[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77ZX6VXJS156KD413M119AN":3,"skill-same:01M77ZX6VXJS156KD413M119AN":59},{"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":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":31,"updatedAt":32,"install":33,"resources":46},"01M77ZX6VXJS156KD413M119AN","cmux-sprint-orchestrator",null,"Use when asked to orchestrate a sprint, produce a sprint plan, run a sprint task, coordinate agents in cmux, or drive a multi-agent implementation, and when a goal spans multiple modules, several days, or more work than one agent turn can gate. Also use when worker panes must be dispatched, polled, or torn down inside cmux.","ArangoGutierrez\u002Fclaude-toolkit","ArangoGutierrez","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F15933089?v=4","01M77ZX10BNKRH5SRK782MGZQR","Carlos Eduardo Arango Gutierrez","01M77ZX1108AW5SJRFWGFKPKTA",4,"MIT","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fcmux-sprint-orchestrator\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: cmux-sprint-orchestrator\ndescription: Use when asked to orchestrate a sprint, produce a sprint plan, run a sprint task, coordinate agents in cmux, or drive a multi-agent implementation, and when a goal spans multiple modules, several days, or more work than one agent turn can gate. Also use when worker panes must be dispatched, polled, or torn down inside cmux.\n---\n\n# cmux Sprint Orchestrator\n\n## Overview\n\nDrives sprint-scale work from a single root surface by dispatching bounded tasks to\nheterogeneous agent panes in cmux, then converging their output behind one verification gate.\n\n**Core principle: the plan is the product of Phase 1, and code is not.** An orchestrator that\nedits files while it is still deciding what to build has no plan, only a diff. Splitting the\ndecision from the execution is what makes the work reviewable.\n\nMeasured baseline: three fresh agents given a three-item sprint under deadline pressure all\nbegan editing source immediately. None produced a plan artifact and none paused for approval.\nOne rewrote production source before naming a single invariant. That is the failure this skill\nexists to prevent.\n\n## The Iron Gate\n\n```\nPHASE 1 IS READ-ONLY. NO FILE IN THE TARGET REPO CHANGES BEFORE `APPROVED`.\n```\n\nIn Phase 1 you may read, search, list, run existing test suites, and write exactly one file:\n`SPRINT_PLAN.md`. Nothing else.\n\n**Forbidden in Phase 1, without exception:**\n\n- Editing, creating, or deleting any application, test, config, or build file.\n- Running a formatter, linter, or codemod that writes to the tree.\n- `git commit`, `git checkout -b`, `git stash`, or any index mutation.\n- Creating a cmux split, pane, or workspace.\n- Sending a single character to any worker CLI.\n- \"Scaffolding\", \"a stub so the DAG typechecks\", or \"one obvious fix while I read\".\n\nViolating the letter of this gate is violating the spirit of it. If you changed a file before\n`APPROVED`, revert it (`git checkout -- \u003Cpath>`, delete untracked files), say so plainly, and\nrestart Phase 1.\n\n### Rationalizations, and what is actually true\n\n| Rationalization | Reality |\n|---|---|\n| \"The deadline is tomorrow, planning is ceremony.\" | The baseline agents that skipped the plan produced three uncoordinated diffs that no reviewer could gate. The plan is the cheapest artifact in the sprint. |\n| \"The lead already signed off on the direction.\" | Sign-off on direction is not approval of a task graph. `APPROVED` gates the DAG, the invariants, and the routing, none of which existed when the lead spoke. |\n| \"It's one small fix, I'm already in the file.\" | Phase 1 has no small fixes. An unplanned edit is an untracked task with no owner, no verification, and no line in the report. |\n| \"I need a stub to validate the dependency order.\" | Validate order by reading imports and the test suite. A stub is code. |\n| \"I'll write the plan after, from what I learned building.\" | A plan written after the diff documents the diff. It cannot change a decision that is already spent. |\n| \"Spawning the workers now saves wall-clock, I'll plan while they run.\" | Workers dispatched without a DAG collide on shared files. Concurrency without a plan is a merge conflict with extra steps. |\n| \"The user obviously wants working code, not a document.\" | The user wants working code they can verify. Phase 3 is where code is judged; Phase 1 is where it is made judgeable. |\n| \"Reading is slow, I'll learn faster by changing something.\" | Then change it in a scratch clone outside the target repo, and cite it as evidence in the plan. |\n\n### Red flags: stop and restart Phase 1\n\n- `git status` in the target repo is not clean and you have not printed `APPROVED`.\n- You are composing a `cmux send` payload and `SPRINT_PLAN.md` does not exist.\n- You catch yourself writing \"while I'm here\" or \"quick\".\n- You are about to ask the user to approve a plan you have already half-built.\n\n## Prerequisites\n\nRun this first. cmux talks over a unix domain socket, which the agent sandbox denies, so\n**every `cmux` call in this skill runs with the sandbox disabled.** An\n`Operation not permitted (errno 1)` from `cmux` is the sandbox, not a broken socket.\n\n```bash\nfor b in cmux codex claude jq; do\n  command -v \"$b\" >\u002Fdev\u002Fnull 2>&1 || { echo \"MISSING: $b\"; exit 1; }\ndone\ncmux ping || { echo \"cmux socket unreachable\"; exit 1; }\nROOT_SURFACE=$(cmux identify | jq -r '.caller.surface_ref')\nWORKSPACE=$(cmux identify | jq -r '.caller.workspace_ref')\necho \"root=$ROOT_SURFACE workspace=$WORKSPACE\"\n```\n\n`cmux ping` prints `PONG`. Surface refs are 1-indexed: the root orchestrator is `surface:1`,\nand there is no `surface:0`. The socket defaults to `~\u002F.local\u002Fstate\u002Fcmux\u002Fcmux.sock` and may be\nuid-suffixed (`cmux-\u003Cuid>.sock`); override with `CMUX_SOCKET_PATH`. It is not `\u002Ftmp\u002Fcmux.sock`.\n\n## Phase 1: architectural formulation (read-only)\n\n### 1.1 Map the ground truth\n\nEvery count, path, and command that reaches `SPRINT_PLAN.md` comes from a command you ran in\nthis session and pasted. A remembered module list is not evidence.\n\n```bash\ngit -C \"$REPO\" status --short          # must be clean, and stay clean\ngit -C \"$REPO\" log --oneline -10\n\n# Scratch dir, self-ignoring so the census and logs never dirty git status.\nmkdir -p \"$REPO\u002F.sprint\" && printf '*\\n' > \"$REPO\u002F.sprint\u002F.gitignore\"\n\n# File census. `rg --files` alone SKIPS HIDDEN DIRECTORIES, so a repo that keeps\n# its content under .claude\u002F, .github\u002F or .cursor\u002F reports a fraction of itself.\ngit -C \"$REPO\" ls-files > \"$REPO\u002F.sprint\u002Fcensus.txt\"                   # tracked files\n# not a git repo? use: rg --files --hidden --glob '!.git' \"$REPO\"\nwc -l \u003C \"$REPO\u002F.sprint\u002Fcensus.txt\"\n\n# Language mix. The `grep '\\.'` guard matters: an extensionless file (LICENSE,\n# Makefile) would otherwise have its whole path counted as an extension.\ngrep '\\.' \"$REPO\u002F.sprint\u002Fcensus.txt\" | sed 's|.*\\.||' | sort | uniq -c | sort -rn | head\n\n# Dependency shape, over the census rather than a fresh blind walk.\nrg -n '^(import|from|require|use )' --hidden --glob '!.git' \"$REPO\" --stats | tail -5\nls \"$REPO\"\u002F{Makefile,justfile,package.json,pyproject.toml,go.mod} 2>\u002Fdev\u002Fnull\n```\n\nProve the census can see before you trust what it does not show. Pick a file you\nalready know exists, ideally a deep or hidden one, and confirm it is in the list:\n\n```bash\ngrep -c 'some\u002Ffile\u002Fyou\u002Fknow\u002Fexists' \"$REPO\u002F.sprint\u002Fcensus.txt\"   # must be >= 1\n```\n\nMeasured on a real repo, the difference is not marginal: `rg --files` reported 56\nfiles where `git ls-files` reported 383, and it listed exactly 0 of the 246 files\nunder `.claude\u002F`. A census that cannot see a directory reports no work there and\nno risk there, and the DAG you build on it inherits both blind spots.\n\nThen run the existing suite once, unchanged, and record the result. A green baseline is the\nonly thing that makes a Phase 3 regression legible.\n\n```bash\n( cd \"$REPO\" && make test ) > \"$REPO\u002F.sprint\u002Fbaseline.log\" 2>&1; echo \"baseline rc=$?\"\ntail -5 \"$REPO\u002F.sprint\u002Fbaseline.log\"\n```\n\nIf a repo has a code knowledge graph (`graphify-out\u002Fgraph.json`), orient with\n`graphify query \"\u003Cquestion>\"` before grepping.\n\nIdentify high-risk integration zones explicitly: files imported by three or more modules,\nanything in an auth or crypto path, and any file two planned tasks would both touch. The last\ncategory is the one that decides your DAG edges.\n\n### 1.2 Confirm the target repo is dispatch-ready\n\nA dispatched worker runs inside **your** hook stack, not a clean one. Hooks you wrote for\nyour own repositories will fire against the target repo, and when one blocks a worker the\norchestrator sees nothing but a stall: no error, no signal, just a timeout.\n\nObserved on the first live run of this skill. A worker created its own worktree, ran the\nbaseline suite green, wrote correct tests, and every write was then refused by a PreToolUse\nhook demanding a linked worktree. Three dispatches burned about five minutes each and the\nroot surface could not tell that from a slow worker. Adding the marker the hook asked for\nmade the same brief succeed unchanged.