[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6MP5MVBHK7WQJSEERXT910J":3,"skill-same:01M6MP5MVBHK7WQJSEERXT910J":85},{"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":23,"publishedAt":44,"linkOnly":45,"markdownContent":46,"markdownBody":47,"frontMatter":48,"refName":51,"commitSha":52,"skillDirectory":53,"forksCount":54,"repositoryHtmlUrl":55,"repositoryHomepageUrl":56,"ownerHtmlUrl":57,"ownerBlogUrl":58,"ownerDescription":59,"updatedAt":60,"install":61,"resources":74},"01M6MP5MVBHK7WQJSEERXT910J","work",null,"Execute a work plan or specification systematically -- read the plan, set up a branch, implement tasks with continuous testing, commit incrementally, and ship a PR. Use when you have a plan file, spec, or task list ready to execute. --auto answers the routine gates (branch, final commit, PR handoff, workspace cleanup, third failed fix attempt) in advance and stops only for a blocker, a Critical finding, or a merge conflict.","hashgraph-online\u002Fawesome-codex-plugins","hashgraph-online","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F160371962?v=4","01M6CNB6NRJVNNGN14TR01Q6D3","HOL (Hashgraph Online)","01M6JT1HSKKG9V68KQAQGBA7XP",1575,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fyagizdo\u002Fquiver\u002Fskills\u002Fwork\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量","SCENARIO","使用场景",[5,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43],"list","routine","for","you","step by step","conflict","specification","ready","auto","use","fix","testing","read","branch","plan","ship","start","file","run","cleanup","2026-11-10 10:12:30",false,"---\nname: work\ndescription: \"Execute a work plan or specification systematically -- read the plan, set up a branch, implement tasks with continuous testing, commit incrementally, and ship a PR. Use when you have a plan file, spec, or task list ready to execute. --auto answers the routine gates (branch, final commit, PR handoff, workspace cleanup, third failed fix attempt) in advance and stops only for a blocker, a Critical finding, or a merge conflict.\"\nargument-hint: \"\u003Cplan file path or task description> [--auto]\"\nwhen-to-use: \"user wants to execute a saved plan or implement tasks step by step -- '\u002Fwork', '\u002Fwork --auto', 'start implementation', 'execute the plan', 'work through these tasks', 'run the plan without stopping to ask'\"\n---\n\n# Gather Context\n\n```\n!`git rev-parse --is-inside-work-tree 2>\u002Fdev\u002Fnull || echo \"NO_GIT\"`\n```\n\n```\n!`git branch --show-current 2>\u002Fdev\u002Fnull || echo \"NO_GIT\"`\n```\n\n```\n!`git log --oneline -5 2>\u002Fdev\u002Fnull || echo \"NO_COMMITS\"`\n```\n\n```\n!`git status --short 2>\u002Fdev\u002Fnull || echo \"NO_GIT\"`\n```\n\n---\n\n# Work\n\nExecute a work plan, specification, or task list systematically. The focus is on shipping complete features by understanding requirements quickly, following existing patterns, and maintaining quality throughout.\n\n**Announce:** \"Using the work skill to execute the plan.\"\n\n**Arguments.** `--auto` anywhere in `$ARGUMENTS` sets **auto mode** for this run. Strip it before Phase 1 reads the rest as a path or description. Auto mode pre-answers the routine gates and nothing else: the branch question in Phase 2 (continue on the current branch), the third failed fix attempt in Phase 3 and in the orchestrator's FAILED handling (accepted and listed; an orchestrated failure's dependents stay paused), the final commit in 5a (committed without the prompt), the handoff in 5b (no review, no PR -- the commands are printed as text), and the workspace in 5c-bis (kept). It answers no question whose answer changes what gets built or deletes something: a plan contradiction (1e), a blocker (Phase 3), an unaddressed Critical finding or unmet acceptance criterion (4c), and a merge conflict stop and ask exactly as they do without the flag. `\u002Fship` passes it so a run that answered every question up front is not stopped for a routine one; a user can pass it directly for the same effect. Every `AskUserQuestion` site below names its auto-mode branch; a prompt added any other way stalls an auto run.\n\n**Before starting Phase 1**, use the Glob tool to gather plan context silently (do not show results to the user):\n1. `.claude\u002Fplans\u002F*.md` -- existing plans\n2. `**\u002Fplans\u002F*.md` (max depth 4) -- plans in other locations\n\nTreat empty Glob results as \"no plans found\". Proceed regardless.\n\n**Use when:** a plan file exists, the user provides a spec or task list, or the user says \"execute\", \"build this\", \"implement the plan\". Use `\u002Fquiver:plan` first if no plan exists yet. Use `\u002Fquiver:review` for code review.\n\n---\n\n## Workflow\n\n```\n+----------+     +---------+     +---------+     +---------+     +--------+\n| 1. LOAD  | --> | 2. SETUP| --> | 3. BUILD| --> | 4. CHECK| --> | 5. SHIP|\n| Read plan|     | Branch  |     | Execute |     | Quality |     | PR     |\n+----------+     +---------+     +---------+     +---------+     +--------+\n```\n\n### Phase 0: Git Availability\n\nIf any gather-context block above returned `NO_GIT`, this directory is not a git repository.\nPrint: `> No git repository detected -- skipping branch\u002Fcommit context.`\nProceed to Phase 1. Treat all git-sourced fields (branch, log, diff, status) as empty. Skip branch creation in Phase 2, commit steps in Phase 3, and git-dependent actions in Phase 5.\n\n### Phase 1: Load and Clarify\n\n#### 1a -- Load the plan\n\n**Case A: Path provided.** If `$ARGUMENTS` is a file path (ends in `.md` or contains `\u002F`):\n- Read that file as the work plan.\n- Print: `> Executing plan: {plan filename} ({step count} steps) on branch {branch}`\n- Proceed directly to Phase 1b. The user chose this plan explicitly -- no confirmation needed.\n\n**Case B: No arguments.** If `$ARGUMENTS` is empty, collect all `.md` files from every `plans\u002F` directory discovered by the Glob block above, plus the `.claude\u002Fplans\u002F` listing.\n\n- If exactly one plan exists across all directories, read it and use `AskUserQuestion`:\n  > Found one plan: `{relative path}` (in `{directory}`)\n  > **Goal:** {goal from plan}\n  > **Steps:** {step count}\n  Buttons: `[\"Execute this plan\", \"Other -- I'll provide a path or description\"]`\n- If multiple plans exist, present them via `AskUserQuestion` with full relative paths as buttons (most recent first), plus an `\"Other -- I'll provide a path or description\"` option. Show the directory in each label so the user can distinguish plans in different locations.\n- If the user picks \"Other\", ask for a path or task description.\n\nIf no plans found:\n> No plans found in any `plans\u002F` directory. Usage:\n> - `\u002Fwork \u003Cpath-to-plan.md>` -- execute a specific plan\n> - `\u002Fwork \u003Cplan-name>` -- search for a plan by name\n> - `\u002Fwork \u003Ctask description>` -- work on a task directly\n> - `\u002Fplan \u003Ctask>` -- create a plan first\n\n**Stop here.**\n\n**Case C: Name or description provided.** If `$ARGUMENTS` is not a file path and is not empty:\n\n1. **Discover plan files.** Read every `plans\u002F` directory discovered by the Glob block. Combine with `.claude\u002Fplans\u002F` listing.\n2. **Match by filename.** Compare `$ARGUMENTS` against each plan file's stem (without `.md` extension):\n   - **Exact match:** stem equals `$ARGUMENTS` (case-insensitive).\n   - **Partial match:** stem contains `$ARGUMENTS` as a substring (case-insensitive).\n3. **Rank and select.** Rank exact matches above partial matches.\n   - **1 match** -- Read the plan file, print `> Executing plan: {filename} from {directory}`, proceed to Phase 1b. No confirmation needed.\n   - **Multiple matches** -- Use `AskUserQuestion` to present candidates with full relative paths as button labels. Add a `\"None of these -- treat as task description\"` button.\n   - **0 matches** -- Treat `$ARGUMENTS` as an inline task description. Print `> No matching plan found for \"{$ARGUMENTS}\". Treating as task description.` Proceed to Phase 1b with the description as the work specification.\n\n#### 1b -- Review references\n\nIf the plan links to files, patterns, or prior research -- read those now. Understanding context before coding prevents rework.\n\nIf the plan carries a `## Global Constraints` section, read it as binding context for every task in this run, not as background prose -- it states what this work must not do.\n\n#### 1c -- Detect review-fix plan\n\nCheck if this is a plan created to address review findings:\n- **Primary**: Check plan YAML frontmatter for a `review_source` field (e.g., `review_source: .claude\u002Freports\u002Freview-2026-03-10_14-30-00.md`).\n- **Fallback**: Scan plan content for paths matching `.claude\u002Freports\u002Freview-*.md` or `review-*_*-*-*.md`.