Changes (newest first)

2026-10-07 wf res queue coalesces gates: a queued locked job on commit C (--commit REV, default a hex word of a locked job's title, gate 1a2b3c4) supersedes same-lock queued jobs on ancestors of C (they end superseded by r-N, it keeps the earliest turn, ledger covers); one on an ancestor of a queued job is covered, not queued. A batch of N commits → ≤ 1 queued gate per lock. wf res wait follows superseded ids; a red coalesced job names the range and bisect. Gate red procedure mentions it. Projects: a gate script that greps its own log by sha should wf res wait the printed id (covered/superseded logs are never written).
2026-10-07 wf res: a job started from inside a live job (by.batch, e.g. a batch worker's gate) borrows the parent's claim (mem + cpus, lent once) when fitting, so a batch waiting on its own queued gate no longer deadlocks; a queued entry whose batch ended is dropped at prune (r-N dropped (batch r-M gone)), never heads the queue. Projects: nothing.
2026-10-07 Lane cloud: one stuck task no longer stops it: wf orch post <id> cloud with awaiting / handback / lost parks only that task (awaiting: park as local lanes; handback / lost: Cloud: no + status cleared + note, local lanes may retry with pull's Recovery note), prints alert: …; lane cloud keeps picking, then the next cloud pick (never the same id). Projects: nothing.
2026-10-07 Owner-absent default: workers never wait; owner-only step → wf add -s human + wf set <id> --after <h-id> + status clear, result needs-owner <h-id>. wf orch post parks it (After: the h-task, claim cleared, alert), lane keeps picking; owner's wf done <h-id> makes it pickable again. Worker def, shared CLAUDE.md, wf-orchestrate, wf-pilot, batch prompt, docs. Projects: nothing.
2026-10-07 One stuck task no longer freezes its lane: wf orch post with awaiting / handback / post-check-red (local lanes; lane cloud unchanged: pull re-queues) parks only that task (blocked on the a-id named in the result, else Sessions: owner + an orch … note; handback also clears its claim), commits, prints alert: <id> <outcome> → tell the owner (…); lane <lane> keeps picking, then the lane's next pick. stop lane only for none / stop file / push-failed / wip / no report. Batch prompt, wf-orchestrate, wf-pilot, docs: keep picking, never wait on the owner while tasks are pickable. Projects: nothing.
2026-10-07 wf orch post done: leftover TASKS/archive changes in the main tree are committed (<id> done (code <sha>) in a split project, else <id> done (orchestrator)); split: a --commit sha not in the code repo (stale private HEAD) is replaced by the code repo's master sha (noted). wf-worker: the report's commit = first sha of report: commit …, never a looked-up HEAD. Projects: nothing.
2026-10-07 wf start refuses an existing worktree of another repo; .wf-home ignored by finish/merge stray checks. Projects: nothing.
2026-10-07 split projects 6/6: wf check split rules (cloud error; warns: no push, .wf-home not git-ignored, merged code worktrees); docs. Projects: split projects: push key, .wf-home in the code repo's .gitignore (migration task in the workflow project).
2026-10-07 split projects 5/6: wf orch pick/post use the code repo (worktrees, branch/merge checks, commit sha); .wf/push-failed after a done worker stops the lane (push-failed). Projects: nothing.
2026-10-07 split projects 4/6: workflow.toml push (cwd root, env WF_MAIN), wf push, .wf/push-failed marker (wf next shows it); finish report gains push-failed <rc>. Projects: split projects set push (follow-up migration task).
2026-10-07 split projects 3/6: wf finish / wf merge / wf wip across code + private repo: paths under the private root are committed there (with TASKS/archive, message <msg> (code <sha>)), code paths on the branch; no code paths → no code commit; report commit <code|-> books <private>; wf done merge advice names the private books. Projects: nothing.
2026-10-07 split projects 2/6: wf start makes code worktrees (branch in the code repo, .wf-home pointing at the private project); wf next multi-session advice names wf start in the code repo. Projects: nothing.
2026-10-07 split projects 1/6: Config.split, push key, .wf-home discovery. Projects: nothing.
2026-10-07 Shared push rule: in a split project the public code_root repo is pushed only via the project's publish command, never git push home --all (that pushes local review branches unscanned). Projects: split projects name their publish command in their CLAUDE.md.