\n\nBefore dispatching anything, check what your own hooks require of a repo:\n\n```bash\nls \"$REPO\"\u002F{AGENTS.md,CLAUDE.md} 2>\u002Fdev\u002Fnull      # project instruction files the hooks read\ngrep -rl 'PreToolUse' ~\u002F.claude\u002Fsettings*.json 2>\u002Fdev\u002Fnull\n```\n\nThen satisfy the requirement in the target repo, or hand each worker a properly linked\nworktree. Either is fine; discovering which one your hooks want after three silent stalls is\nnot. A worker blocked by a policy hook has not failed the task, and re-briefing it will not\nhelp, so this belongs in Phase 1 rather than in the failure policy.\n\nClean up after a blocked worker: it may have created a git worktree of its own before being\nstopped. `git -C \"$REPO\" worktree list` shows them and they outlive the surface you closed.\n\n### 1.3 Decide whether to fan out at all\n\nDo this before writing the DAG, and be willing to answer no. Three independent agents given a\nthree-item sprint all reached the same conclusion unprompted: two of the items rewrote the same\nfile, so any two workers would have collided and one would have had to rebase onto the other.\nTheir words: \"coordination overhead would have exceeded the work.\"\n\nFan out only when every one of these holds:\n\n- Two or more tasks share no path in their `outputs` column.\n- Those tasks have no `depends_on` edge between them.\n- Each task is large enough that a worker's startup and your reconciliation cost less than\n  doing it inline.\n\nIf none of that holds, say so in section 3, run the sprint serially from the root surface, and\nskip Phase 2. A serial sprint executed against an approved plan is a success, not a fallback.\nThe most common seam that does parallelize is tests-after-interface: land the interface change\nfirst, freeze it, then fan out the test sweep across the settled API.\n\nAn orchestrator that dispatches three workers onto one file has not parallelized the sprint,\nit has scheduled a merge conflict.\n\n### 1.4 Write SPRINT_PLAN.md\n\nWrite it to the repo root. It has exactly these four sections, with these headings:\n\n```markdown\n## 1. High-Level Objectives & Anti-Goals\n\nObjectives: numbered, each one observable. \"Store layer exposes async get\u002Fput and the suite\npasses\" is observable. \"Improve the store\" is not.\n\nAnti-Goals: what this sprint will not touch, named as paths or subsystems. This is the\nboundary that keeps a worker from wandering. Every anti-goal is a path a reviewer can check.\n\n## 2. Invariants & Security Boundaries\n\nContracts that hold before and after every single task, not just at the end. Each invariant\nstates the command that falsifies it, and each command has been shown to discriminate.\n\n- Public HTTP handler signatures unchanged: `rg -n 'def (get|put)_item' src\u002Fapi\u002F`\n- No credential literal enters the tree: `rg -n 'token\\s*=\\s*\"' src\u002F`\n- Baseline tests keep passing: `make test`\n\n## 3. Worker Task Graph (DAG)\n\nOne row per task. `depends_on` is what makes this a graph rather than a list, and tasks with\nno path between them are the only ones that may run concurrently.\n\n| id | task | agent | depends_on | inputs (paths) | outputs (paths) | done when |\n|----|------|-------|------------|----------------|-----------------|-----------|\n| T1 | async store layer | Opus | - | src\u002Fstore\u002Fdb.py | src\u002Fstore\u002Fdb.py | `make test` green, get\u002Fput awaitable |\n| T2 | extract auth middleware | Opus | - | src\u002Fapi\u002Fhandlers.py | src\u002Fapi\u002Fmiddleware.py, src\u002Fapi\u002Fhandlers.py | no inline token compare in handlers |\n| T3 | async + middleware tests | GPT-5.6-Sol | T1, T2 | tests\u002F | tests\u002Ftest_store.py, tests\u002Ftest_middleware.py | new tests fail when T1\u002FT2 reverted |\n\nTwo tasks that list the same path in `outputs` must have a `depends_on` edge. If they do not,\nthe DAG is wrong and the workers will overwrite each other.\n\n## 4. Verification Harness\n\nThe exact commands Phase 3 runs, copied from what you ran in 1.1, plus what the new work adds.\n\n- Suite: `make test`\n- Lint: `make lint`\n- Invariant sweep: the commands from section 2\n- Mutation spot-check: revert T1's source, confirm T3's tests go red\n```\n\nProve every invariant command discriminates before it enters the plan. A check that cannot\ngo red is not a check, and both failure directions were observed in testing:\n\n- **Too broad.** A `rg 'get_item'` invariant also matched the import line, so it would keep\n  passing after the definition it guards was deleted. Run it and read the matches, do not\n  just read the exit code.\n- **Too narrow.** A lint target with hardcoded filenames never compiles a module a worker\n  adds, so the invariant passes vacuously for exactly the new code it was meant to cover.\n  Any task whose outputs include a new file owns updating the harness that must see it.\n\nThe cheap proof is to run the command against the current tree and confirm it returns the\npositive it is supposed to find. A pattern that matches nothing today will also match nothing\nafter a worker breaks the thing it guards, and you will read that silence as success.\n\n### 1.5 The execution gate\n\nPrint a summary of the plan: objective count, task count, the concurrency-eligible set, and\nthe verification commands. Then stop and ask, with `AskUserQuestion`, whether to execute.\nOffer approval as one option and revision as the other; never end a turn with a bare prose\nquestion, because a prose choice loses the picker and the review panel both.\n\n**You are blocked until the user answers.** Proceed to Phase 2 only on an explicit approval:\nthe literal string `APPROVED`, or the approve option chosen in the picker. Silence is not\napproval. \"Looks good\" on a different question is not approval. A user asking a clarifying\nquestion about task T2 is not approval of T1.\n\n## Phase 2: heterogeneous cmux dispatch\n\n### Verified command contract\n\nEvery form below was run against a live cmux socket. The corrections matter: three plausible\nvariants of these commands do not exist.\n\n| Need | Command | Behaviour |\n|---|---|---|\n| Create worker | `cmux new-split right --focus false` | prints `OK surface:11 workspace:1` as **plain text**. There is no `--json` flag. |\n| Capture its id | `... \\| cut -d' ' -f2` | yields `surface:11` |\n| Send a command | `cmux send --surface \"$S\" \"$CMD\"'\\n'` | the trailing `\\n` is interpreted by cmux and submits the line |\n| Send a bare key | `cmux send-key --surface \"$S\" enter` | use when a TUI swallowed the newline |\n| Read the pane | `cmux read-screen --surface \"$S\" --lines 50` | add `--scrollback` for history |\n| Block on completion | `cmux wait-for \"$TOK\" --timeout 900` | rc=0 when signalled, rc=1 on timeout |\n| Signal from worker | `cmux wait-for -S \"$TOK\"` | the worker's last action |\n| Agent state | `cmux top --all --format tsv` | tag rows carry `Running` or `Needs input` |\n| Close worker | `cmux close-surface --surface \"$S\"` | see the warning below |\n| Enumerate | `cmux tree` | the authoritative list of live surfaces |\n\n**`close-surface` prints a misleading confirmation.** Closing `surface:13` printed\n`OK surface:14 workspace:1`, an id that never existed. The number in that line is not the\nsurface you closed. Never parse it. Confirm a close against the tree:\n\n```bash\ncmux close-surface --surface \"$S\" >\u002Fdev\u002Fnull 2>&1\ncmux tree | grep -q \"surface $S \" && echo \"STILL OPEN: $S\" || echo \"closed: $S\"\n```\n\n**`list-pane-surfaces` only lists the caller's own pane**, so it will not show workers you\nsplit off. Use `cmux tree` to enumerate.\n\n### Surface lifecycle\n\nThe root orchestrator stays on its own surface under `--agent chief-operator` and never\nexecutes task work itself. **Maximum three concurrent worker splits.** The cap is about\nsocket contention and your own attention: four panes of scrollback exceeds what one root turn\ncan reconcile. Dispatch the DAG in topological waves, at most three wide.\n\n### Deliver the brief through a file, never through nested quotes\n\nThe quoting hazard is the most common way this dispatch breaks. A task payload containing an\napostrophe, a double quote, a backtick, or a `$` corrupts a nested-quote command line, and\nthe worker silently receives a truncated brief. Write the brief to a file in the repo and let\nthe worker CLI read it from stdin. Verified byte-exact against a brief containing all four\nhostile characters.\n\nUse a repo-relative path, not `$TMPDIR`: `$TMPDIR` resolves differently for a sandboxed\norchestrator (`\u002Ftmp\u002Fclaude-NNN\u002F...`) than for the unsandboxed worker shell\n(`\u002Fvar\u002Ffolders\u002F...\u002FT\u002F...`), so a brief written to one is missing in the other.\n\nEvery brief states: the task id, the exact paths the worker may touch, the paths it must not,\nthe invariants from section 2, the done-when condition from the DAG, a clause waiving the\nworker's own approval reflex, and this last line:\n\n```\nYou are running non-interactively. Nobody can answer a question, and asking one ends\nyour turn with the task untouched. The approval gate for this sprint already happened\nat the plan level, so you have authority to implement this brief. Where the brief\nleaves a design choice open, make it, record it in a docstring or comment, and carry\non. Do not ask for confirmation.