\n- If detected, read the review report file. If the file exists, note this as a **review-fix plan** and carry the parsed findings forward to Phase 4. If the file does not exist, warn: \"Review report not found at {path}. Proceeding without review-aware verification.\"\n- If no review reference detected, proceed normally.\n- **Iteration tracking**: Read the plan's `review_iteration` frontmatter field (default: `1` if absent). If `review_iteration >= 2`, this is the **final iteration** -- Phase 4c verification is the only quality gate, and Phase 4d agent review is skipped unconditionally.\n\n#### 1d -- Flag ambiguities\n\nIf the plan contains genuine contradictions (e.g., two steps that conflict, a referenced file that does not exist), note them. If the plan is clear, skip this step entirely -- do not invent ambiguities.\n\n#### 1e -- Proceed or clarify\n\n- **No plan found:** Ask the user what to work on.\n- **Plan has contradictions** flagged in 1d: Ask about those specific contradictions only.\n- **Plan is clear:** Move directly to Phase 2. Do NOT ask \"should I proceed?\", \"are you sure?\", \"shall I start?\", or any variation of confirmation. The user invoked `\u002Fwork` with a plan -- that IS the approval. Summarizing the plan back and asking to continue is wasted time.\n\n### Phase 2: Setup Environment\n\n**If git is NOT available (Phase 0 detected `NO_GIT`):** Skip this phase entirely -- proceed to Phase 2.5.\n\n**If the repository has no commit yet** (git is available and the `git log` gather block printed `NO_COMMITS`), no branch exists until something is committed. `git checkout -b` then only renames the commit-less branch, so the default branch never comes into existence, and a later `\u002Fquiver:review --base \u003Cdefault branch>` stops because that branch is not found. Settle it before the rules below, with an empty first commit: `git commit --allow-empty --only -m \"chore: initial commit\"`. `--only` with no paths leaves anything already staged out of it, so the commit holds no file; that is why this skill makes it without asking whether to commit.\n\n- On `main` or `master`: make the commit, print `> This repository had no commit yet, so I made an empty first commit on \u003Cbranch> for the new branch to start from.`, and create the branch as below.\n- On any other branch there is no main branch to compare against, and its name is the user's choice, so ask, in auto mode too. One `AskUserQuestion`, with `main` first unless `git config --get init.defaultBranch` prints `master`:\n  > This repository has no commit yet, and you are on `\u003Cbranch>`. A review compares this branch against a main branch, and there is none yet. Which one should I create?\n\n  - \"Create main\" -- An empty first commit becomes `main`, and the work continues on `\u003Cbranch>` from it.\n  - \"Create master\" -- An empty first commit becomes `master`, and the work continues on `\u003Cbranch>` from it.\n  - \"Create neither\" -- Nothing is committed now; a review will have no main branch to compare this branch against.\n\n  On \"Create main\" or \"Create master\", make the commit, then run `git branch \u003Cname>`, and continue on `\u003Cbranch>`. On \"Create neither\", make no commit.\n\n**If already on a feature branch** (not main\u002Fmaster), use `AskUserQuestion`:\n> You're on `{branch}`. Continue here or create a new branch?\nButtons: `[\"Continue on {branch}\", \"Create new branch\"]`\nIf continuing, move to Phase 2.5. In auto mode, continue on the current branch without asking.\n\n**If on the default branch**, create a new branch by default: `git checkout -b \u003Cmeaningful-name>` using a descriptive name (e.g., `feat\u002Fuser-auth`, `fix\u002Femail-validation`). Never commit to the default branch without explicit user confirmation.\n\n### Phase 2.5: Orchestration Decision\n\nParse the plan and count top-level tasks (one deliverable = one task, regardless of markup format). Announce before proceeding:\n```\nStrategy: {sequential | parallel orchestration} ({N} tasks found)\nReason: {why}\n```\n\n#### Resolve the verification command\n\nRead `skills\u002Fverification\u002FSKILL.md` and follow its `## Command Resolution` section. This runs for every plan, sequential and orchestrated alike, once per run, before the strategy split below, so both paths carry the value. Print one line under the strategy line:\n\n```\nVerification: test \u003Ccommand | none (\u003Creason>)>, build \u003Ccommand | none (\u003Creason>)> (source: docs | stack | none)\n```\n\nThe orchestrator copies the test value into every brief's `Test command:` line. `none` is printed as-is and carried forward -- it is never replaced by a guess.\n\nThen choose the path; the first line that matches wins:\n\n- **No commit yet** (the `git log` gather block printed `NO_COMMITS`, which it also prints outside a git repository, and Phase 2 made no empty first commit): Sequential, whatever the task count. The Reason line says the repository has no commit yet, or that there is no git repository when Phase 0 detected `NO_GIT`. The working branch does not resolve until a first commit exists, and outside a repository there is no branch at all, so every worktree subagent's first step -- switching to that branch -- would come back BLOCKED. That leaves a run outside git and one where the user chose to create no main branch; every other run got its first commit in Phase 2 and takes the rules below.\n- **1-2 tasks:** Sequential. Proceed to Phase 3 with the verification command resolved above. The ledger below is orchestration-path-only; the sequential path keeps TodoWrite unchanged and writes nothing to disk.\n- **3+ tasks:** Parallel orchestration. Resolve the workspace and check for a ledger (below), then follow `skills\u002Fwork\u002Forchestrator.md`. Skip Phase 3 entirely -- orchestration replaces it.\n\n#### Resolve the workspace\n\n`\u003Cplan-basename>` is the loaded plan file's name without `.md`. The workspace is `.claude\u002Fwork\u002F\u003Cplan-basename>\u002F` and the ledger is `.claude\u002Fwork\u002F\u003Cplan-basename>\u002Fprogress.md`.\n\nAny path that reaches Phase 2.5 without a plan file has no plan basename -- Phase 1 Case B's \"Other -- I'll provide a path or description\" and Case C with 0 matches both do. Derive one instead of proceeding without it: slugify the task description to at most 40 characters and use `.claude\u002Fwork\u002Fadhoc-\u003Cslug>\u002F`. The workspace and ledger are otherwise identical, and the identity line names the task description in place of a plan file path. The orchestrator requires a workspace for every run -- every dispatch step writes to one. Name the derived workspace in the strategy line.\n\n#### Check for a ledger\n\nRead `.claude\u002Fwork\u002F\u003Cplan-basename>\u002Fprogress.md`. Its first two lines are the identity header:\n\n```\n# work ledger -- plan: \u003Cfull plan file path>\n# run: \u003CISO-8601 start timestamp>\n```\n\nThe `run:` line names the run that owns the workspace, and Phase 5c-bis checks it before deleting anything -- the plan path alone cannot separate this run from another session working the same plan.\n\nApply these rules, which restate `skills\u002Fwork\u002Forchestrator.md` Section 0:\n\n| State | Action |\n|-------|--------|\n| No file, or a file whose first two lines are not a `# work ledger -- plan:` line followed by a `# run:` line | Fresh run. Create the directory if needed, overwrite the file with the identity header. Do not suffix -- a directory with no identity claims no plan. |\n| First line names this plan file, and every completion line's task number and title match the plan | Resumable. Do not re-dispatch any task carrying a `Task \u003CN> [\u003Ctask title>]: complete (branch \u003Cbranch>, commits \u003Cbase7>..\u003Chead7>)` line. Merge a completed task's branch unless a `Task \u003CN>: merged` line also exists for it -- a `complete` line records that the work was done, not that it landed. A complete line reading `commits none` has no branch to merge. Resume dispatch at the first task with no matching `complete` line. Overwrite the `# run:` line with this run's start timestamp and read it back -- the workspace now belongs to this run. |\n\nFor a ledger naming a different plan file, or a completion line whose title no longer matches the plan, `skills\u002Fwork\u002Forchestrator.md` Section 0 \"Resume rules\" is authoritative -- follow it there and print the line it requires.\n\n#### Announce\n\nWhen tasks are skipped:\n> Resuming the previous run: skipping Task 1, Task 2 (already complete).\n\nWhen the ledger is stale:\n> Plan changed since the last run -- starting fresh.\n\nWhen a suffixed workspace is used:\n> Another plan's run already uses that folder name -- using .claude\u002Fwork\u002F\u003Cname>-2\u002F instead.