msd-core/gsd-core/workflows/review.md

@~/.claude/gsd-core/references/response-language-directive.md

Cross-AI peer review — invoke external AI CLIs to independently review phase plans. Each CLI gets the same prompt (PROJECT.md context, phase plans, requirements) and produces structured feedback. Results are combined into REVIEWS.md for the planner to incorporate via --reviews flag.

This implements adversarial review: different AI models catch different blind spots. A plan that survives review from 2-3 independent AI systems is more robust.

Check which AI CLIs are available on the system:

_GSD_SHIM_NAME="gsd-tools.cjs"; _GSD_RUNTIME_ROOT="${RUNTIME_DIR:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}"; GSD_TOOLS="${_GSD_RUNTIME_ROOT}/gsd-core/bin/${_GSD_SHIM_NAME}"; _gsd_at() { for _p; do if [ -f "$_p" ]; then GSD_TOOLS="$_p"; return 0; fi; done; return 1; }; if _gsd_at "${_GSD_RUNTIME_ROOT}/gsd-core/bin/${_GSD_SHIM_NAME}" "${_GSD_RUNTIME_ROOT}/.claude/gsd-core/bin/${_GSD_SHIM_NAME}" "${_GSD_RUNTIME_ROOT}/.codex/gsd-core/bin/${_GSD_SHIM_NAME}"; then gsd_run() { node "$GSD_TOOLS" "$@"; }; elif unset -f gsd_run; _G="$(command -v gsd_run)"; then GSD_TOOLS="$_G"; gsd_run() { "$GSD_TOOLS" "$@"; }; elif _gsd_at "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/gsd-core/bin/${_GSD_SHIM_NAME}" "${HERMES_HOME:-$HOME/.hermes}/gsd-core/bin/${_GSD_SHIM_NAME}" "${CURSOR_CONFIG_DIR:-$HOME/.cursor}/gsd-core/bin/${_GSD_SHIM_NAME}" "${CODEX_HOME:-$HOME/.codex}/gsd-core/bin/${_GSD_SHIM_NAME}" "${GEMINI_CONFIG_DIR:-$HOME/.gemini}/gsd-core/bin/${_GSD_SHIM_NAME}" "${COPILOT_CONFIG_DIR:-$HOME/.copilot}/gsd-core/bin/${_GSD_SHIM_NAME}" "${WINDSURF_CONFIG_DIR:-$HOME/.codeium/windsurf}/gsd-core/bin/${_GSD_SHIM_NAME}" "${AUGMENT_CONFIG_DIR:-$HOME/.augment}/gsd-core/bin/${_GSD_SHIM_NAME}" "${TRAE_CONFIG_DIR:-$HOME/.trae}/gsd-core/bin/${_GSD_SHIM_NAME}" "${QWEN_CONFIG_DIR:-$HOME/.qwen}/gsd-core/bin/${_GSD_SHIM_NAME}" "${CODEBUDDY_CONFIG_DIR:-$HOME/.codebuddy}/gsd-core/bin/${_GSD_SHIM_NAME}" "${CLINE_CONFIG_DIR:-$HOME/.cline}/gsd-core/bin/${_GSD_SHIM_NAME}" "${GROK_AGENTS_HOME:-$HOME/.agents}/gsd-core/bin/${_GSD_SHIM_NAME}" "${ANTIGRAVITY_CONFIG_DIR:-$HOME/.gemini/antigravity}/gsd-core/bin/${_GSD_SHIM_NAME}" "${OPENCODE_CONFIG_DIR:-${XDG_CONFIG_HOME:-$HOME/.config}/opencode}/gsd-core/bin/${_GSD_SHIM_NAME}" "${KILO_CONFIG_DIR:-${XDG_CONFIG_HOME:-$HOME/.config}/kilo}/gsd-core/bin/${_GSD_SHIM_NAME}"; then gsd_run() { node "$GSD_TOOLS" "$@"; }; else echo "ERROR: gsd-tools.cjs not found at $GSD_TOOLS and gsd_run is not on PATH. Run: npx -y @opengsd/gsd-core@latest --claude --local" >&2; exit 1; fi; GSD_IDENTITY_STATUS=unverified; case "$(gsd_run runtime-identity --raw 2>/dev/null || true)" in '{"packageName":"@opengsd/gsd-core"'*'}') GSD_IDENTITY_STATUS=ok;; esac; export GSD_IDENTITY_STATUS; [ "$GSD_IDENTITY_STATUS" = ok ] || echo "WARNING: \"$GSD_TOOLS\" did not prove it is @opengsd/gsd-core - it is either a different package or an @opengsd/gsd-core older than the runtime-identity verb. See docs/how-to/diagnose-a-foreign-gsd-tools.md" >&2; if [ -n "${CLAUDE_ENV_FILE:-}" ] && [ -n "${GSD_TOOLS:-}" ]; then printf "export PATH='%s':\"\$PATH\"\n" "${GSD_TOOLS%/*}" >> "$CLAUDE_ENV_FILE" 2>/dev/null || true; fi
# Check each CLI
command -v gemini >/dev/null 2>&1 && echo "gemini:available" || echo "gemini:missing"
command -v claude >/dev/null 2>&1 && echo "claude:available" || echo "claude:missing"
command -v codex >/dev/null 2>&1 && echo "codex:available" || echo "codex:missing"
command -v coderabbit >/dev/null 2>&1 && echo "coderabbit:available" || echo "coderabbit:missing"
command -v opencode >/dev/null 2>&1 && echo "opencode:available" || echo "opencode:missing"
command -v qwen >/dev/null 2>&1 && echo "qwen:available" || echo "qwen:missing"
command -v cursor-agent >/dev/null 2>&1 && echo "cursor:available" || echo "cursor:missing"
command -v agy >/dev/null 2>&1 && echo "antigravity:available" || echo "antigravity:missing"
command -v kimi >/dev/null 2>&1 && echo "kimi-code:available" || echo "kimi-code:missing"

# Check local model servers (OpenAI-compatible HTTP API — no CLI binary required)
OLLAMA_HOST=$(gsd_run query config-get review.ollama_host --raw 2>/dev/null || echo "")
if [ -z "$OLLAMA_HOST" ] || [ "$OLLAMA_HOST" = "null" ]; then OLLAMA_HOST="http://localhost:11434"; fi
curl -s --max-time 2 "${OLLAMA_HOST}/v1/models" >/dev/null 2>&1 && echo "ollama:available" || echo "ollama:missing"