\n```\n\n**Do not ask the worker to signal.** The dispatcher appends the signal to the worker's shell\nchain, after the process exits and after its exit status is recorded. A brief that also tells\nthe worker to run the signal itself makes the token fire from inside a session that is still\nfinishing: observed here, the wait returned while the worker was still printing its diff, and\nthe exit-status file did not exist yet. One signal, owned by the chain, and the token then\nmeans exactly \"the worker process ended\".\n\nThe waiver is not boilerplate. A dispatched worker inherits the operator's own engineering\nstandards, and those usually say \"brainstorm and get the design approved before implementing\".\nThat is right for interactive work and fatal here. Observed twice on one task: the worker\nproposed a sensible design, asked \"Do you approve this design?\", and exited **rc=0** having\nwritten nothing, in 57 and 69 seconds. Both runs looked like clean successes from the outside.\n\nWithout that chain-appended signal the orchestrator has nothing to block on and degrades to\nscreen-scraping.\n\n### Run each worker in a linked worktree\n\n`dispatch-worker.sh` puts each worker in `\u003Crepo>\u002F.sprint\u002Fworktrees\u002F\u003Ctask>` on a `sprint\u002F\u003Ctask>`\nbranch by default. Pass `--in-place` to override, and expect trouble if you do.\n\nThis is a correctness requirement, not hygiene. A Codex worker inherits the operator's\nPreToolUse policy, which permits edits only when `git rev-parse --absolute-git-dir` differs\nfrom `--git-common-dir`. That is true inside a linked worktree and false in a main checkout:\n\n```bash\ncd \"$REPO\"                       && git rev-parse --absolute-git-dir --git-common-dir\ncd \"$REPO\u002F.sprint\u002Fworktrees\u002FT1\"  && git rev-parse --absolute-git-dir --git-common-dir\n```\n\nThe policy inspects the worker's **cwd**, not the path being written. A worker that creates a\nworktree and then patches into it from the main checkout is still refused, which is exactly\nhow three dispatches stalled with no error and no signal. The dispatcher proves the predicate\nholds before spending a worker on it.\n\nIsolation is the second benefit: a worker's edits land on its own branch, so a wrong answer\nnever touches the main checkout and two workers cannot see each other's half-finished state.\n\n### Worker routing matrix\n\n| Task class | Agent | Why |\n|---|---|---|\n| `unit_tests`, `cli_debugging`, `lint_fix`, `localized_patch` | GPT-5.6-Sol via Codex | fast iterative loops inside one file or one test directory |\n| `architecture`, `api_design`, `multi_file_refactor`, `migration` | Claude Opus | judgment across modules, where the brief cannot spell out every edit |\n\nRoute by whether the brief can fully specify the change. If it can, send it to Sol. If the\nworker has to make a design call, send it to Opus.\n\n### Dispatch\n\nThe dispatch itself lives in `scripts\u002Fdispatch-worker.sh` beside this file, not inline\nhere, and that placement is deliberate. See the warning below.\n\n```bash\n# dispatch-worker.sh \u003Csol|opus> \u003Ctask-id> \u003Cbrief-path> \u003Crepo>\n# prints \"\u003Ctask> \u003Cagent> \u003Csurface> \u003Ctoken>\" for your dispatch ledger\nSKILL_DIR=\"$HOME\u002F.claude\u002Fskills\u002Fcmux-sprint-orchestrator\"\n\"$SKILL_DIR\u002Fscripts\u002Fdispatch-worker.sh\" sol  T1 \"$REPO\u002F.sprint\u002Fbriefs\u002FT1.md\" \"$REPO\"\n\"$SKILL_DIR\u002Fscripts\u002Fdispatch-worker.sh\" opus T2 \"$REPO\u002F.sprint\u002Fbriefs\u002FT2.md\" \"$REPO\"\n```\n\n**Never write a shell positional or an awk field reference into this file** (a dollar sign\nfollowed by a digit). When the skill is invoked with arguments, the loader replaces every one\nof them in the SKILL.md body with a word from those arguments before the agent reads it.\nObserved here: the awk print-second-field idiom was delivered with a path from the invocation\nin place of the field reference, so it captured no surface id, and the next\n`cmux send --surface \"\"` targeted the wrong pane. That is the single most dangerous failure\nin this whole document, and the loader causes it.\n\nThis is why the capture idiom above is `cut -d' ' -f2` rather than awk, why the agent-state\nrow is read with `cut -f4,6,7` rather than a field comparison, and why the dispatch function\nsits in `scripts\u002Fdispatch-worker.sh`. A referenced file is read from disk and is never\nrewritten. Keep it that way: if you move that script's body back inline, it breaks silently\nand only when someone passes arguments.\n\n\nThe `\u003C \u002Fdev\u002Fnull` on the Claude worker is not optional. Without it `claude -p` waits three\nseconds for stdin that a PTY never sends, then prints `Warning: no stdin data received in 3s`\ninto the pane. That warning is noise a stall check can misread, and it delays every Opus\nworker you launch.\n\n`codex exec` is the non-interactive subcommand and it takes the prompt as a positional\nargument or on stdin. `codex run` does not exist, and there is no `--prompt` flag; both\nappear in older notes and both fail. Add `--dangerously-bypass-approvals-and-sandbox` only\nwhen a Sol worker must edit files without prompting, and record that you did.\n\nChild Claude workers **never** get `--agent chief-operator`. A worker running this persona\nwould start planning its own sprint and splitting its own panes. Mirror the root's model,\neffort, and permission flags only. The combined\n`--permission-mode auto --dangerously-skip-permissions` form was verified to run.\n\n### Convergence and stall detection\n\nBlock on the token rather than polling text. `wait-for` returns rc=0 the moment the worker\nsignals, and rc=1 at the timeout, which is your stall detector.\n\n**The token proves the worker's process ended. It does not prove the task succeeded.** The\nsignal is the next command in the worker's shell chain, so it fires just as reliably when the\nagent refused, errored, or asked a question and quit. `dispatch-worker.sh` therefore also\nrecords the worker's exit status to `.sprint\u002Fstatus\u002F\u003Ctask>.rc`, but even that is not enough:\nin the observed failure the worker exited **0** having written nothing. Only the task's own\ndone-when condition, re-run by you at the root, separates finished from succeeded:\n\n```bash\ngrep -c 'def slugify' src\u002Fstrings.py      # the done-when from the DAG row, run by YOU\n```\n\nTreat a worker's report the way you would treat any unverified claim.\n\n```bash\nif cmux wait-for \"$TOK\" --timeout 900; then\n  echo \"$TASK signalled\"\nelse\n  echo \"$TASK STALLED after 900s\"\n  cmux read-screen --surface \"$S\" --lines 80        # dump PTY to root context\nfi\n```\n\nBetween waves, or when a wait times out, read agent state:\n\n```bash\ncmux top --all --format tsv | cut -f4,6,7 | grep '^tag'\n```\n\n`Needs input` means the worker is blocked on a prompt, not working. That is a stall: read the\npane, answer it with `cmux send` if the answer is unambiguous, otherwise treat it as a failure\nand re-diagnose. Note the scope limit: this tag is **workspace-scoped, not per-surface**. A\nworkspace holding several agent surfaces reports one aggregated tag whose state can be blank,\nverified on a workspace with seven agent surfaces. So the tag tells you a workspace needs\nattention; `read-screen` per surface tells you which worker. When per-worker state matters\nmore than a single-screen view, give each worker its own workspace with\n`cmux new-workspace --command ...` instead of a split.\n\n### Can workers talk to each other\n\nDefault answer: no, and they should not need to. The root mediates, and two workers that need\nto talk mid-task are a dependency edge you failed to draw in the DAG. Turn the conversation\ninto an edge and the second worker into a later wave.\n\nWorktree isolation makes this concrete rather than a preference. Each worker is on its own\nbranch in its own checkout, so it cannot see another worker's half-finished state even if it\nlooks. That is the point: partial work is not a contract, and a worker that reads another's\nuncommitted file has coupled itself to something that may still be reverted.\n\nThree mechanisms exist, in the order you should reach for them:\n\n| Need | Mechanism | Cost |\n|---|---|---|\n| B needs A's finished output | a `depends_on` edge; B runs in a later wave | none, this is the DAG doing its job |\n| B needs A's interface but not its tests | root merges A's branch, then dispatches B | one extra wave, full attribution |\n| B must block mid-task on A reaching a point | `cmux wait-for` as a named rendezvous | real coupling, use sparingly |\n\nThe rendezvous is a genuine primitive, not a workaround: `cmux wait-for -S \u003Cname>` signals and\n`cmux wait-for \u003Cname> --timeout N` blocks, returning rc=0 when signalled and rc=1 on timeout.\nA worker can therefore block on `interface-frozen` while another signals it. Reach for it only\non the tests-after-interface seam, where the alternative is serialising a long task behind a\nshort one, and name the token in both briefs so the coupling is visible in the plan.