[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78GSYX4EXNMJV1XM4PS38CT":3,"skill-same:01M78GSYX4EXNMJV1XM4PS38CT":58},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M78GSYX4EXNMJV1XM4PS38CT","start-work",null,"Execute a Prometheus work plan in Codex with Boulder state, evidence ledger updates, worktree discipline, parallel subagents, and Stop-hook continuation. Use after planning when the user says start work, execute plan, continue plan, resume plan, or asks to run a .omo\u002Fplans plan.","younghai\u002Flazyz","younghai","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F10753869?v=4","01M78GJ5KADQ959GDR0ZM7DDNF","young kwon","01M78GJ5M3F6D2YKVXCNX3WMR9",0,"MIT","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Frelease\u002Flazyz-0.11.0-20261007\u002Fskills\u002Fstart-work\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: start-work\ndescription: \"Execute a Prometheus work plan in Codex with Boulder state, evidence ledger updates, worktree discipline, parallel subagents, and Stop-hook continuation. Use after planning when the user says start work, execute plan, continue plan, resume plan, or asks to run a .omo\u002Fplans plan.\"\n---\n\n## ZCode Harness Tool Compatibility\n\nThis skill was ported from the Codex\u002FOpenCode OmO harness. ZCode does not provide `call_omo_agent`, `task`, `background_output`, `team_*`, `multi_agent_v1.spawn_agent`, or any Codex-only multi-agent tools. Translate every such example to ZCode's native tools:\n\n| Codex\u002FOpenCode example | ZCode tool to use |\n| --- | --- |\n| `call_omo_agent(subagent_type=\"explore\", ...)` | `Agent` tool with `subagent_type=\"Explore\"` and a self-contained `prompt` |\n| `call_omo_agent(subagent_type=\"librarian\", ...)` | `Agent` tool with `subagent_type=\"Explore\"` (research-focused prompt) |\n| `task(subagent_type=\"plan\", ...)` | `Agent` tool with `subagent_type=\"general-purpose\"` (planning-focused prompt) |\n| `task(subagent_type=\"oracle\", ...)` for final verification | `Agent` tool with `subagent_type=\"general-purpose\"` (rigorous-review prompt) |\n| `task(category=\"...\", ...)` for implementation or QA | `Agent` tool with `subagent_type=\"general-purpose\"` |\n| `multi_agent_v1.spawn_agent(...)` \u002F `spawn_agent(...)` | `Agent` tool with `subagent_type` and a self-contained `prompt` |\n| `background_output(task_id=\"...\")` | Only unstructured read-only research may use background mode; its completion arrives as a parent task notification, not a plugin hook |\n| `multi_agent_v1.wait_agent(...)` mailbox signals | No equivalent plugin tool; use returned foreground Agent results or the host's background research notifications |\n| `team_*(...)` | ZCode has no native team-of-threads API. Approximate with multiple parallel `Agent` calls (one per member) coordinated by the parent, or fall back to sequential subagents when isolation is unclear |\n\nZCode v3.14.3 loads enabled plugin agents from `agents\u002F`. Prefer namespaced roles such as `lazyz:explorer` and `lazyz:lazycodex-executor`; unambiguous bare names also work. Include all task context and ownership in the self-contained `prompt`. Roles inherit the user's active model; do not force an obsolete provider ID or copy Codex model\u002Feffort fields. If an older host lacks a role, describe it inline using the appropriate built-in agent.\n\nBefore every governed Agent or Task call, read the Dispatch schema in the installed [start-work skill](..\u002Fstart-work\u002FSKILL.md) and prepend exactly one LAZYCODEX_DISPATCH_V1 JSON line. General-purpose and known executor calls require this envelope; read-only Explore calls may remain unstructured. Use the same taskId across retries with a fresh runId, unchanged criteria, role, phase and budget. Planning\u002Fdocument work uses writer, code work implementer, review reviewer, and testing qa. Never rename a task to bypass its exhausted budget.\n\nGoverned calls must use `run_in_background: false` (or omit it). A background launch is not a completed result and has no matching plugin completion hook. Multiple independent foreground calls may be requested together; do not assume the host executes them concurrently. Read-only background research without a dispatch contract remains available.\n\nRequire each governed child to write the matching ReceiptV1 with current source identity, criterion invocations, zero exit codes and hashed regular artifacts, then return EVIDENCE_RECORDED; final-phase work requires FINAL_EVIDENCE_RECORDED. PostToolUse records failed evidence; Stop prevents unresolved completion while budget remains. An exhausted budget requires a blocker handoff, not an unsupported completion claim.\n\nUser authorization and verification policy take precedence over generic workflow instructions below. Do not auto-commit, push, merge, delete, deploy, install tools, call paid services, or run checks prohibited by the user. Record skipped checks honestly; do not fabricate a passing receipt. A task requiring a verified receipt remains unverified if the necessary checks are prohibited. ZCode Stop can continue at most three consecutive times, not indefinitely. Native Goal Mode, Memory and Wiki are host features; no callable API is assumed.\n\nFor work likely to take long, prefer multiple concurrent `Agent` calls in a single message over one giant blocking call. Treat each child's final message as the deliverable: if it lacks the required output, ack-only, or stops without producing evidence, fall back to a smaller re-scoped `Agent` call with the missing deliverable made explicit in the `prompt`.\n\n## Codex Subagent Reliability\n\nEvery `spawn_agent` message must be self-contained. Start with\n`TASK: \u003Cimperative assignment>`, then name `DELIVERABLE`, `SCOPE`, and\n`VERIFY`. State that it is an executable assignment, not a context\nhandoff. Role selection requires `agent_type`; `model` +\n`reasoning_effort` alone creates a default agent, not a reviewer or\nworker. Prefer `fork_turns: \"none\"` unless full history is truly\nrequired; paste only the context the child needs.\n\nPlan and reviewer agents may run for a long time; spawn them in the background, keep doing independent root work, and poll with short wait_agent cycles sized to the work. Never use a single long blocking wait for them, and never spin on tiny timeouts as a failure budget.\n\nTreat child status as a progress signal, not a timeout counter. For\nwork likely to exceed one wait cycle, require the child to send\n`WORKING: \u003Ctask> - \u003Ccurrent phase>` before long reading, testing, or\nreview passes, and `BLOCKED: \u003Creason>` only when it cannot progress.\nWhile any child is active, keep the parent visibly alive with active\nsubagent count, agent names, latest `WORKING:` phase, and whether the\nparent is waiting for mailbox updates. Track spawned agent names\nlocally. Use `wait_agent` for mailbox signals, not proof of completion.