2026-10-07 Denylist guard at commit time: .githooks/pre-commit (staged lines + paths) and .githooks/commit-msg run scripts/denylist_check.py against the git-ignored publish-denylist.local (worktree top + main tree, plus git config denylist.file lists; word / !token exemption / #, case-insensitive; output word #<line> only; no list → warning). Enable per clone: git config core.hooksPath .githooks; audit: denylist_check.py tree. scripts/publish_snapshot.py removed (history is public now). Projects: nothing.
2026-10-07 Tests use generic fixtures (/h/u, 10.0.0.x) instead of real home-dir and LAN paths (publish-scan clean). Projects: nothing.
2026-10-06 Tool path is now /projects/public/workflow (was /projects/workflow): docs, shared CLAUDE.md, skills, wf-worker agent, messages. wf projects / wf usage --report scan the grandparent when the parent holds no project. Projects: point ~/.claude symlinks/settings, wf-res.service and any script calling wf.py at the new path.
2026-10-06 New scripts/publish_snapshot.py [DEST]: scrubbed fresh-history snapshot of the master tree into a separate local repo (default ~/src/wf-public; drops inbox.md, .worktrees/, pycache/, out/); refuses on any word of the git-ignored publish-denylist.local in a path or file; one commit per run (new CHANGES lines as message); never pushes or adds a remote (docs/manual.md). Projects: nothing.
2026-10-06 docs/manual.md: owner-facing manual (setup, daily use per session type, writing good tasks, best practices, troubleshooting); linked from the README top (replaces 'no installation guide'). Projects: nothing.
2026-10-06 wf orch post: --duration 0 (orchestrator lacked duration_ms) falls back to the time since the pick like an omitted one (was logged 0m00s, starving the task p90 of wf batch --left/--time-left); a post-check-red post keeps the orch record, so the re-post after the fix logs the real duration. Projects: nothing.
2026-10-06 wf res entry project = main tree name (git common dir's parent), not the lane worktree dir (was 'slow'/'fast-2': history kinds split per worktree); wf res hist, history hints, wf batch --status use it too. Older history records keep their worktree name. Projects: nothing.
2026-10-06 wf batch K --left DURATION: K' = min(K, floor(left / task p90)) (p90 of done/handback durations in out/wf-orch.log, local lanes, > 0; < 3 runs → 30 min), prints fit: K' of K tasks (…), K' = 0 → nothing started (rc 0); --for defaults to left. New wf batch --time-left DEADLINE → time left X, task p90 Y (n runs): spawn|stop; batch prompt runs it before each round (deadline = start + ceiling) and stops with 'stopped: deadline (…)'. wf-pilot launches with --left <time left> (was --for <time left, ≥ 1h>: a pilot of a few hours closed ~1h early). Projects: nothing.
2026-10-06 wf res reservations from history: finished runs (rc=0, peak) go to ~/.local/state/wf/resources-history.jsonl when pruned from the ledger; new wf res hist [--project P] = per project and title kind (title minus trailing sha/hex words and (…)) n, request, peak p50/p95, estimate, duration p50/p90, suggestion (≥ 3 runs: p95 peak ×1.15 ≥ 0.2 GB, p90 duration ×1.5 ≥ 5 min); run/note asking > 2× it print one hint: history says ~X GB / Y min (n runs) line on stderr (no auto-resize); wf batch default --mem/--for = the project's wf-batch suggestion when it exists (ceiling stays 6h; not --prep). Projects: nothing.
2026-10-06 wf batch cloud tail: when the orchestrator exits with .wf/cloud records left, the batch sidecar keeps pulling every 10 min (wf orch post … --no-pick, no picks/sends) until none remain, out/wf-batch.stop or 24 h (batch-sidecar --cap); the job's ledger entry shrinks to 0.2 GB / 1 cpu (via WF_RES_ID; unit MemoryMax unchanged); summary line cloud tail ended: <why>; <n> pulls; left: …. Cause: tasks sent at batch end stayed unpulled. Projects: nothing.
2026-10-06 wf batch on a cloud = true project pulls cloud tasks itself: the job runs wf_res.py batch-sidecar around claude -p, every 10 min while it lives (until the stop file) wf cloud pull --all → per ended task wf orch post <id> cloud --result <state> [--commit <sha>] --no-pick + a (sidecar pull) summary line; batch prompt: the orchestrator only picks cloud, never pulls (pulls starved while it blocked on foreground workers). wf cloud pull: flock on the task record, a second concurrent pull of one id prints <id>: pulled by another wf cloud pull, skipped (rc 0) instead of branch … exists (local WIP?). Projects: nothing.