\n\nWhat you should not do is let workers `cmux send` into each other's surfaces. It works, and it\nproduces a sprint whose failures are unattributable: when the result is wrong you cannot tell\nwhich worker's instruction caused it, the root's ledger no longer describes what happened, and\nthe three-strikes policy has nothing coherent to re-brief. Keep the star topology. The root is\nthe only writer of instructions.\n\n### Failure policy\n\nThree strikes per task, and the strike that matters most is the quiet one.\n\nThe dangerous worker is not the one that errors. It is the one that exits **0**, leaves its own\nsuite **green**, and is wrong anyway. Observed: a task under-specified its tie-breaking rule, a\nCodex worker reached for the standard library exactly as the brief suggested, documented the\ntie-to-even behaviour it had chosen in a docstring, exited 0 with 24 tests passing, and failed\nhalf the sprint's acceptance cases. Nothing in the pane, the exit status or the worker's suite\nsaid so. Only the root's own acceptance check did.\n\nThat is the whole argument for re-running done-when yourself. A worker's green suite tests what\nthe worker decided to build.\n\n**Re-briefing is a new dispatch, not a conversation.** The worker process has already exited, so\nyou cannot reply to it. Re-dispatch the same task id: `dispatch-worker.sh` reuses the existing\n`.sprint\u002Fworktrees\u002F\u003Ctask>`, so the worker opens its own previous attempt and fixes it in place\nrather than starting over. A re-brief that lands the fix contains four things, in this order:\n\n1. That this is attempt N of 3, and that the previous attempt is already in the worktree.\n2. The failure output verbatim, passing cases included. Showing which cases passed tells the\n   worker the shape of the rule rather than only that it is wrong.\n3. A one-line diagnosis naming the cause, and saying plainly where the brief was at fault when\n   it was. A worker told its defensible call was reasonable but superseded fixes the code; one\n   told it was simply wrong tends to rewrite more than it needs to.\n4. The requirement, now stated explicitly, with the exact expected values.\n\nThat recipe fixed the rounding task on attempt 2: all eight acceptance cases correct, the\ndocstring rewritten to state the real rule, four tie tests added, and reverting the fix takes\nthe suite red, so the new tests discriminate.\n\nOn the third failure, or on any repeated identical output (a loop):\n\n1. `cmux read-screen --surface \"$S\" --lines 100 --scrollback` and keep the dump.\n2. Close the surface and confirm the close against `cmux tree`.\n3. Revert that task's partial work so the tree returns to a known state.\n4. Re-diagnose at the root. A task that failed three times has a wrong brief or a wrong DAG\n   edge, not an unlucky worker. Fix the plan, and say in the report that you did.\n\nNever dispatch a replacement worker for a task whose brief you have not changed. Re-running an\nunchanged brief is how three strikes get spent on one bug.\n\n## Phase 3: consolidation and verification\n\nRun from the root surface. Paste the output; a remembered pass is not a pass.\n\n```bash\ncd \"$REPO\"\nmkdir -p .sprint && printf '*\\n' > .sprint\u002F.gitignore   # scratch stays out of git status\nmake test > .sprint\u002Ftest.log 2>&1; test_rc=$?\ntail -20 .sprint\u002Ftest.log; echo \"test rc=$test_rc\"\nmake lint > .sprint\u002Flint.log 2>&1; lint_rc=$?\ntail -10 .sprint\u002Flint.log; echo \"lint rc=$lint_rc\"\n[ \"$test_rc\" -eq 0 ] || echo \"SUITE RED: do not report this sprint as done\"\n```\n\nGate on the command, not on a pipe. `cmd | tail && commit` inspects tail's exit code and will\nhappily commit a red suite. Do not reach for `${PIPESTATUS[0]}` as the fix either: the root\nsurface runs zsh, where that expands to the empty string, so the rc check prints nothing and\npasses vacuously. Verified: the same pipeline reports `7` in bash and an empty value in zsh.\nRedirect to a file, capture `$?` on the next line, then read the file.\n\nThen check the invariants and the boundary, one command per line of section 2:\n\n```bash\ngit status --short\ngit diff --stat\ngit diff --name-only | while read -r f; do\n  grep -qx \"$f\" .sprint-allowed-paths || echo \"BOUNDARY VIOLATION: $f\"\ndone\n```\n\nAny file changed that no task listed as an output is an invariant violation. Investigate it\nbefore reporting, and do not launder it into \"incidental cleanup\".\n\nVerify the new tests discriminate. A test suite that stays green when its subject is deleted\nis theater:\n\n```bash\ncp src\u002Fstore\u002Fdb.py \u002Ftmp\u002Fdb.bak && git checkout HEAD~1 -- src\u002Fstore\u002Fdb.py\nmake test 2>&1 | tail -3      # MUST be red\ncommand cp -f \u002Ftmp\u002Fdb.bak src\u002Fstore\u002Fdb.py\n```\n\nCollect each task branch before you verify the whole. The worker leaves its output\nuncommitted in its own worktree, so the root commits it on the task branch and merges:\n\nWorkers differ in whether they commit. Opus workers committed their own output, in RED then\nGREEN pairs; Codex workers left it uncommitted in the worktree. Collection must tolerate both\nor it fails on whichever kind you did not expect:\n\n```bash\nW=\"$REPO\u002F.sprint\u002Fworktrees\u002F$TASK\"\nif [ -n \"$(git -C \"$W\" status --porcelain)\" ]; then\n  git -C \"$W\" add -A\n  git -C \"$W\" commit -m \"feat(scope): $TASK\"\nfi\ngit -C \"$REPO\" merge --no-ff --no-edit \"sprint\u002F$TASK\"\ngit -C \"$REPO\" worktree remove \"$W\" --force && git -C \"$REPO\" branch -d \"sprint\u002F$TASK\"\n```\n\nMerge one task at a time and run the suite after each, so a conflict or a regression is\nattributable to a single worker. `git -C \"$REPO\" worktree list` afterwards must show only the\nmain checkout: a worktree that outlives its task is the litter a blocked worker leaves behind.\n\nClose every remaining worker surface and confirm against the tree.\n\nEmit `SPRINT_COMPLETION_REPORT.md` with: per-task status against the DAG (`DONE` \u002F\n`DONE_WITH_CONCERNS` \u002F `BLOCKED`), the routing breakdown (how many tasks to Sol, how many to\nOpus, how many re-briefed), stall and retry counts, files changed against files planned, the\npasted verification output, and every deviation from `SPRINT_PLAN.md` with its reason.\n\n## Common mistakes\n\n| Mistake | Consequence | Fix |\n|---|---|---|\n| `cmux new-split right --json \\| jq -r '.surface_id'` | empty `SURFACE_ID`; the next `cmux send` with an empty `--surface` targets the wrong pane | `--json` does not exist; `cut -d' ' -f2` the plain-text `OK` line |\n| `codex run --model X --prompt '...'` | no such subcommand, no such flag | `codex exec -m X - \u003C brief` |\n| Parsing the id out of `close-surface` output | you believe a surface closed that is still running | confirm with `cmux tree` |\n| Nesting the brief in quotes inside the send payload | truncated brief on any apostrophe; the worker builds the wrong thing | brief file plus stdin redirect |\n| Passing `--agent chief-operator` to a worker | recursive orchestrator spawning its own splits | root persona stays on the root surface |\n| Polling `read-screen` as the completion signal | you close a pane mid-write, or wait forever on a finished one | `cmux wait-for` token, screen reads for diagnosis |\n| Briefs under `$TMPDIR` | worker cannot find the file across the sandbox boundary | repo-relative brief paths |\n| Running `cmux` sandboxed | `Operation not permitted (errno 1)` misread as a dead socket | disable the sandbox for cmux calls |\n| `make test \\| tail; echo ${PIPESTATUS[0]}` in the root shell | empty rc under zsh, so a red suite reports as passing | redirect to a log, capture `$?` on the next line |\n| An invariant grep that matches the import line, or a lint target with hardcoded filenames | the gate passes vacuously and a broken invariant reads as green | run each check and read its matches; a new output file owns updating the harness |\n| More than three concurrent splits | socket contention and unreconcilable scrollback | topological waves, width 3 |\n","# cmux Sprint Orchestrator\n\n## Overview\n\nDrives sprint-scale work from a single root surface by dispatching bounded tasks to\nheterogeneous agent panes in cmux, then converging their output behind one verification gate.\n\n**Core principle: the plan is the product of Phase 1, and code is not.** An orchestrator that\nedits files while it is still deciding what to build has no plan, only a diff. Splitting the\ndecision from the execution is what makes the work reviewable.\n\nMeasured baseline: three fresh agents given a three-item sprint under deadline pressure all\nbegan editing source immediately. None produced a plan artifact and none paused for approval.\nOne rewrote production source before naming a single invariant. That is the failure this skill\nexists to prevent.\n\n## The Iron Gate\n\n```\nPHASE 1 IS READ-ONLY. NO FILE IN THE TARGET REPO CHANGES BEFORE `APPROVED`.\n```\n\nIn Phase 1 you may read, search, list, run existing test suites, and write exactly one file:\n`SPRINT_PLAN.md`. Nothing else.\n\n**Forbidden in Phase 1, without exception:**\n\n- Editing, creating, or deleting any application, test, config, or build file.\n- Running a formatter, linter, or codemod that writes to the tree.\n- `git commit`, `git checkout -b`, `git stash`, or any index mutation.\n- Creating a cmux split, pane, or workspace.\n- Sending a single character to any worker CLI.\n- \"Scaffolding\", \"a stub so the DAG typechecks\", or \"one obvious fix while I read\".