LM_STUDIO_HOST=$(gsd_run query config-get review.lm_studio_host --raw 2>/dev/null || echo "")
if [ -z "$LM_STUDIO_HOST" ] || [ "$LM_STUDIO_HOST" = "null" ]; then LM_STUDIO_HOST="http://localhost:1234"; fi
curl -s --max-time 2 "${LM_STUDIO_HOST}/v1/models" >/dev/null 2>&1 && echo "lm_studio:available" || echo "lm_studio:missing"

LLAMA_CPP_HOST=$(gsd_run query config-get review.llama_cpp_host --raw 2>/dev/null || echo "")
if [ -z "$LLAMA_CPP_HOST" ] || [ "$LLAMA_CPP_HOST" = "null" ]; then LLAMA_CPP_HOST="http://localhost:8080"; fi
curl -s --max-time 2 "${LLAMA_CPP_HOST}/v1/models" >/dev/null 2>&1 && echo "llama_cpp:available" || echo "llama_cpp:missing"

# jq prerequisite (#2589). The config/model/budget lookups in this workflow no
# longer need jq — they use the native --raw/--pick flags. But the lanes listed
# under "jq-dependent reviewer lanes" below parse structured JSON that gsd-tools
# does not emit (OpenAI-compatible /v1/chat/completions responses, opencode's
# JSONL event stream, agy's conversation cache), so they cannot run without jq.
# Probe it here rather than letting each lane swallow exit 127 into empty output.
command -v jq >/dev/null 2>&1 && echo "jq:available" || echo "jq:missing"

jq-dependent reviewer lanes. jq is a production prerequisite for the ollama, lm_studio, llama_cpp, opencode, and antigravity lanes only. If detect_clis reports jq:missing, treat those five as undetected — they follow the same "known-but-undetected" path as a missing CLI. Which path that is depends on how the lane was selected (see the precedence rules below): reached through review.default_reviewers or --all it is an info note and the lane is ignored; named by an explicit flag it is an error, because the user asserted that lane. Tell the user to install jq:

NOTE: jq is not on PATH — the ollama, lm_studio, llama_cpp, opencode, and
antigravity reviewer lanes are unavailable. Install jq (https://jqlang.org/download/)
or select a lane that does not require it (--gemini, --claude, --codex,
--coderabbit, --qwen, --cursor).

The remaining lanes (gemini, claude, codex, coderabbit, qwen, cursor) do not require jq and must stay selectable on a jq-less host.

Parse flags from $ARGUMENTS:

• --gemini → include Gemini
• --claude → include Claude
• --codex → include Codex
• --coderabbit → include CodeRabbit
• --opencode → include OpenCode
• --qwen → include Qwen Code
• --cursor → include Cursor
• --agy or --antigravity → include Antigravity CLI
• --kimi-code → include Kimi CLI
• --ollama → include Ollama (local server, OpenAI-compatible)
• --lm-studio → include LM Studio (local server, OpenAI-compatible)
• --llama-cpp → include llama.cpp (local server, OpenAI-compatible)
• --all → include all available (CLIs + running local servers)
• No flags → if review.default_reviewers is set, include only configured reviewers that are detected; otherwise include all available

Reviewer-selection precedence:

1. Individual reviewer flags (--gemini, --codex, etc.)
2. --all
3. review.default_reviewers
4. No key + no flags → all detected reviewers

Explicit reviewer flags are an assertion, not a preference (ADR-2782 D4). A lane the user named on the command line and that cannot run is an error, surfaced and non-silent — even when other named lanes did run. Do not proceed with a thinner reviewer set and report success: --gemini --qwen on a host without qwen fails, it does not quietly become a Gemini-only review. This applies however the lane became unavailable — binary missing, prerequisite jq absent, or a local server not reachable.

The asymmetry is deliberate: not finding a lane nobody asked for is normal; failing to run a lane somebody asked for is an error. A user who wants "whatever is available" has --all; a user who wants a preferred set has review.default_reviewers. Both stay lenient below.

review.default_reviewers behavior:

• Value must be a non-empty array of slug strings (configured via gsd config-set review.default_reviewers '["gemini","codex"]')
• Unknown slugs warn and are ignored
• Known-but-undetected slugs emit an info note and are ignored — a configured default is a preference evaluated across many hosts, so a subset being present is expected, not an error
• If all configured reviewers are unavailable, fail with an actionable message

If section_manifest is null or "reviewer-instances-note-1" is in its included list: read and execute gsd-core/workflows/review/steps/reviewer-instances-note-1.md. Otherwise skip — do not read the file.

If no CLIs are available:

No external AI CLIs found. Install at least one:
- gemini: https://github.com/google-gemini/gemini-cli
- codex: https://github.com/openai/codex
- claude: https://github.com/anthropics/claude-code
- opencode: https://opencode.ai (leverages GitHub Copilot subscription models)
- qwen: https://github.com/nicepkg/qwen-code (Alibaba Qwen models)
- cursor: https://cursor.com (Cursor IDE agent mode)
- agy: curl -fsSL https://antigravity.google/cli/install.sh | bash (Antigravity CLI — free with Google credentials)

Then run /gsd:review again.

Exit.

Determine which CLI to skip based on the current runtime environment:

# Environment-based runtime detection (priority order)
if [ "$ANTIGRAVITY_AGENT" = "1" ]; then
  # Antigravity is a separate client — all CLIs are external, skip none
  SELF_CLI="none"
elif [ -n "$CURSOR_SESSION_ID" ]; then
  # Running inside Cursor agent — skip cursor for independence
  SELF_CLI="cursor"
elif [ -n "$CLAUDE_CODE_ENTRYPOINT" ]; then
  # Running inside Claude Code CLI — skip claude for independence
  SELF_CLI="claude"
else
  # Other environments (Gemini CLI, Codex CLI, etc.)
  # Fall back to AI self-identification to decide which CLI to skip
  SELF_CLI="auto"
fi

Rules:

• If SELF_CLI="none" → invoke ALL available CLIs (no skip)
• If SELF_CLI="claude" → skip claude, use gemini/codex
• If SELF_CLI="auto" → the executing AI identifies itself and skips its own CLI
• At least one DIFFERENT CLI must be available for the review to proceed.

Collect phase artifacts for the review prompt:

INIT=$(gsd_run query init.review "${PHASE_ARG}")
if [[ "$INIT" == @file:* ]]; then INIT=$(cat "${INIT#@file:}"); fi

# #2358: ONE run-scoped temp dir (portable via ${TMPDIR:-/tmp}) so overlapping
# runs never collide or read each other's stale files.
RUN_DIR=$(mktemp -d "${TMPDIR:-/tmp}/gsd-review-XXXXXX")
echo "RUN_DIR=$RUN_DIR"

Read from init: phase_dir, phase_number, padded_phase.