2026-10-06 Cloud routing per task: wf add --cloud yes|no (like wf set --cloud); Cloud: yes still needs Model opus (sonnet/haiku never fit, was: yes overrode Model) and skips only the regex list; wf orch pick cloud order = Cloud: yes first, then prio, then larger effort (was: opus first). shared/CLAUDE.md "Lanes and models": Cloud bullet (classify at add); wf-worker follow-ups get --cloud in cloud projects. Projects: with cloud = true, mark runner-ready opus tasks Cloud: yes|no.
2026-10-06 wf finish no longer half-finishes: a PATH outside the repo (e.g. a tool/other-project file), missing, unchanged (lane worktree) or the worktree's own TASKS/archive copy is refused before done; an id already archived resumes (commit + merge) instead of unknown id; a git add/commit failure after done says rerun resumes. wf done/wf finish in the main tree refuse a task in progress: <branch> checked out in a linked worktree (run it there, or wf status <id> clear first). wf-worker.md: paths = worktree files only, never hand-edit/commit the books. Cause: workers passed other-repo paths → done ran, commit failed, retry unknown id → books hand-edited on the branch / left uncommitted in the main tree. Projects: nothing.
2026-10-06 wf res run --lock KEY: one queued/running job per KEY and main tree (lane worktrees share it); titles starting with gate lock gate unasked, so overlapping gates in one shared gate checkout are refused (exit 3 busy: lock 'gate' held by r-N …, also with --force) or wait with --queue (queued lock-blocked entries are skipped by the FIFO, not blocking others). Ledger field lock. The 'ambiguous argument %s' in a-game-project-builder gate logs r-671/r-672 = nested \"%h %s\" quoting in a hand-written bash -c copy of gate-bg.sh (no wf/systemd expansion; test pins verbatim argv). Projects: nothing (gate scripts titled gate … get the lock).
2026-10-06 wf cloud pull archives each session it ends (done/awaiting/handback/lost) in the claude.ai app via the CLI's internal POST /v1/code/sessions/<sid>/archive (login token read in-process, never printed); failure → one archive <sid> failed: …; archive by hand (wf cloud archive --ended) line, pull's exit code unchanged. New wf cloud archive <sid>|--ended. Ledger rows get archived: true. Projects: nothing.
2026-10-06 wf cloud pull: no ● WF-RESULT but the last assistant message ends in a bare WF-PATCH-END (session dropped the key words) → handback bad result: no WF-RESULT key (usage charged, patch kept when a sha256=… bytes=… header decodes) instead of running until 24h lost. Projects: nothing.
2026-10-06 Cloud prompt: final-message lines written literally (WF-RESULT <…>, …; the [WF-RESULT R] form made a live session drop every key word → pull saw no result), the usage script prints the WF-USAGE line and the patch command prints the WF-PATCH-BEGIN sha256=… bytes=… / base64 / WF-PATCH-END block to copy; commit only needed paths (no pycache/bin/obj). wf cloud send --dry-run snapshots into out/cloud/<id>.dry-run: it no longer deletes the live snapshot of a sent task. Projects: nothing.
2026-10-06 wf orch pick cloud [--id ID]: virtual lane cloud = stop file, ledger check (stop lane cloud: ledger|max parallel (…)), first task across all lanes that is runner-ready, not in progress and fits the cloud rules (opus first; none → stop lane cloud: none fit (…)) → wf cloud send, orch record lane cloud (task_lane, branch); wf orch post <id> cloud --result <state> after wf cloud pull --all (branch check uses the recorded <task lane>/<id>). Local lanes already skip in progress: cloud:*. docs/orchestrator.md "Cloud lane", wf-orchestrate skill. Projects: opt in with cloud = true.
