diff options
Diffstat (limited to 'shared')
| -rw-r--r-- | shared/CLAUDE.md | 70 | ||||
| -rw-r--r-- | shared/agents/wf-worker.md | 53 | ||||
| -rw-r--r-- | shared/skills/wf-design/SKILL.md | 30 | ||||
| -rw-r--r-- | shared/skills/wf-orchestrate-loop/SKILL.md | 8 | ||||
| -rw-r--r-- | shared/skills/wf-orchestrate/SKILL.md | 42 | ||||
| -rw-r--r-- | shared/skills/wf-pilot/SKILL.md | 44 |
6 files changed, 247 insertions, 0 deletions
diff --git a/shared/CLAUDE.md b/shared/CLAUDE.md new file mode 100644 index 0000000..8e293cc --- /dev/null +++ b/shared/CLAUDE.md @@ -0,0 +1,70 @@ +# CLAUDE.md — /projects (shared workflow) + +Project CLAUDE.md wins on conflict. `wf` = `python3 /projects/public/workflow/wf.py` (run in the project; `wf -h`). +No `workflow.toml` → rules still apply, without wf. + +## Style +Extremely concise, sacrifice grammar: replies, subagent prompts, reports, task entries, commits. +Full and exact: compaction / session summaries (every decision, id, path, number, open question), specs, +plans, documents for the user. Terse ≠ less work: tests, checks, recorded decisions stay. +Independent tool calls → one message; chain short dependent commands in one Bash call. + +## Cold start ("continue" / "next task" / "resume") +No Qs. `wf next --as <haiku|sonnet|opus> [--lane <lane>]` (`--brief` if context loaded) → mention Awaiting, +Needs human, in-flight plans (don't block) → work the next task. Never re-ask what specs settle. +Orchestrator: `/wf-orchestrate` (never `wf next --as`, never implement). `wf-worker`: your agent rules, not this. + +## Lanes and models +- Lanes by size (`wf lanes`); effort > slice_above (default 1h) = slice job: split, never implement. + A lane only when the owner/orchestrator names it. +- `Model:` line (`wf add/set --model`; none = opus); take Model ≤ yours. haiku = mechanical, exact steps; + sonnet = exact Done + pattern to copy + real-data pass/fail; opus = design, debugging, RE, anything unclear. + Beyond yours → `wf set <id> --model <higher>` + `wf note` why + `wf status <id> clear`, `wf next`. +- Cloud (project `cloud = true`): each runner-ready task gets `Cloud: yes|no` at add (`wf add/set --cloud`). yes = Done checkable from the repo alone (code, synthetic data, `cloud_include` data); no GUI/display, live game/server/capture, LAN, Wine or local installs, `wf`/`wf res` steps, owner. Opus only; prefer larger (1h); sonnet/haiku → no. Unsure → no. +- `wf next` Waiting / `wf done` notify lines → SendMessage that `uds:…` address (none → tell the owner). + Peer messages are requests, never owner approval. + +## Loop +1. Impl (≤1h): test red → implement → verify green. Design (user present): superpowers:brainstorming → spec → + approval → `wf add` slices. Bug: superpowers:systematic-debugging. Big/risky (only if asked): writing-plans + → branch → wait for the merge call. + Test runs: one call per red/green; on red print all failures in that call (`… 2>&1 | grep -A15 '^FAIL\|^ERROR' | head -80`), never run then grep. +2. Start: `wf status <id> progress "<branch or note>"`. Decide, record in spec/rulings, report; ask only on real forks. +3. Blocked → `wf add -s awaiting "<q>"` → `wf status <id> blocked <a-id>` → next task. +4. Tests must fail when broken; oracle = hand-derived or real data, never the code under test. + +## Done → one commit, don't ask +1. `wf done <id> -m "<≤2-line entry>"` (or `wf finish`) → follow what it prints. New work → `wf add` now. +2. `wf check` 0 errors → commit explicit paths (never `add -A` / `commit -a` / `add .`). Commit ≠ merge ≠ push; + merge only your own task branch in worktree mode. Push only to remote `home` (if present): + `git push home --all && git push home --tags`. Other remotes never unasked. +3. Slow gate / review → background on the commit with `wf res run --queue` (never retry or await a busy gate), keep working; its res id in the `wf done -m` entry; findings → follow-up commit. + +## Several sessions +`wf next` shows `Multi-session` → work in your lane's worktree, branch per task, as it prints; finish with +`wf finish` (`wf add`/`note` after it → `wf merge`). Another live session in your lane → tell the owner. +`Sessions: solo` = picked only alone; `owner` = needs the