Capture RUN_DIR above (created ONCE) and thread it into every {run_dir} placeholder and $RUN_DIR/${RUN_DIR} reference within a bash block. Do NOT re-run mktemp -d later — every block must resolve to this same directory, or build_prompt's writes and invoke_reviewers' reads split.

Then read:

1. .planning/PROJECT.md (first 80 lines — project context)
2. Phase section from .planning/ROADMAP.md
3. All *-PLAN.md files in the phase directory
4. *-CONTEXT.md if present (user decisions)
5. *-RESEARCH.md if present (domain research)
6. .planning/REQUIREMENTS.md (requirements this phase addresses)

Build a structured review prompt:

# Cross-AI Plan Review Request

You are reviewing implementation plans for a software project phase.
Provide structured feedback on plan quality, completeness, and risks.

## Project Context
{first 80 lines of PROJECT.md}

## Phase {N}: {phase name}
### Roadmap Section
{roadmap phase section}

### Requirements Addressed
{requirements for this phase}

### User Decisions (CONTEXT.md)
{context if present}

### Research Findings
{research if present}

### Plans to Review
For each `*-PLAN.md` in the phase directory, in glob order, include its full content preceded by a `####` header carrying the plan's **repo-relative path** (e.g. `#### .planning/phases/<phase>/<NN>-PLAN.md`). The path header is the citable anchor for findings about the plan itself — cite it as `<repo-relative plan path>:<line>` (name the heading in prose beside the citation if it helps the reader); reserve `path:line` for repo files the plan references.
{per-plan: `#### <repo-relative plan path>` + full plan contents}

## Review Instructions

**Verify against source — do not review the plan text in isolation.** The plans reference real files, migrations, routes, and tests in this repo.
1. Open the referenced files and check each claim against the actual code.
2. For every strength or concern, cite concrete `path/to/file:line` evidence plus the mechanism.
3. When a plan asserts a mechanism works (a guard, a query filter, a test that exercises a path), trace whether it actually does what is claimed — do not take the plan's word for it.
4. If you cannot read the repo (no file access), say so and downgrade that finding to an open question rather than asserting it.

Findings citing `file:line` evidence are weighted far more heavily than impressionistic ones; a review that only restates the plan's own claims has low value.

**Plan coverage is mandatory (#3301).** The exact list of plan ids and the total plan count for
this review are given in the "## Plan Coverage Manifest" section below. Give **every** listed id
its own `##`-level section headed with that id **verbatim** (e.g. `## 12.6-01`) before writing any
cross-plan comparison, an overall risk assessment, or a consensus-style summary. A review that
stops before every id has its own section is an incomplete review, not a summary — if you must
stop early, say so explicitly and name which ids you did not reach.

Analyze each plan and provide:

1. **Summary** — One-paragraph assessment
2. **Strengths** — What's well-designed (bullet points)
3. **Concerns** — Potential issues, gaps, risks (bullet points with severity: HIGH/MEDIUM/LOW)
4. **Suggestions** — Specific improvements (bullet points)
5. **Risk Assessment** — Overall risk level (LOW/MEDIUM/HIGH) with justification

Focus on:
- Missing edge cases or error handling
- Dependency ordering issues
- Scope creep or over-engineering
- Security considerations
- Performance implications
- Whether the plans actually achieve the phase goals

Output your review in markdown format.

Write to a temp file: {run_dir}/gsd-review-prompt.md

Also write individual section files so the budget tool can re-trim per reviewer:

# #2962: zsh aborts the block on an unmatched for-list glob (nomatch); bash passes it through. nullglob both.
shopt -s nullglob 2>/dev/null; setopt NULL_GLOB 2>/dev/null

RUN_DIR="{run_dir}"   # from gather_context

# Write individual section files for per-reviewer budget trimming
# These are always written so reviewers with a budget can invoke prompt-budget
cp "$INSTRUCTIONS_BLOCK_FILE" "${RUN_DIR}/gsd-review-instructions.md"
cp "$ROADMAP_SECTION_FILE" "${RUN_DIR}/gsd-review-roadmap.md"

# Plan files: copy each PLAN.md to a predictable path named after its source
# plan id (#3959: a bare padded index discards provenance — the budget tool's
# per-plan `### <file>` header then renders a run-dir artifact name no reviewer
# or consensus step can resolve. The plan id keeps the gsd-review-plan-*.md glob
# prepare_trimmed_prompt_for_reviewer consumes.)
for PLAN_FILE in "${PHASE_DIR}"/*-PLAN.md; do
  PLAN_BASENAME=$(basename "$PLAN_FILE")
  PLAN_ID="${PLAN_BASENAME%-PLAN.md}"
  cp "$PLAN_FILE" "${RUN_DIR}/gsd-review-plan-${PLAN_ID}.md"
done

# #3301: plan coverage manifest — tell reviewers exactly which plan ids exist and
# how many there are, so a review that silently covers 6 of 7 plans is no longer
# indistinguishable from one that covers all 7. The id is the plan file's own
# basename with the `-PLAN.md` suffix stripped (e.g. `12.6-01-PLAN.md` ->
# `12.6-01`) — NOT the `plan:` frontmatter key, which holds only the bare
# in-phase sequence number ("01") and can never reconstruct the phase-qualified
# id reviewers need to cite. The filename is guaranteed present for every copied
# plan, so no plan is ever dropped from the manifest for lacking a key.
# Count and bullets are both derived directly from the glob loop below, never
# from re-splitting an accumulated string: bash word-splits an unquoted `$VAR`
# on IFS by default, but zsh does not (no `setopt SH_WORD_SPLIT` here), so a
# prior `PLAN_IDS="$PLAN_IDS $id"` + `for x in $PLAN_IDS` round-trip silently
# collapsed every id onto one iteration under zsh whenever there were 2+ plans
# (gsd-core#4099). Direct glob iteration (`for f in "${PHASE_DIR}"/*-PLAN.md`,
# same pattern as the copy loop above) is identical under both shells, so this
# never needs word-splitting at all.
PLAN_COUNT=0
PLAN_ID_BULLETS=""
for PLAN_FILE in "${PHASE_DIR}"/*-PLAN.md; do
  PLAN_BASENAME=$(basename "$PLAN_FILE")
  PLAN_ID="${PLAN_BASENAME%-PLAN.md}"
  PLAN_COUNT=$((PLAN_COUNT + 1))
  PLAN_ID_BULLETS="${PLAN_ID_BULLETS}- ${PLAN_ID}
"
done

