plan

Analysis-only. Produces structured plan, no code. Use to understand scope, risks, effort before /develop:feature, /develop:fix, /develop:refactor.

NOT for: code/tests (use develop mode); .claude/ config (use /foundry:manage (requires foundry plugin)).

• non-Python-only projects (JS/TS/Go/Rust with no Python source) — downstream develop skills assume pytest; planning analysis language-agnostic but downstream implementation needs language-native toolchain
• mixed refactor+feature tasks — run /develop:refactor first, then /develop:feature

Agent Resolution

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
_DEV_SHARED=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_develop}/bin/dev_shared_resolve.py" 2>/dev/null)  # timeout: 5000
[ -z "$_DEV_SHARED" ] && _DEV_SHARED="plugins/cc_develop/skills/_shared"
echo "$_DEV_SHARED" > "${TMPDIR:-/tmp}/dev-shared-${CSID}"  # cold resolve — every later block warm-reads this
cat "$_DEV_SHARED/agent-resolution.md"

Contains: foundry check + fallback table. If foundry not installed: substitute each foundry:X with general-purpose per table. Agents this skill uses: foundry:sw-engineer, foundry:qa-specialist, foundry:challenger.

Checkpoint: plan single-pass — .plans/active/<slug> file existence = implicit resume signal. No .developments/ checkpoint needed; if interrupted, re-run /develop:plan to regenerate (no code changes made).

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r _DEV_SHARED < "${TMPDIR:-/tmp}/dev-shared-${CSID}" 2>/dev/null || _DEV_SHARED=""  # timeout: 5000
[ -z "$_DEV_SHARED" ] && _DEV_SHARED="plugins/cc_develop/skills/_shared"
cat "$_DEV_SHARED/task-hygiene.md"

Flag parsing

Parse flags into shell variables (not prose) so downstream blocks see correct values. Persist to per-invocation namespaced temp dir for cross-block access (bash state lost between Bash() calls). Namespace by PID to prevent collision when two /develop:plan invocations run concurrently:

# timeout: 5000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
PLAN_NS="${TMPDIR:-/tmp}/dev-plan-$$-${CSID}"
mkdir -p "$PLAN_NS"
echo "$PLAN_NS" > "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}"  # downstream blocks read back namespace
python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_develop}/bin/dev_parse_args.py" --skill plan --write-files "$ARGUMENTS"
# written to ${TMPDIR:-/tmp}/dev-plan-<flag>-${CSID} (legacy paths: SKILL_SPECS["plan"])
cp "${TMPDIR:-/tmp}/dev-challenge-enabled-${CSID}"  "$PLAN_NS/challenge-enabled" 2>/dev/null || echo "true"  > "$PLAN_NS/challenge-enabled"
cp "${TMPDIR:-/tmp}/dev-codemap-raw-${CSID}"        "$PLAN_NS/codemap-raw"       2>/dev/null || echo "auto"  > "$PLAN_NS/codemap-raw"
cp "${TMPDIR:-/tmp}/dev-semble-enabled-${CSID}"     "$PLAN_NS/semble-enabled"    2>/dev/null || echo "false" > "$PLAN_NS/semble-enabled"
cp "${TMPDIR:-/tmp}/dev-plan-max-depth-${CSID}"     "$PLAN_NS/max-depth"         2>/dev/null || echo "3"     > "$PLAN_NS/max-depth"

Downstream blocks recover namespace then read back, e.g. IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""; IFS= read -r CODEMAP_ENABLED < "$PLAN_NS/codemap-enabled" 2>/dev/null || CODEMAP_ENABLED=false.

Unsupported flag check — after all supported flags extracted, scan $ARGUMENTS for remaining --<token> tokens not in the supported list below. If found: print ! Unknown flag(s): `--<token>`. Supported: `--no-challenge`, `--codemap`, `--no-codemap`, `--semble`, `--max-depth`. then invoke AskUserQuestion — (a) Abort (stop, re-invoke with correct flags) · (b) Continue ignoring (skip unknown flags, proceed). On Abort: stop.