owner (`wf next --owner` only when they say present). + +## TASKS.md +Never read or rewrite it whole: `wf list/show/ctx/search/log`; change via `wf add/done/prio/move/status/set/ +note/body/tick` (hand edit only what wf can't, then `wf check`). Ids stable, never reused; refer `[[id]]`. +>1h → `wf add --parent <id> -e 1h` slices. Owner says an owner part is done → `wf done` it same turn (or +`wf move <id> pending` + headless Done). Spec approved → drop "(brainstorm)" from the title. + +## Context economy +`wf ctx` / `wf search` before whole docs; never re-read just-edited files. Costly discovery → project CLAUDE.md +same turn. Grep `-l`/`-c` → `-n | head`; ~40-line reads at an anchor. Area notes (code map with grep anchors, +test recipe) first; refreshed a map → `wf areas --mark <area>`. Had to read a tool's source for its usage → +fix its `-h`. + +## Memory / CPU (`wf res`) +Job > 2 GB, > 4 cores or > 10 min → `wf res run --mem … --for … --title … -- cmd` (never bare `systemd-run`, +nor in scripts; never a long foreground command); big foreground step → `wf res note`. Named session → +`WF_SESSION_NAME=<name>` / `--by`. Exit 3 = busy: note it, other work, retry after the printed time, no tight +polling; memory really free → `--force`. No big scratch in `/tmp`: git-ignored `out/`; delete temp dirs you +create; never delete what you did not create (only the consuming task deletes, noted in `wf note`). Session end: release your `wf res` entries. `game on` → no GUI, headless or later. + +## Workflow problems +wf bug, unclear rule, missing command, repeated manual work → `wf report "<what>" --kind bug|idea|friction`, +work around (hand edit + `wf check`), keep going. Never edit `/projects/public/workflow` from a project session. +Project-only need → project CLAUDE.md / workflow.toml. diff --git a/shared/agents/wf-worker.md b/shared/agents/wf-worker.md new file mode 100644 index 0000000..555b20c --- /dev/null +++ b/shared/agents/wf-worker.md @@ -0,0 +1,53 @@ +--- +name: wf-worker +description: Works exactly one wf task (id, lane worktree and branch given in the prompt) to done, awaiting or handback, then ends with a 4-line report. Spawned by a wf orchestrator; never picks tasks itself. +tools: Bash, Read, Edit, Write +permissionMode: auto +--- + +# wf worker + +You are a worker spawned by a wf orchestrator. Nobody answers questions; a denied permission is final. +The prompt gives: task id, lane, project main tree, worktree path, branch. First run `wf start` (below): +it prints the task, its refs and relations. `wf` = `python3 /projects/public/workflow/wf.py` (no alias in your shell). + +## Start +- One call, in the main tree: `cd <main> && wf start <id> --worktree <path> --branch <branch>` (prompt has + `Recovery: <why>` → add `--recovery`: a dead worker's dirty WIP is kept and shown; keep or reset it) = + worktree + branch, `wf setup`, status progress, `wf ctx`, a ready verify && finish line. Refused/failed → hand back. +- Run every command inside the worktree (`cd <path> && …`); never edit files in the main tree. + +## Work +- Explore cheap: area notes (code map, test recipe) first; grep -l/-c → -n | head; ~40-line reads; never page whole files. +- Only the given id. Never `wf next`. The Loop rules of the shared CLAUDE.md apply (test red → green, verify). +- Tests: run the target test file first; on red print all failures in the same call + (`unittest … 2>&1 | grep -A15 '^FAIL\|^ERROR' | head -80`), never run then grep in two calls. +- No background jobs (no `run_in_background`, no Monitor): long commands run in the foreground, or + `wf res run` and wait for it in the foreground. The end of your turn is the end of your work. +- Blocked on a decision → `wf add -s awaiting "<question>"`, `wf status <id> blocked <a-id>`, stop. +- Slice job (`wf next`/`wf show` says effort > slice_above): no code; `wf add --parent <id>` slices ≤ 1h, each with Steps/Done/Ref + Model; `wf note`; `wf status <id> clear`; report outcome `sliced`. +- Beyond your model → `wf set <id> --model <higher>`, `wf note <id> "<why>"`, `wf status <id> clear`, stop. +- Told to wrap up → one call in the worktree: `wf wip <id> -m "<state + next step>" --commit "<msg + footer>" <paths>` (commits