# Named to avoid BOTH existing RUN_DIR globs: `gsd-review-*.md` (reviewer
# reports, invoke_reviewers) and `gsd-review-plan-*.md` (the plan copies just
# above) — a manifest matching either would be picked up as a report or as a
# plan to review.
{
  echo ""
  echo "## Plan Coverage Manifest"
  echo ""
  echo "Total plans in this review: ${PLAN_COUNT}"
  echo ""
  echo "Plan ids (give each one its own \`##\`-level section, headed verbatim):"
  printf '%s' "$PLAN_ID_BULLETS"
} > "${RUN_DIR}/.plans-manifest.md"

# Optional section files (only if content was included in the combined prompt)
if [ -f ".planning/PROJECT.md" ]; then
  cp .planning/PROJECT.md "${RUN_DIR}/gsd-review-project.md"
fi
_CTX=( "${PHASE_DIR}"/*-CONTEXT.md )
if [ ${#_CTX[@]} -gt 0 ]; then
  cat "${_CTX[@]}" > "${RUN_DIR}/gsd-review-context.md"
fi
_RESEARCH=( "${PHASE_DIR}"/*-RESEARCH.md )
if [ ${#_RESEARCH[@]} -gt 0 ]; then
  cat "${_RESEARCH[@]}" > "${RUN_DIR}/gsd-review-research.md"
fi
if [ -f ".planning/REQUIREMENTS.md" ]; then
  cp .planning/REQUIREMENTS.md "${RUN_DIR}/gsd-review-requirements.md"
fi

# #3301: append the manifest to BOTH files reviewers actually read — the
# per-lane budget-trimmed instructions file (descriptor lanes get
# `--instructions-file`) and the full combined prompt (combined-prompt lanes
# read the whole file). The `instructions` fragment is in prompt-budget's
# `minimumFor` floor set and is never trimmed, so this survives per-lane
# budget trimming intact.
cat "${RUN_DIR}/.plans-manifest.md" >> "${RUN_DIR}/gsd-review-instructions.md"
cat "${RUN_DIR}/.plans-manifest.md" >> "${RUN_DIR}/gsd-review-prompt.md"

Note: INSTRUCTIONS_BLOCK_FILE, ROADMAP_SECTION_FILE, and PHASE_DIR come from prompt assembly; RUN_DIR is the run-scoped dir from gather_context (#2358) re-assigned from {run_dir} above. Copy the temp files written during prompt assembly to these section paths (or write each section here if the prompt was built inline).

Every reviewer lane is **declared data** (ADR-2782). This step iterates the lanes the selection resolved; it does not enumerate them. Adding a reviewer is a capability manifest, not an edit here.

Do not re-add a per-CLI block. A <!-- reviewer-lane: … --> marker anywhere in this step now FAILS the parity gate (checkReviewerLaneParity → bespoke_leg_present). Lane divergence is declared in the manifest — timeout floor, probe, prompt/output channel, empty-output policy — and behaviour that data genuinely cannot express is a named first-party handler (ADR-2782 D6), never a bespoke block here.

Effort and model resolution (#4255). A lane's reasoning effort and model each resolve through their own declared key, and the resolution order is inspectable rather than implicit:

piece	order, highest first
model	pinned reviewer-instance --model → the lane's modelConfigKey (review.models.<slug>) → the CLI's own default
effort	the lane's effortConfigKey (review.effort.<slug>) → the lane's declared defaultEffort → nothing emitted, so the CLI's own configuration applies

Both come from the LANE. Effort in particular is never read from an agent's execution settings: until #4255 it was resolved by querying gsd-plan-checker, so every prompt-fed lane ran at that verifier's low and, because the rendered argument is a CLI config override, it silently beat the effort the operator had configured for the reviewer CLI itself. A lane that declares no effort emits no argument at all — a value borrowed from an unrelated agent is worse than no value.

Timeout guidance (#2194): prompt-fed source-grounded reviews are slow — measured ~570 s for Codex at xhigh effort and ~525 s for headless Claude on a large plan set. Each lane declares its own timeoutFloorMs and the runner enforces it internally, but the Bash tool call wrapping the loop below must still be given a high timeout: — at least 900000, and 1200000 when Codex or headless Claude are in the selection — or the host kills the whole loop mid-lane. On Claude Code, raise the host cap via BASH_MAX_TIMEOUT_MS if a review can exceed it.

A silent empty output after a long run is a timeout kill, not a crash — the Codex 0xc0000142 misdiagnosis persisted for exactly this reason, because an empty result cannot distinguish the two on its own. Treat an empty result on a slow lane as a dropped lane and re-run with more time rather than diagnosing a CLI or sandbox failure. A cross-AI review that silently drops a lane is blind in one eye.

No hook-trust bypass (#2479): no lane passes a hook-trust bypass flag and none runs a capability probe for one. That flag only bypasses persisted hook trust (a first-run condition) and flagless invocations work in steady state, while host-harness safety classifiers deny commands carrying it. An environment that genuinely hits an untrusted-hook prompt surfaces through the .err capture and the empty-output stub as a dropped lane with diagnosable stderr, not silent attrition. Do not reintroduce the flag (even spelled out in prose — a regression test bans the literal file-wide).

If section_manifest is null or "reviewer-instances-note-2" is in its included list: read and execute gsd-core/workflows/review/steps/reviewer-instances-note-2.md. Otherwise skip — do not read the file.

Lanes run sequentially by default — concurrent invocation trips provider rate limits, and a lane lost to one is a cross-AI review that quietly went blind in one eye. A project whose providers can accept the concurrency opts in with review.parallel_lanes: true (#3034): the selected lanes are dispatched together and all joined before aggregation. The default is unchanged, and convergence cycles stay sequential either way — only the lanes within one pass overlap.

# #2962: zsh aborts the block on an unmatched for-list glob (nomatch); bash passes it through. nullglob both.
shopt -s nullglob 2>/dev/null; setopt NULL_GLOB 2>/dev/null

RUN_DIR="{run_dir}"
REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
# SELECTED_REVIEWERS is the comma-separated result of reviewer selection (ADR-0011 precedence:
# explicit flags > --all > review.default_reviewers > all detected). Unchanged by this phase.