Codemap auto-detection — normalize CODEMAP_RAW to true/false; strict mode hard-fails when codemap unavailable:

# timeout: 5000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
# resolves PLAN_NS itself — plan keeps its flags in a run-namespace dir, not TMPDIR sentinels
CODEMAP_ENABLED=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_develop}/bin/dev_codemap_gate.py" plan) || exit 1

loads: codemap-gates.md

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r _DEV_SHARED < "${TMPDIR:-/tmp}/dev-shared-${CSID}" 2>/dev/null || _DEV_SHARED=""  # timeout: 5000
[ -z "$_DEV_SHARED" ] && _DEV_SHARED="plugins/cc_develop/skills/_shared"
cat "$_DEV_SHARED/codemap-gates.md"

Follow Gate A and Gate B.

Preflight — runs only when --semble was passed; the block re-reads SEMBLE_ENABLED from the namespace because bash state is lost between Bash() calls:

# timeout: 5000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
IFS= read -r SEMBLE_ENABLED < "$PLAN_NS/semble-enabled" 2>/dev/null || SEMBLE_ENABLED=false
if [ "$SEMBLE_ENABLED" = "true" ]; then
    IFS= read -r _DEV_SHARED < "${TMPDIR:-/tmp}/dev-shared-${CSID}" 2>/dev/null || _DEV_SHARED=""
    [ -z "$_DEV_SHARED" ] && _DEV_SHARED="plugins/cc_develop/skills/_shared"
    cat "$_DEV_SHARED/preflight-helpers.md"
else
    echo "→ --semble not passed — skipping semble preflight"
fi

When the block printed preflight-helpers.md, execute the semble preflight it describes; when it printed the skip line, proceed. Codemap validation handled by auto-detect block above.

Step 1: Classify and scope

Determine task type and affected surface.

Codemap target derivation — when goal names explicit target as module.path or module.path::function, pre-set TARGET_MODULE/TARGET_FN so codemap-context.md runs caller-impact queries (rdeps, fn-rdeps) instead of only central baseline. Goal with no explicit target → both empty → only central runs (correct: affected surface unknown until agent searches):

# timeout: 5000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
if [[ "$ARGUMENTS" == *"::"* ]]; then
    _QNAME=$(printf '%s\n' "$ARGUMENTS" | grep -oE '[A-Za-z_][A-Za-z0-9_.]*::[A-Za-z_][A-Za-z0-9_]*' | head -1)
    TARGET_MODULE="${_QNAME%%::*}"
    TARGET_FN="${_QNAME##*::}"               # bare fn — codemap-context.md builds module::fn
elif [[ "$ARGUMENTS" =~ ([A-Za-z_][A-Za-z0-9_]*(\.[A-Za-z_][A-Za-z0-9_]*)+) ]]; then
    TARGET_MODULE="${BASH_REMATCH[1]}"       # dotted module target
    TARGET_FN=""
else
    TARGET_MODULE=""
    TARGET_FN=""
fi
export TARGET_MODULE TARGET_FN
echo "$TARGET_MODULE" > "$PLAN_NS/target-module"   # persist — bash state lost between Bash() calls
echo "$TARGET_FN"     > "$PLAN_NS/target-fn"

Structural context — runs only when codemap or semble is enabled; both values re-read from the namespace (bash state lost between Bash() calls):