2026-10-06 wf cloud pull <id> | --all [--export FILE] [--no-push]: teleport + /export, parse the final ● WF-RESULT message (wrapped header/base64 rejoined, sha256 + bytes checked, gunzip); done → patch paths checked (outside the project / TASKS / archive / .wf / .worktrees / out / cloud_include refused), git am --3way on <lane>/<id> from the recorded master in a free lane worktree, worktree_setup, quick_gate, wf done -m "<WF-REPORT> (cloud <sid>)" + wf merge, report: commit <sha>; awaiting → wf add -s awaiting + blocked; handback / red gate / bad patch → note Recovery: cloud attempt <sid> — <why>, status clear, patch kept in out/cloud/<id>.patch; no result → running (exit 4), > 24h → lost. Every end charges the ledger and deletes snapshot + record; branch exists / worktree setup failed → exit 1, session kept. wf.lane_worktree shared with orch pick. Projects: nothing.
2026-10-06 wf cloud send <id> [--dry-run] [--project DIR]: snapshot of the project's master (git archive minus TASKS/archive/.wf/.worktrees/out, plus cloud_include) in out/cloud/<id> as a one-commit repo, 90 MB packed cap, prompt templates/cloud-prompt.md (task + area notes + cloud_note, WF-RESULT/REPORT/USAGE/PATCH final-message rules, usage script), ledger reserve (refused: exit 3), claude --cloud, claim in progress: cloud:<sid>, record .wf/cloud/<id>.json. cloud.final_message(export) = the marker parser (only an assistant ● WF-RESULT message). New workflow.toml keys cloud, cloud_include, cloud_note. Projects: nothing (opt in with cloud = true).
2026-10-06 Cloud pty driver (library, no command yet): wf_cloud.send(folder, prompt) -> sid runs claude --cloud on a pty (pre-accepts the trust dialog in ~/.claude.json, Down+Enter fallback), wf_cloud.export(folder, sid, out) -> path = --teleport → /export (Enter retries) → /exit, teleport retried once; cloud.strip_ansi/parse_sid/trust pure. Env WF_CLAUDE, WF_CLAUDE_JSON. Projects: nothing.
2026-10-05 wf wip <id> -m NOTE [--commit MSG] [PATHS] (lane worktree): wrap-up in one call = commit paths, note, status clear; wf-worker prompt uses it. Projects: nothing.
2026-10-05 Lane worktree config: wf run inside a linked worktree whose copy of the project has a workflow.toml reads worktree_setup, quick_gate, areas, area_stale_commits, area_ignore from that copy, and the areas file (default CLAUDE.md) resolves there: a branch can test its setup/gate/area edits before merge; wf areas --mark writes the worktree copy (commit it on the branch). Tasks/archive/ledgers/other keys stay the main tree's. New config.load_at / find_local. Projects: nothing.
2026-10-05 wf orch pick <lane> [--id ID] [--recovery WHY] (stop file, runner pick skipping in-progress, claim, free worktree .worktrees/<lane>/-2… (no live session cwd, no other picked worker, clean detached or own branch), prints the wf-worker prompt; record .wf/orch/<id>.json) and wf orch post <id> <lane> --result LINE [--commit] [--agent] [--duration] [--no-pick] (post-check: archive line, branch gone, worktree clean + in master else wf merge, wf check; leftover commit otherwise; out/wf-orch.log + wf-cost.log lines; model-raised detection; next pick or stop lane). wf-orchestrate 'One worker', docs/orchestrator.md, batch prompt use them. Projects: nothing.
2026-10-05 wf res ledger fix: entries vanished (ledger moved to .bad, restarted empty at r-1) when a process running older code (a wf res wait started before the by-field release) read entries with a key it did not know. Unknown entry keys are now ignored; a corrupt ledger restarts with its salvaged id counter; starting a job removes stale logs/<id>.rc/.peak. Projects: nothing.
2026-10-05 wf finish prints report: commit <sha> [tool <sha>] (+ --tool-commit); wf-worker.md Report says copy it. Projects: nothing.
2026-10-05 wf ctx <id> ends with the notes (code map, test recipe…) of every area the task names (area name, code-map anchor or Paths entry, whole word). workflow.toml code_root: areas' anchors / Paths / Checked refer to another repo (wf areas, --mark, check, stale t-map tasks; no uncovered nudge). Projects: nothing.
2026-10-05 /wf-orchestrate-loop retired (duplicate of /wf-pilot): the skill only points to /wf-pilot for one release, then is removed; docs/orchestrator.md section dropped. .gitattributes: CHANGES.md merge=union (parallel releases no longer conflict on the top line). Projects: nothing.