# #3034: opt-in concurrent lane dispatch. STRICT equality on "true" is deliberate — "1", "yes" and
# "TRUE" must NOT opt in, so a mistyped config gets the conservative behaviour rather than firing
# concurrent requests at a rate-limited provider. Note the `|| echo "false"` fallback is the
# OPPOSITE polarity from the commit_docs guard, which fails OPEN: there, failing open preserves the
# user's intent; here it would fire exactly the requests the default exists to prevent.
PARALLEL_LANES=$(gsd_run query config-get review.parallel_lanes --raw 2>/dev/null || echo "false")

# Shared budget-trim helper. Was defined inside the Ollama leg; it is lane-agnostic, so it is
# hoisted here now that any lane may declare a promptBudgetKey. Returns non-zero when the budget
# is too small for the minimum review set (prompt-budget exit 2 / 11).
prepare_trimmed_prompt_for_reviewer() {
  REVIEWER_KEY="$1"; REVIEWER_BUDGET="$2"; OUTPUT_PROMPT="$3"; OUTPUT_META="$4"

  PLAN_FILE_ARGS=""
  for p in "$RUN_DIR"/gsd-review-plan-*.md; do
    [ -f "$p" ] && PLAN_FILE_ARGS="$PLAN_FILE_ARGS --plan-file $p"
  done
  PROJECT_ARG=""
  [ -f "$RUN_DIR/gsd-review-project.md" ] && PROJECT_ARG="--project-file $RUN_DIR/gsd-review-project.md"
  CONTEXT_ARG=""
  [ -f "$RUN_DIR/gsd-review-context.md" ] && CONTEXT_ARG="--context-file $RUN_DIR/gsd-review-context.md"
  RESEARCH_ARG=""
  [ -f "$RUN_DIR/gsd-review-research.md" ] && RESEARCH_ARG="--research-file $RUN_DIR/gsd-review-research.md"
  REQUIREMENTS_ARG=""
  [ -f "$RUN_DIR/gsd-review-requirements.md" ] && REQUIREMENTS_ARG="--requirements-file $RUN_DIR/gsd-review-requirements.md"

  gsd_run query prompt-budget \
    --budget "$REVIEWER_BUDGET" \
    --instructions-file "$RUN_DIR/gsd-review-instructions.md" \
    --roadmap-file "$RUN_DIR/gsd-review-roadmap.md" \
    $PLAN_FILE_ARGS $PROJECT_ARG $CONTEXT_ARG $RESEARCH_ARG $REQUIREMENTS_ARG \
    --output-prompt "$OUTPUT_PROMPT" \
    --output-metadata "$OUTPUT_META"
  return $?
}

gsd_run query review-lane plan \
  --selected "$SELECTED_REVIEWERS" --run-dir "$RUN_DIR" --repo-root "$REPO_ROOT" --json \
  > "$RUN_DIR/gsd-review-lanes.json"

# One lane, start to finish. Hoisted into a function so the sequential and concurrent paths share
# ONE body: two dispatch bodies kept in sync by hand is the generative-fix divergence ADR-2782 spent
# a phase deleting, and it is what let #2494/#2605 be filed twice as the same defect.
#
# The result goes to a SLUG-SCOPED file, never a shared append. Concurrent O_APPEND is atomic only
# below PIPE_BUF (4096 on Linux, 512 on some platforms), so a lane result above that bound could
# interleave — and write_reviews parses this JSONL to render the models:/model_sources: frontmatter,
# so a torn line is a broken REVIEWS.md, not a cosmetic log defect.
run_review_lane() {
  # `local` is hygiene, not a live fix: each `&`-dispatched call already forks its own subshell, so
  # concurrent lanes cannot share these today. Scoped anyway so the isolation is a property of this
  # function rather than of the dispatch mechanism happening to fork.
  local SLUG LANE_BUDGET PROMPT_ARG TRIMMED
  SLUG="$1"
  # Per-lane prompt budget. The lane declares its own `promptBudgetKey`; `plan` resolved it,
  # applying #2797's sentinel rule (-1 = unset → fall back to the global budget; 0 legitimately
  # means "do not trim this lane"). Trimming itself stays in prompt-budget, which owns it.
  LANE_BUDGET=$(gsd_run query review-lane plan --selected "$SLUG" --run-dir "$RUN_DIR" \
                  --repo-root "$REPO_ROOT" --json 2>/dev/null \
                | sed -n 's/.*"promptBudget": *\([0-9-]*\).*/\1/p' | head -1)
  PROMPT_ARG=""
  if [ -n "$LANE_BUDGET" ] && [ "$LANE_BUDGET" != "null" ] && [ "$LANE_BUDGET" -gt 0 ] 2>/dev/null; then
    TRIMMED="$RUN_DIR/gsd-review-prompt-$SLUG.md"
    if prepare_trimmed_prompt_for_reviewer "$SLUG" "$LANE_BUDGET" "$TRIMMED" \
         "$RUN_DIR/gsd-review-prompt-$SLUG.metadata.json"; then
      PROMPT_ARG="--prompt-file $TRIMMED"
    else
      # A budget too small for the minimum review set drops the lane just as silently as an empty
      # response used to (#2605), so leave the skip visible in the review output, not only on stderr.
      echo "$SLUG review skipped: prompt budget (${LANE_BUDGET} tokens) too small for the minimum review set." \
        > "$RUN_DIR/gsd-review-$SLUG.md"
      # Was `continue` when this was a loop body. Inside a function that keyword is not the loop
      # control it looks like — `return 0` is what skips this lane and leaves it with no result line.
      return 0
    fi
  fi

  # One invocation, whatever the lane's transport, prompt channel, output channel or handler.
  # `--explicit` marks a lane the user NAMED: ADR-2782 D4 — not finding a lane nobody asked for is
  # normal, failing to run one somebody asked for is an error.
  gsd_run query review-lane invoke --slug "$SLUG" \
    --run-dir "$RUN_DIR" --repo-root "$REPO_ROOT" $PROMPT_ARG $EXPLICIT_FLAG --json \
    > "$RUN_DIR/gsd-review-lane-result-$SLUG.json"
}

# Split ONCE, de-duplicated, and reuse for both loops below. Two reasons, and the second is
# load-bearing: a slug repeated in SELECTED_REVIEWERS would put TWO concurrent background jobs on
# `> "$RUN_DIR/gsd-review-lane-result-$SLUG.json"` — the same file, both truncating. The shared-append
# form this replaced could not corrupt itself that way, so de-duping is what keeps the concurrent
# path no worse than the sequential one. Selection de-dupes today (the roster is a Set;
# review.default_reviewers normalizes lowercase-unique), but reachability analysis is not a contract
# and the next caller should not have to redo it.
#
# A plain string accumulator, not an array: zsh and bash disagree on array indexing and this block
# runs under both (see the nullglob/NULL_GLOB pairing above).
DISPATCH_SLUGS=""
for SLUG in $(echo "$SELECTED_REVIEWERS" | tr ',' ' '); do
  case " $DISPATCH_SLUGS " in
    *" $SLUG "*) continue ;;
  esac
  DISPATCH_SLUGS="$DISPATCH_SLUGS $SLUG"
done