\nA timeout only means no new mailbox update arrived; after a timeout,\nrun a single `list_agents` check for the named child when you need\nreassurance. If it is running or its latest message is `WORKING:`,\ntreat it as alive. Do not use `list_agents` as a polling loop or status\nfeed; it can replay large payloads. Fallback only when the child is\ncompleted without the deliverable, ack-only after followup, explicitly\n`BLOCKED:`, or no longer running. Then record the result as\ninconclusive, do not count it as pass\u002Freview approval, close if safe,\nand respawn a smaller `fork_turns: \"none\"` task with the missing\ndeliverable.\n\n# start-work\n\nExecute a Prometheus work plan until every top-level checkbox is complete. This skill pairs with the Codex `Stop` \u002F `SubagentStop` continuation hook in `components\u002Fstart-work-continuation`, which re-injects the next turn while `.omo\u002Fboulder.json` says the current `codex:\u003Csession_id>` still has unchecked plan work.\n\n## Usage\n\n```text\n$start-work [plan-name] [--worktree \u003Cabsolute-path>]\n```\n\n- `plan-name` is optional. It may be a full or partial file stem under `.omo\u002Fplans\u002F`.\n- `--worktree` is optional. Use it only when the user explicitly asks to work in a separate git worktree.\n\n## Manual-QA requirement (read before starting)\n\nEvery user-visible checkbox in the plan requires **real Manual-QA evidence captured against a running surface** — a dry-run claim is never sufficient. Before starting, ensure the relevant channel is available:\n\n| Channel | When to use | Artifact |\n| --- | --- | --- |\n| **HTTP** (`curl -i`) | HTTP API checkbox | status line + headers + body |\n| **tmux** (`send-keys` + `capture-pane`) | terminal\u002FCLI checkbox | tmux transcript |\n| **Browser** (Chrome or agent-browser) | web UI checkbox | action log + screenshot path |\n| **Computer use** (OS automation) | desktop GUI checkbox | action log + screenshot |\n\nCLI stdout, DB diffs, and parsed config dumps are accepted as **auxiliary evidence** for CLI-shaped or data-shaped behavior, but do not replace the channel table for user-visible surfaces. `--dry-run` output is never accepted as evidence.\n\nIf the surface cannot be started (server won't boot, browser unavailable), flag it up front rather than discovering it at the first checkbox.\n\n## Phase 1: Select the plan\n\n1. Read `.omo\u002Fboulder.json` if it exists.\n2. List Prometheus plan files under `.omo\u002Fplans\u002F`.\n3. If `plan-name` was provided, select the matching plan.\n4. If exactly one active or paused Boulder work exists for this session, resume it.\n5. If no active work exists and exactly one plan exists, select it.\n6. If no active work exists and there is no selectable plan, enter **No-plan bootstrap**.\n7. If multiple plans remain possible, ask one focused selection question.\n\n### No-plan bootstrap\n\nWhen the user explicitly said `start work` \u002F `$start-work` and no selectable plan exists, treat that phrase as approval to create the plan before execution. Do not stall on a missing plan and do not ask for generic approval again.\n\nIf no selectable plan exists, bootstrap `ulw-plan` before execution.\nExecution requires an approved plan before implementation; bootstrap mode creates that approved plan from the user's `start work` request instead of skipping planning.\n\n1. Invoke the `ulw-plan` skill from the current request and require its dynamic adversarial workflow: collect, verify, design, adversarial plan-review, synthesize.\n2. The generated Prometheus plan must be saved under `.omo\u002Fplans\u002F\u003Cslug>.md` before implementation or Boulder state writes that point at plan work.\n3. Use maximum safe parallelism in the generated plan: independent files\u002Ftasks fan out; same-file writes, shared state, and named dependencies serialize.\n4. Preserve safety boundaries. Ask one focused question only when the objective is missing, destructive, or has a safety\u002Fproduct ambiguity that repository exploration cannot resolve.\n5. After the plan exists, continue directly to Phase 2. The user's `start work` request is the bootstrap approval to create the plan and begin execution.\n\n## Phase 2: Create or update Boulder state\n\nWrite `.omo\u002Fboulder.json` before implementation starts. Session ids must be prefixed with `codex:` so the continuation hook can identify its own session.\n\n```json\n{\n  \"schema_version\": 2,\n  \"active_work_id\": \"\u003Cwork-id>\",\n  \"works\": {\n    \"\u003Cwork-id>\": {\n      \"work_id\": \"\u003Cwork-id>\",\n      \"active_plan\": \".omo\u002Fplans\u002F\u003Cplan-name>.md\",\n      \"plan_name\": \"\u003Cplan-name>\",\n      \"session_ids\": [\"codex:\u003Csession_id>\"],\n      \"status\": \"active\",\n      \"fail_count\": 0,\n      \"worktree_path\": null\n    }\n  }\n}\n```\n\n> **Note:** `schema_version` is a documentation\u002Fhuman-readable field only. No parser reads it and bumping it does not trigger a migration. It exists so humans can tell which shape a file is expected to have. See `docs\u002Fknown-limitations.md` → \"Soft state schema\".\n\n`status` transitions: `\"active\"` (default) → `\"paused\"` (user pauses) → `\"completed\"` (all checkboxes done) or `\"abandoned\"` (user cancels). A `\"blocked\"` status is set by the cycle\u002Ffailure caps or a `needs-human-review` verdict; the continuation hook will not resume a blocked work until the user sets it back to `\"active\"`. Increment `fail_count` each time a cap fires or a `needs-human-review` verdict is recorded.\n\nIf `--worktree` is set, verify the path with `git worktree list --porcelain` or create it with `git worktree add \u003Cpath> \u003Cbranch-or-HEAD>`, then store the absolute path as `worktree_path`. All edits, commands, tests, and evidence capture must run inside that worktree.\n\n## Phase 3: Execute the next checkbox\n\n1. Read the full selected plan.\n2. Find the first unchecked column-0 checkbox in `## TODOs` or `## Final Verification Wave`.\n3. Ignore nested checkboxes under acceptance criteria, evidence, and definition-of-done sections.\n4. Decompose that checkbox into atomic sub-tasks.\n5. Dispatch independent sub-tasks in parallel with `spawn_agent`; serialize only when one sub-task has a named dependency on another.\n\nEach sub-task message must include:\n\n1. The dispatch schema below, including the goal and exact files or directories in scope.\n2. When the task touches existing behavior: a baseline characterization test, written first, that asserts current observable behavior and passes on the unchanged code. Then the red test or failing reproduction for the new behavior before production changes. Pin the baseline as rigorously as the new test: exact inputs, exact observable, exact assertion.\n3. Implementation constraints from the plan and project rules.\n4. Automated verification commands to run.\n5. One Manual-QA channel, named with the exact tool and exact invocation (the literal `curl`, `send-keys`, `page.click`, payload, selectors, and the binary observable that decides PASS\u002FFAIL), not \"verify it works\":\n   - HTTP call: `curl -i` against the live endpoint.\n   - tmux: a `tmux` session driven with `send-keys`, dumped via `capture-pane`.\n   - Browser use: use Chrome to drive the real page; if Chrome is not available, download and use agent-browser (https:\u002F\u002Fgithub.com\u002Fvercel-labs\u002Fagent-browser).