# timeout: 5000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
IFS= read -r CODEMAP_ENABLED < "$PLAN_NS/codemap-enabled" 2>/dev/null || CODEMAP_ENABLED=false
IFS= read -r SEMBLE_ENABLED  < "$PLAN_NS/semble-enabled"  2>/dev/null || SEMBLE_ENABLED=false
echo "CODEMAP_ENABLED=$CODEMAP_ENABLED SEMBLE_ENABLED=$SEMBLE_ENABLED"
if [ "$CODEMAP_ENABLED" = "true" ] || [ "$SEMBLE_ENABLED" = "true" ]; then
    IFS= read -r _DEV_SHARED < "${TMPDIR:-/tmp}/dev-shared-${CSID}" 2>/dev/null || _DEV_SHARED=""
    [ -z "$_DEV_SHARED" ] && _DEV_SHARED="plugins/cc_develop/skills/_shared"
    cat "$_DEV_SHARED/codemap-context.md"
else
    echo "→ codemap and semble both off — skipping structural context"
fi

Follow enabled sections per the values the block echoed (codemap block if CODEMAP_ENABLED=true, semble companion if SEMBLE_ENABLED=true). Nothing printed beyond the skip line → proceed.

Effort sizing (codemap-py) — when CODEMAP_ENABLED=true, derive blast-radius tier table from reverse dependencies so complexity estimate structural, not guessed. Degrade silently when codemap-py absent — plan works unchanged, sizing falls back to agent's file-count heuristic. Run Extended scan (--source=diff when partial diff exists, e.g. re-planning after abandoned work; otherwise per-target rdeps when TARGET_MODULE known):

# timeout: 15000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
IFS= read -r CODEMAP_ENABLED < "$PLAN_NS/codemap-enabled" 2>/dev/null || CODEMAP_ENABLED=false
IFS= read -r TARGET_MODULE < "$PLAN_NS/target-module" 2>/dev/null || TARGET_MODULE=""
if [ "$CODEMAP_ENABLED" = "true" ] && command -v codemap-py >/dev/null 2>&1; then
    if [ -n "$(git diff HEAD --name-only 2>/dev/null | grep '\.py$')" ]; then
        python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_develop}/bin/codemap_scan.py" --source=diff > "$PLAN_NS/sizing-rdeps" 2>/dev/null || true
    elif [ -n "$TARGET_MODULE" ]; then
        codemap-py query --timeout 5 rdeps "$TARGET_MODULE" --top 10 --exclude-tests > "$PLAN_NS/sizing-rdeps" 2>/dev/null || true
        codemap-py query --timeout 5 coupled --top 10 >> "$PLAN_NS/sizing-rdeps" 2>/dev/null || true
    fi
fi

Interpret $PLAN_NS/sizing-rdeps (skip when empty — codemap absent or no target): each rdeps block's caller count sets per-module blast tier; coupled output lists co-change pairs. Tiers match develop's convention:

• >= 5 rdeps → HIGH blast radius — cross-module reach; nudges complexity toward large and adds a Risks entry
• 1–4 rdeps → MODERATE — note affected importers in plan
• 0 rdeps → LOW — self-contained; proceed

Fold highest tier across affected modules into complexity assessment below (HIGH tier or ≥3 affected modules → large), and pass tier table + coupled pairs to sw-engineer spawn as ## Structural blast radius block so scope estimate accounts for downstream callers rather than file count alone.

Spawn foundry:sw-engineer agent with full goal text from $ARGUMENTS. Agent should:

• Classify task as feature, fix, refactor, or debug
  • debug: root cause unknown — symptoms present but cause unclear, investigation needed before fix scoped; when classified debug, recommend running /develop:debug first, then re-run /develop:plan once root cause identified to produce fix plan
  • WARNING: debug classification recommends /develop:debug; once root cause found, /develop:debug's own output tells the user to re-run /develop:plan — a user-mediated plan→debug→plan cycle, not automatic re-invocation (/develop:debug never calls /develop:plan itself). Caller tracks cycle depth via shared checkpoint file to cap repeated cycles (not a CLI flag — /develop:debug does not accept --max-depth). Max depth = $MAX_DEPTH (default 3, CLAUDE.md safety break). Before invoking /develop:debug, execute depth-checkpoint bash block below:

# anti-loop guard  # timeout: 3000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
IFS= read -r MAX_DEPTH < "$PLAN_NS/max-depth" 2>/dev/null || MAX_DEPTH=3
DEPTH_FILE="${TMPDIR:-/tmp}/dev-plan-depth-checkpoint-${CSID}"
IFS= read -r CURRENT_DEPTH < "$DEPTH_FILE" 2>/dev/null || CURRENT_DEPTH="$MAX_DEPTH"
if [ "$CURRENT_DEPTH" -le 0 ]; then
    echo "! depth limit ($MAX_DEPTH) reached — stopping plan→debug→plan loop"
    # don't invoke /develop:debug — proceed to AskUserQuestion below
else
    NEXT_DEPTH=$(( CURRENT_DEPTH - 1 ))
    echo "$NEXT_DEPTH" > "$DEPTH_FILE"
    echo "→ invoking /develop:debug (depth remaining: $NEXT_DEPTH)"
fi

At depth 0: stop, report current plan state, invoke AskUserQuestion — (a) Accept plan as-is · (b) Re-scope with reduced depth requirement.

• Identify affected files and modules (search codebase — no guessing)
• Assess complexity: small (1-3 files, self-contained), medium (4-8 files or 1-2 modules), large (cross-module, API changes, or 3+ modules). When effort-sizing block produced tier table, let structural reach override file count: any HIGH blast module (≥5 rdeps) or ≥3 affected modules → large, regardless of raw file count.
• Return two separate structured fields (not merged into flat risks list):
  • breaking_changes: list of changes affecting public API only — see criteria below; empty list when none
  • risks: non-breaking concerns (missing tests, unclear requirements, external dependencies, internal coupling); when effort-sizing tier table flags HIGH/MODERATE modules or coupled pairs, add each as concrete risk (e.g. "changing <mod> reaches N downstream callers", "<a>/<b> co-change coupling")
• Note complexity smells: ambiguous goal, scope creep risk, missing reproduction case, directory-wide refactor without explicit goal

Agent returns findings inline (no file handoff — output short).

Breaking change gate: gate triggers only when breaking_changes non-empty — items in risks do NOT trigger this gate. Stop before writing plan. Call AskUserQuestion per breaking change (group only when logically one atomic change). State: what worked before, what breaks, why needed. Options: (a) Accept breaking change — proceed with plan as-is · (b) Revise to non-breaking — return to Step 1 with constraint to avoid this breaking change · (c) Abort — stop immediately. Proceed only on explicit user selection of (a). Prose question in response body does NOT count — AskUserQuestion mandatory per communication.md. If user selects (b) or (c): stop immediately — do not proceed to Step 2 or subsequent steps.

Breaking change criteria — change is breaking when it affects public API (exported from __init__.py, documented in README, or stable interface used by external consumers) and any of these apply: removed public API (function, class, method, or module), changed function signatures (parameter names, types, order, or defaults), changed config key names or schema, changed output format (return type, serialization structure, CLI output shape). Internal/private signature changes (functions prefixed _, classes not exported) do NOT count as breaking — list under risks instead.

Step 2: Structured plan

Derive filename slug from goal: first 4-5 meaningful words, lowercase, hyphen-separated (e.g. "improve caching in data loader" -> plan_improve-caching-data-loader.md). If .plans/active/<slug> already exists, append counter suffix (-2, -3, etc.) before writing — never silently overwrite. Store full path as PLAN_FILE — used in Steps 3 and Final output.