# Rewrapped through unquoted command substitution, not consumed as a bare
# `$DISPATCH_SLUGS`: bash word-splits an unquoted scalar on IFS by default,
# but zsh does not, so a bare re-split collapsed every slug onto one
# iteration under zsh whenever 2+ reviewers were selected (gsd-core#4109).
# Unquoted `$(...)` re-splits identically under both shells regardless of
# `SH_WORD_SPLIT` — same reason the accumulator-building loop above already
# works under both.
for SLUG in $(printf '%s' "$DISPATCH_SLUGS"); do
  if [ "$PARALLEL_LANES" = "true" ]; then
    run_review_lane "$SLUG" &
  else
    run_review_lane "$SLUG"
  fi
done

# Join every dispatched lane. A bare `wait` with no background jobs returns 0, so the sequential
# path needs no guard around it. NOTHING below this line may run before every lane has finished —
# write_reviews renders REVIEWS.md and the consensus summary from the aggregate below, and a review
# assembled from a partial set looks complete while silently missing a reviewer.
wait

# Aggregate in SELECTED_REVIEWERS order, NOT completion order, so the JSONL a concurrent run
# produces is byte-identical to the one a sequential run produces. This is post-join and therefore
# single-threaded, so `>>` here is safe. A lane that was budget-skipped, or that never started,
# leaves no result file and correctly contributes no line.
# Rewrapped through unquoted command substitution (gsd-core#4109) — see the
# dispatch loop above for why a bare `$DISPATCH_SLUGS` collapses under zsh.
for SLUG in $(printf '%s' "$DISPATCH_SLUGS"); do
  LANE_RESULT="$RUN_DIR/gsd-review-lane-result-$SLUG.json"
  if [ -f "$LANE_RESULT" ]; then
    cat "$LANE_RESULT" >> "$RUN_DIR/gsd-review-lane-results.jsonl"
  fi
done

Each lane leaves {run_dir}/gsd-review-<slug>.md — its review, or a diagnostic stub carrying the captured stderr (and, for an OpenAI-compatible lane, the raw response body, where such a server puts its error JSON on an HTTP 4xx/5xx while still exiting 0). A stub is never mistaken for a clean review: it keeps its "failed or returned empty output" header (#2494/#2605/#2794).

A lane that will not run reports a typed reason rather than an empty file — missing_binary, probe_failed, probe_timeout, missing_required_binary, host_unreachable, egress_host_changed, unknown_handler, budget_too_small. egress_host_changed means the lane was consented to send plans to one destination and .planning/config.json now names another; it is blocked, not silently redirected (ADR-2782 D5).

Display progress:

### GSD ► CROSS-AI REVIEW — Phase {N}

◆ Reviewing with {CLI}... done ✓
◆ Reviewing with {CLI}... done ✓

**#3352 (ADR-3473 §8.5): no artifact from failed inputs.** Before rendering anything, gate on whether any lane actually produced a result — "every lane failed" is exactly "the aggregate JSONL has zero lines" (§`invoke_reviewers`'s aggregation loop already builds this file as a byproduct; a lane that never started or was budget-skipped contributes no line either way).

RUN_DIR="{run_dir}"
JSONL="$RUN_DIR/gsd-review-lane-results.jsonl"
LANE_LINES=0
[ -f "$JSONL" ] && LANE_LINES=$(wc -l < "$JSONL" | tr -d ' ')

TOTAL_LANE_FAILURE="false"
ALL_LANES_SKIPPED="false"
if [ "${LANE_LINES:-0}" -eq 0 ]; then
  # Zero lines means every dispatched lane left no result JSON — either every one
  # was budget-skipped (N5: a skip is not a failure) or every one actually failed
  # to run. Re-derive the dispatched-slug set the same way invoke_reviewers did
  # (SELECTED_REVIEWERS is a shell block boundary — recompute, do not assume the
  # earlier step's local DISPATCH_SLUGS variable survived into this block).
  DISPATCH_SLUGS=""
  for SLUG in $(echo "$SELECTED_REVIEWERS" | tr ',' ' '); do
    case " $DISPATCH_SLUGS " in
      *" $SLUG "*) continue ;;
    esac
    DISPATCH_SLUGS="$DISPATCH_SLUGS $SLUG"
  done
  # Distinguish by whether every dispatched slug's stub markdown says "skipped":
  # a skip stub always does (see run_review_lane's budget branch, which writes
  # this exact text before returning without ever invoking the lane); a real
  # failure stub does not. If a slug has no stub at all, it is not a skip.
  DISPATCHED_COUNT=0
  SKIPPED_COUNT=0
  # Rewrapped through unquoted command substitution (gsd-core#4109): a bare
  # `$DISPATCH_SLUGS` word-splits under bash but not zsh, collapsing every
  # slug onto one iteration there whenever 2+ reviewers were selected.
  for SLUG in $(printf '%s' "$DISPATCH_SLUGS"); do
    DISPATCHED_COUNT=$((DISPATCHED_COUNT + 1))
    STUB="$RUN_DIR/gsd-review-$SLUG.md"
    if [ -f "$STUB" ] && grep -q "review skipped: prompt budget" "$STUB" 2>/dev/null; then
      SKIPPED_COUNT=$((SKIPPED_COUNT + 1))
    fi
  done
  if [ "$DISPATCHED_COUNT" -gt 0 ] && [ "$SKIPPED_COUNT" -eq "$DISPATCHED_COUNT" ]; then
    ALL_LANES_SKIPPED="true"
  else
    TOTAL_LANE_FAILURE="true"
  fi
fi

• If ALL_LANES_SKIPPED=true: do NOT write REVIEWS.md and do NOT run the commit below — there is nothing to review. Report to the user that every selected lane was budget-skipped (not a failure) and stop; do not proceed to present_results' summary claiming a review ran.
• If TOTAL_LANE_FAILURE=true: do NOT write REVIEWS.md and do NOT run the commit below. Report the total lane failure to the user (name the lanes that were dispatched and point at their .err/stub files preserved under .review-diagnostics/ by present_results) and stop.
• Otherwise (at least one lane produced a result — R1, unchanged): proceed exactly as below.