2026-10-05 wf start <id> --worktree PATH --branch B [--recovery] (main tree): worker setup in one call = worktree add/switch (dirty refused; --recovery keeps a dirty WIP on B and prints it with master..HEAD), worktree_setup, status progress B, wf ctx, a ready cd wt && (verify) && wf finish … line. wf-worker.md Start section → that one call. Projects: nothing.
2026-10-05 shared/CLAUDE.md trimmed 7.1 → 4.9 KB (same rules, less wording; TASKS item format left to wf check; new line: batch independent tool calls). Projects: nothing.
2026-10-05 wf res entries name their owner: ledger field by (name: --by on run/note, else WF_SESSION_NAME, else <lane> session from .wf/sessions by CLAUDE_PID; task: WF_TASK, else <lane>/<id> branch; batch: WF_RES_ID, now set in every job unit; address: uds:<messaging socket>). wf res status lines end [by … , message uds:…], throttle warning ends ; started by …. Projects: nothing.
2026-10-05 wf res run/note --force: fit against MemAvailable − user reserve − headroom only (ledger claims + queue ignored; entry still recorded); busy line ends with '; X GB really free beyond the reserve: --force starts it past the ledger (…)' when --force would fit; refused → 'busy even with --force: …' exit 3. wf batch --prep / wf prep default --mem 1G (was 4-8G). Projects: nothing.
2026-10-05 Batch stop at once: batch prompt, wf-orchestrate-loop, wf-orchestrate 'One worker' check out/wf-batch.stop before EVERY spawn (not per round; running workers finish, latency ≤ the running tasks); batch prompt gets the absolute stop path; wf lanes --wait exits 2 ('stop requested') as soon as the stop file exists; pilot/loop treat exit 2 as stop. Projects: nothing.
2026-10-05 Uncovered-area nudge: when a project's areas use - Paths:, wf done adds t-map-<folder> (≤3, once per id, also not when archived) for folders (≤2 levels) the task's diff (branch since merge-base with master/main + uncommitted + untracked; bookkeeping files skipped) touches outside every area's Paths (an area without Paths: folders of files holding its Code map anchors); wf areas prints uncovered: <folder> (n files) lines. New workflow.toml area_ignore (default tests/test/docs/doc; dot folders and root files always skipped). Projects: nothing (optional area_ignore).
2026-10-05 wf finish <id> -m ENTRY [--commit MSG PATH…] = quick_gate → done → commit explicit paths → wf merge in one call (main tree: TASKS/archive join the commit); refuses before done on gate red / stray uncommitted files. wf ctx <id> prints a Verify (run before wf finish): line. wf-worker.md Done uses finish. wf usage --explore: git worktree list / git branch --list no longer count as bookkeeping. Projects: nothing.
2026-10-05 Prep batch: wf batch K --prep [--lanes …] = headless run (templates/prep-prompt.md) where sonnet subagents write Done (+Model) for up to K pending tasks without Done (not owner-bound, no status/open slices; pickable first); unclear → awaiting + blocked; summary lines id → Done: … | id → awaiting a-…; none → 'nothing started'. New wf set <id> --done TEXT ("" removes). wf-pilot / wf-orchestrate-loop run a prep batch when idle with not-runner-ready tasks (once per idle spell). Projects: nothing.
2026-10-05 wf-pilot / wf-orchestrate-loop: every PushNotification also logged as 'ALERT ...' line in out/wf-orch.log + summary; docs/orchestrator.md names the push setting (agentPushNotifEnabled). Projects: nothing.
2026-10-05 wf batch writes the out/wf-batch-*.md "# wf-batch <project> <time> (N, lanes)" header itself before launch (also --dry-run); prompt only appends. Projects: nothing.
2026-10-05 wf merge on a detached-HEAD lane worktree with commits not in master now rebases + ff-merges them (message 'bookkeeping', output 'merged detached HEAD'), never only the bookkeeping; dirty worktree refused in both modes; wf done there warns 'detached HEAD has N commits not in master'. Orchestrator post-check (wf-orchestrate, docs/orchestrator.md): worktree HEAD ancestor of master. Projects: nothing. (wf check unaffected.)
2026-10-05 wf usage --log ... --duration <s> writes dur= (agent wall time); wf usage --report adds med_dur column; wf-orchestrate / loop pass duration_ms/1000. Projects: nothing.
