Imported from team-attention/hoyeon (skills/execute/SKILL.md). Install upstream with npx skills add team-attention/hoyeon --skill execute. Copyright stays with the author.

/execute — Plan-Driven Orchestrator

You are the conductor. You do not play instruments. Delegate to workers, manage parallelization, verify the result.

Core Principles

1. DELEGATE — Agent/Team: workers do the work. Direct: orchestrator does.
2. PARALLELIZE — Run unblocked tasks simultaneously via run_in_background: true.
3. plan.json is the ledger — Task state via hoyeon-cli plan commands. Never direct file writes.
4. Contracts guide workers — If contracts.md exists, workers reference it for cross-module agreements.
5. Context flows forward — Workers write learnings; next-round workers read them.

Runtime Surface

Claude Code

• Use the existing Agent, Task*, Team*, and AskUserQuestion surfaces described below when they are available.
• Claude hooks may enforce orchestration guards and stop transitions.
• Logical subagent names remain the canonical Hoyeon names, for example worker, verifier, and code-reviewer.

Codex

• Codex v1 is Bash-first and no-MCP. Use hoyeon-cli for all plan state.
• Do not edit plan.json directly; use hoyeon-cli plan task for every status transition.
• Use logical Hoyeon subagent names in the charter, mapped to Codex adapters when installed:
  • worker -> hoyeon-worker
  • verifier -> hoyeon-verifier
  • code-reviewer -> hoyeon-code-reviewer
• In Codex, translate every logical Agent(...) dispatch in the references to the native spawn_agent tool:
  • Agent(subagent_type="worker", ...) -> spawn_agent(agent_type="hoyeon-worker", ...)
  • Agent(subagent_type="verifier", ...) -> spawn_agent(agent_type="hoyeon-verifier", ...)
  • Agent(subagent_type="code-reviewer", ...) -> spawn_agent(agent_type="hoyeon-code-reviewer", ...) Pass the worker charter as the agent message, and keep the charter's path-and-ID-only contract unchanged.
• Treat TaskCreate, TaskUpdate, TaskOutput, and TeamCreate examples in the reference recipes as Claude Code protocol notes, not literal Codex calls. Codex execute state is tracked through hoyeon-cli plan task plus returned subagent messages.
• If the current Codex session has not loaded the adapter names, fall back to direct single-worker execution and keep the same charter/output contract.
• Do not rely on hooks, TeamCreate, or automatic stop transitions in Codex v1.
• Parallel Codex worker dispatch is allowed for disjoint parallel_safe tasks when the hoyeon-worker adapter is prompt-visible. Use spawn_agent for dispatch, hoyeon-cli plan task for state, and returned subagent messages for evidence.
• scripts/codex-execute-smoke.sh validates only single-worker plan state transitions. It does not prove parallel subagent behavior; verify parallel changes with a bounded live spawn_agent smoke.

Phase 0: Initialize

Phase 0 is plan-first. The orchestrator resolves an input to a valid plan.json, asks two questions (dispatch + verify depth), and prepares a worker charter template. Phase 0 never reads requirements.md or contracts.md body — only plan.json structural fields (INV-3).

0.1 Parse Input & Resolve

/execute [<spec_dir>] [--work worktree|branch|no-commit]

raw_path = $1  (may be empty)

# (a) No argument → virtual plan path (R-F1.3)
IF raw_path is empty:
  input_mode = "virtual"
  spec_dir   = null   # resolved later in 0.2
  GOTO 0.2

# (b) Argument provided → must exist (R-F1.4)
IF NOT exists(raw_path):
  ERROR: "No such path: {raw_path}"
  guidance: "Provide a directory containing plan.json or requirements.md,
             or call /execute with no argument to synthesize a virtual plan."
  ABORT

spec_dir = raw_path if is_dir(raw_path) else dirname(raw_path)

# (c) Classify what we found
IF exists(spec_dir/plan.json):
  input_mode = "plan"          # R-F1.1
ELIF exists(spec_dir/requirements.md):
  input_mode = "requirements"  # R-F1.2
ELSE:
  ERROR: "{spec_dir} has neither plan.json nor requirements.md"
  guidance: "Run /blueprint on requirements.md, or call /execute without arguments
             for a session-synthesized virtual plan."
  ABORT

0.2 Resolve to plan.json

Handle each input_mode — the goal is to end this sub-phase with a validated plan.json on disk and spec_dir set.