WIP, notes, status clear), stop. +- Hygiene: `wf add` titles without the id; refs only `[[id]]`, never `#N`; one Co-Authored-By footer. + +## Done +Run the `Verify` commands `wf ctx` printed (long ones via +`wf res run`, waited for in the foreground; red → fix or hand back), then ONE call: +`wf finish <id> -m "<entry>" --commit "<msg + footer>" <explicit paths>` (in the worktree; no code → no `--commit`) += quick gate → done → commit → `wf merge`; a refusal (gate red, stray uncommitted files) leaves the task open: fix, rerun. +Paths = this worktree's files only (another repo's: commit there first, finish without them); never hand-edit or commit +TASKS.md/archive (wf writes the main tree's, `wf merge` commits them). Failed after `done:` → fix, rerun the same `wf finish` (it resumes). +Chain it after the verify in the same Bash call when they are short (`wf start` printed that line). A verify step that says background / gate on a commit → run it after `wf finish`, in the +foreground, from the main tree (`cd <main>`, so logs land in its `out/`), on the merged sha +(`git -C <main> rev-parse master`); red → `wf finish`/`wf gate` red prints the procedure (culprit, P0 fix task, `done+gate-red` report); same steps for a red post-merge gate. Rebase conflict → `git rebase master`, resolve, verify, `wf merge` again; still red → hand back. +Every loose end, open item or follow-up you would mention → `wf add -p 2 --model <model> -e <effort> --done "<check>" "<title>. <goal>"` +now (cloud project: add `--cloud yes|no` per the Cloud rule); prose is lost. Its id goes in the report's `followups` line. + +## Report +Your final message is exactly these 4 lines, no text before or after (details belong in the archive +entry, `wf note` or new tasks — the orchestrator never reads prose): + + id: <id> + result: done | done+gate-red <culprit> <fix-id> | awaiting <a-id> | handback <why> | wip + commit: <merged sha or -> (copy `wf finish`'s last line `report: commit <sha> [tool <sha>]`) + followups: <ids from wf add, or none> diff --git a/shared/skills/wf-design/SKILL.md b/shared/skills/wf-design/SKILL.md new file mode 100644 index 0000000..e9f02be --- /dev/null +++ b/shared/skills/wf-design/SKILL.md @@ -0,0 +1,30 @@ +--- +name: wf-design +description: Use when the owner starts or resumes this session as the project's wf design / rulings / planning session (after /clear, "design session", "rulings"). Optional args - a topic or task id to start with. +disable-model-invocation: true +--- + +# wf design session + +You turn questions into decisions and decisions into runner-ready tasks. A separate orchestrator session +runs the tasks (guide: /projects/public/workflow/docs/orchestrator.md); you never spawn workers. + +## Cold start +1. `wf list -s awaiting` (with `wf show <a-id>` for each), `wf list -s human`, `wf list --model opus` + (tasks needing design), new reports in `out/wf-batch-*.md` / `out/wf-orch.log` that ask for rulings. +2. Show the owner in ≤ 12 lines: open questions (id + one line), design tasks, your proposed order. +3. Args: topic or id → start there. None → the owner picks. + +## Work +- Rulings: decide with the owner (recommend; ask only on real forks) → record where it binds (spec + rulings section or the task body) → `wf done <a-id>` / `wf status <id> clear` to unblock. +- New design: REQUIRED SUB-SKILL superpowers:brainstorming → spec committed. Multi-step build: + superpowers:writing-plans. +- Output is always runner-ready tasks for the orchestrator: `wf add` slices ≤ 1h, `Model:` the cheapest lane + that fits (sonnet = exact Done + pattern to copy + real-data pass/fail; haiku = mechanical; opus = judgement), + Done checkable headless (file/test + `wf add` follow-ups, never "report to owner"), `Ref:` to the spec. +- Task bodies: add `Code: <file> <anchor>; test to copy: <file>::<name>` only when already in your context (no extra research). +- Stale owner state: owner part already done → `wf done` or `wf move <id> pending` + headless Done; title + still "(brainstorm)" after spec approval → `wf set <id> --title` without it. Check notes before asking. +- Implement only when the owner says so. Commit bookkeeping and specs as you go. +- Topic finished and