Yesloop

---
name: yesloop
description: Autonomous task loop — analyze, plan, execute, verify, review, finish. Runs as visible TUI agent in git worktree. Use when user says "yesloop", "loop", "run it autonomously".
user-invocable: true
disable-model-invocation: false
allowed-tools: Bash(git *) Bash(go *) Bash(make *) Bash(npm *) Bash(docker *) Bash(python3 *) Bash(sqlite3 *) Bash(curl *) Bash(find *) Bash(ls *) Bash(cat *) Bash(rg *) Bash(mv *) Bash(cp *) Bash(rm *) Bash(mkdir *) Bash(test *) Grep Glob Read Write Edit todowrite Task
---

# Autonomous Task Loop

You are an autonomous agent. Work through tasks without asking the user.

## CRITICAL: Always spawn in a worktree

When invoked from an interactive session, **do NOT execute the pipeline yourself.** Instead, spawn an isolated agent in a git worktree:

1. **Create worktree:** `git worktree add -b yesloop/<task-slug> <repo>/.worktrees/yesloop-<task-slug>`
   - Isolates agent changes from your working directory
   - Multiple agents can run in parallel without conflicts
   - Easy cleanup: delete worktree + branch if abandoned
2. Write the task to scratchpad with worktree path and skill-read mandate:
   `scratchpad_write(project="<project>", section="yesloop-<task-slug>", content="YOU ARE A YESLOOP-AGENT. MANDATORY - YOU MUST READ THE YESLOOP SKILL AND YOU MUST FOLLOW THE INSTRUCTIONS IN THIS SKILL. Load the yesloop skill: use your platform's skill tool, or read ~/.claude/skills/yesloop/SKILL.md, or <repo-root>/skills/bundled-skills/yesloop/SKILL.md. Follow the 6-phase pipeline (ANALYZE → PLAN → EXECUTE → VERIFY → REVIEW → FINISH) FORMALLY AND SUBSTANTIVELY with ATTENTION.\n\nWorktree: <path>\nTask: <description>\n\nReport to scratchpad.")`
   The spawned agent has NOT auto-loaded this skill — the mandate prefix is mandatory.
3. **Spawn TUI agent (MANDATORY — never run the pipeline inline):** `yesmem_spawn_agent(project="<project>", section="yesloop-<task-slug>", backend="opencode", work_dir="<worktree-path>")`
4. Wait 15s for opencode TUI to load + PTY injection to deliver the startup prompt
5. **Relay kick (backup):** `yesmem_relay_agent(to="<agent-id>", content="Read scratchpad section 'yesloop-<task-slug>' AND load the yesloop skill (skill tool or ~/.claude/skills/yesloop/SKILL.md). MANDATORY - YOU MUST READ THE YESLOOP SKILL AND YOU MUST FOLLOW THE INSTRUCTIONS IN THIS SKILL. Begin the 6-phase pipeline FORMALLY AND SUBSTANTIVELY with ATTENTION.\n")` — backup if PTY is slow
6. Confirm: "Agent spawned in worktree — visible in terminal."

**Only run inline if:** the user EXPLICITLY says `--inline`. There is NO trivial-task exception — "< 2 min" is not a license to skip the worktree. If you'd commit it, it runs in the worktree. Create the worktree BEFORE the first edit.

**"Pipeline" includes exploratory work.** Writing tests, spiking an approach, "just quickly trying the fix on main to see if it works" — if you'd commit it, it belongs in the worktree. Create the worktree BEFORE the first edit, not after you've verified the fix works.

## Model selection (opencode backend)

`yesmem_spawn_agent(model=...)` accepts three forms, resolved by `resolveSpawnModel` in `internal/daemon/spawn_model.go`:

- **Empty** → returns "" — lets the opencode CLI pick its default from its own configuration. No model is hardcoded in the daemon.
- **Slash-qualified** (e.g. `zai/glm-5.2`, `zai-coding-plan/glm-5.2`) → passed through verbatim.
- **Bare model name** (e.g. `glm-5.2`, `deepseek-v4-pro`) → resolved against the auto-discovered provider map (built from `~/.cache/opencode/models.json` + `~/.config/opencode/opencode.json`). The daemon prepends the matching providerID, e.g. `glm-5.2` becomes `zai/glm-5.2`.

**Coding-optimized variants** (any provider whose ID contains "coding", e.g. `zai-coding-plan`) are included in the auto-discovery map like any other provider. When multiple providers carry the same bare modelID, the alphabetically-first ProviderID wins deterministically; resolve ambiguity explicitly via `model="provider/model"`. Rule of thumb: bare names resolve to whatever the active opencode config exposes; if only `zai-coding-plan` carries `glm-5.2`, bare `glm-5.2` resolves to `zai-coding-plan/glm-5.2`.

Unknown bare names log a warning and pass through unchanged (best-effort, no error).

## Orchestrator Contract (what to prescribe vs. delegate)

When you spawn a yesloop agent, the scratchpad task defines the working relationship. Over-prescribing takes ANALYZE+PLAN away from the agent; under-prescribing sends it into orientation loops. Get the level right.

**Prescribe (orchestrator's job):**
- **Skill-Read-Mandate** — The scratchpad task MUST start with 'YOU ARE A YESLOOP-AGENT.' and instruct the agent to load the yesloop skill AND read it with: MANDATORY - YOU MUST READ THE YESLOOP SKILL AND YOU MUST FOLLOW THE INSTRUCTIONS IN THIS SKILL. The agent MUST follow this skill FORMALLY AND SUBSTANTIVELY — every phase, block format, guardrail, and rule — with ATTENTION and MANDATORY. The spawned agent has NOT auto-loaded this skill; without the mandate it works without pipeline context. Path fallbacks: platform skill tool, `~/.claude/skills/yesloop/SKILL.md`, or `<repo-root>/skills/bundled-skills/yesloop/SKILL.md`.
- **Goal** in 1-2 sentences — what success looks like, not how to get there
- **Dense context** — facts, file paths, relevant learning IDs, what's already been tried, what failed. The agent starts cold; you don't.
- **Hard constraints** — schema-breaking changes, backfills, new dependencies, destructive ops, files off-limits
- **Escalation triggers** — decisions that belong to the user (product direction, API shape changes, breaking compatibility)

**Delegate (agent's job):**
- **ANALYZE** — reading code, grepping, finding the call sites, identifying edge cases
- **PLAN** — commit structure, file selection, test strategy, step ordering, milestone breakdown
- **EXECUTE/VERIFY/REVIEW** — the full 6-phase pipeline on the agent's plan, not yours

**Calibration test:** Before writing the scratchpad, ask: "If the agent came back with a plan, would I be surprised?" If yes for the right reasons (better approach), you prescribed right. If yes for wrong reasons (misunderstood goal), add context. If bored (matches what you'd write), over-prescribed — shorten.

## Relay-to-Resume Timing (paused agents)

Relay delivery to a `paused` agent is **asynchronous**: the daemon heartbeats the inject socket at ~1s intervals, the opencode TUI may be mid-turn, and the PTY bridge buffers. A relay is NOT reflected in agent status or scratchpad within seconds.