# persist — bash state lost between Bash() calls  # timeout: 3000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
mkdir -p ".plans/active"
SLUG=$(printf '%s\n' "$ARGUMENTS" | tr '[:upper:]' '[:lower:]' | grep -oE '[a-z0-9]+' | grep -vE '^(a|an|the|in|on|of|to|for|and|or|is|with|at|by|from)$' | head -5 | paste -sd- -)
PLAN_FILE=".plans/active/plan_${SLUG}.md"
N=2
while [ -e "$PLAN_FILE" ]; do
    PLAN_FILE=".plans/active/plan_${SLUG}-${N}.md"
    N=$((N+1))
done
echo "$PLAN_FILE" > "$PLAN_NS/plan-file"

# Plan: <goal>

## Brief

*[Generated after agent review — see below]*

---

## Full Plan

**Classification**: feature | fix | refactor
**Complexity**: small | medium | large
**Date**: <YYYY-MM-DD>

### Goal

<One-paragraph restatement of goal in concrete terms — what changes, what doesn't.>

### Affected files

- `path/to/file.py` — reason
- `path/to/other.py` — reason

### Risks

- <risk 1>
- <risk 2>

### Suggested approach

1. <Step 1>
2. <Step 2>
3. <Step 3>
...

Step 3: Agent feasibility review

Spawn ONE foundry:sw-engineer feasibility agent covering both role perspectives in a single pass (merged spawn unit — two role-labelled verdicts, one file read, ~300 bytes of JSON; two spawns would pay ~120,851 tok of fixed overhead each for the same read):

• feature / fix / refactor: one spawn, roles sw-engineer (implementation feasibility) + qa-specialist (testability, coverage feasibility)
• debug: skip feasibility review — no implementation plan to review; proceed directly to Final output with debug recommendation

foundry:linting-expert intentionally excluded — its role post-implementation static analysis (ruff/mypy), not pre-plan architectural feasibility. Including it produces noise (trivial ok: true) or false blockers on linting-config concerns. Surface lint-specific notes (e.g. "target module has no type annotations — mypy will flag everything") in Final output advisory notes section instead. The two surviving roles are dimensions of one review, not cross-checks — merging them into one spawn preserves both checklists.

The agent receives only plan file path and both role checklists — no conversation history, no unrelated context. Prompt (substitute <PLAN_FILE>):

"Read <PLAN_FILE>. Review the plan twice, once from each perspective: (1) as sw-engineer — implementation feasibility, architectural risks, blockers; (2) as qa-specialist — testability, coverage feasibility, verification blockers. For each role: flag domain-specific concerns, risks, or blockers you see; can that role execute its part autonomously without further user input? Return only a JSON array with exactly two elements, one per role, nothing else: [{\"a\":\"sw-engineer\",\"ok\":true|false,\"blockers\":[\"...\"],\"q\":[\"...\"],\"concerns\":[\"...\"]},{\"a\":\"qa-specialist\",\"ok\":true|false,\"blockers\":[\"...\"],\"q\":[\"...\"],\"concerns\":[\"...\"]}]"

Parse-failure handling: agent responses may not be valid JSON (especially fallback general-purpose agents that wrap JSON in prose). Before processing:

1. Attempt to extract the JSON array: prefer echo "$RESPONSE" | jq -c '.' 2>/dev/null for parseable input. For mixed prose+JSON, per-role recovery: echo "$RESPONSE" | grep -oE '\{[^{}]*(\{[^{}]*\}[^{}]*)?\}' | jq -c '.' 2>/dev/null — extracts each balanced JSON object (one nesting level; breaks on strings containing { or }); match each to its role via the "a":"<ROLE>" anchor. If jq not available or both jq attempts fail, fallback: echo "$RESPONSE" | python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_develop}/bin/extract_json_field.py" . — recovers outermost balanced JSON from arbitrary prose+JSON text; pass specific field name (e.g. ok, a) instead of . to extract just that field.
2. If extraction succeeds: use extracted objects (one per role)
3. If extraction fails entirely for a role: log ⚠ non-JSON plan response — falling back to prose extraction; treat that role as {"a":"<ROLE>","ok":false,"blockers":["agent returned non-JSON response"],"q":[],"concerns":[]} and enter resolution loop with re-query

Verdicts return inline (~300 bytes total — no file handoff). Collect both role results:

• All ok: true, empty blockers, q, concerns -> note ✓ agents ready in final output and proceed
• Any ok: false, non-empty blockers or q -> enter internal resolution loop below before surfacing to user
• Non-empty concerns with ok: true -> surface as advisory notes in final output (not blockers, domain-specific flags user should know before starting)

Internal resolution loop (max 3 iterations)

The loop body is model-driven (codebase search, agent re-query), so the counter lives in a file rather than a shell variable — bash state is lost between Bash() calls and an in-prose ITER would never be assigned.