\n\nViolating the letter of this gate is violating the spirit of it. If you changed a file before\n`APPROVED`, revert it (`git checkout -- \u003Cpath>`, delete untracked files), say so plainly, and\nrestart Phase 1.\n\n### Rationalizations, and what is actually true\n\n| Rationalization | Reality |\n|---|---|\n| \"The deadline is tomorrow, planning is ceremony.\" | The baseline agents that skipped the plan produced three uncoordinated diffs that no reviewer could gate. The plan is the cheapest artifact in the sprint. |\n| \"The lead already signed off on the direction.\" | Sign-off on direction is not approval of a task graph. `APPROVED` gates the DAG, the invariants, and the routing, none of which existed when the lead spoke. |\n| \"It's one small fix, I'm already in the file.\" | Phase 1 has no small fixes. An unplanned edit is an untracked task with no owner, no verification, and no line in the report. |\n| \"I need a stub to validate the dependency order.\" | Validate order by reading imports and the test suite. A stub is code. |\n| \"I'll write the plan after, from what I learned building.\" | A plan written after the diff documents the diff. It cannot change a decision that is already spent. |\n| \"Spawning the workers now saves wall-clock, I'll plan while they run.\" | Workers dispatched without a DAG collide on shared files. Concurrency without a plan is a merge conflict with extra steps. |\n| \"The user obviously wants working code, not a document.\" | The user wants working code they can verify. Phase 3 is where code is judged; Phase 1 is where it is made judgeable. |\n| \"Reading is slow, I'll learn faster by changing something.\" | Then change it in a scratch clone outside the target repo, and cite it as evidence in the plan. |\n\n### Red flags: stop and restart Phase 1\n\n- `git status` in the target repo is not clean and you have not printed `APPROVED`.\n- You are composing a `cmux send` payload and `SPRINT_PLAN.md` does not exist.\n- You catch yourself writing \"while I'm here\" or \"quick\".\n- You are about to ask the user to approve a plan you have already half-built.\n\n## Prerequisites\n\nRun this first. cmux talks over a unix domain socket, which the agent sandbox denies, so\n**every `cmux` call in this skill runs with the sandbox disabled.** An\n`Operation not permitted (errno 1)` from `cmux` is the sandbox, not a broken socket.\n\n```bash\nfor b in cmux codex claude jq; do\n  command -v \"$b\" >\u002Fdev\u002Fnull 2>&1 || { echo \"MISSING: $b\"; exit 1; }\ndone\ncmux ping || { echo \"cmux socket unreachable\"; exit 1; }\nROOT_SURFACE=$(cmux identify | jq -r '.caller.surface_ref')\nWORKSPACE=$(cmux identify | jq -r '.caller.workspace_ref')\necho \"root=$ROOT_SURFACE workspace=$WORKSPACE\"\n```\n\n`cmux ping` prints `PONG`. Surface refs are 1-indexed: the root orchestrator is `surface:1`,\nand there is no `surface:0`. The socket defaults to `~\u002F.local\u002Fstate\u002Fcmux\u002Fcmux.sock` and may be\nuid-suffixed (`cmux-\u003Cuid>.sock`); override with `CMUX_SOCKET_PATH`. It is not `\u002Ftmp\u002Fcmux.sock`.\n\n## Phase 1: architectural formulation (read-only)\n\n### 1.1 Map the ground truth\n\nEvery count, path, and command that reaches `SPRINT_PLAN.md` comes from a command you ran in\nthis session and pasted. A remembered module list is not evidence.\n\n```bash\ngit -C \"$REPO\" status --short          # must be clean, and stay clean\ngit -C \"$REPO\" log --oneline -10\n\n# Scratch dir, self-ignoring so the census and logs never dirty git status.\nmkdir -p \"$REPO\u002F.sprint\" && printf '*\\n' > \"$REPO\u002F.sprint\u002F.gitignore\"\n\n# File census. `rg --files` alone SKIPS HIDDEN DIRECTORIES, so a repo that keeps\n# its content under .claude\u002F, .github\u002F or .cursor\u002F reports a fraction of itself.\ngit -C \"$REPO\" ls-files > \"$REPO\u002F.sprint\u002Fcensus.txt\"                   # tracked files\n# not a git repo? use: rg --files --hidden --glob '!.git' \"$REPO\"\nwc -l \u003C \"$REPO\u002F.sprint\u002Fcensus.txt\"\n\n# Language mix. The `grep '\\.'` guard matters: an extensionless file (LICENSE,\n# Makefile) would otherwise have its whole path counted as an extension.\ngrep '\\.' \"$REPO\u002F.sprint\u002Fcensus.txt\" | sed 's|.*\\.||' | sort | uniq -c | sort -rn | head\n\n# Dependency shape, over the census rather than a fresh blind walk.\nrg -n '^(import|from|require|use )' --hidden --glob '!.git' \"$REPO\" --stats | tail -5\nls \"$REPO\"\u002F{Makefile,justfile,package.json,pyproject.toml,go.mod} 2>\u002Fdev\u002Fnull\n```\n\nProve the census can see before you trust what it does not show. Pick a file you\nalready know exists, ideally a deep or hidden one, and confirm it is in the list:\n\n```bash\ngrep -c 'some\u002Ffile\u002Fyou\u002Fknow\u002Fexists' \"$REPO\u002F.sprint\u002Fcensus.txt\"   # must be >= 1\n```\n\nMeasured on a real repo, the difference is not marginal: `rg --files` reported 56\nfiles where `git ls-files` reported 383, and it listed exactly 0 of the 246 files\nunder `.claude\u002F`. A census that cannot see a directory reports no work there and\nno risk there, and the DAG you build on it inherits both blind spots.\n\nThen run the existing suite once, unchanged, and record the result. A green baseline is the\nonly thing that makes a Phase 3 regression legible.\n\n```bash\n( cd \"$REPO\" && make test ) > \"$REPO\u002F.sprint\u002Fbaseline.log\" 2>&1; echo \"baseline rc=$?\"\ntail -5 \"$REPO\u002F.sprint\u002Fbaseline.log\"\n```\n\nIf a repo has a code knowledge graph (`graphify-out\u002Fgraph.json`), orient with\n`graphify query \"\u003Cquestion>\"` before grepping.\n\nIdentify high-risk integration zones explicitly: files imported by three or more modules,\nanything in an auth or crypto path, and any file two planned tasks would both touch. The last\ncategory is the one that decides your DAG edges.\n\n### 1.2 Confirm the target repo is dispatch-ready\n\nA dispatched worker runs inside **your** hook stack, not a clean one. Hooks you wrote for\nyour own repositories will fire against the target repo, and when one blocks a worker the\norchestrator sees nothing but a stall: no error, no signal, just a timeout.\n\nObserved on the first live run of this skill. A worker created its own worktree, ran the\nbaseline suite green, wrote correct tests, and every write was then refused by a PreToolUse\nhook demanding a linked worktree. Three dispatches burned about five minutes each and the\nroot surface could not tell that from a slow worker. Adding the marker the hook asked for\nmade the same brief succeed unchanged.\n\nBefore dispatching anything, check what your own hooks require of a repo:\n\n```bash\nls \"$REPO\"\u002F{AGENTS.md,CLAUDE.md} 2>\u002Fdev\u002Fnull      # project instruction files the hooks read\ngrep -rl 'PreToolUse' ~\u002F.claude\u002Fsettings*.json 2>\u002Fdev\u002Fnull\n```\n\nThen satisfy the requirement in the target repo, or hand each worker a properly linked\nworktree. Either is fine; discovering which one your hooks want after three silent stalls is\nnot. A worker blocked by a policy hook has not failed the task, and re-briefing it will not\nhelp, so this belongs in Phase 1 rather than in the failure policy.\n\nClean up after a blocked worker: it may have created a git worktree of its own before being\nstopped. `git -C \"$REPO\" worktree list` shows them and they outlive the surface you closed.\n\n### 1.3 Decide whether to fan out at all\n\nDo this before writing the DAG, and be willing to answer no. Three independent agents given a\nthree-item sprint all reached the same conclusion unprompted: two of the items rewrote the same\nfile, so any two workers would have collided and one would have had to rebase onto the other.\nTheir words: \"coordination overhead would have exceeded the work.\"\n\nFan out only when every one of these holds:\n\n- Two or more tasks share no path in their `outputs` column.\n- Those tasks have no `depends_on` edge between them.\n- Each task is large enough that a worker's startup and your reconciliation cost less than\n  doing it inline.\n\nIf none of that holds, say so in section 3, run the sprint serially from the root surface, and\nskip Phase 2. A serial sprint executed against an approved plan is a success, not a fallback.\nThe most common seam that does parallelize is tests-after-interface: land the interface change\nfirst, freeze it, then fan out the test sweep across the settled API.\n\nAn orchestrator that dispatches three workers onto one file has not parallelized the sprint,\nit has scheduled a merge conflict.\n\n### 1.4 Write SPRINT_PLAN.md\n\nWrite it to the repo root. It has exactly these four sections, with these headings:\n\n```markdown\n## 1. High-Level Objectives & Anti-Goals\n\nObjectives: numbered, each one observable. \"Store layer exposes async get\u002Fput and the suite\npasses\" is observable. \"Improve the store\" is not.\n\nAnti-Goals: what this sprint will not touch, named as paths or subsystems. This is the\nboundary that keeps a worker from wandering. Every anti-goal is a path a reviewer can check.\n\n## 2. Invariants & Security Boundaries\n\nContracts that hold before and after every single task, not just at the end. Each invariant\nstates the command that falsifies it, and each command has been shown to discriminate.