\n\n### Phase 3: Build\n\nBreak the plan into TodoWrite tasks (specific, dependency-ordered, with testing tasks included). For each task: mark `in_progress`, read referenced files, match existing patterns, then follow the cycle in `skills\u002Ftdd\u002FSKILL.md`: write the test the task's test step names, run the resolved test command, and print its `red:` line (this run's exit code, the new test's name, its first error line); implement; run the resolved command again and print its pass line. Fix failures before moving on, within three runs after the implementation exists -- the red run is not one of them. A task still failing after the third run is a blocker under the rule below; in auto mode it is accepted instead: print `Task {N}: still failing after 3 attempts -- accepted`, list it in the 5d summary, and continue to the next task. The budget is never extended on its own. When the command resolved to `none`, print `skipped: test command none (\u003Creason>)` once for the run and skip the red step on every task; when a task produces no testable behavior, print `skipped: no testable behavior (\u003Cwhat the change is>)` for that task. Then mark `completed` and update plan checkboxes if present.\n\nEvery task on this path is subject to the plan's `## Global Constraints` when the plan carries one: a task that cannot be completed without violating a constraint is a blocker, handled by the Blockers rule below.\n\n**Commits:** After each logical unit, commit if the evidence line is a pass and the change is meaningful (`git add \u003Cspecific files>` -- never `git add .`). Wait if tests fail or the message would say \"WIP\". Heuristic: \"Can I write a message describing a complete, valuable change?\" Write each message by the rules in the `# Commit Message Generation` section of `skills\u002Fcommit\u002FSKILL.md`, read once per run at the first commit.\n\n**Blockers:** Stop immediately. Note in TodoWrite. Ask the user. Do not proceed until resolved.\n\n### Phase 4: Quality Check\n\nBefore shipping, verify the work meets standards.\n\n#### 4a -- Core checks (always run)\n\n1. **Tests pass.** Run the test command resolved in Phase 2.5 and quote the evidence line from this run -- exit code and the runner's summary line, per the `## Evidence Rule` section of `skills\u002Fverification\u002FSKILL.md`. A claim without that line is not a pass. When the command resolved to `none`, this item is a WARNING, not a block: print `Tests: skipped -- \u003Creason>` here and again in the 5d summary.\n2. **Linting passes.** Run the project's lint command if one exists.\n3. **All TodoWrite tasks marked completed.** No tasks left in progress.\n4. **Code follows existing patterns.** No new conventions introduced unless the plan explicitly called for them.\n5. **No uncommitted changes** that belong to this work.\n6. **Plan checkboxes updated** (if applicable).\n7. **Post-merge test suite passes** (when orchestration was used), using the same resolved command and quoting its evidence line; all worktree branches merged with no unresolved conflicts.\n\n**These are BLOCKING -- fix all before proceeding to Phase 5.**\n\n#### 4b -- System-wide impact check\n\nFor non-trivial changes, pause and consider:\n\n| Question | Action |\n|----------|--------|\n| What else fires when this runs? (callbacks, middleware, observers, hooks) | Trace two levels out from your change. Read the actual code. |\n| Do tests exercise the real chain? | If every dependency is mocked, add at least one integration test using real objects. |\n| Can failure leave orphaned state? | If state is persisted before an external call, test the failure path. |\n| What other interfaces expose this? | Grep for the method\u002Fbehavior in related classes. Add parity if needed. |\n\n**Skip this check for:** leaf-node changes with no callbacks, no state persistence, no parallel interfaces. Purely additive changes (new helper, new partial) need only a quick scan.\n\n**WARNING (review but do not block):** linting warnings present; no integration tests for changes touching callbacks or middleware; plan acceptance criteria not explicitly verified (NOTE: for review-fix plans, acceptance criteria are BLOCKING per Phase 4c step 6).\n\n#### 4c -- Review finding verification (review-fix plans only)\n\nIf Phase 1 identified this as a review-fix plan and the review report was successfully loaded:\n\n1. **Parse findings.** Extract all non-filtered findings from the review report, grouped by severity (Critical, High, Medium, Low). Skip the `## Filtered Findings` section entirely.\n\n2. **Map findings to plan steps.** Match each finding to a plan task by file path or title\u002Fdescription. Findings with no matching plan step are marked \"Not in scope.\" When the review report carries any `Verification:` line, a finding without `Verification: confirmed` matches a plan task only by its ID or title, never by file path alone: the review could not show it holds against the code, so a change to its file does not address it. A plan that names such a finding keeps it in scope, and an unfixed Critical among them still blocks in item 5.\n\n3. **Check addressed status.** For each in-scope finding, verify the referenced file was modified (`git diff`) and the corresponding TodoWrite task is `completed`. Both conditions met = **Addressed**; otherwise = **Not addressed**.\n\n4. **Cross-reference check.** Flag when a file modified to fix finding A is also referenced by finding B (potential regression area). Present these as notes, not blockers.\n\n5. **Present verification summary** as a table with columns: ID, Severity, Finding, Status, Notes. Then apply gates:\n   - **BLOCKING**: Any Critical finding marked \"Not addressed.\" Use `AskUserQuestion`:\n     > Critical review finding not addressed: {finding title}\n     > Original finding: {finding text}\n     Buttons: `[\"I've verified this is fixed\", \"Fix it now\", \"Skip -- not applicable\"]`\n   - **WARNING**: Any High\u002FMedium\u002FLow finding marked \"Not addressed.\" List in summary but do not block.\n   - **INFO**: Findings marked \"Not in scope.\" Listed for awareness only.\n\n6. **Acceptance criteria check.** If the plan has an `Acceptance Criteria` section, mark each criterion as **Met** or **Not met**. If any are **Not met**, use `AskUserQuestion`:\n   > Acceptance criterion not met: {criterion text}\n   Buttons: `[\"Fix it now\", \"Skip -- criterion is outdated\", \"Mark as met (I've verified manually)\"]`\n   All criteria must be **Met** to proceed to Phase 5.\n\n7. **Convergence verdict.** If all in-scope findings are **Addressed** AND all acceptance criteria are **Met**, the review-fix cycle is **COMPLETE** -- proceed to Phase 5, do NOT trigger another review. If `review_iteration >= 2`, the cycle is **COMPLETE** regardless of remaining Low\u002FMedium warnings; only unaddressed Critical findings can block. Print:\n   ```\n   Review-Fix Cycle Status: Iteration {review_iteration} | Findings {addressed}\u002F{total_in_scope} | Criteria {met}\u002F{total_criteria} | COMPLETE\n   ```\n\n\u003C!-- SYNC: This verification parses the report format defined in skills\u002Freview\u002FSKILL.md, section `### Synthesized report structure`. If the report structure changes, update the parsing logic here. It reads each finding's severity, file path, title, and text, plus each finding's `Verification:` line for the item 2 scope rule only, and nothing else: the What's Working Well, Recommended Fix Order and Senior Assessment sections, each finding's `Disposition:` line, the `Dispositions:` line in Review Context, the `**Deferred:**` line, and the `Fix corrected:`, `Fix flagged:` and `No verified fix` notes are invisible to it. A Critical marked Not addressed blocks in item 5 whatever its disposition. -->\n\n#### 4d -- Optional: Agent-assisted review\n\n**SKIP this phase entirely if this is a review-fix plan.** The Phase 4c verification is the quality gate for review-fix work. Dispatching review agents on review-fix changes creates infinite loops -- the agents will always find new issues that weren't in the original scope.\n\nFor **non-review-fix plans** with large, risky, or security-sensitive changes, consider dispatching review agents. Discover available review agents by scanning `agents\u002Freview\u002F*.md` and dispatch them in parallel -- one `Agent` call per agent, all in a single response, the way `skills\u002Freview\u002FSKILL.md` Step 2 does it. The scan skips `finding-verifier`, which needs a finding and a reviewed commit that this phase never supplies.\n\n**Do not use review agents by default.** Tests + linting + pattern-following is sufficient for most work. Reserve agent reviews for:\n- Large refactors (10+ files) that are NOT review-fix plans\n- Security-sensitive changes (auth, permissions, data access)\n- Complex business logic or algorithms\n\n### Phase 5: Ship\n\n**If git is NOT available (Phase 0 detected `NO_GIT`):** Skip commit and PR steps. Summarize what was completed and remaining follow-ups, then stop.\n\n**CRITICAL: Every git action in this phase requires explicit user confirmation via `AskUserQuestion`. NEVER commit, push, or create a PR without asking first.