2026-10-05 Stopped lane (handback/awaiting) resumes only after the task/fix that stopped it is done, or the owner says so: wf-orchestrate-loop, wf-pilot, batch prompt text. Projects: nothing.
2026-10-05 wf add --done "<text>" writes the Done line (follow-ups born runner-ready); wf add -p 0 without Done warns on stderr naming --done; wf-worker adds follow-ups / gate-fix tasks with --done. Projects: nothing.
2026-10-05 wf merge run again in the (detached) worktree after a merge commits TASKS.md/archive changes made since (late notes, follow-ups; message 'bookkeeping'); nothing dirty → still 'nothing to merge'. Workers: run wf merge once more after post-merge wf add/note. Projects: nothing.
2026-10-05 Gate-red outcome: new optional workflow.toml quick_gate = [...] + wf gate (runs it in the current tree, red = exit 1); wf-worker runs wf gate before wf done. Async slow gate red not caused by the worker's diff → worker adds a runner-ready P0 fix task and reports done+gate-red <culprit> <fix-id>; orchestrate / loop / batch prompt treat it as done and pick the fix next (lane not stopped); wf usage --report counts it as done. Projects: optional — set quick_gate to a fast regression check (e.g. contract tests).
2026-10-05 wf lanes / lanes --wait count runner-pickable (= list --runner --lane: needs a usable Done); not-ready ones shown as "· N not runner-ready (no Done)". wf next lanes block unchanged. Projects: nothing.
2026-10-05 New skill /wf-pilot [--batch 4] [--lanes a,b] [--for 8h]: nested overnight pilot running wf batch K in a loop (wait via wf res wait, idle-poll wf lanes --wait, stop via wf batch --stop, PushNotification on stop/awaiting/end); docs/orchestrator.md 'Pilot (nested)'. Install: ln -s /projects/workflow/shared/skills/wf-pilot ~/.claude/skills/. Projects: nothing.
2026-10-05 wf batch --stop: creates out/wf-batch.stop; batch checks it before each round, stops ("stopped: stop file"), removes it; --status shows it. Projects: nothing. (wf check unaffected: no change to check code.)
2026-10-05 New skill /wf-orchestrate-loop [N] [--lanes a,b] [--for 10h] (replaces the loop N argument of wf-orchestrate, same release day): piloted loop, typed into any idle project session (e.g. phone Remote Control); docs/orchestrator.md 'Piloted loop session'. Install: ln -s /projects/workflow/shared/skills/wf-orchestrate-loop ~/.claude/skills/. Projects: nothing.
2026-10-05 /wf-orchestrate loop N [--lanes a,b] [--for 10h]: piloted overnight loop mode (background workers, never asks, idle-poll via wf lanes --wait, summary in out/wf-batch-*.md); docs/orchestrator.md 'Piloted batch session'. Projects: nothing (skill symlink picks it up at next session start).
2026-10-05 wf lanes --wait SECS: block until any lane has a pickable task (exit 0) or timeout (exit 1); prints the lanes line either way. Projects: nothing.
2026-10-05 wf body -h names stdin, kept lines (Model/Sessions/After/Ref), heredoc example; wf body ID text errors with stdin hint. Projects: nothing.
2026-10-05 wf list --runner no longer lists sliced-out parents (effort > slice_above, slices all archived); row set matches wf lanes pickable. Projects: nothing.
2026-10-05 Model-lane leftovers removed: wf batch --lanes <model> warning (a model name is now just an unknown lane), old {haiku,sonnet,opus}.json registry cleanup. docs/model-lanes.md folded into docs/lanes.md (now a current reference: Model line §0, registry/claims §4). Projects: nothing.
2026-10-05 Speed lanes: lanes by size instead of model ([lanes.*] in workflow.toml; default fast = <1h unblock-first + slice jobs, slow = 1h priority-first, falls back to fast), slice_above (default 1h: bigger = slice job, split never implement), sessions take Model ≤ their own (wf next --as M [--lane L]), wf lanes, orchestrator/worker/batch by lane, wf areas (code-map anchors, staleness, --mark) + wf done auto-adds t-map-<area> refresh tasks. Old: --lanes <model> warns; old model registry files dropped. Spec: docs/lanes.md. Projects: rollout — area sections get Paths: + first wf areas --mark, remove merged model worktrees.