#3301: plan coverage check. For each dispatched lane that produced a real review (not a stub, not budget-skipped, not empty), check whether its output mentions every plan id from .plans-manifest.md — the same manifest build_prompt gave the reviewer, so the expected-id list here can never diverge from what the reviewer was actually told. This is diagnostic only: it never blocks the workflow, never fails a lane, and never changes the TOTAL_LANE_FAILURE/ ALL_LANES_SKIPPED gate above.

CodeRabbit is excluded — it is a diff-only lane that never receives the source-grounding prompt (and therefore never receives the manifest or the per-id section instruction either), the same fact that already excludes it from grounded-review weighting in the Consensus Summary below.

The match is intentionally lenient about where an id appears (a ##-headed section is asked for, but plain prose mentioning the id still counts as coverage — grading only the letter of the formatting instruction would produce false INCOMPLETE verdicts against a reviewer that cited real evidence correctly). It is strict about what counts as a match: the id is regex-escaped (a decimal phase like 12.6 must not let 12X6-01 satisfy 12.6-01 through an unescaped .), and a -/word character immediately before or after the candidate match does not count as a boundary (so a threat id like T-04-07 elsewhere in the review must not register as covering plan 04-07).

RUN_DIR="{run_dir}"
MANIFEST="$RUN_DIR/.plans-manifest.md"

# Recompute — a shell variable does not survive across separate fenced blocks
# (each is its own process), so DISPATCH_SLUGS from the gate-check block above
# cannot be assumed to still be set here. Same recomputation as that block and
# as invoke_reviewers.
DISPATCH_SLUGS=""
for SLUG in $(echo "$SELECTED_REVIEWERS" | tr ',' ' '); do
  case " $DISPATCH_SLUGS " in
    *" $SLUG "*) continue ;;
  esac
  DISPATCH_SLUGS="$DISPATCH_SLUGS $SLUG"
done

# Rewrapped through unquoted command substitution (gsd-core#4109): a bare
# `$DISPATCH_SLUGS` word-splits under bash but not zsh, collapsing every
# slug onto one iteration there whenever 2+ reviewers were selected.
for SLUG in $(printf '%s' "$DISPATCH_SLUGS"); do
  [ "$SLUG" = "coderabbit" ] && continue
  REVIEW_FILE="$RUN_DIR/gsd-review-$SLUG.md"
  [ -f "$REVIEW_FILE" ] || continue
  [ -s "$REVIEW_FILE" ] || continue
  grep -q "review skipped: prompt budget" "$REVIEW_FILE" 2>/dev/null && continue
  grep -q "failed or returned empty output" "$REVIEW_FILE" 2>/dev/null && continue

  node -e '
    const fs = require("fs");
    const { escapeRegex } = require("./gsd-core/bin/lib/pattern.cjs");
    const manifest = fs.readFileSync(process.argv[1], "utf8");
    const review = fs.readFileSync(process.argv[2], "utf8");
    const ids = manifest.split("\n")
      .filter((l) => l.startsWith("- "))
      .map((l) => l.slice(2).trim())
      .filter(Boolean);
    const missing = ids.filter((id) => {
      const re = new RegExp("(?<![\\w-])" + escapeRegex(id) + "(?![\\w-])");
      return !re.test(review);
    });
    process.stdout.write(JSON.stringify({ complete: missing.length === 0, missing_ids: missing, total: ids.length }));
  ' "$MANIFEST" "$REVIEW_FILE" > "$RUN_DIR/.plan-coverage-$SLUG.json"
done

Each ${RUN_DIR}/.plan-coverage-<slug>.json carries {complete, missing_ids, total} for one graded lane. Collect these into a plan_coverage frontmatter block — only when at least one graded lane has complete: false (mirrors the existing trimmed_reviewers precedent: present only when there is something to report):

plan_coverage:        # only present if at least one graded lane is incomplete
  <slug>:
    total: 7
    missing: ["12.6-07"]

Combine all review responses into {phase_dir}/{padded_phase}-REVIEWS.md:

Capture only the existing conflict entry bytes after the exact ## Plan-Revision Conflicts heading and before the end of the first exact <!-- gsd:plan-revision-conflicts:begin --> / <!-- gsd:plan-revision-conflicts:end --> pair immediately after the artifact title, if present, as {preserved_plan_revision_conflict_entries}. Ignore identical headings or delimiters in reviewer output: reviewers do not own blocking state. Restore the captured bytes at the explicit slot below.

After all reviewers complete, collect trim metadata files written during the run. For each reviewer that was trimmed (i.e. a .metadata.json file exists and hardFailed or omitted is non-empty, or projectMdShrunk is true, or planTruncationPct > 0), include a trimmed_reviewers block in the frontmatter. Omit the key entirely if no reviewer was trimmed.

Reviewer instances (#1517, optional): when instances ran, frontmatter records their names, each gets its own ## <Adapter> Review (<instance>) section, and ≥2 same-cli instances print a one-line shared-adapter caveat. Format in gsd-core/references/reviewer-instances.md.

Resolved model (#2295): each lane's review-lane invoke --json line in {run_dir}/gsd-review-lane-results.jsonl carries a model object; render models: from its value and model_sources: from its source. Both maps carry exactly one entry per reviewer that appears in reviewers: — the two key sets always match. Write the literal unknown rather than omitting a key: an omitted key is indistinguishable from the feature not having run, and a reader must be able to tell no model recorded from nothing to look at. Emit every models:/model_sources: value as a DOUBLE-QUOTED YAML scalar — a legitimate model id can contain : (llama3:70b, qwen2.5:7b), which is unquotable as a bare scalar; a control character is already refused at the recording seam, so quoting is what closes the remaining :/#/leading-- cases. When GSD applied a reasoning effort to a lane, its value already carries a (reasoning=<level>) suffix (e.g. gpt-5.6-sol (reasoning=high)) — render it as-is, without re-deriving or re-formatting it.