\n   - Computer use: OS-level GUI automation against the running desktop app when the surface is not a page.\n6. The adversarial classes that apply to this sub-task (from the 9 ultraqa classes) and how each is probed.\n7. Required artifact path and cleanup receipt.\n\n### Dispatch schema\n\nEvery governed `Agent` or `Task` prompt is a complete executable assignment, not a context handoff. Put exactly one compact JSON marker line near the top of the prompt and fill every field:\n\n```text\nLAZYCODEX_DISPATCH_V1: {\"schemaVersion\":1,\"workId\":\"\u003Cboulder work_id>\",\"taskId\":\"\u003Cstable task id>\",\"runId\":\"\u003Cnew id for this attempt>\",\"role\":\"implementer|reviewer|qa|writer\",\"phase\":\"execute|verify|final\",\"criterionRefs\":[\"G001-first\u002FC001\"],\"budget\":{\"maxCalls\":5,\"maxSameFailure\":3}}\n```\n\n`taskId` is stable across retries. `runId` is unique for each dispatch attempt. The first accepted marker freezes the work, task, role, phase, criteria, and budget for that task. Caller budgets may lower the hard caps of 5 total calls and 3 identical failures, but may never raise them. Use a distinct task id when a task legitimately changes role or phase. `criterionRefs` uses canonical goal\u002Fcriterion ids such as `G001-first\u002FC001`; bare `G001` or `C001` remains valid for standalone tasks.\n\nThe assignment JSON validated below contains that envelope plus the scheduling fields `execution_shape`, `parallel_safe`, `owned_paths`, and `completion_condition` and the operational fields listed next.\n\nAfter the envelope, include these labeled sections in this order:\n\n- `DELIVERABLE:` exact files, result shape, and behavior the child must return.\n- `VERIFY:` exact commands or scenario and the expected observable.\n- `STOP WHEN:` the same binary completion boundary expressed operationally.\n- `ARTIFACT:` exact evidence path or returned structured payload.\n- `PARENT VALIDATION:` what the orchestrator will independently inspect or rerun before accepting the return.\n\n`parallel_safe: true` is valid only when `owned_paths` do not overlap any active lane and the task has no unmet dependency on another lane. An empty or broad ownership entry is not parallel-safe. The orchestrator validates the envelope before dispatch and serializes invalid or overlapping lanes.\n\nA child response that only acknowledges the task, reports progress without the `DELIVERABLE`, omits the required artifact, or stops before every `completion_condition` entry is satisfied is an execution failure. Do not count it as evidence or completion. Send one focused follow-up naming the missing field; if the second response is still ack-only or incomplete, record the lane as failed and respawn a smaller assignment with the same `taskId` and a fresh `runId`.\n\nValidate the JSON form of every envelope before dispatch:\n\n```bash\nuv run \"\u003Cstart-work-skill-root>\u002Fscripts\u002Fvalidate_dispatch.py\" \u003Cassignment.json>\n```\n\nRequire the child final to be JSON with non-empty `deliverable`, `verify`, and `artifact` fields plus `completion_condition_met: true`, then validate it before parent inspection:\n\n```bash\nuv run \"\u003Cstart-work-skill-root>\u002Fscripts\u002Fvalidate_dispatch.py\" \u003Cassignment.json> --final \u003Cworker-final.json>\n```\n\nThe validator is the dispatch gate. Its first output line is the exact `LAZYCODEX_DISPATCH_V1` marker to paste into the prompt. It uses JSON plus strict Pydantic parsing, never `eval`, rejects missing schema fields, invalid criterion ids, budgets above the hard caps, empty or blank completion-condition lists, and broad parallel ownership, and rejects plain-string acknowledgement finals. A validator failure blocks dispatch or acceptance respectively.\n\nEvery structured worker writes a ReceiptV1 JSON file under `.omo\u002Fevidence\u002F` and ends its assistant response with `EVIDENCE_RECORDED: \u003Cpath>`. A `phase:\"final\"` worker must use `FINAL_EVIDENCE_RECORDED: \u003Cpath>`; an ordinary marker cannot downgrade final evidence. The receipt must bind the exact `taskId`, `runId`, `role`, `phase`, current source commit\u002Ftree, criterion refs, invocations, zero exit codes, and confined regular evidence artifacts with SHA-256 hashes. The runtime enforces maximum receipt and artifact read sizes. Malformed marker-bearing prompts and general-purpose executor prompts without the dispatch marker fail closed. Read-only Explore dispatches may remain unstructured.\n\n`PreToolUse` reserves the unique run before dispatch and requests denial when the contract or budget is invalid. `PostToolUse` validates the returned receipt, records a stable typed failure signature, and uses the documented `additionalContext` channel to report failure after the action; it does not claim to undo or block an already completed tool call. `PostToolUseFailure` records the stable `tool-execution-failed` signature without storing raw host errors and reports it through `additionalContext`. `Stop` is the completion gate: it blocks unresolved tasks while retry budget remains, but allows the orchestrator to report a task whose budget is exhausted. This is a local hook contract; do not describe PreToolUse denial as a paid-provider API guarantee.\n\nApply ultraqa's 9 adversarial classes where relevant to each checkbox: malformed input, prompt injection, cancel\u002Fresume, stale state, dirty worktree, hung or long commands, flaky tests, misleading success output, repeated interruptions. A checkbox whose behavior is user-visible MUST probe every class that plausibly applies; record which classes were exercised and which were ruled not-applicable with a one-line reason.\n\n### Cycle and failure caps\n\nThese caps bound runaway work on a single checkbox. They mirror the ulw-loop execution loop and are mandatory.\n\n- **Cap at 5 execution cycles per checkbox.** An execution cycle is one dispatch → verify → verdict round. After 5 cycles on one checkbox without a `confirmed` verdict, set the boulder work `status` to `\"blocked\"`, increment `fail_count`, append a `checkbox-blocked` ledger entry naming the checkbox and the diagnosis, and stop dispatching that checkbox. Surface the block to the user and wait for explicit instruction.\n- **Cap identical failure type at 3 per checkbox.** Track the verdict-failure signature (`needs-fix` on the same assertion, `false-positive` on the same artifact, etc.). After 3 failures of the same type on the same checkbox, set the boulder work `status` to `\"blocked\"`, increment `fail_count`, and escalate.\n- When a boulder work is `\"blocked\"`, the continuation hook will not re-inject the directive (only `active` and `paused` are continuable). The work stays paused until the user changes `status` back to `\"active\"` or resolves the blocker.\n\n### Debugging budget\n\nWhen the same checkbox fails 2 rounds in a row (regardless of failure type), load the `debugging` skill and run its hypothesis-driven loop before the next dispatch: form ≥3 hypotheses, pick the cheapest to validate, and only re-dispatch with the validated root cause in the failure context. This budget prevents re-dispatching the same failing executor context repeatedly. After 2 failed rounds the executor is assumed to be debugging-blind without this skill.