**Rule: after `relay_agent` on a paused agent, wait ≥ 2 minutes before re-checking status or concluding the relay failed.** A status that still reads `paused` 10 seconds after a relay is expected behavior, not a failure — the relay hasn't been picked up yet.

Anti-pattern (observed): orchestrator relays approval → checks status 5s later → still `paused` → concludes "relay did nothing, must resume manually" → calls `resume_agent`. This is redundant (paused PID is alive, PTY bridge open — resume is a no-op dressed up as action) and confusing. See Learning #81175: for `paused` agents, `relay_agent` is the correct action; `resume_agent` is only for genuinely dead processes.

Correct sequence after relaying to a paused agent:
1. `relay_agent(to=<id>, content="...")` — approval/instruction delivered
2. **Wait ≥ 2 minutes** — the agent needs to pick up the relay on its next tick, process it, write scratchpad, and call `update_agent_status`. Shorter windows produce false negatives.
3. THEN re-check: `get_agent` status + `scratchpad_read(project, section)` for new evidence (phase update, milestone, DONE marker)
4. If after 2 min there is NO new scratchpad activity AND status unchanged → read the `progress` field. If it shows an escalation/waiting state, the agent is correctly paused — relay again or escalate higher. Only consider `resume_agent` if PID is dead or `progress` indicates a real stall (crash, timeout without escalation line).

## Monitoring discipline (running agents)

Intervention requires stale **activity signals**, not just a stale status label. A running agent's status field can lag behind actual work — the agent may be mid-turn, processing a long tool call, or writing scratchpad while `update_agent_status` hasn't fired yet.

**Rule: only intervene when ALL of these are stale for several minutes:**
- `last_activity_at` — no new activity timestamp
- `turns_used` — turn count not advancing
- `stream_bytes` — SSE stream not producing output

A status label alone is not a trigger. If activity signals are advancing, the agent is working — let it run. The monitoring job (heartbeat) is the next check, not a manual poke.

Anti-pattern: orchestrator sees status `paused` or `running` with no recent context → assumes stall → relays/resumes/kills. This disrupts agents that are mid-work. Check the activity signals first. See Learning #82126.

## Temp file discipline

NEVER write to `/tmp/` or `~/.claude/yesmem/tmp/` in autonomous operations. `/tmp/` is unreliable across sandbox/container contexts (logs vanish, see #72938). Global paths collide when multiple agents run in parallel.

Use `<worktree>/.yesmem/tmp/` instead. Project-local, always writable from any agent context, cleaned up with worktree removal. `mkdir -p .yesmem/tmp` at session start.

## Execution Modes

**tui-agent** (DEFAULT) — Spawned via yesmem_spawn_agent → gnome-terminal, visible, non-blocking.
**inline** — User explicitly requested inline. Run in current context.
**scheduled** — User asked for recurring work. Use yesmem_schedule.

## Definition of DONE

DONE is a contract, not a feeling. Status is DONE only when ALL six phase sections exist in scratchpad and each carries `**Status:** COMPLETE`.

**Valid Status values per phase:**
- `IN PROGRESS` — actively working
- `COMPLETE` — phase finished, evidence present
- `BLOCKED — <reason>` — cannot proceed, needs orchestrator input

**Phase status update is mandatory at every transition:**
```
update_agent_status(phase="Phase N/6 NAME")
update_agent_status(phase="Phase N/6 NAME (milestone M/K)")   # EXECUTE with milestones
update_agent_status(phase="Phase N/6 NAME (blocked: <why>)")
```

Free-form string, canonical format. Single source of truth for live progress.

**Orchestrator validates DONE:**
- DONE-guard (auto, every 30s) checks phase block compliance via regex
- Manual fallback: Read scratchpad section
- Check `### Phase 1` through `### Phase 6` headers exist
- Each has `**Status:** COMPLETE` (or `BLOCKED — <reason>` for genuinely blocked phases)
- Phase 5 has `self-review: done` plus evidence-backed `cold-review: ses_<id>`, `consequence: ses_<id>`, `security: ses_<id>` dispatch lines
- Phase 6 has `Deploy executed: yes` or `Deploy required: no`

Missing evidence → guard fires automatically (freeze + notify orchestrator).

## Automated DONE-Guard (v3)

Since yesmem v2.1.21, the heartbeat scheduler enforces the DONE contract **automatically** every 30 seconds. The DONE-guard (`internal/daemon/done_gate.go`) uses deterministic regex validation — no LLM, no `claude -p`, no false positives from model drift.

**What the guard checks:**

1. All 6 phase blocks (`### Phase 1` through `### Phase 6`) must be present in scratchpad
2. Each block must have `**Status:**` on its own line with valid value (COMPLETE|BLOCKED|IN PROGRESS)
3. Per-phase required fields (see Phase Pipeline below)
4. Phase 5 must contain `self-review: done` AND a `cold-review: ses_<id>` dispatch line
5. Phase 4 must contain `**Regression baseline:**` with a non-empty value and `**RED proof:**`
6. Phase 5 must contain `consequence: ses_<id>` AND `security: ses_<id>` dispatch lines — after the format check, the daemon verifies each reported id against opencode's session DB (row exists, parent session matches the agent's own session, agent type allowed per sub-phase, ids pairwise distinct)
7. Phase 6 must contain `send_to orchestrator:` and deploy evidence
8. Phase 5 must contain `**Stufe:**` (value 1-3 with a same-line rationale) and a non-empty `**Modules:**` line — the risk-based-code-review classification from the 5.2 cold reviewer

**When the guard fires:**
- Agent is paused with reason `DONE-GUARD: phase validation failed — <errors>`
- Orchestrator receives `⛔ DONE-GUARD:` notification via `send_to`
- Agent must fix phase blocks and be resumed

Before pausing, the guard relays the concrete missing fields and re-fires up to 2 times at 90s intervals. If a DONE-GUARD-paused agent's scratchpad becomes compliant, the guard unpauses it automatically (RECOVERY) and notifies the orchestrator.

**Phases missing from scratchpad = still in progress.** The guard only escalates when all 6 phase blocks exist but their content is non-compliant.

### DISMISSAL-PATTERN

These patterns cause the guard to reject the DONE claim:

| Pattern | Detection | Why it fails |
|---|---|---|
| Status on same line as header | `### Phase N: → **Status:**` inline | Status must be on its own `**Status:**` line |
| Old-format Phase 5 | Retired `task() dispatched:` / `consequence dispatched:` lines, new dispatch lines missing | Hard cut — the relay names the new field (`cold-review: ses_<id>`) |
| Missing self-review dispatch | No `self-review: done` in Phase 5 | Sub-phase 5.1 is mandatory |
| Missing cold-review dispatch | No `cold-review: ses_<id>` in Phase 5 | Sub-phase 5.2 is mandatory (supersedes the old Stage 2 header check) |
| Missing consequence dispatch | No `consequence: ses_<id>` in Phase 5 | Sub-phase 5.3 is mandatory |
| Missing security dispatch | No `security: ses_<id>` in Phase 5 | Sub-phase 5.4 is mandatory |
| Missing risk classification | No `**Stufe:**` line with value 1-3 + same-line rationale in Phase 5 | risk-based-code-review classification is required 5.2 evidence |
| Missing module selection | No `**Modules:**` value in Phase 5 | active modules per trigger table are required 5.2 evidence |
| Invalid evidence session ids | Reported id absent from opencode.db, parent session mismatch, duplicated id, or agent type not in the allowed set | Daemon evidence layer verifies substance, not just format (internal/daemon/done_gate_evidence.go) |
| Missing regression baseline | No `**Regression baseline:**` with a non-empty value in Phase 4 | Deterministic base-vs-head test diff is mandatory (quality stage-2) |
| Missing orchestrator notification | No `send_to orchestrator:` in Phase 6 | Orchestrator must be notified for DONE |
| Partial fields in Phase 5 | No issue breakdown, no findings per sub-phase | Self-Review must be structured |
| No deploy evidence in Phase 6 | No `Deploy executed:` or `Deploy required:` | Must document deployment outcome |
| Missing status line | Phase block exists but no `**Status:**` | Every phase needs a status |
| Missing session id note | Phase 1 has no `**Session id:**` field and no "session id missing" note | `whoami()` must be called at Phase 1 start; backend session id (or retry-failure note) must appear in Phase 1 block |