** Auto mode is the one exception, for the commit alone: the caller's approval stands in for the 5a prompt. Push and PR are never automatic.\n\n**NO ATTRIBUTION: Do not add `Co-Authored-By`, `Generated with Claude`, `Built with AI`, or any similar attribution lines to commit messages or PR descriptions. Only add attribution if the user explicitly requests it.**\n\n#### 5a -- Final commit\n\nIf there are uncommitted changes after Phase 4, stage the relevant files (specific files only -- never `git add .`) then invoke the `commit` skill. If the user cancels, do not re-ask or proceed to 5b. In auto mode, commit directly instead of delegating to the prompt, writing the message by the same `# Commit Message Generation` section of `skills\u002Fcommit\u002FSKILL.md` that Phase 3 commits under.\n\n#### 5b -- Create PR\n\nAfter committing (or if all commits were already made during Phase 3), print the review command as text:\n> Review before merging: \u002Fquiver:review --base \u003Cdefault branch> --run-checks\nAdd `--plan \u003Cplan path>` to that line when the run had a plan file; `\u003Cdefault branch>` is `main` or `master`, whichever the repository has. `--run-checks` lets the review run the project's tests to check its findings, which a branch review does only on request because the branch could be someone else's; this one is the run's own work. The run starts no review: a review reads the finished branch and is the user's step before merging, so the command is printed as text and is never a button here. This is the whole review story for `\u002Fwork`: the run itself dispatches no review agents, because `\u002Fquiver:review` on the finished branch sees every task's change together with the synthesis filters a per-task pass would not have.\n\nThen use `AskUserQuestion`:\n> All work is committed on `{branch_name}`. What would you like to do next?\nButtons: `[\"Create a pull request\", \"Done -- I'll review or handle the rest\"]`\n\n- **Create a pull request** -- invoke the `create-pr` skill.\n- **Done** -- stop here. Move to 5c.\n\nIn auto mode, skip the question and act as **Done**: the review line is already printed; print `\u002Fcreate-pr` beside it, carry both as the last line of the 5d summary, and invoke neither.\n\n`\u002Fcreate-pr` owns the pull request for the rest of the session, in both modes. When the user asks for one after the run -- 'push and open a PR', 'create the PR' -- invoke the `create-pr` skill; never run `gh pr create` from this skill or from the conversation it leaves behind. A description drafted during the run (a rationale document the user asked to put in the PR, a summary of the branch) is input to that skill, which reads it as something the user said in this conversation, not a substitute for it. Only an explicit instruction to skip the skill -- 'open it with gh directly', 'don't use create-pr' -- overrides this.\n\n#### 5c -- Update plan status\n\nIf the work document has YAML frontmatter with a `status` field, update it:\n```\nstatus: active  -->  status: completed\n```\n\n#### 5c-bis -- Clean up the orchestration workspace\n\nRuns only when orchestration was used, every task reached DONE, and Phase 4a check 7 passed. A run that ends blocked, failed, or cancelled skips this step entirely -- the surviving directory is what makes the retry cheap, and deleting it throws away the resume.\n\n1. Resolve `\u003Cworkspace-dir>` to the directory this run actually used -- the suffixed one (`-2`, `-3`) if the orchestrator's suffix-retry rule fired, otherwise `.claude\u002Fwork\u002F\u003Cplan-basename>\u002F`. Confirm `\u003Cworkspace-dir>` exists, that its `progress.md` first line names this plan file, and that its `# run:` line is this run's start timestamp. If any check fails, delete nothing and say so -- the directory belongs to a different run, and a `run:` line naming another run means a concurrent session owns it. Never re-derive the path from `\u003Cplan-basename>` after this step.\n2. Name `\u003Cworkspace-dir>` and its file count, then gate the delete on `AskUserQuestion`:\n   > Orchestration finished and the work is committed. Delete the run workspace at `\u003Cworkspace-dir>` ({N} files)?\n   Buttons: `[\"Delete it\", \"Keep it\"]`\n   On \"Keep it\", print one line saying it was kept and continue to 5d. In auto mode, keep it without asking and print the same line.\n3. On \"Delete it\", remove `\u003Cworkspace-dir>` -- the exact path confirmed in step 1 and shown in step 2, no other -- then re-list `.claude\u002Fwork\u002F` to confirm it is gone, and name what was deleted in the 5d summary.\n\n#### 5d -- Notify user\n\nSummarize:\n- What was completed\n- Link to the PR (if one was created)\n- Any follow-up work needed or remaining tasks, including every task accepted after three failed attempts and every discovered edit the group announcements printed (file and reason), so a change outside the plan's file lists is visible before the PR\n- The test-first tally, one line: `TDD: \u003Cn> red-verified, \u003Cm> skipped (\u003Cdistinct reasons>)`. The sequential path counts the lines Phase 3 printed; the orchestration path counts the `TDD` lines the group-completion announcements printed. Include this line whenever the run implemented anything.\n- The Phase 2.5 lines, verbatim: `Strategy:`, `Reason:`, and `Verification:` as resolved. They are the only record of which path ran and whether the test command came from the rule or a guess, and a prose restatement (\"no test script\") drops the source. Include them whenever the run reached Phase 2.5.\n\nThis summary ends `\u002Fwork`. It does not end the turn when another skill loaded `\u002Fwork` through the Skill tool: check whether a skill invoked `\u002Fwork` and follow its continuation step.\n\n---\n\n## Anti-Patterns\n\n- **Don't** skip Phase 1 clarification -- ask now, not after building the wrong thing.\n- **Don't** ignore plan references -- the plan has file paths and pattern links for a reason.\n- **Don't** save all testing for the end -- test continuously or suffer compounding failures.\n- **Don't** use `git add .` -- stage specific files to avoid accidental inclusions.\n- **Don't** commit with \"WIP\" messages -- wait until a logical unit is complete.\n- **Don't** force through blockers -- stop, note the issue, ask the user.\n- **Don't** over-review simple changes -- save agent reviews for genuinely complex or risky work.\n- **Don't** leave TodoWrite tasks unfinished -- track progress or lose track of what is done.\n- **Don't** move on at 80% -- finish the feature before starting something new.\n- **Don't** commit directly to the default branch without explicit user permission.\n- **Don't** commit, push, or create a PR without asking the user first via `AskUserQuestion` -- every git action in Phase 5 requires explicit confirmation.\n- **Don't** add AI attribution to commits or PRs (`Co-Authored-By`, `Generated with Claude`, etc.) unless the user explicitly asks for it.\n- **Don't** trigger another review cycle after completing a review-fix plan -- Phase 4c verification is the terminal quality gate. Dispatching review agents on review-fix work creates infinite loops.\n- **Don't** spawn subagents for 1-2 task plans -- the overhead exceeds the benefit. Use sequential execution.\n- **Don't** skip dependency resolution -- check both explicit `blockedBy` and file overlap before dispatching parallel agents; stop dispatching dependent tasks when a dependency fails.\n- **Don't** attempt automatic merge conflict resolution -- report conflicts to the user and stop.\n- **Don't** answer a blocker, a plan contradiction, a Critical finding, or a merge conflict on the user's behalf in auto mode -- auto covers the routine gates named at the top and nothing else.\n- **Don't** extend the three-attempt fix budget on your own, in either mode. A loop that grants itself more attempts has no cap.\n","# Gather Context\n\n```\n!`git rev-parse --is-inside-work-tree 2>\u002Fdev\u002Fnull || echo \"NO_GIT\"`\n```\n\n```\n!`git branch --show-current 2>\u002Fdev\u002Fnull || echo \"NO_GIT\"`\n```\n\n```\n!`git log --oneline -5 2>\u002Fdev\u002Fnull || echo \"NO_COMMITS\"`\n```\n\n```\n!`git status --short 2>\u002Fdev\u002Fnull || echo \"NO_GIT\"`\n```\n\n---\n\n# Work\n\nExecute a work plan, specification, or task list systematically. The focus is on shipping complete features by understanding requirements quickly, following existing patterns, and maintaining quality throughout.\n\n**Announce:** \"Using the work skill to execute the plan.\"\n\n**Arguments.