---
phase: {N}
reviewers: [gemini, claude, codex, coderabbit, opencode, qwen, cursor, antigravity, ollama, lm_studio, llama_cpp]  # populate at runtime with only the reviewers actually invoked
reviewed_at: {ISO timestamp}
plans_reviewed: [{list of PLAN.md files}]
models:                   # resolved model per reviewer; `unknown` when not recoverable
  codex: "gpt-5.6-sol (reasoning=low)"
  antigravity: "unknown"
model_sources:            # how each value above was determined
  codex: "banner"
  antigravity: "unknown"
trimmed_reviewers:        # only present if at least one reviewer was trimmed
  ollama:
    budget: 6000
    effective_budget: 5400
    estimated_tokens: 5380
    omitted: [context, research]
    project_md_shrunk: true
    plan_truncation_pct: 22
    hard_failed: false
    note_injected: true
plan_coverage:            # only present if at least one graded lane is incomplete (#3301)
  ollama:
    total: 7
    missing: ["12.6-07"]
---

# Cross-AI Plan Review — Phase {N}

<!-- gsd:plan-revision-conflicts:begin -->
## Plan-Revision Conflicts
{preserved_plan_revision_conflict_entries}
<!-- gsd:plan-revision-conflicts:end -->

<!-- Sections are RENDERED from each lane's declared `reviewsSection`, in descriptor order.
     There is deliberately no hardcoded per-reviewer heading list here any more: a hand-maintained
     list is exactly the drift #2781 was filed about, and it silently disagreed with the roster.
     `gsd_run query review-lane sections --selected "$SELECTED_REVIEWERS"` emits
     `<slug><TAB><reviewsSection>` in order; for each row, emit:

         ## <reviewsSection> Review

         {contents of {run_dir}/gsd-review-<slug>.md}

         ---

     Two headings must NOT be generated from this list, because they are not lanes:
       * `## <Adapter> Review (<instance>)` — an ADR-1517 reviewer INSTANCE resolves THROUGH a lane
         and is rendered from the instance list, not the lane list (ADR-2782 D8).
       * `## Consensus Summary` — not a review section at all.

     A lane whose `evidenceClass` is `diff-only` (CodeRabbit) carries its caveat from data: it never
     received the source-grounding prompt, so its verdict is folded in as a diff observation and is
     not weighted as a grounded plan review. -->

## Consensus Summary

{synthesize common concerns across all reviewers. CodeRabbit is a diff-only reviewer (it never received the source-grounding prompt), so do not weight its verdict as a grounded plan review — fold in its diff findings, but base plan-level consensus on the prompt-fed reviewers. A reviewer output carrying the `[reviewed-without-repo-access]` marker (or beginning with `REVIEWED-WITHOUT-REPO-ACCESS`) ran without repo access (#2176) — treat it the same way: note its concerns, but do not count its verdict at full consensus weight. A reviewer output carrying the `[reviewed-without-source-citations]` marker (#3194) declared source-grounded evidence but cited no `file:line` evidence, so it reviewed the plan text only — treat it the same way: note its concerns, but do not count its verdict at full consensus weight.}

### Agreed Strengths
{strengths mentioned by 2+ reviewers}

### Agreed Concerns
{concerns raised by 2+ reviewers — highest priority}

### Divergent Views
{where reviewers disagreed — worth investigating}

Commit (only reached when TOTAL_LANE_FAILURE and ALL_LANES_SKIPPED are both false — the gate above):

gsd_run query commit "docs: cross-AI review for phase {N}" --files {phase_dir}/{padded_phase}-REVIEWS.md

**If `write_reviews` set `TOTAL_LANE_FAILURE=true` or `ALL_LANES_SKIPPED=true`, skip the success summary below entirely** — no `REVIEWS.md` was written or committed, so there is nothing to present as complete. Report instead:

### GSD ► REVIEW FAILED

Phase {N}: every selected reviewer lane {failed to produce a result|was budget-skipped} — no
REVIEWS.md was written.

{If the preserve+cleanup block below reports success: "Diagnostics preserved:
{phase_dir}/.review-diagnostics/". If it reports failure: relay its own warning verbatim —
it names the intact run directory holding the un-preserved evidence instead.}

Otherwise (at least one lane succeeded), display summary:

### GSD ► REVIEW COMPLETE

Phase {N} reviewed by {count} AI systems.

Consensus concerns:
{top 3 shared concerns}

Full review: {padded_phase}-REVIEWS.md

To incorporate feedback into planning:
  /gsd:plan-phase {N} --reviews

#3352 (ADR-3473 §8.5, R3): preserve per-lane evidence before destroying it. Regardless of which branch above ran, the run's temp directory is the only record that a lane failed at all — copy it beside the phase's artifacts BEFORE cleanup. A lane that produced no output at all (L4) leaves nothing to preserve; that is a smaller diagnostics folder, not a fabricated one, and is NOT a preservation failure. This copy is deliberately NOT part of the commit above (N6) — that step names only {padded_phase}-REVIEWS.md explicitly, never a directory glob, so .review-diagnostics/ is never swept into it.

Preservation and cleanup MUST run in the same fenced block below (a shell variable cannot survive across separate fences — each is its own process). mkdir -p and every cp are exit-status checked; rm -rf "$RUN_DIR" runs ONLY if nothing was preserved (nothing to preserve is not a failure) or everything that needed preserving was copied successfully. If preservation fails partway, $RUN_DIR is left intact and a message names it as the location of the un-preserved evidence — a leftover temp directory is far cheaper than destroyed evidence:

shopt -s nullglob 2>/dev/null; setopt NULL_GLOB 2>/dev/null

RUN_DIR="{run_dir}"
DIAG_DIR="{phase_dir}/.review-diagnostics"

_DIAG_MD=( "$RUN_DIR"/gsd-review-*.md )
_DIAG_ERR=()
for f in "$RUN_DIR"/gsd-review-*.err; do
  [ -s "$f" ] && _DIAG_ERR+=("$f")
done

_PRESERVE_OK=true
if [ ${#_DIAG_MD[@]} -gt 0 ] || [ ${#_DIAG_ERR[@]} -gt 0 ]; then
  if mkdir -p "$DIAG_DIR"; then
    if [ ${#_DIAG_MD[@]} -gt 0 ] && ! cp "${_DIAG_MD[@]}" "$DIAG_DIR/"; then
      _PRESERVE_OK=false
    fi
    if [ ${#_DIAG_ERR[@]} -gt 0 ] && ! cp "${_DIAG_ERR[@]}" "$DIAG_DIR/"; then
      _PRESERVE_OK=false
    fi
  else
    _PRESERVE_OK=false
  fi
fi

if [ "$_PRESERVE_OK" = "true" ]; then
  rm -rf "$RUN_DIR"
else
  echo "WARNING: evidence preservation to $DIAG_DIR failed — leaving the un-preserved run directory intact at: $RUN_DIR" >&2
fi

<success_criteria>

• At least one external CLI invoked successfully
• REVIEWS.md written with structured feedback
• Consensus summary synthesized from multiple reviewers
• Temp files cleaned up
• User knows how to use feedback (/gsd:plan-phase --reviews) </success_criteria>