### Findings-Table

When the DONE-guard rejects a claim, its findings are structured as:

| Phase | Missing/Invalid Field | Detail |
|---|---|---|
| 5 | `cold-review:` | required field not found in phase block |
| 6 | `send_to orchestrator:` | required field not found in phase block |

**DISMISSAL → DONE is rejected. Fix fields → resume agent.**

## Phase Pipeline

Each phase writes ONE structured block to scratchpad. Block format is mandatory — prose-only entries are not valid phase completion.

### CRITICAL — scratchpad project string + session id

All `scratchpad_write` and `scratchpad_read` calls MUST use the `project` value returned by `whoami()`. Call `whoami()` as the FIRST action at Phase 1 start and use its `project` field verbatim for every scratchpad call afterwards. Do NOT guess, do NOT use the worktree basename, do NOT hardcode a short name. The orchestrator and the idle/done-verify state machines read from `agent.Project` (set by `yesmem_spawn_agent`); using a different string means your writes and reads land in a different scope and the orchestrator never sees your progress.

`whoami()` also returns the backend session id (`opencode_session_id` for opencode, `codex_session_id` for codex). This id is required for `resume_agent` after crashes. The daemon polls it asynchronously after spawn, so the first `whoami()` call may return it empty. Retry up to 3 times with 5 seconds sleep between calls. If still empty after 3 retries, note "session id missing — resume may fail" in Phase 1 scratchpad and proceed. Do not block beyond 15 seconds.

**SCRATCHPAD DISCIPLINE (MANDATORY — prevents briefing clobber):**

`scratchpad_write` is an UPSERT — it REPLACES the entire section. **The FIRST action after spawn MUST be `scratchpad_read`.** Never call `scratchpad_write` before your first `scratchpad_read` — it will overwrite the orchestrator's briefing (#80063).

For intermediate progress updates (milestone complete, phase done, status markers), use **`scratchpad_append`** instead of `scratchpad_write`. `scratchpad_append` adds content WITH A DIVIDER, preserving the original briefing.

|| Action | Tool |
|---|---|---|
|| First thing after spawn | `scratchpad_read` |
|| Progress update (milestone/phase done) | `scratchpad_append` |
|| Full DONE report (all 6 phases) | `scratchpad_write` |
|| DRIFT alert / STUCK notice | `scratchpad_append` (preserves briefing) |

### Phase 1: ANALYZE
```
update_agent_status(phase="Phase 1/6 ANALYZE")

### Phase 1: ANALYZE
**Status:** COMPLETE
**Goal understood:** <1 sentence>
**Task type:** <debug|feature|chore|docs>
**Session id:** <opencode_session_id or codex_session_id from whoami(), or "missing after 3 retries">
**Codebase explored:** <files/packages inspected>
**Constraints identified:** <list>
**Dense context from memory:** <learning IDs relevant, or "none">
**Risks:** <list>
**Open questions:** none (must be inferrable, no user questions)
```

Information gathering is automatic — never ask the user. Code questions → search_code_index, grep, graph_traverse. API/docs → docs_search, hybrid_search. Errors → deep_search. Unknown concepts → WebFetch. NEVER guess — if you don't know, look it up (WebFetch or the indexed docs); infer only from code/context already in your turn.

### Phase 2: PLAN
```
update_agent_status(phase="Phase 2/6 PLAN")

### Phase 2: PLAN
**Status:** COMPLETE
**Plan stored via set_plan:** yes
**Decisions resolved:** all — files, schema, API, blast radius
  - File selection: <which files, why not others>
  - Schema decisions: <data structures, interfaces, types>
  - API surface: <public interface, signature choices>
  - Blast radius: <what else could break, dependencies affected>
**Milestones:** N total (see criteria below)
  - Milestone 1: <name> (<step count> steps)
  - Milestone 2: <name> (<step count> steps)
**Files in scope:** <list>
**Test strategy:** <approach>
**Verification gates between milestones:** <compile/test/deploy checks>
```

**When `**Task type:** debug` — Depth-lock-Disziplin (Pflichtfeld, vom DONE-Guard erzwungen):**
Phase 2 braucht zusätzlich diese Zeile:
`**Depth-lock:** D1 home=<file::function>; D3 deepest=<file::function>; D4 repro=<RED am unpatched Stand VOR Implementierung>`
- **D3 bestimmt die Fix-Ebene:** der Fix muss genau auf der D3-Ebene liegen. Ein Fix auf einer höheren/niederen Ebene als D3 ist ein depth-miss → zurück zum Plan, D1/D3 überarbeiten.
- **D4 muss echt gelaufen sein**, nicht geplant: der RED-Output stammt vom unpatched Stand VOR der Implementierung. Ohne realen RED-Lauf ist der Depth-Lock nicht erfüllt.
- Kompakt halten — Single-Agent-Format, keine Multi-Agent-Overheads (der Autoprompt darf 5 Agenten spannen, yesloop nicht, #73504).

**Milestone criteria (agent decides):**
- ≤5 steps: single milestone
- 6-15 steps: MUST split into 2-3 milestones à 3-6 steps
- 16+ steps: MUST split into 4-6 milestones à 3-5 steps

Milestones are a recovery aid for context-collapse, not a bureaucracy. Small plan (≤5 steps) → one milestone suffices; this is a granularity choice, NOT a license to skip the phase structure or the milestone blocks.

**NO OVERENGINEERING.** If a step looks complicated, ask: "Can I solve this with a boolean flag and 10 lines?" Usually yes. Resist: per-agent configuration, persistence layers, abstract Strategy interfaces, retry mechanisms with backoff. Prefer: global constants, single struct + map, simple if/return.

### Ledger Plan-Schema (set_plan convention — feeds the Live-Ledger)
Shape the stored plan (set_plan) so the live-ledger subscriber (`ledger_nudge`) can render it as a compact, immediately readable system-state block. A plan following this schema gives the orchestrator and the next self a shared proof-ledger, not a wish-list.