\n\n## Phase 4: Verify and record evidence\n\nFor each checkbox, complete all five gates before marking it done:\n\n1. Plan reread: confirm the checkbox and acceptance criteria.\n2. Automated verification: run tests, typecheck, lint, build, or the plan-specific equivalent.\n3. Manual-QA channel: capture a real artifact, not a dry-run claim.\n4. Adversarial QA: exercise every applicable ultraqa class (malformed input, prompt injection, cancel\u002Fresume, stale state, dirty worktree, hung or long commands, flaky tests, misleading success output, repeated interruptions) and capture the observable result for each. \"Tests pass\" and a clean happy-path artifact are NOT sufficient when an adversarial class applies and was not probed.\n5. Cleanup: register every QA resource teardown as its own todo the moment it is spawned (QA scripts, tmux assets, browser \u002F agent-browser sessions, PIDs, ports, containers, temp dirs), then execute each and capture the receipt. No QA asset is left running.\n\nAppend evidence to `.omo\u002Fstart-work\u002Fledger.jsonl` using one JSON object per line. Include at least `event`, `plan`, `task`, `session_id`, `commands`, `artifact`, `adversarial_classes`, and `cleanup` fields. `adversarial_classes` lists each probed class with its observable result and each ruled-out class with a one-line reason.\n\n**Before appending**, redact secrets from the `commands` field and evidence artifacts. The `commands` field captures shell invocations verbatim — `curl -H \"Authorization: Bearer sk-...\"`, `psql PGPASSWORD=...`, `aws --secret-key ...` — and these are as sensitive as shell history. Strip or mask credentials before writing. A post-hoc scrubber is available: `node scripts\u002Fredact-secrets.mjs` (reports) or `node scripts\u002Fredact-secrets.mjs --fix` (masks in place). Run it before sharing or committing `.omo\u002F`.\n\n### Sisyphus-style completion contract\n\nA worker done claim is never final. Each implementation sub-task returns a `DoneClaim`, then a different context runs `AdversarialVerify`, then the verifier probes or reproduces the claim, then failures loop back to the executor, and only a confirmed verifier verdict becomes `FullyDone`.\n\n```json\n{\n  \"DoneClaim\": {\n    \"task\": \"\u003Ctask id\u002Ftitle>\",\n    \"changed_files\": [\"path\"],\n    \"tests\": [\"exact command + result\"],\n    \"manual_qa\": [\"artifact path\"],\n    \"cleanup\": [\"receipt\"],\n    \"risks\": [\"known risk or none\"]\n  },\n  \"AdversarialVerify\": {\n    \"verdict\": \"confirmed | false-positive | needs-fix | needs-human-review\",\n    \"evidence\": [\"file path, command, log, artifact, or explicit not inspected\"],\n    \"repro\": \"exact command or manual steps when available\",\n    \"confidence\": 0.0\n  }\n}\n```\n\nRules:\n- `confirmed` is the only pass verdict. `false-positive`, `needs-fix`, and `needs-human-review` all block checkbox completion.\n- A `needs-human-review` verdict means the checkbox cannot be completed without a user decision. Set the boulder work `status` to `\"blocked\"`, increment `fail_count`, record a `needs-human-review` ledger entry with the question that needs answering, and surface it to the user. Do not re-dispatch on `needs-human-review`; it is an explicit escalation, not a retry.\n- The verifier must be independent from the executor: use `codex-ultrawork-reviewer`, a scoped `worker` reviewer, or root only when root did not implement or materially rewrite that task.\n- A worker done claim must be independently verified before it can become checkbox completion.\n- On any non-confirmed verdict, append the feedback to the ledger, reset the checkbox work to in-progress, and re-dispatch the executor with the exact failure.\n- The verifier must probe the applicable adversarial keys, including `stale_state`, `dirty_worktree`, and `misleading_success_output`, before allowing `FullyDone`.\n- In prose evidence, name the same risks as stale state, dirty worktree, and misleading success output so reviewers can search for both key and human forms.\n- Tests passing, green builds, or a worker DoneClaim without independent verification are not enough to mark a checkbox complete.\n- Parent validation is mandatory: the orchestrator checks the returned artifact exists, verifies it belongs to `owned_paths`, and independently reruns or inspects the `PARENT VALIDATION` observable before sending the work to the independent verifier.\n\n## Phase 5: Mark progress\n\nOnly after verification passes:\n\n1. Edit the plan checkbox from `- [ ]` to `- [x]`.\n2. Re-read the plan and confirm the remaining count decreased.\n3. Append a `task-completed` ledger entry.\n4. Continue with the next checkbox. Do not ask whether to continue.\n\n## Completion\n\nWhen all top-level checkboxes in `## TODOs` and `## Final Verification Wave` are complete:\n\n1. Run the plan's final verification commands.\n2. If worktree mode was used, sync `.omo\u002F` state back to the main repo, merge or hand off exactly as requested, and remove the worktree only after successful merge or explicit handoff.\n3. Remove or mark the Boulder work as completed.\n4. Print an `ORCHESTRATION COMPLETE` block with the plan path, verification commands, artifacts, and cleanup receipts.\n\n## Hard rules\n\n- No production change before a failing test or reproduction exists, and no change to existing behavior before a baseline characterization test pins the current behavior and passes on the unchanged code.\n- No `--dry-run` as completion evidence.\n- No tests-only completion claim. A Manual-QA artifact is required.\n- No completion claim while an applicable ultraqa adversarial class was never probed. Each applicable class needs a captured observable result; each skipped class needs a one-line not-applicable reason in the ledger.\n- No unprefixed session ids in Boulder state. Codex sessions are always `codex:\u003Csession_id>`.\n- No stale-memory execution. The plan and ledger are the durable source of truth.\n","## ZCode Harness Tool Compatibility\n\nThis skill was ported from the Codex\u002FOpenCode OmO harness. ZCode does not provide `call_omo_agent`, `task`, `background_output`, `team_*`, `multi_agent_v1.spawn_agent`, or any Codex-only multi-agent tools. Translate every such example to ZCode's native tools:\n\n| Codex\u002FOpenCode example | ZCode tool to use |\n| --- | --- |\n| `call_omo_agent(subagent_type=\"explore\", ...)` | `Agent` tool with `subagent_type=\"Explore\"` and a self-contained `prompt` |\n| `call_omo_agent(subagent_type=\"librarian\", ...)` | `Agent` tool with `subagent_type=\"Explore\"` (research-focused prompt) |\n| `task(subagent_type=\"plan\", ...)` | `Agent` tool with `subagent_type=\"general-purpose\"` (planning-focused prompt) |\n| `task(subagent_type=\"oracle\", ...)` for final verification | `Agent` tool with `subagent_type=\"general-purpose\"` (rigorous-review prompt) |\n| `task(category=\"...