IF input_mode == "plan":                             # R-F1.1
  # Direct use — NO user confirm.
  Bash("hoyeon-cli plan validate {spec_dir}")

ELIF input_mode == "requirements":                   # R-F1.2
  # Try /blueprint first; fall back to inline planning if unavailable.
  IF Skill(blueprint, args="{spec_dir}") succeeds AND exists(spec_dir/plan.json):
    Bash("hoyeon-cli plan validate {spec_dir}")
  ELSE:
    # Inline planning — lightweight alternative to /blueprint.
    # Read requirements.md directly (exception to INV-3 — only during init).
    reqs_content = Read("{spec_dir}/requirements.md")

    # Extract frontmatter (type, goal, non_goals) + sub-requirements
    meta_type = parse_frontmatter(reqs_content).type
    sub_reqs  = parse_sub_reqs(reqs_content)  # [{id, given, when, then}]

    # Check pre-work section (if present)
    pre_work = parse_pre_work(reqs_content)   # see "Pre-work Gate" below

    # Generate task graph inline (no contracts, no journeys)
    draft_tasks = []
    FOR EACH logical group of sub_reqs (by parent R-X):
      draft_tasks.append({
        id: "T{n}", action: "<what to build>",
        fulfills: [sub_req.id for sub_req in group],
        depends_on: [], parallel_safe: true
      })

    # Generate basic verify_plan (gate 1+2 for all)
    draft_verify = []
    FOR EACH sub_req in sub_reqs:
      draft_verify.append({ target: sub_req.id, type: "sub_req", gates: [1, 2] })

    # Preview (same pattern as blueprint Step 2.3)
    print("[execute] Inline Plan (no /blueprint)")
    print_task_table(draft_tasks)
    print(f"Verify: {len(draft_verify)} entries (all G1+G2)")

    choice = AskUserQuestion(
      question: "Proceed with this inline plan?",
      options: [
        { label: "Proceed", description: "Write plan.json and execute" },
        { label: "Edit",    description: "Revise tasks before proceeding" },
        { label: "Abort",   description: "Stop — run /blueprint first for a detailed plan" }
      ]
    )
    IF choice == "Abort": HALT
    IF choice == "Edit": draft_tasks = interactive_edit(draft_tasks)

    # Write via cli
    Bash("hoyeon-cli plan init {spec_dir} --type {meta_type}")
    write_json_to_tmp({tasks: draft_tasks, verify_plan: draft_verify}) → /tmp/plan-inline.json
    Bash("hoyeon-cli plan merge {spec_dir} --json \"$(cat /tmp/plan-inline.json)\"")
    Bash("hoyeon-cli plan validate {spec_dir}")

ELIF input_mode == "virtual":                        # R-F1.3
  # Session-context synthesis with user confirm.
  timestamp = Bash("date +%Y%m%d-%H%M%S").trim()
  spec_dir  = ".hoyeon/specs/adhoc-{timestamp}"
  Bash("mkdir -p {spec_dir}")

  # Synthesize a minimal plan from recent user messages + cwd state.
  # Keep it in memory first; do NOT write until the user confirms.
  draft_plan = synthesize_virtual_plan(
    recent_user_messages,
    cwd_state = Bash("ls -la").trim()
  )
  # draft_plan shape: { meta, tasks[{id, action, fulfills:[], depends_on:[], parallel_safe}], verify_plan:[] }