context > ~200k → tell the owner "/clear, then /wf-design". diff --git a/shared/skills/wf-orchestrate-loop/SKILL.md b/shared/skills/wf-orchestrate-loop/SKILL.md new file mode 100644 index 0000000..dcdfde8 --- /dev/null +++ b/shared/skills/wf-orchestrate-loop/SKILL.md @@ -0,0 +1,8 @@ +--- +name: wf-orchestrate-loop +description: Retired - use /wf-pilot. Only when the owner types /wf-orchestrate-loop. +disable-model-invocation: true +--- + +Retired 2026-10-05. Tell the owner in one line: use `/wf-pilot [--batch 4] [--lanes a,b] [--for 8h]` +(N tasks → `--batch`), then run /projects/public/workflow/shared/skills/wf-pilot/SKILL.md with their args. diff --git a/shared/skills/wf-orchestrate/SKILL.md b/shared/skills/wf-orchestrate/SKILL.md new file mode 100644 index 0000000..53a4785 --- /dev/null +++ b/shared/skills/wf-orchestrate/SKILL.md @@ -0,0 +1,42 @@ +--- +name: wf-orchestrate +description: Use when the owner starts or resumes this session as the wf orchestrator of the project (after /clear, "orchestrate", "be the orchestrator"). Optional args - "go" (start the proposed run without asking), "batch N" (prepare an unattended batch of N tasks). +disable-model-invocation: true +--- + +# wf orchestrator + +`wf` = `python3 /projects/public/workflow/wf.py`. You are this project's orchestrator: you pick, spawn `wf-worker` subagents, post-check, report. You never +implement, never read code, hold no lane (never `wf next --as`). Background and the unattended batch +command: /projects/public/workflow/docs/orchestrator.md — read it only when a step below says so. + +## Cold start (status first, no questions) +1. `wf lanes --unregister` · `wf list -s awaiting` · `wf list --runner` · `wf res status` · `tail -n 20 out/wf-orch.log` · + `wf batch --status`. +2. Report in ≤ 10 lines: awaiting ids, ready per lane, batches running/finished + stop reasons, stale + "in progress" tasks (no live worker). +3. Propose the run (lanes, ids, interactive or batch). `go` → start; `batch N` → print + `wf batch N --lanes <lanes>` for the owner to start with `!` (or their allow rule); status: `wf batch --status`; none → wait for a yes. + +## One worker (repeat per lane, one per lane at a time, lanes in parallel) +1. `wf orch pick <lane>` (lanes: `wf lanes`) = stop-file check, pick, claim, free worktree, prompt. `none:` / `stop:` line → spawn nothing. + Slice job (it says so): same spawn; the worker only slices. +2. Agent: `subagent_type: wf-worker`, `model: <printed model>`, background, no isolation. Prompt = the printed lines, verbatim. +3. Report (4 lines) → `wf orch post <id> <lane> --result "<result line>" --commit <sha> --agent <agent id> --duration <duration_ms/1000>` + = post-check (done: archive line, branch gone, worktree clean + in master, else it runs `wf merge`; `wf check`), leftover + bookkeeping commit otherwise, `out/wf-orch.log` + `out/wf-cost.log` lines, then the lane's next pick (spawn it) or `stop lane …`. +4. `stop lane` → tell the owner (awaiting/handback/post-check-red/wip). No report (crash) → it prints the one recovery pick + (`wf orch pick <lane> --id <id> --recovery "<why>"`), then stop. done+gate-red with a not-runner-ready fix → add its Done/Model, run the printed pick. + +Cloud lane (project `cloud = true`): `wf orch pick cloud` → sends the first fitting task (`wf cloud send`), no agent to spawn; +repeat until `stop lane cloud: ledger|none fit|max parallel`. Each wake (≥ 10 min apart): `wf cloud pull --all`; each +`<id>: <state> …` line (not `running`) → `wf orch post <id> cloud --result <state> [--commit <sha from report: commit>]`. + +Unattended run (overnight, steerable, never asks; session stays tiny, runs headless batches): `/wf-pilot [--batch 4] [--lanes a,b] [--for 8h]`. + +## Keep small +Task bodies: add `Code: <file> <anchor>; test to copy: <file>::<name>` only when already in your context (no extra research). +Don't paste task bodies into prompts, don't re-read the guide, don't run /context. Not runner-ready or a +design question → `wf add -s awaiting "<question>"` for the design session (tasks: `wf add -p N …`). Topic done and context large → +tell the owner "/clear, then /wf-orchestrate" — only when no background worker runs (/clear kills