\", ...)` for implementation or QA | `Agent` tool with `subagent_type=\"general-purpose\"` |\n| `multi_agent_v1.spawn_agent(...)` \u002F `spawn_agent(...)` | `Agent` tool with `subagent_type` and a self-contained `prompt` |\n| `background_output(task_id=\"...\")` | Only unstructured read-only research may use background mode; its completion arrives as a parent task notification, not a plugin hook |\n| `multi_agent_v1.wait_agent(...)` mailbox signals | No equivalent plugin tool; use returned foreground Agent results or the host's background research notifications |\n| `team_*(...)` | ZCode has no native team-of-threads API. Approximate with multiple parallel `Agent` calls (one per member) coordinated by the parent, or fall back to sequential subagents when isolation is unclear |\n\nZCode v3.14.3 loads enabled plugin agents from `agents\u002F`. Prefer namespaced roles such as `lazyz:explorer` and `lazyz:lazycodex-executor`; unambiguous bare names also work. Include all task context and ownership in the self-contained `prompt`. Roles inherit the user's active model; do not force an obsolete provider ID or copy Codex model\u002Feffort fields. If an older host lacks a role, describe it inline using the appropriate built-in agent.\n\nBefore every governed Agent or Task call, read the Dispatch schema in the installed [start-work skill](..\u002Fstart-work\u002FSKILL.md) and prepend exactly one LAZYCODEX_DISPATCH_V1 JSON line. General-purpose and known executor calls require this envelope; read-only Explore calls may remain unstructured. Use the same taskId across retries with a fresh runId, unchanged criteria, role, phase and budget. Planning\u002Fdocument work uses writer, code work implementer, review reviewer, and testing qa. Never rename a task to bypass its exhausted budget.\n\nGoverned calls must use `run_in_background: false` (or omit it). A background launch is not a completed result and has no matching plugin completion hook. Multiple independent foreground calls may be requested together; do not assume the host executes them concurrently. Read-only background research without a dispatch contract remains available.\n\nRequire each governed child to write the matching ReceiptV1 with current source identity, criterion invocations, zero exit codes and hashed regular artifacts, then return EVIDENCE_RECORDED; final-phase work requires FINAL_EVIDENCE_RECORDED. PostToolUse records failed evidence; Stop prevents unresolved completion while budget remains. An exhausted budget requires a blocker handoff, not an unsupported completion claim.\n\nUser authorization and verification policy take precedence over generic workflow instructions below. Do not auto-commit, push, merge, delete, deploy, install tools, call paid services, or run checks prohibited by the user. Record skipped checks honestly; do not fabricate a passing receipt. A task requiring a verified receipt remains unverified if the necessary checks are prohibited. ZCode Stop can continue at most three consecutive times, not indefinitely. Native Goal Mode, Memory and Wiki are host features; no callable API is assumed.\n\nFor work likely to take long, prefer multiple concurrent `Agent` calls in a single message over one giant blocking call. Treat each child's final message as the deliverable: if it lacks the required output, ack-only, or stops without producing evidence, fall back to a smaller re-scoped `Agent` call with the missing deliverable made explicit in the `prompt`.\n\n## Codex Subagent Reliability\n\nEvery `spawn_agent` message must be self-contained. Start with\n`TASK: \u003Cimperative assignment>`, then name `DELIVERABLE`, `SCOPE`, and\n`VERIFY`. State that it is an executable assignment, not a context\nhandoff. Role selection requires `agent_type`; `model` +\n`reasoning_effort` alone creates a default agent, not a reviewer or\nworker. Prefer `fork_turns: \"none\"` unless full history is truly\nrequired; paste only the context the child needs.\n\nPlan and reviewer agents may run for a long time; spawn them in the background, keep doing independent root work, and poll with short wait_agent cycles sized to the work. Never use a single long blocking wait for them, and never spin on tiny timeouts as a failure budget.\n\nTreat child status as a progress signal, not a timeout counter. For\nwork likely to exceed one wait cycle, require the child to send\n`WORKING: \u003Ctask> - \u003Ccurrent phase>` before long reading, testing, or\nreview passes, and `BLOCKED: \u003Creason>` only when it cannot progress.\nWhile any child is active, keep the parent visibly alive with active\nsubagent count, agent names, latest `WORKING:` phase, and whether the\nparent is waiting for mailbox updates. Track spawned agent names\nlocally. Use `wait_agent` for mailbox signals, not proof of completion.\nA timeout only means no new mailbox update arrived; after a timeout,\nrun a single `list_agents` check for the named child when you need\nreassurance. If it is running or its latest message is `WORKING:`,\ntreat it as alive. Do not use `list_agents` as a polling loop or status\nfeed; it can replay large payloads. Fallback only when the child is\ncompleted without the deliverable, ack-only after followup, explicitly\n`BLOCKED:`, or no longer running. Then record the result as\ninconclusive, do not count it as pass\u002Freview approval, close if safe,\nand respawn a smaller `fork_turns: \"none\"` task with the missing\ndeliverable.\n\n# start-work\n\nExecute a Prometheus work plan until every top-level checkbox is complete. This skill pairs with the Codex `Stop` \u002F `SubagentStop` continuation hook in `components\u002Fstart-work-continuation`, which re-injects the next turn while `.omo\u002Fboulder.json` says the current `codex:\u003Csession_id>` still has unchecked plan work.\n\n## Usage\n\n```text\n$start-work [plan-name] [--worktree \u003Cabsolute-path>]\n```\n\n- `plan-name` is optional. It may be a full or partial file stem under `.omo\u002Fplans\u002F`.\n- `--worktree` is optional. Use it only when the user explicitly asks to work in a separate git worktree.\n\n## Manual-QA requirement (read before starting)\n\nEvery user-visible checkbox in the plan requires **real Manual-QA evidence captured against a running surface** — a dry-run claim is never sufficient. Before starting, ensure the relevant channel is available:\n\n| Channel | When to use | Artifact |\n| --- | --- | --- |\n| **HTTP** (`curl -i`) | HTTP API checkbox | status line + headers + body |\n| **tmux** (`send-keys` + `capture-pane`) | terminal\u002FCLI checkbox | tmux transcript |\n| **Browser** (Chrome or agent-browser) | web UI checkbox | action log + screenshot path |\n| **Computer use** (OS automation) | desktop GUI checkbox | action log + screenshot |\n\nCLI stdout, DB diffs, and parsed config dumps are accepted as **auxiliary evidence** for CLI-shaped or data-shaped behavior, but do not replace the channel table for user-visible surfaces. `--dry-run` output is never accepted as evidence.\n\nIf the surface cannot be started (server won't boot, browser unavailable), flag it up front rather than discovering it at the first checkbox.