\n\n- Public HTTP handler signatures unchanged: `rg -n 'def (get|put)_item' src\u002Fapi\u002F`\n- No credential literal enters the tree: `rg -n 'token\\s*=\\s*\"' src\u002F`\n- Baseline tests keep passing: `make test`\n\n## 3. Worker Task Graph (DAG)\n\nOne row per task. `depends_on` is what makes this a graph rather than a list, and tasks with\nno path between them are the only ones that may run concurrently.\n\n| id | task | agent | depends_on | inputs (paths) | outputs (paths) | done when |\n|----|------|-------|------------|----------------|-----------------|-----------|\n| T1 | async store layer | Opus | - | src\u002Fstore\u002Fdb.py | src\u002Fstore\u002Fdb.py | `make test` green, get\u002Fput awaitable |\n| T2 | extract auth middleware | Opus | - | src\u002Fapi\u002Fhandlers.py | src\u002Fapi\u002Fmiddleware.py, src\u002Fapi\u002Fhandlers.py | no inline token compare in handlers |\n| T3 | async + middleware tests | GPT-5.6-Sol | T1, T2 | tests\u002F | tests\u002Ftest_store.py, tests\u002Ftest_middleware.py | new tests fail when T1\u002FT2 reverted |\n\nTwo tasks that list the same path in `outputs` must have a `depends_on` edge. If they do not,\nthe DAG is wrong and the workers will overwrite each other.\n\n## 4. Verification Harness\n\nThe exact commands Phase 3 runs, copied from what you ran in 1.1, plus what the new work adds.\n\n- Suite: `make test`\n- Lint: `make lint`\n- Invariant sweep: the commands from section 2\n- Mutation spot-check: revert T1's source, confirm T3's tests go red\n```\n\nProve every invariant command discriminates before it enters the plan. A check that cannot\ngo red is not a check, and both failure directions were observed in testing:\n\n- **Too broad.** A `rg 'get_item'` invariant also matched the import line, so it would keep\n  passing after the definition it guards was deleted. Run it and read the matches, do not\n  just read the exit code.\n- **Too narrow.** A lint target with hardcoded filenames never compiles a module a worker\n  adds, so the invariant passes vacuously for exactly the new code it was meant to cover.\n  Any task whose outputs include a new file owns updating the harness that must see it.\n\nThe cheap proof is to run the command against the current tree and confirm it returns the\npositive it is supposed to find. A pattern that matches nothing today will also match nothing\nafter a worker breaks the thing it guards, and you will read that silence as success.\n\n### 1.5 The execution gate\n\nPrint a summary of the plan: objective count, task count, the concurrency-eligible set, and\nthe verification commands. Then stop and ask, with `AskUserQuestion`, whether to execute.\nOffer approval as one option and revision as the other; never end a turn with a bare prose\nquestion, because a prose choice loses the picker and the review panel both.\n\n**You are blocked until the user answers.** Proceed to Phase 2 only on an explicit approval:\nthe literal string `APPROVED`, or the approve option chosen in the picker. Silence is not\napproval. \"Looks good\" on a different question is not approval. A user asking a clarifying\nquestion about task T2 is not approval of T1.\n\n## Phase 2: heterogeneous cmux dispatch\n\n### Verified command contract\n\nEvery form below was run against a live cmux socket. The corrections matter: three plausible\nvariants of these commands do not exist.\n\n| Need | Command | Behaviour |\n|---|---|---|\n| Create worker | `cmux new-split right --focus false` | prints `OK surface:11 workspace:1` as **plain text**. There is no `--json` flag. |\n| Capture its id | `... \\| cut -d' ' -f2` | yields `surface:11` |\n| Send a command | `cmux send --surface \"$S\" \"$CMD\"'\\n'` | the trailing `\\n` is interpreted by cmux and submits the line |\n| Send a bare key | `cmux send-key --surface \"$S\" enter` | use when a TUI swallowed the newline |\n| Read the pane | `cmux read-screen --surface \"$S\" --lines 50` | add `--scrollback` for history |\n| Block on completion | `cmux wait-for \"$TOK\" --timeout 900` | rc=0 when signalled, rc=1 on timeout |\n| Signal from worker | `cmux wait-for -S \"$TOK\"` | the worker's last action |\n| Agent state | `cmux top --all --format tsv` | tag rows carry `Running` or `Needs input` |\n| Close worker | `cmux close-surface --surface \"$S\"` | see the warning below |\n| Enumerate | `cmux tree` | the authoritative list of live surfaces |\n\n**`close-surface` prints a misleading confirmation.** Closing `surface:13` printed\n`OK surface:14 workspace:1`, an id that never existed. The number in that line is not the\nsurface you closed. Never parse it. Confirm a close against the tree:\n\n```bash\ncmux close-surface --surface \"$S\" >\u002Fdev\u002Fnull 2>&1\ncmux tree | grep -q \"surface $S \" && echo \"STILL OPEN: $S\" || echo \"closed: $S\"\n```\n\n**`list-pane-surfaces` only lists the caller's own pane**, so it will not show workers you\nsplit off. Use `cmux tree` to enumerate.\n\n### Surface lifecycle\n\nThe root orchestrator stays on its own surface under `--agent chief-operator` and never\nexecutes task work itself. **Maximum three concurrent worker splits.** The cap is about\nsocket contention and your own attention: four panes of scrollback exceeds what one root turn\ncan reconcile. Dispatch the DAG in topological waves, at most three wide.\n\n### Deliver the brief through a file, never through nested quotes\n\nThe quoting hazard is the most common way this dispatch breaks. A task payload containing an\napostrophe, a double quote, a backtick, or a `$` corrupts a nested-quote command line, and\nthe worker silently receives a truncated brief. Write the brief to a file in the repo and let\nthe worker CLI read it from stdin. Verified byte-exact against a brief containing all four\nhostile characters.\n\nUse a repo-relative path, not `$TMPDIR`: `$TMPDIR` resolves differently for a sandboxed\norchestrator (`\u002Ftmp\u002Fclaude-NNN\u002F...`) than for the unsandboxed worker shell\n(`\u002Fvar\u002Ffolders\u002F...\u002FT\u002F...`), so a brief written to one is missing in the other.\n\nEvery brief states: the task id, the exact paths the worker may touch, the paths it must not,\nthe invariants from section 2, the done-when condition from the DAG, a clause waiving the\nworker's own approval reflex, and this last line:\n\n```\nYou are running non-interactively. Nobody can answer a question, and asking one ends\nyour turn with the task untouched. The approval gate for this sprint already happened\nat the plan level, so you have authority to implement this brief. Where the brief\nleaves a design choice open, make it, record it in a docstring or comment, and carry\non. Do not ask for confirmation.\n```\n\n**Do not ask the worker to signal.** The dispatcher appends the signal to the worker's shell\nchain, after the process exits and after its exit status is recorded. A brief that also tells\nthe worker to run the signal itself makes the token fire from inside a session that is still\nfinishing: observed here, the wait returned while the worker was still printing its diff, and\nthe exit-status file did not exist yet. One signal, owned by the chain, and the token then\nmeans exactly \"the worker process ended\".\n\nThe waiver is not boilerplate. A dispatched worker inherits the operator's own engineering\nstandards, and those usually say \"brainstorm and get the design approved before implementing\".\nThat is right for interactive work and fatal here. Observed twice on one task: the worker\nproposed a sensible design, asked \"Do you approve this design?\", and exited **rc=0** having\nwritten nothing, in 57 and 69 seconds. Both runs looked like clean successes from the outside.\n\nWithout that chain-appended signal the orchestrator has nothing to block on and degrades to\nscreen-scraping.\n\n### Run each worker in a linked worktree\n\n`dispatch-worker.sh` puts each worker in `\u003Crepo>\u002F.sprint\u002Fworktrees\u002F\u003Ctask>` on a `sprint\u002F\u003Ctask>`\nbranch by default. Pass `--in-place` to override, and expect trouble if you do.\n\nThis is a correctness requirement, not hygiene. A Codex worker inherits the operator's\nPreToolUse policy, which permits edits only when `git rev-parse --absolute-git-dir` differs\nfrom `--git-common-dir`. That is true inside a linked worktree and false in a main checkout:\n\n```bash\ncd \"$REPO\"                       && git rev-parse --absolute-git-dir --git-common-dir\ncd \"$REPO\u002F.sprint\u002Fworktrees\u002FT1\"  && git rev-parse --absolute-git-dir --git-common-dir\n```\n\nThe policy inspects the worker's **cwd**, not the path being written. A worker that creates a\nworktree and then patches into it from the main checkout is still refused, which is exactly\nhow three dispatches stalled with no error and no signal. The dispatcher proves the predicate\nholds before spending a worker on it.\n\nIsolation is the second benefit: a worker's edits land on its own branch, so a wrong answer\nnever touches the main checkout and two workers cannot see each other's half-finished state.\n\n### Worker routing matrix\n\n| Task class | Agent | Why |\n|---|---|---|\n| `unit_tests`, `cli_debugging`, `lint_fix`, `localized_patch` | GPT-5.6-Sol via Codex | fast iterative loops inside one file or one test directory |\n| `architecture`, `api_design`, `multi_file_refactor`, `migration` | Claude Opus | judgment across modules, where the brief cannot spell out every edit |\n\nRoute by whether the brief can fully specify the change. If it can, send it to Sol. If the\nworker has to make a design call, send it to Opus.