  # Summary preview (stdout, not a file)
  print("Virtual plan synthesized ({len(draft_plan.tasks)} tasks):")
  for t in draft_plan.tasks: print("  {t.id}  {t.action}")

  choice = AskUserQuestion(
    question: "Proceed with this virtual plan?",
    options: [
      { label: "Proceed", description: "Write plan.json to {spec_dir} and execute" },
      { label: "Edit",    description: "Open an interactive edit loop to revise tasks" },
      { label: "Abort",   description: "Discard and exit" }
    ]
  )
  IF choice == "Abort": HALT
  IF choice == "Edit":
    draft_plan = interactive_edit(draft_plan)  # loop until user says proceed

  # Persist via cli (INV-5). Write, then validate.
  write_json_to_tmp(draft_plan) → /tmp/plan-virtual.json
  Bash("hoyeon-cli plan init {spec_dir} --type {draft_plan.meta.type}")
  Bash("hoyeon-cli plan merge {spec_dir} --json \"$(cat /tmp/plan-virtual.json)\"")
  Bash("hoyeon-cli plan validate {spec_dir}")

At the end of 0.2 we have: spec_dir/plan.json (valid) + input_mode + (optionally) spec_dir/contracts.md.

0.2b Pre-work Gate (optional)

If requirements.md exists and contains a ## Pre-work section, parse it and gate on blocking items before proceeding.

IF exists("{spec_dir}/requirements.md"):
  pre_work = scan for "## Pre-work" section → parse checkbox items
  # Expected format: "- [ ] action (blocking)" or "- [ ] action (non-blocking)"
  blocking = [item for item in pre_work if "blocking" in item]

  IF len(blocking) > 0:
    print("Pre-work items requiring completion:")
    FOR EACH item in blocking:
      print("  - {item.action}")

    AskUserQuestion(
      question: "Have you completed all blocking pre-work items?",
      options: [
        { label: "Yes, all done", description: "Continue to execution" },
        { label: "Not yet",       description: "Abort — complete pre-work first" }
      ]
    )
    IF answer == "Not yet": HALT

If there is no ## Pre-work section, skip silently.

0.3 Load plan.json (structural fields only)

INV-3: read only structural fields from plan.json. Do NOT open requirements.md or contracts.md body here.

plan = Read("{spec_dir}/plan.json") → JSON.parse

# Structural metrics used by 0.4 prompts
task_count     = len(plan.tasks)
parallel_count = count(t for t in plan.tasks if t.parallel_safe)
parallel_ratio = parallel_count / task_count if task_count else 0

# Gate distribution across verify_plan (for the verify-depth hint)
gate_hist = {1:0, 2:0, 3:0, 4:0}
FOR vp in plan.verify_plan:
  FOR g in vp.gates: gate_hist[g] += 1
max_gate = max(g for g,c in gate_hist.items() if c > 0) if any else 0

# contracts.md path — path only, never the body (INV-2)
contracts_path = "{spec_dir}/contracts.md" if exists(spec_dir/contracts.md) else null

0.4 User Configuration (dispatch + verify depth)

Two AskUserQuestion calls in order. Both include a structural hint computed in 0.3.

# --- Dispatch mode (R-F2.1) -----------------------------------------
# Recommendation from task count + parallel_safe ratio:
IF task_count <= 3:          recommended = "Direct"
ELIF parallel_ratio >= 0.6:  recommended = "Team"      # many independent tasks
ELSE:                        recommended = "Agent"

dispatch = AskUserQuestion(
  question: "Dispatch mode? ({task_count} tasks, {parallel_count} parallel_safe → recommended: {recommended})",
  options: [
    { label: "Direct", description: "Orchestrator executes tasks sequentially in its own context (best for ≤3 tasks)" },
    { label: "Agent",  description: "Spawn worker subagents per module group, round-level commit" },
    { label: "Team",   description: "TeamCreate persistent workers claim tasks (best for high parallel_safe ratio)" }
  ]
)