**Schema keys (Goal/Core/Verified/Open/Next):**
- **Goal:** the one aim of this session, 1-2 lines
- **Core:** fixed decisions / constraints that are NO LONGER up for debate
- **Verified:** what is already proven — every item carries a proof line:
  - `<claim>` — `proof: <command> → <result>` (e.g. `proof: go test ./internal/... → exit 0, 18/18`)
- **Open:** open questions / alternatives still to be decided
- **Next:** the immediate next step

**Proof-line requirement:** only an item with a `proof:` line counts as Verified. A claim without a proof pointer belongs in **Open**, not **Verified**. The ledger is an evidence ledger — claims must cite the command and its result.

### Phase 3: EXECUTE
```
update_agent_status(phase="Phase 3/6 EXECUTE")
# or with milestones:
update_agent_status(phase="Phase 3/6 EXECUTE (milestone 1/3)")

### Phase 3: EXECUTE
**Status:** IN PROGRESS
**Plan items:** N total

#### Milestone 1: <name>
**Status:** COMPLETE
**Commits:** <hash> <msg>, <hash> <msg>
**Steps completed:** N/N
**todowrite synced:** yes

#### Milestone 2: <name>
**Status:** IN PROGRESS
**Current step:** N/M
**Last commit:** <hash> <msg>

**Total progress:** M/N milestones, X/Y steps
**Uncommitted work:** none | <list with reason>
```

**Per-milestone discipline:**
- `update_plan(completed=["milestone M"])` after each milestone finishes — persistence for collapse recovery
- `scratchpad_write` per milestone, not per step — controls bloat
- DRIFT CHECK before each milestone (not each step): re-read goal, compare scope

**Context-collapse recovery:** `get_plan()` → see current milestone → `update_agent_status` → read milestone block from scratchpad → continue.

**Commit cadence:** one commit per logical unit. Never leave uncommitted work across phase boundary without documenting why.

### Phase 4: VERIFY
```
update_agent_status(phase="Phase 4/6 VERIFY")

### Phase 4: VERIFY
**Status:** COMPLETE
**Tests run:** <command> → exit <code>, last 5 lines: <output>
**Regression baseline:** base=<origin/main hash> failures=<none|list>; head failures=<none|list>; diff=<none|new failures>
**RED proof:** <neuer Test RED vor Fix, GREEN nach Fix> | none — docs-only
**Lint/type-check:** <command> → <result>
**Build:** <command> → <binary mtime if applicable>
**Coverage gaps:** none | <list>
**VERIFY cycles used:** N/5
```

**Regression baseline (mandatory, deterministic):**
1. In Phase 2 (PLAN), note the merge-base: `git merge-base HEAD origin/main`
2. In Phase 4, create a temp checkout of origin/main INSIDE your worktree:
   `git worktree add .yesmem/tmp/base-checkout origin/main` (project-local, never /tmp)
3. Run the relevant test suite there (same commands as head), record pass/fail sets
4. Remove it after use: `git worktree remove .yesmem/tmp/base-checkout --force`
5. `diff=` lists ONLY test failures that are green on base but fail on head — each is a
   REGRESSION and must be fixed before REVIEW exit (or proven pre-existing on base)

If issues found → fix and re-verify (max 5 cycles, see CONVERGENCE GATE).

### Phase 5: REVIEW (Checklist: 5.1 Self + 5.2 Cold + 5.3 Consequence + 5.4 Security)

Phase 5 is a checklist of four separately checkable sub-phases, each reviewed with
its own task() subagent (5.1 by the agent itself). The daemon's DONE-guard verifies
every reported subagent session id against opencode's own session DB — the id is
the `ses_<id>` the task tool returns. Old lines `task() dispatched:` and
`consequence dispatched:` are retired (hard cut): blocks using them fail the gate.
`blocked` is NOT a valid value anywhere in Phase 5. If task() is truly unavailable,
`send_to` the orchestrator `BLOCKED Phase 5.x: <reason>` and stay in Phase 5 — never
claim DONE.

| # | Sub-phase | Execution | Mandatory line in Phase 5 block | Gate check |
|---|---|---|---|---|
| 5.1 | Self-Review | agent itself | `self-review: done` | format |
| 5.2 | Cold Code Review | task(reviewer or silent-bob) | `cold-review: ses_<id>` | row exists in opencode.db, parent_id == agent session, agent in {reviewer, silent-bob} |
| 5.3 | Consequence & Intent | task(general or reviewer) | `consequence: ses_<id>` | like 5.2, agent in {general, reviewer}, id distinct from 5.2 |
| 5.4 | Security Review | task(reviewer); the subagent loads the security-review skill | `security: ses_<id>` | like 5.2, agent in {reviewer}, id distinct from 5.2 and 5.3 |

```
update_agent_status(phase="Phase 5/6 REVIEW")

### Phase 5: REVIEW
**Status:** COMPLETE

**Sub-phase 5.1: Self-Review** (agent itself)
self-review: done
- Strengths: <list with file:line refs>
- Issues: Critical (N) / Important (N) / Minor (N), with file:line + reasoning
- Recommendations: <list>
- Assessment: Yes | With fixes | No

**Sub-phase 5.2: Cold Code Review**
cold-review: ses_<id>
**Stufe:** <1|2|3> — <one-line rationale per risk-based-code-review>
**Modules:** <active modules A–R> | N/A: <excluded, reason>
**Findings:** <list or "none">
**Merged assessment:** Yes | With fixes | No — <reasoning>
**Fix commits:** <hashes or "none needed">

**Sub-phase 5.3: Consequence & Intent**
consequence: ses_<id>
**Findings:** <intended-vs-actual list or "none">

**Sub-phase 5.4: Security Review**
security: ses_<id>
**Findings:** <list with NEW/MODIFIED distinction, or "none — diff reviewed, no findings">

**REVIEW→VERIFY cycles used:** N/5
```

**Stage 1 — Self-Review (Sub-phase 5.1)** (catches mechanical issues): Get full delta `git diff origin/main` + `git log origin/main..HEAD --oneline`. Checklist: Plan alignment, Code quality, Architecture, Testing, Production readiness, **security (dispatched in 5.4)**, Second-order effects ("if this ships, what happens next? trace 2+ levels"), Assumption surfacing ("what must be TRUE for this to work?").

**Sub-phase 5.4 doctrine — Security (MANDATORY):** The 5.4 subagent (never the loop agent itself) INVOKES the `security-review` skill via the Skill tool when the diff contains ANY executable code (`.go/.py/.js/.ts/.tsx/.jsx/.java/.rs/.php/.rb`). Apply the NEW/MODIFIED doctrine: for NEW code (diff-added), fix ALL findings HIGH/MEDIUM/LOW; for MODIFIED code (existing function touched), fix issues the diff introduces and document pre-existing issues as OUT OF SCOPE with a Learning reference. Even for docs/config-only diffs the 5.4 subagent is dispatched and verifies docs-only — no self-issued `security:` line. Every finding line carries either a fix-commit-hash, a "not exploitable because X" note, or an OUT-OF-SCOPE annotation.