\n\n## Phase 1: Select the plan\n\n1. Read `.omo\u002Fboulder.json` if it exists.\n2. List Prometheus plan files under `.omo\u002Fplans\u002F`.\n3. If `plan-name` was provided, select the matching plan.\n4. If exactly one active or paused Boulder work exists for this session, resume it.\n5. If no active work exists and exactly one plan exists, select it.\n6. If no active work exists and there is no selectable plan, enter **No-plan bootstrap**.\n7. If multiple plans remain possible, ask one focused selection question.\n\n### No-plan bootstrap\n\nWhen the user explicitly said `start work` \u002F `$start-work` and no selectable plan exists, treat that phrase as approval to create the plan before execution. Do not stall on a missing plan and do not ask for generic approval again.\n\nIf no selectable plan exists, bootstrap `ulw-plan` before execution.\nExecution requires an approved plan before implementation; bootstrap mode creates that approved plan from the user's `start work` request instead of skipping planning.\n\n1. Invoke the `ulw-plan` skill from the current request and require its dynamic adversarial workflow: collect, verify, design, adversarial plan-review, synthesize.\n2. The generated Prometheus plan must be saved under `.omo\u002Fplans\u002F\u003Cslug>.md` before implementation or Boulder state writes that point at plan work.\n3. Use maximum safe parallelism in the generated plan: independent files\u002Ftasks fan out; same-file writes, shared state, and named dependencies serialize.\n4. Preserve safety boundaries. Ask one focused question only when the objective is missing, destructive, or has a safety\u002Fproduct ambiguity that repository exploration cannot resolve.\n5. After the plan exists, continue directly to Phase 2. The user's `start work` request is the bootstrap approval to create the plan and begin execution.\n\n## Phase 2: Create or update Boulder state\n\nWrite `.omo\u002Fboulder.json` before implementation starts. Session ids must be prefixed with `codex:` so the continuation hook can identify its own session.\n\n```json\n{\n  \"schema_version\": 2,\n  \"active_work_id\": \"\u003Cwork-id>\",\n  \"works\": {\n    \"\u003Cwork-id>\": {\n      \"work_id\": \"\u003Cwork-id>\",\n      \"active_plan\": \".omo\u002Fplans\u002F\u003Cplan-name>.md\",\n      \"plan_name\": \"\u003Cplan-name>\",\n      \"session_ids\": [\"codex:\u003Csession_id>\"],\n      \"status\": \"active\",\n      \"fail_count\": 0,\n      \"worktree_path\": null\n    }\n  }\n}\n```\n\n> **Note:** `schema_version` is a documentation\u002Fhuman-readable field only. No parser reads it and bumping it does not trigger a migration. It exists so humans can tell which shape a file is expected to have. See `docs\u002Fknown-limitations.md` → \"Soft state schema\".\n\n`status` transitions: `\"active\"` (default) → `\"paused\"` (user pauses) → `\"completed\"` (all checkboxes done) or `\"abandoned\"` (user cancels). A `\"blocked\"` status is set by the cycle\u002Ffailure caps or a `needs-human-review` verdict; the continuation hook will not resume a blocked work until the user sets it back to `\"active\"`. Increment `fail_count` each time a cap fires or a `needs-human-review` verdict is recorded.\n\nIf `--worktree` is set, verify the path with `git worktree list --porcelain` or create it with `git worktree add \u003Cpath> \u003Cbranch-or-HEAD>`, then store the absolute path as `worktree_path`. All edits, commands, tests, and evidence capture must run inside that worktree.\n\n## Phase 3: Execute the next checkbox\n\n1. Read the full selected plan.\n2. Find the first unchecked column-0 checkbox in `## TODOs` or `## Final Verification Wave`.\n3. Ignore nested checkboxes under acceptance criteria, evidence, and definition-of-done sections.\n4. Decompose that checkbox into atomic sub-tasks.\n5. Dispatch independent sub-tasks in parallel with `spawn_agent`; serialize only when one sub-task has a named dependency on another.\n\nEach sub-task message must include:\n\n1. The dispatch schema below, including the goal and exact files or directories in scope.\n2. When the task touches existing behavior: a baseline characterization test, written first, that asserts current observable behavior and passes on the unchanged code. Then the red test or failing reproduction for the new behavior before production changes. Pin the baseline as rigorously as the new test: exact inputs, exact observable, exact assertion.\n3. Implementation constraints from the plan and project rules.\n4. Automated verification commands to run.\n5. One Manual-QA channel, named with the exact tool and exact invocation (the literal `curl`, `send-keys`, `page.click`, payload, selectors, and the binary observable that decides PASS\u002FFAIL), not \"verify it works\":\n   - HTTP call: `curl -i` against the live endpoint.\n   - tmux: a `tmux` session driven with `send-keys`, dumped via `capture-pane`.\n   - Browser use: use Chrome to drive the real page; if Chrome is not available, download and use agent-browser (https:\u002F\u002Fgithub.com\u002Fvercel-labs\u002Fagent-browser).\n   - Computer use: OS-level GUI automation against the running desktop app when the surface is not a page.\n6. The adversarial classes that apply to this sub-task (from the 9 ultraqa classes) and how each is probed.\n7. Required artifact path and cleanup receipt.\n\n### Dispatch schema\n\nEvery governed `Agent` or `Task` prompt is a complete executable assignment, not a context handoff. Put exactly one compact JSON marker line near the top of the prompt and fill every field:\n\n```text\nLAZYCODEX_DISPATCH_V1: {\"schemaVersion\":1,\"workId\":\"\u003Cboulder work_id>\",\"taskId\":\"\u003Cstable task id>\",\"runId\":\"\u003Cnew id for this attempt>\",\"role\":\"implementer|reviewer|qa|writer\",\"phase\":\"execute|verify|final\",\"criterionRefs\":[\"G001-first\u002FC001\"],\"budget\":{\"maxCalls\":5,\"maxSameFailure\":3}}\n```\n\n`taskId` is stable across retries. `runId` is unique for each dispatch attempt. The first accepted marker freezes the work, task, role, phase, criteria, and budget for that task. Caller budgets may lower the hard caps of 5 total calls and 3 identical failures, but may never raise them. Use a distinct task id when a task legitimately changes role or phase. `criterionRefs` uses canonical goal\u002Fcriterion ids such as `G001-first\u002FC001`; bare `G001` or `C001` remains valid for standalone tasks.\n\nThe assignment JSON validated below contains that envelope plus the scheduling fields `execution_shape`, `parallel_safe`, `owned_paths`, and `completion_condition` and the operational fields listed next.\n\nAfter the envelope, include these labeled sections in this order:\n\n- `DELIVERABLE:` exact files, result shape, and behavior the child must return.\n- `VERIFY:` exact commands or scenario and the expected observable.\n- `STOP WHEN:` the same binary completion boundary expressed operationally.\n- `ARTIFACT:` exact evidence path or returned structured payload.\n- `PARENT VALIDATION:` what the orchestrator will independently inspect or rerun before accepting the return.