# --- Verify depth (R-F2.2) ------------------------------------------
# Hint shows gate distribution from plan.verify_plan.
hint = "gates present: " + join([f"{g}={gate_hist[g]}" for g in [1,2,3,4] if gate_hist[g] > 0], ", ")
IF max_gate == 0: hint = "no verify_plan entries"

verify = AskUserQuestion(
  question: "Verify depth? ({hint})",
  options: [
    { label: "Light",    description: "Gate 1 only — build/lint/typecheck (caps all sub_reqs at gate ≤ 1)" },
    { label: "Standard", description: "Gates 1-2 — build + sub_req double-review (caps all sub_reqs at gate ≤ 2)" },
    { label: "Thorough", description: "All gates — no cap; runs gate 3 (qa-verifier) where planned" }
  ]
)

# --- Work mode (flag or prompt) -------------------------------------
IF --work flag provided:
  work = flag_value
ELSE:
  work = AskUserQuestion(
    question: "Work mode?",
    options: [
      { label: "Worktree",           description: "Isolated worktree, commit per round" },
      { label: "New Branch + Commit", description: "Create feat/ branch from current, commit per round" },
      { label: "Branch + Commit",     description: "Current branch as-is, commit per round" },
      { label: "No Commit",           description: "No git commits" }
    ]
  )

0.5 Setup — session state + charter template

# (a) Session state
STATE_FILE="$HOME/.hoyeon/$CLAUDE_SESSION_ID/state.json"
Bash: jq -n \
  --arg dispatch "{dispatch}" \
  --arg verify   "{verify}" \
  --arg work     "{work}" \
  --arg spec_dir "{spec_dir}" \
  --arg input    "{input_mode}" \
  --arg contracts "{contracts_path or ""}" \
  '{dispatch:$dispatch, verify:$verify, work:$work, spec_dir:$spec_dir,
    input_mode:$input, contracts_path: ($contracts|select(length>0))}' \
  > $STATE_FILE

# (b) Branch/Worktree setup
IF work == "Worktree":
  spec_dir       = Bash("realpath {spec_dir}").trim()
  contracts_path = Bash("realpath {contracts_path}").trim() if contracts_path
  EnterWorktree(name=basename(spec_dir))

IF work == "New Branch + Commit":
  branch_name = "feat/{spec_name}"   # derived from spec_dir basename
  Bash: git checkout -b {branch_name}
  audit_append("BRANCH_CREATE {branch_name} from {current_branch}")

# (c) Context files (next to plan.json)
CONTEXT_DIR = spec_dir
Bash: [ -s {CONTEXT_DIR}/learnings.json ] || echo '[]' > {CONTEXT_DIR}/learnings.json   # initialize to [] if new
Bash: [ -s {CONTEXT_DIR}/issues.json ]    || echo '[]' > {CONTEXT_DIR}/issues.json      # initialize to [] if new
Bash: touch {CONTEXT_DIR}/audit.md

# (d) Worker charter template — paths and IDs only (INV-2, R-N15.1)
# NEVER inline GWT, requirements prose, or contracts body into this template.
CHARTER_TEMPLATE = {
  task_id:              "<injected per task>",
  plan_path:            "{spec_dir}/plan.json",
  contracts_path:       contracts_path,          # path only — see R-F2.3
  contracts_directive:  (contracts_path != null)
                          ? "Read contracts.md before coding — it defines the
                             cross-module surface you must respect."
                          : null,
  sub_req_ids:          "<injected per task from plan.tasks[].fulfills>",
  round:                1,
  prior_failure_context: null
}

CHARTER_TEMPLATE is consumed by dispatch references (direct/agent/team) and by the worker charter recipe (T7). The contracts_directive field fulfills R-F2.3: if contracts.md exists, its path is embedded in every charter together with a "read before coding" instruction; its body is never inlined (INV-2).

Orchestrator Boundaries (INV-3, R-N15.2)

The orchestrator is a router over structure. It must NOT read the body of any spec prose. Enforced everywhere in this skill and all dispatch/verify references.