\n\n### Dispatch\n\nThe dispatch itself lives in `scripts\u002Fdispatch-worker.sh` beside this file, not inline\nhere, and that placement is deliberate. See the warning below.\n\n```bash\n# dispatch-worker.sh \u003Csol|opus> \u003Ctask-id> \u003Cbrief-path> \u003Crepo>\n# prints \"\u003Ctask> \u003Cagent> \u003Csurface> \u003Ctoken>\" for your dispatch ledger\nSKILL_DIR=\"$HOME\u002F.claude\u002Fskills\u002Fcmux-sprint-orchestrator\"\n\"$SKILL_DIR\u002Fscripts\u002Fdispatch-worker.sh\" sol  T1 \"$REPO\u002F.sprint\u002Fbriefs\u002FT1.md\" \"$REPO\"\n\"$SKILL_DIR\u002Fscripts\u002Fdispatch-worker.sh\" opus T2 \"$REPO\u002F.sprint\u002Fbriefs\u002FT2.md\" \"$REPO\"\n```\n\n**Never write a shell positional or an awk field reference into this file** (a dollar sign\nfollowed by a digit). When the skill is invoked with arguments, the loader replaces every one\nof them in the SKILL.md body with a word from those arguments before the agent reads it.\nObserved here: the awk print-second-field idiom was delivered with a path from the invocation\nin place of the field reference, so it captured no surface id, and the next\n`cmux send --surface \"\"` targeted the wrong pane. That is the single most dangerous failure\nin this whole document, and the loader causes it.\n\nThis is why the capture idiom above is `cut -d' ' -f2` rather than awk, why the agent-state\nrow is read with `cut -f4,6,7` rather than a field comparison, and why the dispatch function\nsits in `scripts\u002Fdispatch-worker.sh`. A referenced file is read from disk and is never\nrewritten. Keep it that way: if you move that script's body back inline, it breaks silently\nand only when someone passes arguments.\n\n\nThe `\u003C \u002Fdev\u002Fnull` on the Claude worker is not optional. Without it `claude -p` waits three\nseconds for stdin that a PTY never sends, then prints `Warning: no stdin data received in 3s`\ninto the pane. That warning is noise a stall check can misread, and it delays every Opus\nworker you launch.\n\n`codex exec` is the non-interactive subcommand and it takes the prompt as a positional\nargument or on stdin. `codex run` does not exist, and there is no `--prompt` flag; both\nappear in older notes and both fail. Add `--dangerously-bypass-approvals-and-sandbox` only\nwhen a Sol worker must edit files without prompting, and record that you did.\n\nChild Claude workers **never** get `--agent chief-operator`. A worker running this persona\nwould start planning its own sprint and splitting its own panes. Mirror the root's model,\neffort, and permission flags only. The combined\n`--permission-mode auto --dangerously-skip-permissions` form was verified to run.\n\n### Convergence and stall detection\n\nBlock on the token rather than polling text. `wait-for` returns rc=0 the moment the worker\nsignals, and rc=1 at the timeout, which is your stall detector.\n\n**The token proves the worker's process ended. It does not prove the task succeeded.** The\nsignal is the next command in the worker's shell chain, so it fires just as reliably when the\nagent refused, errored, or asked a question and quit. `dispatch-worker.sh` therefore also\nrecords the worker's exit status to `.sprint\u002Fstatus\u002F\u003Ctask>.rc`, but even that is not enough:\nin the observed failure the worker exited **0** having written nothing. Only the task's own\ndone-when condition, re-run by you at the root, separates finished from succeeded:\n\n```bash\ngrep -c 'def slugify' src\u002Fstrings.py      # the done-when from the DAG row, run by YOU\n```\n\nTreat a worker's report the way you would treat any unverified claim.\n\n```bash\nif cmux wait-for \"$TOK\" --timeout 900; then\n  echo \"$TASK signalled\"\nelse\n  echo \"$TASK STALLED after 900s\"\n  cmux read-screen --surface \"$S\" --lines 80        # dump PTY to root context\nfi\n```\n\nBetween waves, or when a wait times out, read agent state:\n\n```bash\ncmux top --all --format tsv | cut -f4,6,7 | grep '^tag'\n```\n\n`Needs input` means the worker is blocked on a prompt, not working. That is a stall: read the\npane, answer it with `cmux send` if the answer is unambiguous, otherwise treat it as a failure\nand re-diagnose. Note the scope limit: this tag is **workspace-scoped, not per-surface**. A\nworkspace holding several agent surfaces reports one aggregated tag whose state can be blank,\nverified on a workspace with seven agent surfaces. So the tag tells you a workspace needs\nattention; `read-screen` per surface tells you which worker. When per-worker state matters\nmore than a single-screen view, give each worker its own workspace with\n`cmux new-workspace --command ...` instead of a split.\n\n### Can workers talk to each other\n\nDefault answer: no, and they should not need to. The root mediates, and two workers that need\nto talk mid-task are a dependency edge you failed to draw in the DAG. Turn the conversation\ninto an edge and the second worker into a later wave.\n\nWorktree isolation makes this concrete rather than a preference. Each worker is on its own\nbranch in its own checkout, so it cannot see another worker's half-finished state even if it\nlooks. That is the point: partial work is not a contract, and a worker that reads another's\nuncommitted file has coupled itself to something that may still be reverted.\n\nThree mechanisms exist, in the order you should reach for them:\n\n| Need | Mechanism | Cost |\n|---|---|---|\n| B needs A's finished output | a `depends_on` edge; B runs in a later wave | none, this is the DAG doing its job |\n| B needs A's interface but not its tests | root merges A's branch, then dispatches B | one extra wave, full attribution |\n| B must block mid-task on A reaching a point | `cmux wait-for` as a named rendezvous | real coupling, use sparingly |\n\nThe rendezvous is a genuine primitive, not a workaround: `cmux wait-for -S \u003Cname>` signals and\n`cmux wait-for \u003Cname> --timeout N` blocks, returning rc=0 when signalled and rc=1 on timeout.\nA worker can therefore block on `interface-frozen` while another signals it. Reach for it only\non the tests-after-interface seam, where the alternative is serialising a long task behind a\nshort one, and name the token in both briefs so the coupling is visible in the plan.\n\nWhat you should not do is let workers `cmux send` into each other's surfaces. It works, and it\nproduces a sprint whose failures are unattributable: when the result is wrong you cannot tell\nwhich worker's instruction caused it, the root's ledger no longer describes what happened, and\nthe three-strikes policy has nothing coherent to re-brief. Keep the star topology. The root is\nthe only writer of instructions.\n\n### Failure policy\n\nThree strikes per task, and the strike that matters most is the quiet one.\n\nThe dangerous worker is not the one that errors. It is the one that exits **0**, leaves its own\nsuite **green**, and is wrong anyway. Observed: a task under-specified its tie-breaking rule, a\nCodex worker reached for the standard library exactly as the brief suggested, documented the\ntie-to-even behaviour it had chosen in a docstring, exited 0 with 24 tests passing, and failed\nhalf the sprint's acceptance cases. Nothing in the pane, the exit status or the worker's suite\nsaid so. Only the root's own acceptance check did.\n\nThat is the whole argument for re-running done-when yourself. A worker's green suite tests what\nthe worker decided to build.\n\n**Re-briefing is a new dispatch, not a conversation.** The worker process has already exited, so\nyou cannot reply to it. Re-dispatch the same task id: `dispatch-worker.sh` reuses the existing\n`.sprint\u002Fworktrees\u002F\u003Ctask>`, so the worker opens its own previous attempt and fixes it in place\nrather than starting over. A re-brief that lands the fix contains four things, in this order:\n\n1. That this is attempt N of 3, and that the previous attempt is already in the worktree.\n2. The failure output verbatim, passing cases included. Showing which cases passed tells the\n   worker the shape of the rule rather than only that it is wrong.\n3. A one-line diagnosis naming the cause, and saying plainly where the brief was at fault when\n   it was. A worker told its defensible call was reasonable but superseded fixes the code; one\n   told it was simply wrong tends to rewrite more than it needs to.\n4. The requirement, now stated explicitly, with the exact expected values.\n\nThat recipe fixed the rounding task on attempt 2: all eight acceptance cases correct, the\ndocstring rewritten to state the real rule, four tie tests added, and reverting the fix takes\nthe suite red, so the new tests discriminate.\n\nOn the third failure, or on any repeated identical output (a loop):\n\n1. `cmux read-screen --surface \"$S\" --lines 100 --scrollback` and keep the dump.\n2. Close the surface and confirm the close against `cmux tree`.\n3. Revert that task's partial work so the tree returns to a known state.\n4. Re-diagnose at the root. A task that failed three times has a wrong brief or a wrong DAG\n   edge, not an unlucky worker. Fix the plan, and say in the report that you did.\n\nNever dispatch a replacement worker for a task whose brief you have not changed. Re-running an\nunchanged brief is how three strikes get spent on one bug.