**Sub-phase 5.2 — Cold Code Review via task()** (fresh eyes, catches architectural blind spots):
- Dispatch focused task()-subagent with code-reviewer template (superpowers requesting-code-review)
- Input: `git diff origin/main` + Phase 2 plan only (no exploration, for speed)
- **RISK CLASSIFICATION (MANDATORY):** The reviewer subagent MUST load the `risk-based-code-review` skill (platform skill tool, `~/.claude/skills/risk-based-code-review/SKILL.md`, or `<repo-root>/skills/bundled-skills/risk-based-code-review/SKILL.md`) BEFORE writing findings, then: (1) classify the change as Stufe 1-3 per its severity ladder — impact over diff size, few lines never lower the Stufe, (2) select the active modules per its trigger table (only fachlich unberührte modules are N/A), (3) report in the skill's Abnahmeformat. The `**Stufe:**` and `**Modules:**` lines in the 5.2 block are part of the required evidence. The Stufe steers review DEPTH, not the existence of any sub-review — 5.1-5.4 all run regardless.
- **MANDATORY — Cold Review is NOT optional.** Phase 5 is only complete when 5.2 has actually run. Empirically verified (2026-06-20, Learning #75412): agents skip review subagents silently if framed as additive. Required evidence: `cold-review: ses_<id>` — the daemon checks the row in opencode.db.
- **ALL subagent findings must be PROVEN ON CODE.** Every finding requires code-anchored evidence: file line refs, a call path (graph_traverse output), or a reproducing command/failing test. A finding without a code anchor is INVALID and must be reworked by the subagent.
- **If task() truly fails:** send_to orchestrator `BLOCKED Phase 5.2: <error>` and stay in Phase 5. Orchestrator spawns separate TUI reviewer as fallback.

**Sub-phase 5.3 — Consequence & Intent Check via task()** (fresh subagent, MANDATORY):
- Dispatch a focused task()-subagent (separate session from 5.2's reviewer) with:
  `git diff origin/main`, the Phase 2 **Decisions resolved** block, and this mandate:
  (a) trace second-order consequences: callers, exposed contracts, defaults, schema,
  dependent behavior — what breaks or changes that the diff does not show?
  (b) intent fidelity: does the landed behavior match what Phase 2 decided?
  Anything drifting into an unwanted direction?
- **Scope via module trigger table:** the 5.3 subagent loads `risk-based-code-review` and walks its trigger table for second-order suspects — state (E), transactions (F), tests with DB/time/random (G), schema/migrations (L), deploy/config (M) are the usual misses.
- **ALL 5.3 findings must be PROVEN ON CODE** — same evidence bar as 5.2:
  file line refs, call path, or a reproducing command/failing test. Unanchored
  findings are INVALID and must be reworked by the subagent.
- Evidence in the Phase 5 block: `consequence: ses_<id>`, `**Findings:**`
- **If task() truly fails:** send_to orchestrator `BLOCKED Phase 5.3: <error>` and stay in Phase 5 — never claim DONE.

**Sub-phase 5.4 — Security Review via task()** (fresh reviewer subagent, MANDATORY):
- Dispatch a focused task(reviewer)-subagent (distinct session from 5.2 and 5.3) that loads the `security-review` skill and reviews `git diff origin/main`.
- **Scope check:** the 5.4 subagent also loads `risk-based-code-review` and verifies its trigger table against the diff — security-relevant modules (Auth/Inputs H, Network I, Secrets J, Destructive K, Schema L, Deploy M) may not be silently dropped. The security-review skill remains the primary doctrine for finding analysis.
- Evidence in the Phase 5 block: `security: ses_<id>`, `**Findings:**` with NEW/MODIFIED distinction per the doctrine above.
- The agent may NOT skip by writing findings itself — the gate verifies a real reviewer session.

**Double-loop exit criteria (REVIEW→VERIFY):** the loop exits only when ALL of
these hold: Sub-phases 5.1-5.4 complete with daemon-verified evidence, and Phase 4
Regression baseline shows `diff=none` (or all new failures fixed + re-run). If a
check that must be clean stays red after 2 fix cycles → REVIEW BLOCKED
escalation.

**Merging & Resolution:**
- Merge findings from both stages (deduplicate, preserve highest severity)
- **ALL findings MUST be fixed autonomously** -- no user feedback, no escalation for fixable issues
- Critical, Important, medium, mittel → MUST fix immediately
- Minor also MUST fix immediately
- After fixes → loop back to Phase 4 VERIFY, then re-review (max 5 REVIEW→VERIFY cycles total)
- Assessment "No" (unfixable) → STOP, escalate via send_to: "REVIEW BLOCKED: <reasons>"
- **Never leave fixable issues for the user** -- review is a work phase, not advisory
- After `git show`/`git diff` always `Read` the current file — diffs show what changed, not what's there NOW

### Behauptungs-Inventar (proof-poverty check for the DONE-Report)
Every coverage / success claim in the DONE-Report must carry a proof pointer (command + result) on the same line; a claim without one is marked UNVERIFIED explicitly. This complements the Phase 4 `**Tests run:**` line — that records the overall run, here each individual claim is bound to its own evidence.

**Format per claim:**
- `<claim>` → `proof: <command> → <result>` (e.g. `proof: bun test plugins/opencode-yesmem → exit 0, 24/24`)
- not verifiable now → `<claim>` → `UNVERIFIED: <why not verified>`

Corollary from the ledgere schema: don't claim success you cannot back. Move unproven items to **Open** in the plan instead of asserting them in **Verified**.

### Phase 6: FINISH
```
update_agent_status(phase="Phase 6/6 FINISH")

### Phase 6: FINISH
**Status:** COMPLETE
**Deploy required:** yes (binary/skill/cap/code change) | no (docs/config only)
**Deploy executed:** yes — <evidence: make build output, md5sum diff>
  | no — <REASON. Default: BLOCKED, escalate to orchestrator>
**Branch pushed:** yes — <git log origin/<branch> -1 output> | no — <reason>
**PR created:** yes — <url> | no — left for orchestrator
**Merged to main:** NO — agents do not merge (merge is orchestrator-only)
**Worktree:** kept (pending merge confirmation)
**send_to orchestrator:** yes — <timestamp>
**send_to payload:** `[DONE] [DEPLOY: yes|no|skipped|failed] [COMMIT: <hash>] [BRANCH: <name>] [MERGE: no|pending-PR|blocked-PR] <summary>`
**set_plan complete:** yes
```

- **send_to marker convention (mandatory in Phase 6):** Orchestrator parses these markers — keep them leading in the `send_to` payload, summary free-text after.
  - `DEPLOY`: `yes` (deployed), `no` (docs-only, not required), `skipped` (required but skipped per instruction/contract), `failed` (attempted, failed — escalate)
  - `MERGE`: `no` (default — agents never merge), `pending-PR` (PR created, awaits orchestrator), `blocked-PR` (PR has blockers)

- **Merge policy: Agents NEVER merge to main themselves.**
  - Default: push the worktree branch to origin, create a PR (or leave for the orchestrator to create one), then `send_to caller_session: "[DONE] [DEPLOY: <status>] [COMMIT: <hash>] [BRANCH: <name>] [MERGE: pending-PR] <summary>"`. Markers are parsed by orchestrator — keep leading, free-text summary after.
  - `--merge` flag on the /yesloop command means the **orchestrator** requested auto-merge after PR checks pass. Even then the agent only pushes the branch and reports the PR URL — it does not run `git merge main` / `git checkout main && git merge <branch>` itself. The orchestrator handles the merge.
  - Running `git checkout main` + `git merge <branch>` inside the agent's Phase 6 is ALWAYS wrong, regardless of flag, confidence, or test results. The agent's git scope is its worktree branch. Main is the orchestrator's scope.
  - Rationale: the orchestrator owns release decisions, rollback, commit-message control, and PR review. A self-merge bypasses all four.
- **Do NOT delete worktree** — keep until the orchestrator confirms the merge is done.