MAY read (structural, via hoyeon-cli plan get or Read on plan.json only):

• plan.json fields: tasks, journeys, verify_plan, meta, contracts.artifact, context
• Session state in $HOME/.hoyeon/$CLAUDE_SESSION_ID/state.json
• Worker output JSON returned by dispatch (WorkerOutput payload)
• audit.md, learnings.json, issues.json for round-to-round context

MUST NOT read:

• requirements.md body (GWT, behavior prose, decisions, context text)
• contracts.md body (invariants, interfaces, data shapes — only the path flows through charters)

That body-read responsibility belongs to workers via the self-read pattern in references/worker-charter.md (§1.3, §2 Self-Read). If the orchestrator ever needs to influence behavior based on a requirement's body, route through a worker or through the contracts-patch recipe — never inline-read.

Resume Behavior (R-F8.2 idempotent restart)

On any re-entry (compaction recovery, session restart, /execute re-invocation on an existing spec_dir), the orchestrator MUST treat plan.json as the source of truth and skip any task whose status == "done". This rule applies uniformly across all three dispatch modes.

# Runs at the top of every dispatch recipe, before any worker is spawned.
plan       = Read("{plan_path}") → JSON.parse
done_ids   = { t.id for t in plan.tasks if t.status == "done" }   # INV-9: monotonic
pending    = [ t for t in plan.tasks if t.status != "done" ]

# Downstream readiness uses done_ids to satisfy depends_on.
ready = [ t for t in pending if all(d in done_ids for d in t.depends_on) ]

Additional guarantees:

• Verify (references/verify.md) builds its coverage matrix from done_tasks only, so a partial run resumes verification on exactly the tasks that completed.
• hoyeon-cli plan task --status X=done is idempotent (INV-5, INV-9); re-issuing it on a task that is already done is a no-op, never a re-transition.
• The orchestrator NEVER rewrites plan.json directly — resume reads are through hoyeon-cli plan get (INV-5).

Concurrency Rules (INV-4 — no sleep, no polling)

Every parallel burst MUST be emitted as a single message with all background dispatches at once. Results arrive via notifications (TaskOutput, SendMessage, run_in_background:true completion events), never via a sleep / re-read poll loop.

Hard bans — enforced across SKILL.md and every dispatch / verify reference:

• No sleep <n> between dispatches.
• No while not done: ... sleep over plan.json or worker state.
• No hoyeon-cli plan get called in a loop to poll for status == "done".
• No serial Bash calls in separate messages where a single-message parallel burst would work (e.g. per-round worker fan-out, gate-1 toolchain fan-out).

Already-enforced sites (for reference):

• direct.md — sequential by design, no polling (one task per loop iteration).
• agent.md — Phase C notification handler, explicitly await_round_notifications(round_id) with the comment "NO SLEEP" (R-N16.1 fulfilled).
• team.md — Phase 2 lead loop uses wait_for_SendMessage() events; standing-by workers wait on inbound SendMessage rather than polling plan.json (INV-4).
• verify.md — INV-4 listed in invariants block; gate-1/gate-2 fan-out runs as a single-message burst.

If you feel tempted to add sleep, you are wrong. Use background dispatch + notification handler instead.

Dispatch Routing

The orchestrator picks one of three recipes based on dispatch from Phase 0.4. All three consume the same Phase 0 state and delegate worker construction to the canonical charter in references/worker-charter.md — no recipe inlines its own charter format.

IF dispatch == "direct":
  Read: ${baseDir}/references/direct.md
  Follow ALL instructions. (sequential — one task at a time in orchestrator context)

ELIF dispatch == "agent":
  Read: ${baseDir}/references/agent.md
  Follow ALL instructions. (round-based parallel — TaskCreate per ready-set group)

ELIF dispatch == "team":
  Read: ${baseDir}/references/team.md
  Follow ALL instructions. (persistent TeamCreate workers claim tasks)

Canonical charter — every dispatch recipe builds worker charters by importing references/worker-charter.md. The charter carries paths and IDs only (INV-2, R-N15.1): task_id, plan_path, contracts_path, sub_req_ids, round, prior_failure_context. No GWT, no requirements prose, no contracts body.