** `--auto` anywhere in `$ARGUMENTS` sets **auto mode** for this run. Strip it before Phase 1 reads the rest as a path or description. Auto mode pre-answers the routine gates and nothing else: the branch question in Phase 2 (continue on the current branch), the third failed fix attempt in Phase 3 and in the orchestrator's FAILED handling (accepted and listed; an orchestrated failure's dependents stay paused), the final commit in 5a (committed without the prompt), the handoff in 5b (no review, no PR -- the commands are printed as text), and the workspace in 5c-bis (kept). It answers no question whose answer changes what gets built or deletes something: a plan contradiction (1e), a blocker (Phase 3), an unaddressed Critical finding or unmet acceptance criterion (4c), and a merge conflict stop and ask exactly as they do without the flag. `\u002Fship` passes it so a run that answered every question up front is not stopped for a routine one; a user can pass it directly for the same effect. Every `AskUserQuestion` site below names its auto-mode branch; a prompt added any other way stalls an auto run.\n\n**Before starting Phase 1**, use the Glob tool to gather plan context silently (do not show results to the user):\n1. `.claude\u002Fplans\u002F*.md` -- existing plans\n2. `**\u002Fplans\u002F*.md` (max depth 4) -- plans in other locations\n\nTreat empty Glob results as \"no plans found\". Proceed regardless.\n\n**Use when:** a plan file exists, the user provides a spec or task list, or the user says \"execute\", \"build this\", \"implement the plan\". Use `\u002Fquiver:plan` first if no plan exists yet. Use `\u002Fquiver:review` for code review.\n\n---\n\n## Workflow\n\n```\n+----------+     +---------+     +---------+     +---------+     +--------+\n| 1. LOAD  | --> | 2. SETUP| --> | 3. BUILD| --> | 4. CHECK| --> | 5. SHIP|\n| Read plan|     | Branch  |     | Execute |     | Quality |     | PR     |\n+----------+     +---------+     +---------+     +---------+     +--------+\n```\n\n### Phase 0: Git Availability\n\nIf any gather-context block above returned `NO_GIT`, this directory is not a git repository.\nPrint: `> No git repository detected -- skipping branch\u002Fcommit context.`\nProceed to Phase 1. Treat all git-sourced fields (branch, log, diff, status) as empty. Skip branch creation in Phase 2, commit steps in Phase 3, and git-dependent actions in Phase 5.\n\n### Phase 1: Load and Clarify\n\n#### 1a -- Load the plan\n\n**Case A: Path provided.** If `$ARGUMENTS` is a file path (ends in `.md` or contains `\u002F`):\n- Read that file as the work plan.\n- Print: `> Executing plan: {plan filename} ({step count} steps) on branch {branch}`\n- Proceed directly to Phase 1b. The user chose this plan explicitly -- no confirmation needed.\n\n**Case B: No arguments.** If `$ARGUMENTS` is empty, collect all `.md` files from every `plans\u002F` directory discovered by the Glob block above, plus the `.claude\u002Fplans\u002F` listing.\n\n- If exactly one plan exists across all directories, read it and use `AskUserQuestion`:\n  > Found one plan: `{relative path}` (in `{directory}`)\n  > **Goal:** {goal from plan}\n  > **Steps:** {step count}\n  Buttons: `[\"Execute this plan\", \"Other -- I'll provide a path or description\"]`\n- If multiple plans exist, present them via `AskUserQuestion` with full relative paths as buttons (most recent first), plus an `\"Other -- I'll provide a path or description\"` option. Show the directory in each label so the user can distinguish plans in different locations.\n- If the user picks \"Other\", ask for a path or task description.\n\nIf no plans found:\n> No plans found in any `plans\u002F` directory. Usage:\n> - `\u002Fwork \u003Cpath-to-plan.md>` -- execute a specific plan\n> - `\u002Fwork \u003Cplan-name>` -- search for a plan by name\n> - `\u002Fwork \u003Ctask description>` -- work on a task directly\n> - `\u002Fplan \u003Ctask>` -- create a plan first\n\n**Stop here.**\n\n**Case C: Name or description provided.** If `$ARGUMENTS` is not a file path and is not empty:\n\n1. **Discover plan files.** Read every `plans\u002F` directory discovered by the Glob block. Combine with `.claude\u002Fplans\u002F` listing.\n2. **Match by filename.** Compare `$ARGUMENTS` against each plan file's stem (without `.md` extension):\n   - **Exact match:** stem equals `$ARGUMENTS` (case-insensitive).\n   - **Partial match:** stem contains `$ARGUMENTS` as a substring (case-insensitive).\n3. **Rank and select.** Rank exact matches above partial matches.\n   - **1 match** -- Read the plan file, print `> Executing plan: {filename} from {directory}`, proceed to Phase 1b. No confirmation needed.\n   - **Multiple matches** -- Use `AskUserQuestion` to present candidates with full relative paths as button labels. Add a `\"None of these -- treat as task description\"` button.\n   - **0 matches** -- Treat `$ARGUMENTS` as an inline task description. Print `> No matching plan found for \"{$ARGUMENTS}\". Treating as task description.` Proceed to Phase 1b with the description as the work specification.\n\n#### 1b -- Review references\n\nIf the plan links to files, patterns, or prior research -- read those now. Understanding context before coding prevents rework.\n\nIf the plan carries a `## Global Constraints` section, read it as binding context for every task in this run, not as background prose -- it states what this work must not do.\n\n#### 1c -- Detect review-fix plan\n\nCheck if this is a plan created to address review findings:\n- **Primary**: Check plan YAML frontmatter for a `review_source` field (e.g., `review_source: .claude\u002Freports\u002Freview-2026-03-10_14-30-00.md`).\n- **Fallback**: Scan plan content for paths matching `.claude\u002Freports\u002Freview-*.md` or `review-*_*-*-*.md`.\n- If detected, read the review report file. If the file exists, note this as a **review-fix plan** and carry the parsed findings forward to Phase 4. If the file does not exist, warn: \"Review report not found at {path}. Proceeding without review-aware verification.\"\n- If no review reference detected, proceed normally.\n- **Iteration tracking**: Read the plan's `review_iteration` frontmatter field (default: `1` if absent). If `review_iteration >= 2`, this is the **final iteration** -- Phase 4c verification is the only quality gate, and Phase 4d agent review is skipped unconditionally.\n\n#### 1d -- Flag ambiguities\n\nIf the plan contains genuine contradictions (e.g., two steps that conflict, a referenced file that does not exist), note them. If the plan is clear, skip this step entirely -- do not invent ambiguities.\n\n#### 1e -- Proceed or clarify\n\n- **No plan found:** Ask the user what to work on.\n- **Plan has contradictions** flagged in 1d: Ask about those specific contradictions only.\n- **Plan is clear:** Move directly to Phase 2. Do NOT ask \"should I proceed?\", \"are you sure?\", \"shall I start?\", or any variation of confirmation. The user invoked `\u002Fwork` with a plan -- that IS the approval. Summarizing the plan back and asking to continue is wasted time.\n\n### Phase 2: Setup Environment\n\n**If git is NOT available (Phase 0 detected `NO_GIT`):** Skip this phase entirely -- proceed to Phase 2.5.\n\n**If the repository has no commit yet** (git is available and the `git log` gather block printed `NO_COMMITS`), no branch exists until something is committed. `git checkout -b` then only renames the commit-less branch, so the default branch never comes into existence, and a later `\u002Fquiver:review --base \u003Cdefault branch>` stops because that branch is not found. Settle it before the rules below, with an empty first commit: `git commit --allow-empty --only -m \"chore: initial commit\"`. `--only` with no paths leaves anything already staged out of it, so the commit holds no file; that is why this skill makes it without asking whether to commit.\n\n- On `main` or `master`: make the commit, print `> This repository had no commit yet, so I made an empty first commit on \u003Cbranch> for the new branch to start from.`, and create the branch as below.\n- On any other branch there is no main branch to compare against, and its name is the user's choice, so ask, in auto mode too. One `AskUserQuestion`, with `main` first unless `git config --get init.defaultBranch` prints `master`:\n  > This repository has no commit yet, and you are on `\u003Cbranch>`. A review compares this branch against a main branch, and there is none yet. Which one should I create?\n\n  - \"Create main\" -- An empty first commit becomes `main`, and the work continues on `\u003Cbranch>` from it.\n  - \"Create master\" -- An empty first commit becomes `master`, and the work continues on `\u003Cbranch>` from it.\n  - \"Create neither\" -- Nothing is committed now; a review will have no main branch to compare this branch against.\n\n  On \"Create main\" or \"Create master\", make the commit, then run `git branch \u003Cname>`, and continue on `\u003Cbranch>`. On \"Create neither\", make no commit.\n\n**If already on a feature branch** (not main\u002Fmaster), use `AskUserQuestion`:\n> You're on `{branch}`. Continue here or create a new branch?\nButtons: `[\"Continue on {branch}\", \"Create new branch\"]`\nIf continuing, move to Phase 2.5. In auto mode, continue on the current branch without asking.