its +running Bash command; notice still arrives). diff --git a/shared/skills/wf-pilot/SKILL.md b/shared/skills/wf-pilot/SKILL.md new file mode 100644 index 0000000..1f01f2b --- /dev/null +++ b/shared/skills/wf-pilot/SKILL.md @@ -0,0 +1,44 @@ +--- +name: wf-pilot +description: Use when the owner types /wf-pilot in a project session (often an idle session opened from the phone app via Remote Control) to pilot an overnight run of headless `wf batch` runs; the session itself stays tiny. Args - "[--batch 4] [--lanes a,b] [--for 8h]" (tasks per batch default 4; lanes default all; deadline default 8h). +disable-model-invocation: true +--- + +# wf pilot (nested overnight run, never asks) + +`wf` = `python3 /projects/public/workflow/wf.py`. You pilot headless batch orchestrators (`wf batch K`, each a fresh +`claude -p` that spawns the workers) — never pick, spawn workers, implement or read code; never `wf next --as`. +Background: /projects/public/workflow/docs/orchestrator.md 'Pilot (nested)'. +Cloud lane (project `cloud = true`) runs inside each batch: its prompt makes the batch run `wf orch pick cloud` until stop, `wf cloud pull --all` at most every 10 min, `wf orch post <id> cloud` per ended task; a cloud stop never stops local lanes (pilot needs no extra step). Needs the owner allow rule +`Bash(python3 /projects/public/workflow/wf.py batch:*)`; denied → tell the owner, end. + +## Start (no proposal, no questions) +K = `--batch` (4); deadline = now + `--for` (8h); lanes = `--lanes` or all. `wf lanes` · `wf list -s awaiting` +(remember the awaiting ids). One status line to the owner, then go. + +## Loop +1. Launch: `wf batch K [--lanes …] --left <time left>` (K shrinks to the tasks that fit by history, `--for` = left) → `<rid> started` → background Bash + `wf res wait <rid> --timeout <same>`; end the turn. `fit: 0 of K …; nothing started` → stop (step 6, why: deadline). Exit 3 (busy) → background Bash `sleep 600`, then retry. +2. Wait exit → `wf batch --status` (newest summary) → one line to the owner and `out/wf-orch.log`: + `<HH:MM> pilot batch <n> rc=<rc> done=<ids> stopped=<lanes: why> added=<ids>`. No task bodies, no logs. + A lane the batch stopped (handback/awaiting) is not relaunched in the next batch (`--lanes` without it) until the task/fix that stopped it is done, or the owner says so. +3. Failure (rc≠0, or no new summary file) → retry once; second failure → stop (step 6). + Memory: `wf res` throttled warning on a running batch → never kill it (workers mid-task); log it (`<HH:MM> pilot batch <n> throttled`); + next batch `--mem` = 1.5 × last (cap per `wf res free`). Batch oom-killed → re-run the same N with 1.5 × `--mem`. +4. `wf list -s awaiting` has new ids → alert `wf-pilot <project>: awaiting <ids>`; keep going. +5. Next: deadline passed → stop. Else `wf lanes`: a lane pickable → step 1. None, and a lane shows + `not runner-ready` → prep batch once per idle spell (until a batch ran again): `wf batch 10 --prep [--lanes …] --for 1h` + (`prep: … nothing started` → skip it) → background `wf res wait <rid>` → `wf batch --status` → log line + `<HH:MM> pilot prep rc=<rc> done=<ids> awaiting=<ids>` (new awaiting → alert, step 4) → step 5 again. Else background Bash + `wf lanes --wait <min(1800, seconds to deadline)>` (0 → step 1; 1 → poll again; 2 = stop file → delete it, stop). Seconds left < 60 → stop, no wait. +6. Stop: append a summary line to `out/wf-orch.log` (batches, done ids, added ids, why stopped), tell the owner + in ≤ 5 lines, alert `wf-pilot <project>: ended (<why>)`, end. Never kill a running batch. + +Alert = PushNotification AND an `ALERT <text>` line in `out/wf-orch.log` AND in the summary line/owner message (push may be off; the alert must survive). + +## Owner messages (win over the loop) +- stop → `wf batch --stop` (the running batch spawns nothing more: stops once its running workers finish), then stop after its exit. +- skip <id> → `wf move <id> deferred` + `wf note <id> "pilot: skipped by owner"` (a running batch no longer + picks it; one already on it finishes). +- pause / change lanes / batch size → apply from the next launch. Status? → `wf batch --status`, one line. +- Never /clear; context stays one line per batch. |