All dispatch references receive these variables from Phase 0:

• plan_path — {spec_dir}/plan.json (workers self-read it, orchestrator structural-reads it)
• spec_dir, CONTEXT_DIR — directory paths
• contracts_path — path to contracts.md, or null
• work — "Worktree" | "Branch + Commit" | "No Commit"
• verify — "Light" | "Standard" | "Thorough" (depth cap for verify.md)
• CHARTER_TEMPLATE — from Phase 0.5, passed through to worker-charter.md

Resume behavior (R-F8.2) is mandatory across all three — see "Resume Behavior" above for the uniform done-skip contract.

Verify Routing

After all plan tasks reach done / failed / blocked, run verification. There is a single verify recipe; verify depth is a parameter that caps gates per sub_req, not a selector between different recipes (T8 design).

Read: ${baseDir}/references/verify.md
Follow ALL instructions; pass `verify` ∈ {"light","standard","thorough"} as the
depth parameter. verify.md reads done tasks from plan.json (structural only) and
evaluates each sub_req / journey against its capped gate set.

Contracts mismatch hook (C4, R-F9.1)

On any worker output that signals a cross-module contract mismatch — during dispatch OR during verify fix loops — the orchestrator invokes the auto-patch recipe before marking the task done or enqueueing a retry:

Read: ${baseDir}/references/contracts-patch.md
Follow ALL instructions. (detect → patch → audit-log → return control)

This hook runs without user confirmation (INV-7), is idempotent per worker output, and routes only through Read / Edit / Write on contracts.md and audit.md — never through hoyeon-cli (INV-5). See references/contracts-patch.md for detection signals (explicit contract_mismatch, contract_issues[], or BLOCKED-with-contract-reason).

Output Artifacts (R-F13.2, R-F13.3, R-F13.4)

execute produces exactly 5 file artifacts during a run, all inside <spec_dir>/. There is no report.md — the final report is stdout-only (INV-8 / R-F13.4 / C5).

audit.md (R-F13.2) — append-only timestamped event log

The orchestrator appends one entry per orchestrator-level event. Format: a markdown list item starting with an ISO-8601 timestamp. Entries are append-only — never edit or delete prior entries. audit.md is initialized as an empty file by Phase 0.5.

The orchestrator writes audit entries via Read + Edit (append) — never via hoyeon-cli (INV-5 / R-N17.1). The contracts-patch recipe writes its own CONTRACTS_PATCH line directly (see references/contracts-patch.md §"audit.md append").

learnings.json + issues.json (R-F13.3) — worker-managed JSON arrays

Both files are flat JSON arrays initialized to [] in Phase 0.5. Workers append entries per references/worker-charter.md §3.5 using Read + Write. The orchestrator only reads these files (round-to-round context, final report).

• learnings.json — appended on worker success. Schema is owned by worker-charter.md.
• issues.json — appended when worker returns BLOCKED or FAILED. Same constraint.

Workers MUST NOT use hoyeon-cli to mutate either file (INV-5).

NO report.md (R-F13.4 / INV-8 / C5)

The orchestrator MUST NOT create a report.md, summary.md, or any other report file. The final report is emitted to stdout only (see "Stdout Report Format" below). This rule applies even if the user asks for a file — surface it via stdout and let the user redirect if they want to capture it.

Stdout Report Format (R-F14.1, R-F14.2, R-F14.3)

After verify completes, the orchestrator prints one final report to stdout — no file is written (INV-8 / R-F13.4). The report has 8 fixed sections in this exact order (R-F14.1):

1. Status — ✅ COMPLETE | ⚠️ PARTIAL | ❌ FAILED
2. Summary — task counts, sub_req coverage counts, contracts patch count
3. Post-Work — action items for the user (directly under Summary, action-first; R-F14.2)
4. Tasks — table: T# / status / commits / round
5. Verify Coverage Matrix — sub_req × gate table with PASS / FAIL / — / MANUAL / GAP cells (R-F14.3)
6. Contracts Auto-Patches — list extracted from audit.md CONTRACTS_PATCH entries
7. Issues Encountered — extracted from issues.json (and audit.md BLOCKED / ABORT entries)
8. Learnings — extracted from learnings.json

Status determination

Section sources

Verify Coverage Matrix (R-F14.3)

Rendered as a markdown table — rows = sub_req, columns = gate, last column = overall result. Cell values: PASS, FAIL, — (gate not applicable at depth), MANUAL (gate=4), GAP (no fulfilling citation).