## Mini Example (abbreviated — field names and headers are guard-compliant, copy them EXACTLY)

```
### Phase 1: ANALYZE
**Status:** COMPLETE
**Goal understood:** Add config.yaml read/write to set_config/get_config
**Session id:** ses_abc123 (from whoami())
**Codebase explored:** internal/daemon/handler_state.go, internal/config/config.go
**Constraints identified:** no breaking MCP API, proxy_state overrides must still work
**Risks:** concurrent writes, type coercion
**Open questions:** none

### Phase 2: PLAN
**Status:** COMPLETE
**Plan stored via set_plan:** yes
**Decisions resolved:** all
  - File selection: config.go, handler_state.go, cmd_config.go, main.go
  - Schema decisions: yaml config struct with Get/Set methods
  - API surface: Save/GetValue/SetValue on Config type
  - Blast radius: proxy_state override precedence, concurrent write
**Milestones:** 2 total
  - Milestone 1: config.go Save/GetValue + tests (4 steps)
  - Milestone 2: handler rewiring + CLI (3 steps)
**Files in scope:** internal/config/config.go, internal/daemon/handler_state.go, cmd_config.go, main.go
**Test strategy:** table-driven for coercion, integration for handler
**Verification gates:** go test ./internal/config/... between milestones

### Phase 3: EXECUTE
**Status:** COMPLETE
**Plan items:** 7 total

#### Milestone 1: config.go Save/GetValue + tests
**Status:** COMPLETE
**Commits:** a9ba499 feat(config): add Save/GetValue/SetValue + 11 tests
**Steps completed:** 4/4
**todowrite synced:** yes

#### Milestone 2: handler rewiring + CLI
**Status:** COMPLETE
**Commits:** a9ba499 feat(config): wire set/get_config to yaml, add CLI
**Steps completed:** 3/3

**Total progress:** 2/2 milestones, 7/7 steps
**Uncommitted work:** none

### Phase 4: VERIFY
**Status:** COMPLETE
**Tests run:** go test ./internal/config/... ./internal/daemon/... → exit 0
**Regression baseline:** base=9e8aa888 failures=none; head failures=none; diff=none
**Build:** go build ./... → success
**VERIFY cycles used:** 1/5

### Phase 5: REVIEW
**Status:** COMPLETE
**Stage 1: Self-Review**
- Strengths: clean API, good test coverage
- Issues: Important (2) / Minor (4), with file:line
- Assessment: With fixes
self-review: done

**Sub-phase 5.2: Cold Code Review**
cold-review: ses_235
**Stufe:** 3 — config write path + persistence (behavior and state risk)
**Modules:** A, B, C, D, E, F, H, Q, R | N/A: G, I, J, K, L, M, N, O, P — fachlich unberührt
**Findings:** same 2 Important + 4 Minor
**Merged assessment:** With fixes — ALL 6 findings (2 Important + 4 Minor) fixed before merge
**Fix commits:** cd2ba04 fix(config): proxy_state dual-write, type coercion + all 4 minors

**Sub-phase 5.3: Consequence & Intent**
consequence: ses_236
**Findings:** none

**Sub-phase 5.4: Security Review**
security: ses_237
**Findings:** none — diff reviewed, no HIGH/MEDIUM/LOW findings

**REVIEW→VERIFY cycles used:** 2/5

### Phase 6: FINISH
**Status:** COMPLETE
**Deploy required:** yes (Go code change)
**Deploy executed:** yes — make build → binary mtime 2026-06-20T00:40
**PR created:** no — merged directly per orchestrator (merge commit 27e35ca)
**send_to orchestrator:** yes — 2026-06-20T00:47
**send_to payload:** `[DONE] [DEPLOY: yes] [COMMIT: 27e35ca] [BRANCH: yesloop/example-task] [MERGE: no] Add config.yaml read/write`
**set_plan complete:** yes
```

## Guardrails (Prevent Agent Drift)

### DRIFT CHECK — Before each MILESTONE (not each step)

1. Re-read the original goal from `scratchpad_read(project, section)` or `get_plan()`
2. Compare current state against original scope:
   - **Still on track?** → proceed
   - **Minor drift?** → MANDATORY fix immediately — correct course NOW, document what drifted and how you corrected it in scratchpad, then continue
   - **Major divergence?** → **STOP.** scratchpad_write("⚠️ DRIFT: <what changed>") + send_to orchestrator: "DRIFT: <details>. Continue or abort?"

What counts as drift:
- Touching files outside the original scope (scope creep)
- Solving a different problem than the one given
- Adding features not requested ("while I'm here, I'll also...")
- Changing architecture without justification

### CONVERGENCE GATE — When progress stalls

| Pattern | Detection | Action |
|---|---|---|
| Edit-Test loop | Same test fails 3+ times with different fixes | **STOP.** scratchpad_write + send_to: "STUCK: <test> fails after 3 attempts." |
| Rewrite loop | Reverting your own changes and trying again | **STOP after 2 rewrites.** |
| Search loop | Searching for the same information repeatedly | Cache in scratchpad. After 3 empty searches, document as unknown. |
| Fix-then-break | Each fix breaks something else | **STOP after 2 cascades.** Code too coupled. |

**Hard limit:** No forward progress (no completed steps in todowrite) for 5 turns → escalate and stop.

### PROOF BURDEN (BEWEISLAST) for scratchpad claims

Before writing "DONE", "completed", "verified" to scratchpad, confirm the artifact matches the claim — not just that the command exited 0.

| Claim | Required independent check |
|---|---|
| "build succeeded" | `go build ./...` exit + binary exists with new mtime |
| "tests green" | show last 5 lines of test output, not just "PASS" |
| "deployed" | curl/GET against running service shows new version, OR md5sum live vs bundled exit 0 |
| "bundle rolled out" | `diff source installed && echo identical` (exit 0) |
| "commit pushed" | `git log origin/<branch> --oneline -1` shows the hash |
| "PR created" | `gh pr view <url>` exit 0 |
| "cold review ran" | task()-subagent ID present in scratchpad |
| "consequence finding" (Stage 3) | file line reference, call path, or repro evidence — unanchored findings are invalid |
| "file edited" | `Read` the file (not just `git show`) |
| "new test tests something" | **RED proof:** RED-Output am unpatched Stand (vor dem Fix), nicht nur geplant — bei docs-only explizit `none — docs-only` |
| "DONE-guard passed" | `ValidatePhaseBlocks(content).Compliant == true` (auto-checked every 30s) |

**Rule:** if verification was skipped, say so explicitly: `<claim> (not independently verified)`. If run, paste observable evidence (one line is enough).

**Apply only to consequential claims:** build, deploy, bundle rollout, merge, migrate, config change, destructive ops. Routine todowrite step-completion does not need this.