Initialize once, before entering the loop:

# timeout: 3000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
echo 0 > "$PLAN_NS/resolution-iter"

Then run this block at the top of every iteration, before doing any resolution work for that pass:

# timeout: 3000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r PLAN_NS < "${TMPDIR:-/tmp}/dev-plan-ns-current-${CSID}" 2>/dev/null || PLAN_NS=""
IFS= read -r ITER < "$PLAN_NS/resolution-iter" 2>/dev/null || ITER=0
if [ "$ITER" -ge 3 ]; then
    echo "! Max feasibility iterations (3) reached — stop the loop and escalate remaining items to user"
else
    echo $(( ITER + 1 )) > "$PLAN_NS/resolution-iter"
    echo "→ resolution pass $(( ITER + 1 )) of 3"
fi

Block printed the ! line → exit the loop immediately and escalate whatever remains unresolved.

For each blocker or open question:

1. Attempt autonomous resolution — search codebase, read relevant files, re-read goal. Fetch primary-source docs for relevant issues (official docs, RFCs, library changelogs, migration guides) via WebFetch — known URLs only; WebFetch fetches specific URL, does not search.
   • Unknown-URL path: if URL needed to resolve blocker unknown (e.g. "what does library X's new API look like?"), do NOT guess or invent URL. Mark blocker requires-user-input and skip WebFetch — escalate to user with note that documentation lookup required.
   • Known URL — mandatory verification gate: after each WebFetch call, before incorporating content into <PLAN_FILE>, perform three-step verification per quality-gates.md link verification: (a) Fetch returned non-error (HTTP 200), (b) Read returned content, (c) Match content against specific blocker — confirm topic alignment. If any step fails: mark URL non-resolving, do not write content to <PLAN_FILE>, escalate to user. Each URL requires its own Fetch+Read+Match pass — no exemption for same-domain or "similar" URLs.
   • If answer determinable from verified source, update <PLAN_FILE> and mark resolved.
2. Re-query raising role — batch ALL items resolved this iteration for that role into ONE re-query (a spawn per item pays ~120,851 tok each): {"a":"<ROLE>","resolved":[{"item":"<item>","answer":"<resolution>"}, ...]}. Agent returns updated ok/blockers for the role; items it accepts drop from the blockers list.
3. After all resolvable items cleared, re-check: if all agents ok: true -> ✓ agents ready.

Plan file coherence: after resolution loop exits (regardless of outcome), annotate <PLAN_FILE>:

• Each resolved blocker: add (resolved ✓) inline
• Each unresolved blocker: add (unresolved — requires user input)
• Update Brief (once it exists): note "N of M blockers resolved autonomously; N require user input" Ensures plan file coherent after partial resolution.

Escalate to user only what cannot be resolved autonomously — blocker requires user input when: depends on business decision, undocumented external constraint, missing credential/secret, or genuine goal ambiguity with two equally valid interpretations.

For each escalated item:

• Issue: one sentence — what blocks or is unclear
• Alternatives: 2-3 concrete options with trade-offs
• Recommendation: which option and why

Do not escalate: items resolvable from codebase, items that are risks (not blockers), items already addressed in plan.

Step 4: Challenger gate

Two states (plan has no diff yet, so no small-diff auto-skip and no --challenge flag — unlike fix/feature/refactor/debug): by default challenger always reviews plan design; --no-challenge (CHALLENGE_ENABLED=false) skips gate entirely.