Concrete example

## ✅ COMPLETE — execute run 2026-04-17T14:32:11Z

### Summary
- Tasks: 13 done / 0 failed / 0 blocked
- Sub-requirements: 50 PASS / 2 MANUAL / 0 FAIL / 0 GAP (52 total)
- Contracts auto-patches: 1
- Dispatch mode: agent · Verify depth: standard · Work mode: branch

### Post-Work
- [ ] MANUAL REVIEW: R-F11.4 (gate=4) — confirm GWT citations match human intent
- [ ] MANUAL REVIEW: R-F12.4 (gate=4) — confirm escalation policy is acceptable

### Tasks
| Task | Status | Commits          | Round |
|------|--------|------------------|-------|
| T1   | done   | a1b2c3d          | 1     |
| T2   | done   | a1b2c3d          | 1     |
| T3   | done   | e4f5g6h          | 2     |
| ...  | ...    | ...              | ...   |

### Verify Coverage Matrix
| sub_req   | Gate 1 | Gate 2 | Gate 3 | Gate 4   | Overall |
|-----------|--------|--------|--------|----------|---------|
| R-F1.1    | PASS   | PASS   | —      | —        | PASS    |
| R-F1.2    | PASS   | PASS   | —      | —        | PASS    |
| R-F11.4   | PASS   | PASS   | —      | MANUAL   | MANUAL  |
| R-F12.4   | PASS   | PASS   | —      | MANUAL   | MANUAL  |
| ...       | ...    | ...    | ...    | ...      | ...     |

### Contracts Auto-Patches
- 2026-04-17T14:18:02Z — `contracts.md:88` — interface `FooAPI.call` signature aligned with worker T7 output (`+ retries: number`)

### Issues Encountered
- (none)

### Learnings
- T7: worker self-read pattern — pulling sub_req IDs from plan.json before reading body avoids re-parsing on round 2
- T8: gate=1 toolchain fan-out via single-message `run_in_background:true` cut wall time ~40% vs serial

The example is illustrative; counts, hashes, and timestamps come from live data. The section order, headings, and matrix shape are normative (R-F14.1, R-F14.2, R-F14.3).

Generic Rules

1. plan.json is the ledger — all task CRUD via hoyeon-cli plan (status/list/get/merge). Never direct file writes (INV-5).
2. Structural reads only — orchestrator reads plan.json fields, never requirements.md / contracts.md body (see Orchestrator Boundaries above; INV-3, R-N15.2).
3. Two-turn task setup — Turn 1: all TaskCreate. Turn 2: all TaskUpdate dependencies.
4. Background for parallel — run_in_background: true for concurrent workers, single-message burst (INV-4; see Concurrency Rules).
5. Contracts are reference — workers receive contracts_path to Read, not inlined content (INV-2).
6. Workers self-read — workers fetch their own task state + context files per references/worker-charter.md §2.
7. Context files — learnings.json, audit.md, issues.json in CONTEXT_DIR. Workers append learnings; orchestrator reads them round-to-round.
8. Compaction recovery — session-compact-hook.sh re-injects state. Use hoyeon-cli plan list to rebuild; done tasks skip on re-entry (R-F8.2; see Resume Behavior above).

Checklist Before Stopping

• Input detected and normalized (plan.json / requirements.md / virtual)
• Plan generated or validated via hoyeon-cli
• Dispatch/verify/work modes selected
• All tasks dispatched and completed (status "done"/"failed"/"blocked" in plan.json)
• Verify recipe ran (single verify.md with selected depth)
• Contracts auto-patch hook invoked on any mismatch signals
• Final report output
• Worktree exited (if entered)