2026-10-05 wf setup (in a lane worktree) runs new workflow.toml key worktree_setup = [cmds] (cwd = worktree, env WF_MAIN = main tree; first non-zero exit stops, exit 1). wf-worker runs it after worktree create/reuse; wf next Multi-session hint appends && wf setup when set. Projects: optional — list commands that provide git-ignored inputs (e.g. link out/… from $WF_MAIN); must be idempotent.
2026-10-05 wf init also seeds CLAUDE.md (templates/CLAUDE.md: Layout · Loop additions · Gotchas · Areas with code map / test recipe / tools / worktree setup); kept if it exists. Projects: nothing.
2026-10-05 wf usage --explore: per subagent of a session, % billed input exploring / before the first edit, median calls before it, result tokens per tool (grep/Read/sed-cat/…), files read by >= 2 agents. Projects: nothing (replaces per-project session-cost scripts).
2026-10-05 re-learning rules: shared CLAUDE.md grep discipline + code-map/test-recipe notes, worker/design/orchestrate lines; wf done always prints a checklist line "re-learned anything…". Projects: keep a code map + test recipe per area in their notes.
2026-10-05 wf usage --effort accepts only task estimates (<1h|1h|5h|10h|100h), not reasoning effort. Projects: nothing; fix old cost-log lines with effort=medium by hand.
2026-10-05 wf usage: subagent output tokens were undercounted (transcripts keep only the stream-start placeholder); now estimated from content (fit ±5% per session), shown ~, cost-log lines get est=N. Projects: nothing; older out/wf-cost.log lines undercount out/usd.
2026-10-05 wf check warns about tool worktrees (wf repo .worktrees/*) whose branch is merged into master (leftover after release; WF_TOOL_ROOT overrides the repo). Projects: nothing.
2026-10-05 wf lanes --unregister drops this session's lane record (orchestrator holds none); wf-orchestrate cold start runs it. Projects: nothing.
2026-10-05 wf list --runner [--model M] = ready + Done line, no Sessions: owner, Done not about owner/report to/confirm, no open slices. wf add of a task without Done prints a stderr hint. Projects: nothing. Orchestrator uses --runner.
2026-10-05 wf list --stale lists in-progress tasks with no live claim (dead holder or none); wf status --clear-stale clears them. Projects: nothing. Orchestrator cold start may use it.
2026-10-05 wf next / wf done: last line context ~Nk tokens (> 100k): ask the owner to /clear, then continue when this session's last prompt is over ctx_hint (workflow.toml, tokens, default 100000, 0 = off). Projects: nothing (optional key).
2026-10-05 wf batch N [--lanes …] [--model] [--for] [--mem] [--dry-run] starts the unattended batch orchestrator (claude -p under wf res run, bg-wait ceiling = --for, prompt in templates/batch-prompt.md); wf batch --status. Owner: allow rule Bash(python3 /projects/workflow/wf.py batch:*) once. Projects: nothing.
2026-10-05 docs: /clear safe only with no background worker (it SIGKILLs the worker's running Bash; notice still arrives). orchestrator.md Context + wf-orchestrate skill. Projects: nothing.
2026-10-05 wf usage: tokens + API-price $ per agent of a session (--session, --agent, --since), deduped per request (transcripts repeat usage per content block: earlier 153 turns / 8.9M cr = 86 / 5.0M); --log TASK OUTCOME --effort E appends to out/wf-cost.log; --report = lane/effort table over all projects. Orchestrator skill step 6 logs the cost line. Projects: nothing to do.
2026-10-04 shared CLAUDE.md compacted (orchestrator → /wf-orchestrate, reflowed); no rule changes.
2026-10-04 wf add --parent no longer chains After to a deferred slice; rules: never delete what you did not create, owner-done parts / approved "(brainstorm)" titles updated same turn, post-merge gates run from the main tree, batch orchestrator wf-adds follow-ups instead of asking.
2026-10-04 Pilot 4 fixes: wf-worker 4-line report with followups (loose ends → wf add), background gates run after wf merge on the merged sha, Recovery: prompt mode, wf spelled out; spawn prompt ends with "Final message: the 4 report lines only."; status-note parentheses error has a hint.
2026-10-04 Skills /wf-orchestrate [go | batch N] and /wf-design [topic | id] (shared/skills/, symlink into ~/.claude/skills/): cold start of an orchestrator or design/rulings session after /clear.