\n\n`parallel_safe: true` is valid only when `owned_paths` do not overlap any active lane and the task has no unmet dependency on another lane. An empty or broad ownership entry is not parallel-safe. The orchestrator validates the envelope before dispatch and serializes invalid or overlapping lanes.\n\nA child response that only acknowledges the task, reports progress without the `DELIVERABLE`, omits the required artifact, or stops before every `completion_condition` entry is satisfied is an execution failure. Do not count it as evidence or completion. Send one focused follow-up naming the missing field; if the second response is still ack-only or incomplete, record the lane as failed and respawn a smaller assignment with the same `taskId` and a fresh `runId`.\n\nValidate the JSON form of every envelope before dispatch:\n\n```bash\nuv run \"\u003Cstart-work-skill-root>\u002Fscripts\u002Fvalidate_dispatch.py\" \u003Cassignment.json>\n```\n\nRequire the child final to be JSON with non-empty `deliverable`, `verify`, and `artifact` fields plus `completion_condition_met: true`, then validate it before parent inspection:\n\n```bash\nuv run \"\u003Cstart-work-skill-root>\u002Fscripts\u002Fvalidate_dispatch.py\" \u003Cassignment.json> --final \u003Cworker-final.json>\n```\n\nThe validator is the dispatch gate. Its first output line is the exact `LAZYCODEX_DISPATCH_V1` marker to paste into the prompt. It uses JSON plus strict Pydantic parsing, never `eval`, rejects missing schema fields, invalid criterion ids, budgets above the hard caps, empty or blank completion-condition lists, and broad parallel ownership, and rejects plain-string acknowledgement finals. A validator failure blocks dispatch or acceptance respectively.\n\nEvery structured worker writes a ReceiptV1 JSON file under `.omo\u002Fevidence\u002F` and ends its assistant response with `EVIDENCE_RECORDED: \u003Cpath>`. A `phase:\"final\"` worker must use `FINAL_EVIDENCE_RECORDED: \u003Cpath>`; an ordinary marker cannot downgrade final evidence. The receipt must bind the exact `taskId`, `runId`, `role`, `phase`, current source commit\u002Ftree, criterion refs, invocations, zero exit codes, and confined regular evidence artifacts with SHA-256 hashes. The runtime enforces maximum receipt and artifact read sizes. Malformed marker-bearing prompts and general-purpose executor prompts without the dispatch marker fail closed. Read-only Explore dispatches may remain unstructured.\n\n`PreToolUse` reserves the unique run before dispatch and requests denial when the contract or budget is invalid. `PostToolUse` validates the returned receipt, records a stable typed failure signature, and uses the documented `additionalContext` channel to report failure after the action; it does not claim to undo or block an already completed tool call. `PostToolUseFailure` records the stable `tool-execution-failed` signature without storing raw host errors and reports it through `additionalContext`. `Stop` is the completion gate: it blocks unresolved tasks while retry budget remains, but allows the orchestrator to report a task whose budget is exhausted. This is a local hook contract; do not describe PreToolUse denial as a paid-provider API guarantee.\n\nApply ultraqa's 9 adversarial classes where relevant to each checkbox: malformed input, prompt injection, cancel\u002Fresume, stale state, dirty worktree, hung or long commands, flaky tests, misleading success output, repeated interruptions. A checkbox whose behavior is user-visible MUST probe every class that plausibly applies; record which classes were exercised and which were ruled not-applicable with a one-line reason.\n\n### Cycle and failure caps\n\nThese caps bound runaway work on a single checkbox. They mirror the ulw-loop execution loop and are mandatory.\n\n- **Cap at 5 execution cycles per checkbox.** An execution cycle is one dispatch → verify → verdict round. After 5 cycles on one checkbox without a `confirmed` verdict, set the boulder work `status` to `\"blocked\"`, increment `fail_count`, append a `checkbox-blocked` ledger entry naming the checkbox and the diagnosis, and stop dispatching that checkbox. Surface the block to the user and wait for explicit instruction.\n- **Cap identical failure type at 3 per checkbox.** Track the verdict-failure signature (`needs-fix` on the same assertion, `false-positive` on the same artifact, etc.). After 3 failures of the same type on the same checkbox, set the boulder work `status` to `\"blocked\"`, increment `fail_count`, and escalate.\n- When a boulder work is `\"blocked\"`, the continuation hook will not re-inject the directive (only `active` and `paused` are continuable). The work stays paused until the user changes `status` back to `\"active\"` or resolves the blocker.\n\n### Debugging budget\n\nWhen the same checkbox fails 2 rounds in a row (regardless of failure type), load the `debugging` skill and run its hypothesis-driven loop before the next dispatch: form ≥3 hypotheses, pick the cheapest to validate, and only re-dispatch with the validated root cause in the failure context. This budget prevents re-dispatching the same failing executor context repeatedly. After 2 failed rounds the executor is assumed to be debugging-blind without this skill.\n\n## Phase 4: Verify and record evidence\n\nFor each checkbox, complete all five gates before marking it done:\n\n1. Plan reread: confirm the checkbox and acceptance criteria.\n2. Automated verification: run tests, typecheck, lint, build, or the plan-specific equivalent.\n3. Manual-QA channel: capture a real artifact, not a dry-run claim.\n4. Adversarial QA: exercise every applicable ultraqa class (malformed input, prompt injection, cancel\u002Fresume, stale state, dirty worktree, hung or long commands, flaky tests, misleading success output, repeated interruptions) and capture the observable result for each. \"Tests pass\" and a clean happy-path artifact are NOT sufficient when an adversarial class applies and was not probed.\n5. Cleanup: register every QA resource teardown as its own todo the moment it is spawned (QA scripts, tmux assets, browser \u002F agent-browser sessions, PIDs, ports, containers, temp dirs), then execute each and capture the receipt. No QA asset is left running.\n\nAppend evidence to `.omo\u002Fstart-work\u002Fledger.jsonl` using one JSON object per line. Include at least `event`, `plan`, `task`, `session_id`, `commands`, `artifact`, `adversarial_classes`, and `cleanup` fields. `adversarial_classes` lists each probed class with its observable result and each ruled-out class with a one-line reason.\n\n**Before appending**, redact secrets from the `commands` field and evidence artifacts. The `commands` field captures shell invocations verbatim — `curl -H \"Authorization: Bearer sk-...\"`, `psql PGPASSWORD=...`, `aws --secret-key ...` — and these are as sensitive as shell history. Strip or mask credentials before writing. A post-hoc scrubber is available: `node scripts\u002Fredact-secrets.mjs` (reports) or `node scripts\u002Fredact-secrets.mjs --fix` (masks in place). Run it before sharing or committing `.omo\u002F`.