## State Recovery (every wake-up)

Your state lives in yesmem, not context. On every wake-up:
1. `get_plan()` — restore active goal and progress
2. `scratchpad_read(project, section)` — restore detailed context
3. **`check_messages`** — poll for orchestrator messages (OpenCode has no push, DB-poll is the only reliable path)
4. Reconstruct current phase + milestone
5. `update_agent_status(phase="Phase N/6 NAME (milestone M/K)")` — re-assert current state
6. Continue where you left off

## TUI Agent Mode (You were spawned in a worktree)

You are a **subagent** — spawned via yesmem_spawn_agent. You run in an isolated git worktree.

### ⛔ WORKTREE GUARDRAIL — Execute BEFORE any file modification

On startup, before ANY other action:
```
1. pwd                           → must match worktree path from scratchpad
2. git rev-parse --show-toplevel → must be the worktree, NOT main repo
3. git branch --show-current     → must be yesloop/<task-slug>, NOT main
4. git status --short            → must be clean
5. Write-Probe                   → mkdir -p .yesmem/tmp && printf probe > .yesmem/tmp/write-probe, dann read-back verifizieren (Inhalt == "probe")
```

**IF ANY CHECK FAILS:** STOP. Do not touch files. `scratchpad_write + send_to orchestrator: "⛔ WORKTREE GUARD FAILED: pwd=<actual>, branch=<actual>. Expected <expected>."` Wait for orchestrator.

**IF ALL PASS:** Identify via `whoami`, `scratchpad_read(project, section)`, `get_plan()`, `update_agent_status(phase="Phase 1/6 ANALYZE")`, then begin pipeline.

Write-Probe-Regel: die Probe-Datei wird NICHT gelöscht (kein rm) — sie bleibt im Worktree liegen und wird mit dem Worktree-Cleanup entfernt.

### Completion — MUST do all three:
1. `scratchpad_write(content="✅ DONE: <summary>. PR: <url>")` — final write with all 6 phase blocks COMPLETE
2. `send_to(target=<caller_session>, content="[DONE] [DEPLOY: <status>] [COMMIT: <hash>] [BRANCH: <name>] [MERGE: <status>] <summary>. PR: <url>")` — markers leading, parsed by orchestrator; summary free-text after
3. `set_plan(...)` mark as completed

**NOTE:** `send_to` stores in DB but push delivery is unreliable for OpenCode targets. Orchestrator polls `check_messages`. Scratchpad is primary completion channel.

**Periodic polling:** Every 5 turns, `check_messages` for new instructions or cancellation.

## Decision Gates (Autonomous)

| Situation | Action |
|---|---|
| Simple fix, clear solution | Execute directly |
| 2-3 approaches, unclear best | Pick one, document rationale, proceed |
| Need to change scope | MANDATORY fix immediately — stop the drift, correct course NOW, document it. Escalate only if the correction itself is impossible or blocking |
| Tests fail repeatedly | Debug max 3 cycles → CONVERGENCE GATE |
| Irreversible action (force-push, drop table) | Pause, request confirmation |
| 3 consecutive ticks with nothing to do | End the loop, report idle |
| Scope creep detected | DRIFT CHECK, correct or escalate |
| Same approach failed 3x | CONVERGENCE GATE — stop and escalate |

## Communication

- **inline mode:** short progress updates to user between phases
- **TUI agent mode:** scratchpad_write for detailed state, send_to for completion
- **scheduled mode:** scratchpad_write for state, broadcast on completion
- Never ask "should I continue?" — just continue unless blocked

## Automated Idle-Trigger

Layer 2 of the yesloop guarantee: **Idle Detection** for yesloop agents with a live PID but inactive SSE stream for >10 minutes. The daemon heartbeat runs an idle state machine (`checkYesloopIdle` in `internal/daemon/yesloop_idle.go`) that pushes agents through a 4-state escalation protocol.

### State Machine

| State | Condition | Action |
|---|---|---|
| 0 WORKING | stream_active=true or idle <10min | No action |
| 1 SELF_CHECK | stream_active=false for 10min, PID alive | Relay 1: ask agent to prove completion via scratchpad marker |
| 2 REMARK_REQUEST | PROVEN marker found | Relay 2: ask agent to mark all 6 phases as done |
| 3 COMMIT_REQUEST | 6 phases marked done | Relay 3: ask agent to commit and send_to orchestrator |
| 4 DONE | send_to evidence in scratchpad | Terminal — notify orchestrator |

### Relay Messages

The daemon sends these exact messages (metachar-free, no markdown, no backticks):

1. **State 0 → 1:** `Have you completed all 6 phases? If not do it now. For each phase prove you have done each, IF you have proven mark each phase in scratchpad with [x] showing it is done. MANDATORY: Make sure that you have also done phase 5 with all code reviews including Sub-phase 5.2 cold review, 5.3 consequence check and 5.4 security review via task subagents. REVIEW BLOCKED without subagent trace is not acceptable. Mandatory: only mark as PROVEN if it IS proven.`
2. **State 1 → 2:** `Mark all 6 phases as done with x in scratchpad.`
3. **State 2 → 3:** `If 1 through 6 are ok commit and send_to to caller.`

### Scratchpad Markers

The idle checker reads the agent's scratchpad section to detect progress:

- **PROVEN** (case-insensitive): signals that the agent has verified all 6 phases are complete → triggers State 1 → 2
- **6x `**Status:** COMPLETE`**: when all 6 phase blocks are completed (detected by `CountCompletedPhases()`) → triggers State 2 → 3
- **send_to orchestrator:** evidence in scratchpad → triggers State 3 → 4

### Re-fire and Escalation

- Each state re-fires the relay after 90 seconds if no progress
- After 2 re-fires (3 total attempts), the agent is **paused** as DEAD_AGENT
- The orchestrator (caller_session) is notified via send_to

### What triggers idle detection

- Section prefix: `yesloop-`
- Agent status: `running`
- PID: alive (`isPIDAlive()` returns true)
- Stream state: `stream_active` false for >10 consecutive minutes

## Automated Done-Verify (v3.2)

Layer 3 of the yesloop guarantee: **Done-Verify** for yesloop agents that emit a DONE-claim (Phase 6 header, `send_to ... DONE`, or `Phase 6/6` marker). The heartbeat runs a 4-state verify machine (`checkYesloopDoneVerify` in `internal/daemon/yesloop_done_verify.go`, called every 5 minutes) that asks the agent to **BEWEISEN** all 6 phases ran — especially Phase 5 Cold Review — before marking DONE.

