Gather, Launch, Triage, Fix advance automatically; only the Done HARD-GATE can pause)
• Never use Bash to print banners — output them as text in your response
• Never use Bash for file reads — use Glob, Read, and Grep tools instead of ls, cat, head, tail, find, or grep
• Never use cd <path> && git — use git -C <path> for git commands in other directories
• Never cd before running bin/flow — it detects the project root internally
• When in autonomous mode, classify tool failures per .claude/rules/autonomous-flow-self-recovery.md — mechanical fixes are in-flow, substantive failures prompt the user
• Never discard uncommitted changes to unblock a workflow step — if any git command fails due to uncommitted changes, show git diff to the user and ask how to proceed