\n\n### Sisyphus-style completion contract\n\nA worker done claim is never final. Each implementation sub-task returns a `DoneClaim`, then a different context runs `AdversarialVerify`, then the verifier probes or reproduces the claim, then failures loop back to the executor, and only a confirmed verifier verdict becomes `FullyDone`.\n\n```json\n{\n  \"DoneClaim\": {\n    \"task\": \"\u003Ctask id\u002Ftitle>\",\n    \"changed_files\": [\"path\"],\n    \"tests\": [\"exact command + result\"],\n    \"manual_qa\": [\"artifact path\"],\n    \"cleanup\": [\"receipt\"],\n    \"risks\": [\"known risk or none\"]\n  },\n  \"AdversarialVerify\": {\n    \"verdict\": \"confirmed | false-positive | needs-fix | needs-human-review\",\n    \"evidence\": [\"file path, command, log, artifact, or explicit not inspected\"],\n    \"repro\": \"exact command or manual steps when available\",\n    \"confidence\": 0.0\n  }\n}\n```\n\nRules:\n- `confirmed` is the only pass verdict. `false-positive`, `needs-fix`, and `needs-human-review` all block checkbox completion.\n- A `needs-human-review` verdict means the checkbox cannot be completed without a user decision. Set the boulder work `status` to `\"blocked\"`, increment `fail_count`, record a `needs-human-review` ledger entry with the question that needs answering, and surface it to the user. Do not re-dispatch on `needs-human-review`; it is an explicit escalation, not a retry.\n- The verifier must be independent from the executor: use `codex-ultrawork-reviewer`, a scoped `worker` reviewer, or root only when root did not implement or materially rewrite that task.\n- A worker done claim must be independently verified before it can become checkbox completion.\n- On any non-confirmed verdict, append the feedback to the ledger, reset the checkbox work to in-progress, and re-dispatch the executor with the exact failure.\n- The verifier must probe the applicable adversarial keys, including `stale_state`, `dirty_worktree`, and `misleading_success_output`, before allowing `FullyDone`.\n- In prose evidence, name the same risks as stale state, dirty worktree, and misleading success output so reviewers can search for both key and human forms.\n- Tests passing, green builds, or a worker DoneClaim without independent verification are not enough to mark a checkbox complete.\n- Parent validation is mandatory: the orchestrator checks the returned artifact exists, verifies it belongs to `owned_paths`, and independently reruns or inspects the `PARENT VALIDATION` observable before sending the work to the independent verifier.\n\n## Phase 5: Mark progress\n\nOnly after verification passes:\n\n1. Edit the plan checkbox from `- [ ]` to `- [x]`.\n2. Re-read the plan and confirm the remaining count decreased.\n3. Append a `task-completed` ledger entry.\n4. Continue with the next checkbox. Do not ask whether to continue.\n\n## Completion\n\nWhen all top-level checkboxes in `## TODOs` and `## Final Verification Wave` are complete:\n\n1. Run the plan's final verification commands.\n2. If worktree mode was used, sync `.omo\u002F` state back to the main repo, merge or hand off exactly as requested, and remove the worktree only after successful merge or explicit handoff.\n3. Remove or mark the Boulder work as completed.\n4. Print an `ORCHESTRATION COMPLETE` block with the plan path, verification commands, artifacts, and cleanup receipts.\n\n## Hard rules\n\n- No production change before a failing test or reproduction exists, and no change to existing behavior before a baseline characterization test pins the current behavior and passes on the unchanged code.\n- No `--dry-run` as completion evidence.\n- No tests-only completion claim. A Manual-QA artifact is required.\n- No completion claim while an applicable ultraqa adversarial class was never probed. Each applicable class needs a captured observable result; each skipped class needs a one-line not-applicable reason in the ledger.\n- No unprefixed session ids in Boulder state. Codex sessions are always `codex:\u003Csession_id>`.\n- No stale-memory execution. The plan and ledger are the durable source of truth.\n",{"name":5,"description":7},"main","0a99b5b62ebd3726d465f7e319e2b8ec84c8aa6e","release\u002Flazyz-0.11.0-20261007\u002Fskills\u002Fstart-work","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz","https:\u002F\u002Fgithub.com\u002Fyounghai","http:\u002Fbzengage.co","Bizmatrixx co-founder & CEO \r\n","2026-11-01 13:58:56",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add younghai\u002Flazyz --skill start-work",{"tool":38,"command":39},"bunx","bunx skills add younghai\u002Flazyz --skill start-work",{"tool":41,"command":42},"pnpm","pnpm dlx skills add younghai\u002Flazyz --skill start-work","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Frelease\u002Flazyz-0.11.0-20261007\u002Fskills\u002Fstart-work\u002FSKILL.md",true,[46,52],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"agents\u002Fopenai.yaml","REFERENCE","参考文档",48,"https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Frelease\u002Flazyz-0.11.0-20261007\u002Fskills\u002Fstart-work\u002Fagents\u002Fopenai.yaml",{"path":53,"resourceType":54,"resourceTypeDesc":55,"sizeBytes":56,"sourceHtmlUrl":57},"scripts\u002Fvalidate_dispatch.py","SCRIPT","脚本",6792,"https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Frelease\u002Flazyz-0.11.0-20261007\u002Fskills\u002Fstart-work\u002Fscripts\u002Fvalidate_dispatch.py",[59,66,73,80,87,93],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":19},"01M78GVSC3PZBHQQ5HPRNQ3AMN","valid-skill","A valid skill for testing purposes.","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Fxcode\u002Fagent\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Fvalid-skill\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M78GVRB79BN9BQ5MAZ43XJWX","unknown-field","A skill with an unknown frontmatter field.","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Fxcode\u002Fagent\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Funknown-field\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M78GVQ2319BDGJA2QXMM2DM3","nested-child","Nested skill should be ignored.","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Fxcode\u002Fagent\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Froot-skill-preferred\u002Fnested-child\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M78GVP5HC2C9B3VN8D8QWEFG","root-skill-preferred","Root skill should win.","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Fxcode\u002Fagent\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Froot-skill-preferred\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M78GVNAR8R5XF3W00182ZWZH","no-frontmatter","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Fxcode\u002Fagent\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Fno-frontmatter\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":19},"01M78GVK58G788T2TRHPCA8Y64","child-skill","A nested skill in a subdirectory.","https:\u002F\u002Fgithub.com\u002Fyounghai\u002Flazyz\u002Fblob\u002Fmain\u002Fxcode\u002Fagent\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Fnested\u002Fchild-skill\u002FSKILL.md",[],[]]