\n\n**If on the default branch**, create a new branch by default: `git checkout -b \u003Cmeaningful-name>` using a descriptive name (e.g., `feat\u002Fuser-auth`, `fix\u002Femail-validation`). Never commit to the default branch without explicit user confirmation.\n\n### Phase 2.5: Orchestration Decision\n\nParse the plan and count top-level tasks (one deliverable = one task, regardless of markup format). Announce before proceeding:\n```\nStrategy: {sequential | parallel orchestration} ({N} tasks found)\nReason: {why}\n```\n\n#### Resolve the verification command\n\nRead `skills\u002Fverification\u002FSKILL.md` and follow its `## Command Resolution` section. This runs for every plan, sequential and orchestrated alike, once per run, before the strategy split below, so both paths carry the value. Print one line under the strategy line:\n\n```\nVerification: test \u003Ccommand | none (\u003Creason>)>, build \u003Ccommand | none (\u003Creason>)> (source: docs | stack | none)\n```\n\nThe orchestrator copies the test value into every brief's `Test command:` line. `none` is printed as-is and carried forward -- it is never replaced by a guess.\n\nThen choose the path; the first line that matches wins:\n\n- **No commit yet** (the `git log` gather block printed `NO_COMMITS`, which it also prints outside a git repository, and Phase 2 made no empty first commit): Sequential, whatever the task count. The Reason line says the repository has no commit yet, or that there is no git repository when Phase 0 detected `NO_GIT`. The working branch does not resolve until a first commit exists, and outside a repository there is no branch at all, so every worktree subagent's first step -- switching to that branch -- would come back BLOCKED. That leaves a run outside git and one where the user chose to create no main branch; every other run got its first commit in Phase 2 and takes the rules below.\n- **1-2 tasks:** Sequential. Proceed to Phase 3 with the verification command resolved above. The ledger below is orchestration-path-only; the sequential path keeps TodoWrite unchanged and writes nothing to disk.\n- **3+ tasks:** Parallel orchestration. Resolve the workspace and check for a ledger (below), then follow `skills\u002Fwork\u002Forchestrator.md`. Skip Phase 3 entirely -- orchestration replaces it.\n\n#### Resolve the workspace\n\n`\u003Cplan-basename>` is the loaded plan file's name without `.md`. The workspace is `.claude\u002Fwork\u002F\u003Cplan-basename>\u002F` and the ledger is `.claude\u002Fwork\u002F\u003Cplan-basename>\u002Fprogress.md`.\n\nAny path that reaches Phase 2.5 without a plan file has no plan basename -- Phase 1 Case B's \"Other -- I'll provide a path or description\" and Case C with 0 matches both do. Derive one instead of proceeding without it: slugify the task description to at most 40 characters and use `.claude\u002Fwork\u002Fadhoc-\u003Cslug>\u002F`. The workspace and ledger are otherwise identical, and the identity line names the task description in place of a plan file path. The orchestrator requires a workspace for every run -- every dispatch step writes to one. Name the derived workspace in the strategy line.\n\n#### Check for a ledger\n\nRead `.claude\u002Fwork\u002F\u003Cplan-basename>\u002Fprogress.md`. Its first two lines are the identity header:\n\n```\n# work ledger -- plan: \u003Cfull plan file path>\n# run: \u003CISO-8601 start timestamp>\n```\n\nThe `run:` line names the run that owns the workspace, and Phase 5c-bis checks it before deleting anything -- the plan path alone cannot separate this run from another session working the same plan.\n\nApply these rules, which restate `skills\u002Fwork\u002Forchestrator.md` Section 0:\n\n| State | Action |\n|-------|--------|\n| No file, or a file whose first two lines are not a `# work ledger -- plan:` line followed by a `# run:` line | Fresh run. Create the directory if needed, overwrite the file with the identity header. Do not suffix -- a directory with no identity claims no plan. |\n| First line names this plan file, and every completion line's task number and title match the plan | Resumable. Do not re-dispatch any task carrying a `Task \u003CN> [\u003Ctask title>]: complete (branch \u003Cbranch>, commits \u003Cbase7>..\u003Chead7>)` line. Merge a completed task's branch unless a `Task \u003CN>: merged` line also exists for it -- a `complete` line records that the work was done, not that it landed. A complete line reading `commits none` has no branch to merge. Resume dispatch at the first task with no matching `complete` line. Overwrite the `# run:` line with this run's start timestamp and read it back -- the workspace now belongs to this run. |\n\nFor a ledger naming a different plan file, or a completion line whose title no longer matches the plan, `skills\u002Fwork\u002Forchestrator.md` Section 0 \"Resume rules\" is authoritative -- follow it there and print the line it requires.\n\n#### Announce\n\nWhen tasks are skipped:\n> Resuming the previous run: skipping Task 1, Task 2 (already complete).\n\nWhen the ledger is stale:\n> Plan changed since the last run -- starting fresh.\n\nWhen a suffixed workspace is used:\n> Another plan's run already uses that folder name -- using .claude\u002Fwork\u002F\u003Cname>-2\u002F instead.\n\n### Phase 3: Build\n\nBreak the plan into TodoWrite tasks (specific, dependency-ordered, with testing tasks included). For each task: mark `in_progress`, read referenced files, match existing patterns, then follow the cycle in `skills\u002Ftdd\u002FSKILL.md`: write the test the task's test step names, run the resolved test command, and print its `red:` line (this run's exit code, the new test's name, its first error line); implement; run the resolved command again and print its pass line. Fix failures before moving on, within three runs after the implementation exists -- the red run is not one of them. A task still failing after the third run is a blocker under the rule below; in auto mode it is accepted instead: print `Task {N}: still failing after 3 attempts -- accepted`, list it in the 5d summary, and continue to the next task. The budget is never extended on its own. When the command resolved to `none`, print `skipped: test command none (\u003Creason>)` once for the run and skip the red step on every task; when a task produces no testable behavior, print `skipped: no testable behavior (\u003Cwhat the change is>)` for that task. Then mark `completed` and update plan checkboxes if present.\n\nEvery task on this path is subject to the plan's `## Global Constraints` when the plan carries one: a task that cannot be completed without violating a constraint is a blocker, handled by the Blockers rule below.\n\n**Commits:** After each logical unit, commit if the evidence line is a pass and the change is meaningful (`git add \u003Cspecific files>` -- never `git add .`). Wait if tests fail or the message would say \"WIP\". Heuristic: \"Can I write a message describing a complete, valuable change?\" Write each message by the rules in the `# Commit Message Generation` section of `skills\u002Fcommit\u002FSKILL.md`, read once per run at the first commit.\n\n**Blockers:** Stop immediately. Note in TodoWrite. Ask the user. Do not proceed until resolved.\n\n### Phase 4: Quality Check\n\nBefore shipping, verify the work meets standards.\n\n#### 4a -- Core checks (always run)\n\n1. **Tests pass.** Run the test command resolved in Phase 2.5 and quote the evidence line from this run -- exit code and the runner's summary line, per the `## Evidence Rule` section of `skills\u002Fverification\u002FSKILL.md`. A claim without that line is not a pass. When the command resolved to `none`, this item is a WARNING, not a block: print `Tests: skipped -- \u003Creason>` here and again in the 5d summary.\n2. **Linting passes.** Run the project's lint command if one exists.\n3. **All TodoWrite tasks marked completed.** No tasks left in progress.\n4. **Code follows existing patterns.** No new conventions introduced unless the plan explicitly called for them.\n5. **No uncommitted changes** that belong to this work.\n6. **Plan checkboxes updated** (if applicable).\n7. **Post-merge test suite passes** (when orchestration was used), using the same resolved command and quoting its evidence line; all worktree branches merged with no unresolved conflicts.\n\n**These are BLOCKING -- fix all before proceeding to Phase 5.**\n\n#### 4b -- System-wide impact check\n\nFor non-trivial changes, pause and consider:\n\n| Question | Action |\n|----------|--------|\n| What else fires when this runs? (callbacks, middleware, observers, hooks) | Trace two levels out from your change. Read the actual code. |\n| Do tests exercise the real chain? | If every dependency is mocked, add at least one integration test using real objects. |\n| Can failure leave orphaned state? | If state is persisted before an external call, test the failure path. |\n| What other interfaces expose this? | Grep for the method\u002Fbehavior in related classes. Add parity if needed. |\n\n**Skip this check for:** leaf-node changes with no callbacks, no state persistence, no parallel interfaces. Purely additive changes (new helper, new partial) need only a quick scan.