### State Machine

| State | Condition | Action |
|---|---|---|
| 0 NOT_DONE | No DONE-claim in scratchpad | Not tracked, cron inactive |
| 1 VERIFY_REQUESTED | DONE-claim detected | Relay: "Hast du wirklich alle Phasen abgehakt..." |
| 2 DONE_VERIFIED | BEWEISEN marker + send_to orchestrator + 6 phases COMPLETE | Terminal — cron stops for this agent |
| 3 DEAD_AGENT_ESCALATION | 3 re-fires without progress | Freeze + notify orchestrator |

### DONE-Claim Indicators (any one triggers)

- `### Phase 6:` header (case-insensitive)
- `send_to ... DONE` pattern
- `Phase 6/6` marker
- `DONE: ` prefix in content

### Relay Message (metachar-free)

`Hast du wirklich alle Phasen abgehakt. Vor allem Phase 5 Code Review. Falls nicht bitte durchfuehren und BEWEISEN dass alle 5 Phasen durch sind und DANN Phase 6 Finish durchfuehren mit commit und send_to. KEIN auto-deploy.`

### Scratchpad Markers

The done-verify checker reads the agent's scratchpad section to detect progress:

- **DONE-claim** (any indicator above): triggers NOT_DONE → VERIFY_REQUESTED
- **BEWEISEN** (case-insensitive) + `send_to orchestrator:` line + 6 phases COMPLETE: triggers VERIFY_REQUESTED → DONE_VERIFIED
- Missing evidence after 3 re-fires: triggers VERIFY_REQUESTED → DEAD_AGENT_ESCALATION

### Re-fire and Escalation

- Re-fires the relay every 5 minutes if no progress
- After 3 re-fires (total 4 relay attempts), the agent is **paused** as DEAD_AGENT
- The orchestrator (caller_session) is notified via send_to

### What triggers done-verify

- Section prefix: `yesloop-`
- Agent status: `running`
- Scratchpad content matches any DONE-claim indicator
- Agent not already in DONE_VERIFIED or DEAD_AGENT_ESCALATION state

### Boundary with other layers

- **Idle Detection (Layer 2)** triggers on stream inactivity; Done-Verify (Layer 3) triggers on DONE-claims. They are complementary and can both fire for the same agent.
- **DONE-Guard (Layer 3 regex validator)** freezes agents that claim DONE with malformed phase blocks. Done-Verify catches the case where phase blocks look valid but Phase 5 Cold Review was silently skipped — the agent must actively BEWEISEN it ran Sub-phase 5.2 (verified subagent session id).

## Automated Skill-Reload Check (Layer 4)

Layer 4 of the yesloop guarantee: **Context-Compaction-Re-Read** gegen Skill-Verlust. Lange yesloop-Läufe können den Skill-Text verlieren, wenn der Kontext kompaktiert wird — der Agent würde dann ohne Pipeline-Vorgaben weitermachen. Der Daemon (`checkYesloopSkillCheck` in `internal/daemon/yesloop_skillcheck.go`, 30s-Heartbeat) beobachtet das kumulierte Arbeitsvolumen je Agent (OutputTokens-Delta seit Reset-Basis ≥ 100.000, Fallback TurnsUsed-Delta ≥ 50) und erinnert per Relay daran, den Skill neu zu laden.

### State Machine

| State | Condition | Action |
|---|---|---|
| 0 TRACKING | Arbeitsvolumen unter Schwelle seit Baseline | No action |
| 1 REMIND | OutputTokens-Delta ≥ 100k ODER Turns-Delta ≥ 50 seit Baseline | Relay: Skill-Ladezustand prüfen, bei Bedarf neu laden, per Marker bestätigen |
| 2 CONFIRMED | **SKILL-RELOAD-<N>:** yes mit aktueller Runden-Nummer im Scratchpad | Baseline auf aktuellen Stand zurücksetzen → zurück zu TRACKING |
| 3 DEAD_AGENT_ESCALATION | 3 Re-Fires ohne Marker | Freeze + notifyOrchestrator |

### Context-Compaction-Re-Read (Agentenseite — MANDATORY)

- Erhält der Agent einen Relay, der auf Context-Compaction / Skill-Verlust hinweist (Absender `yesloop-skillcheck`): DEN SKILL SOFORT NEU LADEN (`skill`-Tool oder `~/.claude/skills/yesloop/SKILL.md` / Bundle-Pfad lesen).
- Danach als Beweis per `scratchpad_append` die EXAKTE Bestätigungszeile aus dem Relay schreiben, inklusive der aktuellen Runden-Nummer: `SKILL-RELOAD-<N>: yes` (z.B. `**SKILL-RELOAD-3:** yes`).
- Der Daemon erkennt den Marker nur, wenn die Runden-Nummer zum aktuellen Fenster passt (case-insensitive, tolerant gegenüber `**`) und setzt dann seine Tracking-Basis zurück — der nächste Reminder feuert erst nach der nächsten akkumulierten Schwelle.
- Die Runden-Nummer ändert sich pro Fenster: ein alter Marker aus einem früheren Fenster bestätigt ein neues Fenster NICHT. Nur eine echte, aktuelle Bestätigung verdient den Reload.
- Ohne echten Reload KEINEN Marker setzen — der Zustand wird nur durch eine tatsächlich geladene Skill-Instanz verdient.

### Re-fire and Escalation

- Re-fires alle 5 Minuten ohne Marker-Progress für das aktuelle Runden-Fenster.
- Nach dem 3. Re-Fire-Intervall ohne Fortschritt (initialer Relay + max 2 weitere gesendete Re-Fires) wird der Agent als DEAD_AGENT gepaust und der Orchester via send_to benachrichtigt.

### Boundary with other layers

- **Compaction-Check (Layer 4)** feuert auf akkumuliertes Arbeitsvolumen eines LAUFENDEN Agents — orthogonal zu Idle (Stream-Inaktivität) und Done-Verify (DONE-Claims). Alle drei können für denselben Agent feuern.

## Automated Stagnation Monitor (Layer 5)

Layer 5 of the yesloop guarantee: **Stagnation Detection** for agents with an **active** stream and a live PID that produce no observable progress (no scratchpad writes, no `update_agent_status` changes) for >2 hours. Complements the Idle guard (which only fires on stream *inactivity*) — catches day-scale runs where the stream stays active but nothing advances. AVO-inspired conditional intervention (arXiv 2603.24517 §3.3): the trigger is fully deterministic, not a permanent supervisor.

**Trigger condition:** running `yesloop-` agent, PID alive, stream `active`, `CountCompletedPhases < 6`, and the progress signal — `stagnationSig(scratchpad content, agent.Progress)` — unchanged for >1h. Any change to the signal resets tracking. Reset happens on PROGRESS only, never on stream blips (#82787 lesson).

### State Machine

| State | Condition | Action |
|---|---|---|
| TRACKING | signal unchanged <2h | no action |
| REFIRING | signal unchanged >1h | relay STAGNATION + up to 3 refires @30min |
| ESCALATION | refires > 2 | `pauseAgent` + `notifyOrchestrator` (DEAD_AGENT) |

Upon receiving a STAGNATION relay: write a status line with current evidence to the scratchpad immediately (scratchpad_append). If genuinely stuck, either change approach or document BLOCKED with the reason. Do not ignore the relay — escalation pauses you after 4 relays.

## Anti-Patterns

- Do NOT write design documents unless the task is architectural
- Do NOT ask clarifying questions — infer from context, document assumptions
- Do NOT scope-creep beyond the given task — see DRIFT CHECK
- Do NOT run endlessly — 3 idle ticks = stop
- Do NOT modify agent config files (.claude/, SYSTEM.md, yesloop.md, etc.)
- Do NOT keep trying the same approach — if 3 attempts fail, approach is wrong
- **⛔ Do NOT work in main.** Always verify worktree before touching files.
- Do NOT report DONE with missing phase sections — report PARTIAL instead
- Do NOT report COMPLETE for a phase without `update_agent_status` having been called for that phase