\n\n## Phase 3: consolidation and verification\n\nRun from the root surface. Paste the output; a remembered pass is not a pass.\n\n```bash\ncd \"$REPO\"\nmkdir -p .sprint && printf '*\\n' > .sprint\u002F.gitignore   # scratch stays out of git status\nmake test > .sprint\u002Ftest.log 2>&1; test_rc=$?\ntail -20 .sprint\u002Ftest.log; echo \"test rc=$test_rc\"\nmake lint > .sprint\u002Flint.log 2>&1; lint_rc=$?\ntail -10 .sprint\u002Flint.log; echo \"lint rc=$lint_rc\"\n[ \"$test_rc\" -eq 0 ] || echo \"SUITE RED: do not report this sprint as done\"\n```\n\nGate on the command, not on a pipe. `cmd | tail && commit` inspects tail's exit code and will\nhappily commit a red suite. Do not reach for `${PIPESTATUS[0]}` as the fix either: the root\nsurface runs zsh, where that expands to the empty string, so the rc check prints nothing and\npasses vacuously. Verified: the same pipeline reports `7` in bash and an empty value in zsh.\nRedirect to a file, capture `$?` on the next line, then read the file.\n\nThen check the invariants and the boundary, one command per line of section 2:\n\n```bash\ngit status --short\ngit diff --stat\ngit diff --name-only | while read -r f; do\n  grep -qx \"$f\" .sprint-allowed-paths || echo \"BOUNDARY VIOLATION: $f\"\ndone\n```\n\nAny file changed that no task listed as an output is an invariant violation. Investigate it\nbefore reporting, and do not launder it into \"incidental cleanup\".\n\nVerify the new tests discriminate. A test suite that stays green when its subject is deleted\nis theater:\n\n```bash\ncp src\u002Fstore\u002Fdb.py \u002Ftmp\u002Fdb.bak && git checkout HEAD~1 -- src\u002Fstore\u002Fdb.py\nmake test 2>&1 | tail -3      # MUST be red\ncommand cp -f \u002Ftmp\u002Fdb.bak src\u002Fstore\u002Fdb.py\n```\n\nCollect each task branch before you verify the whole. The worker leaves its output\nuncommitted in its own worktree, so the root commits it on the task branch and merges:\n\nWorkers differ in whether they commit. Opus workers committed their own output, in RED then\nGREEN pairs; Codex workers left it uncommitted in the worktree. Collection must tolerate both\nor it fails on whichever kind you did not expect:\n\n```bash\nW=\"$REPO\u002F.sprint\u002Fworktrees\u002F$TASK\"\nif [ -n \"$(git -C \"$W\" status --porcelain)\" ]; then\n  git -C \"$W\" add -A\n  git -C \"$W\" commit -m \"feat(scope): $TASK\"\nfi\ngit -C \"$REPO\" merge --no-ff --no-edit \"sprint\u002F$TASK\"\ngit -C \"$REPO\" worktree remove \"$W\" --force && git -C \"$REPO\" branch -d \"sprint\u002F$TASK\"\n```\n\nMerge one task at a time and run the suite after each, so a conflict or a regression is\nattributable to a single worker. `git -C \"$REPO\" worktree list` afterwards must show only the\nmain checkout: a worktree that outlives its task is the litter a blocked worker leaves behind.\n\nClose every remaining worker surface and confirm against the tree.\n\nEmit `SPRINT_COMPLETION_REPORT.md` with: per-task status against the DAG (`DONE` \u002F\n`DONE_WITH_CONCERNS` \u002F `BLOCKED`), the routing breakdown (how many tasks to Sol, how many to\nOpus, how many re-briefed), stall and retry counts, files changed against files planned, the\npasted verification output, and every deviation from `SPRINT_PLAN.md` with its reason.\n\n## Common mistakes\n\n| Mistake | Consequence | Fix |\n|---|---|---|\n| `cmux new-split right --json \\| jq -r '.surface_id'` | empty `SURFACE_ID`; the next `cmux send` with an empty `--surface` targets the wrong pane | `--json` does not exist; `cut -d' ' -f2` the plain-text `OK` line |\n| `codex run --model X --prompt '...'` | no such subcommand, no such flag | `codex exec -m X - \u003C brief` |\n| Parsing the id out of `close-surface` output | you believe a surface closed that is still running | confirm with `cmux tree` |\n| Nesting the brief in quotes inside the send payload | truncated brief on any apostrophe; the worker builds the wrong thing | brief file plus stdin redirect |\n| Passing `--agent chief-operator` to a worker | recursive orchestrator spawning its own splits | root persona stays on the root surface |\n| Polling `read-screen` as the completion signal | you close a pane mid-write, or wait forever on a finished one | `cmux wait-for` token, screen reads for diagnosis |\n| Briefs under `$TMPDIR` | worker cannot find the file across the sandbox boundary | repo-relative brief paths |\n| Running `cmux` sandboxed | `Operation not permitted (errno 1)` misread as a dead socket | disable the sandbox for cmux calls |\n| `make test \\| tail; echo ${PIPESTATUS[0]}` in the root shell | empty rc under zsh, so a red suite reports as passing | redirect to a log, capture `$?` on the next line |\n| An invariant grep that matches the import line, or a lint target with hardcoded filenames | the gate passes vacuously and a broken invariant reads as green | run each check and read its matches; a new output file owns updating the harness |\n| More than three concurrent splits | socket contention and unreconcilable scrollback | topological waves, width 3 |\n",{"name":5,"description":7},"main","26b9a4abffc3b1c9c228a44076c9d3f682ea5b53",".claude\u002Fskills\u002Fcmux-sprint-orchestrator",1,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit","https:\u002F\u002Fgithub.com\u002FArangoGutierrez","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Feduardo-arango\u002F","Senior Systems Software Engineer.\r\n\r\nEnvironmental engineer lost to distributed systems, HPC and cloud computing.","2026-10-13 16:16:20",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add ArangoGutierrez\u002Fclaude-toolkit --skill cmux-sprint-orchestrator",{"tool":39,"command":40},"bunx","bunx skills add ArangoGutierrez\u002Fclaude-toolkit --skill cmux-sprint-orchestrator",{"tool":42,"command":43},"pnpm","pnpm dlx skills add ArangoGutierrez\u002Fclaude-toolkit --skill cmux-sprint-orchestrator","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fcmux-sprint-orchestrator\u002FSKILL.md",true,[47,53],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"README.md","REFERENCE","参考文档",3413,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fcmux-sprint-orchestrator\u002FREADME.md",{"path":54,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":57,"sourceHtmlUrl":58},"scripts\u002Fdispatch-worker.sh","SCRIPT","脚本",3918,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fcmux-sprint-orchestrator\u002Fscripts\u002Fdispatch-worker.sh",[60,67,74,81,88,95],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M7802B5QY9YG45W2HR7ZM17S","update-cursor-settings","Modify Cursor\u002FVSCode user settings in settings.json. Use when you want to change editor settings, preferences, configuration, themes, font size, tab size, format on save, auto save, keybindings, or any settings.json values.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fupdate-cursor-settings\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M780277C2PMVHZQGN6W0VWNB","update-cli-config","View and modify Cursor CLI configuration settings in ~\u002F.cursor\u002Fcli-config.json. Use when the user wants to change CLI settings, configure permissions, switch approval mode, enable vim mode, toggle display options, configure sandbox, or manage any CLI preferences.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fupdate-cli-config\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M78023CDY74JXKCZTDJTYVQF","statusline","Configure a custom status line in the CLI. Use when the user mentions status line, statusline, statusLine, CLI status bar, prompt footer customization, or wants to add session context above the prompt.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fstatusline\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M780209KQZBVMRX43KMZH5TT","split-to-prs","Split current work into small reviewable PRs. Use when the user asks to split a chat, set of changes, branch, or PR.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fsplit-to-prs\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":19},"01M7801XXCWKR3PACKNX01RXMT","shell","Runs the rest of a \u002Fshell request as a literal shell command. Use only when the user explicitly invokes \u002Fshell and wants the following text executed directly in the terminal.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fshell\u002FSKILL.md",[],[],{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":19},"01M7801NF9TC7NWNTZKXQCZXAW","sdk","Guide users building apps, scripts, CI pipelines, or automations on top of the Cursor SDK - TypeScript (`@cursor\u002Fsdk`) or Python (`cursor-sdk` \u002F `cursor_sdk`). Use when the user mentions integrating, installing, or writing code against the Cursor SDK; says `Agent.create`, `Agent.prompt`, `Agent.resume`, `agent.send`, `run.stream`, `run.messages`, `CursorAgentError`, `@cursor\u002Fsdk`, `cursor-sdk`, or `cursor_sdk`; asks to run Cursor agents programmatically from a script, CI\u002FCD pipeline, GitHub Action, backend service, or other code outside the Cursor IDE; wants to pick between local and cloud runtime, configure MCP servers for an SDK agent, or handle streaming, cancellation, or errors; or is wiring Cursor into an automation, bot, or REST `\u002Fv1\u002Fagents` migration. Use eagerly rather than answering from memory; the SDK surface evolves and this skill is the source of truth for the external packages.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fsdk\u002FSKILL.md",[],[]]