\n\n**WARNING (review but do not block):** linting warnings present; no integration tests for changes touching callbacks or middleware; plan acceptance criteria not explicitly verified (NOTE: for review-fix plans, acceptance criteria are BLOCKING per Phase 4c step 6).\n\n#### 4c -- Review finding verification (review-fix plans only)\n\nIf Phase 1 identified this as a review-fix plan and the review report was successfully loaded:\n\n1. **Parse findings.** Extract all non-filtered findings from the review report, grouped by severity (Critical, High, Medium, Low). Skip the `## Filtered Findings` section entirely.\n\n2. **Map findings to plan steps.** Match each finding to a plan task by file path or title\u002Fdescription. Findings with no matching plan step are marked \"Not in scope.\" When the review report carries any `Verification:` line, a finding without `Verification: confirmed` matches a plan task only by its ID or title, never by file path alone: the review could not show it holds against the code, so a change to its file does not address it. A plan that names such a finding keeps it in scope, and an unfixed Critical among them still blocks in item 5.\n\n3. **Check addressed status.** For each in-scope finding, verify the referenced file was modified (`git diff`) and the corresponding TodoWrite task is `completed`. Both conditions met = **Addressed**; otherwise = **Not addressed**.\n\n4. **Cross-reference check.** Flag when a file modified to fix finding A is also referenced by finding B (potential regression area). Present these as notes, not blockers.\n\n5. **Present verification summary** as a table with columns: ID, Severity, Finding, Status, Notes. Then apply gates:\n   - **BLOCKING**: Any Critical finding marked \"Not addressed.\" Use `AskUserQuestion`:\n     > Critical review finding not addressed: {finding title}\n     > Original finding: {finding text}\n     Buttons: `[\"I've verified this is fixed\", \"Fix it now\", \"Skip -- not applicable\"]`\n   - **WARNING**: Any High\u002FMedium\u002FLow finding marked \"Not addressed.\" List in summary but do not block.\n   - **INFO**: Findings marked \"Not in scope.\" Listed for awareness only.\n\n6. **Acceptance criteria check.** If the plan has an `Acceptance Criteria` section, mark each criterion as **Met** or **Not met**. If any are **Not met**, use `AskUserQuestion`:\n   > Acceptance criterion not met: {criterion text}\n   Buttons: `[\"Fix it now\", \"Skip -- criterion is outdated\", \"Mark as met (I've verified manually)\"]`\n   All criteria must be **Met** to proceed to Phase 5.\n\n7. **Convergence verdict.** If all in-scope findings are **Addressed** AND all acceptance criteria are **Met**, the review-fix cycle is **COMPLETE** -- proceed to Phase 5, do NOT trigger another review. If `review_iteration >= 2`, the cycle is **COMPLETE** regardless of remaining Low\u002FMedium warnings; only unaddressed Critical findings can block. Print:\n   ```\n   Review-Fix Cycle Status: Iteration {review_iteration} | Findings {addressed}\u002F{total_in_scope} | Criteria {met}\u002F{total_criteria} | COMPLETE\n   ```\n\n\u003C!-- SYNC: This verification parses the report format defined in skills\u002Freview\u002FSKILL.md, section `### Synthesized report structure`. If the report structure changes, update the parsing logic here. It reads each finding's severity, file path, title, and text, plus each finding's `Verification:` line for the item 2 scope rule only, and nothing else: the What's Working Well, Recommended Fix Order and Senior Assessment sections, each finding's `Disposition:` line, the `Dispositions:` line in Review Context, the `**Deferred:**` line, and the `Fix corrected:`, `Fix flagged:` and `No verified fix` notes are invisible to it. A Critical marked Not addressed blocks in item 5 whatever its disposition. -->\n\n#### 4d -- Optional: Agent-assisted review\n\n**SKIP this phase entirely if this is a review-fix plan.** The Phase 4c verification is the quality gate for review-fix work. Dispatching review agents on review-fix changes creates infinite loops -- the agents will always find new issues that weren't in the original scope.\n\nFor **non-review-fix plans** with large, risky, or security-sensitive changes, consider dispatching review agents. Discover available review agents by scanning `agents\u002Freview\u002F*.md` and dispatch them in parallel -- one `Agent` call per agent, all in a single response, the way `skills\u002Freview\u002FSKILL.md` Step 2 does it. The scan skips `finding-verifier`, which needs a finding and a reviewed commit that this phase never supplies.\n\n**Do not use review agents by default.** Tests + linting + pattern-following is sufficient for most work. Reserve agent reviews for:\n- Large refactors (10+ files) that are NOT review-fix plans\n- Security-sensitive changes (auth, permissions, data access)\n- Complex business logic or algorithms\n\n### Phase 5: Ship\n\n**If git is NOT available (Phase 0 detected `NO_GIT`):** Skip commit and PR steps. Summarize what was completed and remaining follow-ups, then stop.\n\n**CRITICAL: Every git action in this phase requires explicit user confirmation via `AskUserQuestion`. NEVER commit, push, or create a PR without asking first.** Auto mode is the one exception, for the commit alone: the caller's approval stands in for the 5a prompt. Push and PR are never automatic.\n\n**NO ATTRIBUTION: Do not add `Co-Authored-By`, `Generated with Claude`, `Built with AI`, or any similar attribution lines to commit messages or PR descriptions. Only add attribution if the user explicitly requests it.**\n\n#### 5a -- Final commit\n\nIf there are uncommitted changes after Phase 4, stage the relevant files (specific files only -- never `git add .`) then invoke the `commit` skill. If the user cancels, do not re-ask or proceed to 5b. In auto mode, commit directly instead of delegating to the prompt, writing the message by the same `# Commit Message Generation` section of `skills\u002Fcommit\u002FSKILL.md` that Phase 3 commits under.\n\n#### 5b -- Create PR\n\nAfter committing (or if all commits were already made during Phase 3), print the review command as text:\n> Review before merging: \u002Fquiver:review --base \u003Cdefault branch> --run-checks\nAdd `--plan \u003Cplan path>` to that line when the run had a plan file; `\u003Cdefault branch>` is `main` or `master`, whichever the repository has. `--run-checks` lets the review run the project's tests to check its findings, which a branch review does only on request because the branch could be someone else's; this one is the run's own work. The run starts no review: a review reads the finished branch and is the user's step before merging, so the command is printed as text and is never a button here. This is the whole review story for `\u002Fwork`: the run itself dispatches no review agents, because `\u002Fquiver:review` on the finished branch sees every task's change together with the synthesis filters a per-task pass would not have.\n\nThen use `AskUserQuestion`:\n> All work is committed on `{branch_name}`. What would you like to do next?\nButtons: `[\"Create a pull request\", \"Done -- I'll review or handle the rest\"]`\n\n- **Create a pull request** -- invoke the `create-pr` skill.\n- **Done** -- stop here. Move to 5c.\n\nIn auto mode, skip the question and act as **Done**: the review line is already printed; print `\u002Fcreate-pr` beside it, carry both as the last line of the 5d summary, and invoke neither.\n\n`\u002Fcreate-pr` owns the pull request for the rest of the session, in both modes. When the user asks for one after the run -- 'push and open a PR', 'create the PR' -- invoke the `create-pr` skill; never run `gh pr create` from this skill or from the conversation it leaves behind. A description drafted during the run (a rationale document the user asked to put in the PR, a summary of the branch) is input to that skill, which reads it as something the user said in this conversation, not a substitute for it. Only an explicit instruction to skip the skill -- 'open it with gh directly', 'don't use create-pr' -- overrides this.\n\n#### 5c -- Update plan status\n\nIf the work document has YAML frontmatter with a `status` field, update it:\n```\nstatus: active  -->  status: completed\n```\n\n#### 5c-bis -- Clean up the orchestration workspace\n\nRuns only when orchestration was used, every task reached DONE, and Phase 4a check 7 passed. A run that ends blocked, failed, or cancelled skips this step entirely -- the surviving directory is what makes the retry cheap, and deleting it throws away the resume.\n\n1. Resolve `\u003Cworkspace-dir>` to the directory this run actually used -- the suffixed one (`-2`, `-3`) if the orchestrator's suffix-retry rule fired, otherwise `.claude\u002Fwork\u002F\u003Cplan-basename>\u002F`. Confirm `\u003Cworkspace-dir>` exists, that its `progress.md` first line names this plan file, and that its `# run:` line is this run's start timestamp. If any check fails, delete nothing and say so -- the directory belongs to a different run, and a `run:` line naming another run means a concurrent session owns it. Never re-derive the path from `\u003Cplan-basename>` after this step.\n2. Name `\u003Cworkspace-dir>` and its file count, then gate the delete on `AskUserQuestion`:\n   > Orchestration finished and the work is committed. Delete the run workspace at `\u003Cworkspace-dir>` ({N} files)?\n   Buttons: `[\"Delete it\", \"Keep it\"]`\n   On \"Keep it\", print one line saying it was kept and continue to 5d. In auto mode, keep it without asking and print the same line.\n3. On \"Delete it\", remove `\u003Cworkspace-dir>` -- the exact path confirmed in step 1 and shown in step 2, no other -- then re-list `.claude\u002Fwork\u002F` to confirm it is gone, and name what was deleted in the 5d summary.\n\n#### 5d -- Notify user\n\nSummarize:\n- What was completed\n- Link to the PR (if one was created)\n- Any follow-up work needed or remaining tasks, including every task accepted after three failed attempts and every discovered edit the group announcements printed (file and reason), so a change outside the plan's file lists is visible before the PR\n- The test-first tally, one line: `TDD: \u003Cn> red-verified, \u003Cm> skipped (\u003Cdistinct reasons>)`. The sequential path counts the lines Phase 3 printed; the orchestration path counts the `TDD` lines the group-completion announcements printed. Include this line whenever the run implemented anything.\n- The Phase 2.5 lines, verbatim: `Strategy:`, `Reason:`, and `Verification:` as resolved. They are the only record of which path ran and whether the test command came from the rule or a guess, and a prose restatement (\"no test script\") drops the source. Include them whenever the run reached Phase 2.5.\n\nThis summary ends `\u002Fwork`. It does not end the turn when another skill loaded `\u002Fwork` through the Skill tool: check whether a skill invoked `\u002Fwork` and follow its continuation step.\n\n---\n\n## Anti-Patterns\n\n- **Don't** skip Phase 1 clarification -- ask now, not after building the wrong thing.\n- **Don't** ignore plan references -- the plan has file paths and pattern links for a reason.\n- **Don't** save all testing for the end -- test continuously or suffer compounding failures.\n- **Don't** use `git add .` -- stage specific files to avoid accidental inclusions.\n- **Don't** commit with \"WIP\" messages -- wait until a logical unit is complete.\n- **Don't** force through blockers -- stop, note the issue, ask the user.\n- **Don't** over-review simple changes -- save agent reviews for genuinely complex or risky work.\n- **Don't** leave TodoWrite tasks unfinished -- track progress or lose track of what is done.\n- **Don't** move on at 80% -- finish the feature before starting something new.\n- **Don't** commit directly to the default branch without explicit user permission.\n- **Don't** commit, push, or create a PR without asking the user first via `AskUserQuestion` -- every git action in Phase 5 requires explicit confirmation.\n- **Don't** add AI attribution to commits or PRs (`Co-Authored-By`, `Generated with Claude`, etc.) unless the user explicitly asks for it.\n- **Don't** trigger another review cycle after completing a review-fix plan -- Phase 4c verification is the terminal quality gate. Dispatching review agents on review-fix work creates infinite loops.\n- **Don't** spawn subagents for 1-2 task plans -- the overhead exceeds the benefit. Use sequential execution.\n- **Don't** skip dependency resolution -- check both explicit `blockedBy` and file overlap before dispatching parallel agents; stop dispatching dependent tasks when a dependency fails.\n- **Don't** attempt automatic merge conflict resolution -- report conflicts to the user and stop.\n- **Don't** answer a blocker, a plan contradiction, a Critical finding, or a merge conflict on the user's behalf in auto mode -- auto covers the routine gates named at the top and nothing else.\n- **Don't** extend the three-attempt fix budget on your own, in either mode. A loop that grants itself more attempts has no cap.\n",{"name":5,"description":7,"when-to-use":49,"argument-hint":50},"user wants to execute a saved plan or implement tasks step by step -- '\u002Fwork', '\u002Fwork --auto', 'start implementation', 'execute the plan', 'work through these tasks', 'run the plan without stopping to ask'","\u003Cplan file path or task description> [--auto]","main","2b63f856ce19c2a797929cacabd9e0fc936812fa","plugins\u002Fyagizdo\u002Fquiver\u002Fskills\u002Fwork",491,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins","https:\u002F\u002Fhol.org\u002Fplugins\u002Fbest-codex-plugins","https:\u002F\u002Fgithub.com\u002Fhashgraph-online","https:\u002F\u002Fhol.org","Open-source standards and developer tools for interoperable AI agents: discovery, identity, communication, payments, privacy and security.","2026-11-14 14:06:13",{"source":8,"skillName":5,"commands":62,"prompt":72,"zipAvailable":73},[63,66,69],{"tool":64,"command":65},"npx","npx skills add hashgraph-online\u002Fawesome-codex-plugins --skill work",{"tool":67,"command":68},"bunx","bunx skills add hashgraph-online\u002Fawesome-codex-plugins --skill work",{"tool":70,"command":71},"pnpm","pnpm dlx skills add hashgraph-online\u002Fawesome-codex-plugins --skill work","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fyagizdo\u002Fquiver\u002Fskills\u002Fwork\u002FSKILL.md",true,[75,81],{"path":76,"resourceType":77,"resourceTypeDesc":78,"sizeBytes":79,"sourceHtmlUrl":80},"orchestrator.md","REFERENCE","参考文档",32338,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fyagizdo\u002Fquiver\u002Fskills\u002Fwork\u002Forchestrator.md",{"path":82,"resourceType":77,"resourceTypeDesc":78,"sizeBytes":83,"sourceHtmlUrl":84},"TEST-PLAN.md",7419,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fyagizdo\u002Fquiver\u002Fskills\u002Fwork\u002FTEST-PLAN.md",[86,96,104,112,120,128],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":90,"licenseFileUrl":91,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":95},"01M7DV332YQMGD8WTRJ9DZQXTB","worktable-review-with-annotations","Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.","MIT","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-review-with-annotations\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-review-with-annotations\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":97,"name":98,"nameZh":6,"description":99,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":90,"licenseFileUrl":100,"sourceHtmlUrl":101,"categories":102,"keywords":103,"publishedAt":95},"01M7DV2YGHCZVBEB1AHYB6CGHE","worktable-find-and-synthesize-context","Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-find-and-synthesize-context\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-find-and-synthesize-context\u002FSKILL.md",[],[],{"skillNo":105,"name":106,"nameZh":6,"description":107,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":90,"licenseFileUrl":108,"sourceHtmlUrl":109,"categories":110,"keywords":111,"publishedAt":95},"01M7DV2TQN1H11ME2SAF9MXN7V","worktable-create-or-update-docs","Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-update-docs\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-update-docs\u002FSKILL.md",[],[],{"skillNo":113,"name":114,"nameZh":6,"description":115,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":90,"licenseFileUrl":116,"sourceHtmlUrl":117,"categories":118,"keywords":119,"publishedAt":95},"01M7DV2NE452NEA8TXATKSNAJH","worktable-create-or-manage-records","Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-records\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-records\u002FSKILL.md",[],[],{"skillNo":121,"name":122,"nameZh":6,"description":123,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":90,"licenseFileUrl":124,"sourceHtmlUrl":125,"categories":126,"keywords":127,"publishedAt":95},"01M7DV2HENRBR2KSCVM064WH0Y","worktable-create-or-manage-interactive-html","Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-interactive-html\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-interactive-html\u002FSKILL.md",[],[],{"skillNo":129,"name":130,"nameZh":6,"description":131,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":90,"licenseFileUrl":132,"sourceHtmlUrl":133,"categories":134,"keywords":135,"publishedAt":95},"01M7DV2CADR5R7RQFFYPNAGR9Y","worktable-collaborate-threads","Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-collaborate-threads\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-collaborate-threads\u002FSKILL.md",[],[]]