2026-10-04 Orchestrator + workers (docs/orchestrator.md): project lock .wf/lock around every wf write (parallel sessions/workers safe); wf merge in a lane worktree = rebase, ff-merge, TASKS/archive commit, detach, push home, under the lock (wf done there now points to it); wf-worker subagent definition shared/agents/wf-worker.md (install: symlink into ~/.claude/agents/). Shared rule: an orchestrator session holds no lane and spawns workers. Projects: out/ git-ignored for the orchestrator log.
2026-10-04 Sessions field: task line Sessions: parallel|solo|owner [— why] (none = parallel; - Sessions: note form works too; wf add/set --sessions, wf list mark [solo]/[owner], wf check validates). wf next skips solo while another session is live; a solo task in progress → other sessions pick nothing; its wf done notifies them. Owner tasks only with wf next --owner. Header flag , interactive = owner (wf check warns: wf set <id> --sessions owner); --interactive = old spelling, one more release. Projects: nothing required; replace interactive flags when wf check warns.
2026-10-04 Multi-session (docs/multi-session.md): wf status … progress claims the task for the session (.wf/claims/); wf next skips tasks held by another live session and warns when another live session holds your lane. More than one live session → wf next prints a Multi-session block: work in .worktrees/<lane> on a branch <lane>/<task>. wf run in a linked git worktree uses the main tree's project (TASKS.md, archive, docs, .wf). wf done there prints the merge-back steps; ff-merging one's own task branch is allowed. Shared rule: commit explicit paths, never git add -A / commit -a. Projects: .worktrees/ git-ignored.
2026-10-04 Model lanes (docs/model-lanes.md): task line Model: haiku|sonnet|opus (none = opus; wf add/set --model, wf list column + --model, wf check validates). wf next --as <model> picks only that lane (no --as = haiku, with a hint), registers the session in .wf/sessions/ (self-ignored) and shows other lanes' sessions and what waits on them; new wf lanes; wf done names the lanes to notify. Pick order for everyone: inherited priority (a task counts at the best P of all that wait on it), then work other lanes wait on, then more waiting. Shared rule: cold start wf next --as <your model>, own lane only. wf res run keeps the caller's environment (variables the user manager lacks; secret-named ones are not stored). Projects: flag tasks with wf set <id> --model … (plain - Model: sonnet ok notes → check warning).
2026-10-03 shared rule: agents may push to remote home (the owner's private server) after a Done commit or a group of commits; every other remote still only when asked. Projects: nothing (repos without home unchanged).
2026-10-02 wf res: each Claude session in agents.slice is capped at session_mem_gb (default 6 GB, MemoryHigh: throttled, not killed) by the timer tick; status warns when a session sits at its cap. A runaway foreground test can no longer take all RAM. Projects: big work via wf res run (as before).
2026-10-02 wf res automatic clean also sweeps /tmp and /var/tmp litter: own empty dirs idle 2 h, .NET debug pipes of dead processes (thousands accumulated). Files with content are untouched. Projects: tests should still delete their temp dirs.
2026-10-02 wf res run --queue says how to cancel (cancel: wf res release r-N); release -h too. Projects: nothing to do.
2026-10-02 wf res clean: never deletes the scratch dir of a running Claude session, however idle (was: after 2 h, taking its scratchpad and background-task output). Projects: nothing to do.
2026-10-01 Shared rule: /tmp is RAM; temp dirs made by code/tests must be deleted after use. Machine tip: TMPDIR=/var/tmp (disk) in the Claude Code settings env moves agents' temp files off RAM. Projects: nothing to do.
2026-10-01 wf res hook: Claude Code PreToolUse hook; while game mode is on, agent Bash commands run without a display (no windows over the game). Shared rule line added. Machine: add the hook to ~/.claude/settings.json (docs/resource-ledger.md §4.6). Projects: nothing to do.
2026-10-01 tests: the wf res lock race test fails instead of hanging when a child crashes. Projects: nothing to do.
2026-10-01 wf res status: unreserved agent memory says how much of it is reclaimable file cache. Projects: nothing to do.
2026-10-01 wf res status warns about jobs started with bare systemd-run (transient services outside agents.slice); shared rule: also not from project scripts. Projects: a script that calls systemd-run itself → start it through wf res run.
2026-10-01 wf res: run, note, status and -h say what --for means: a job is never killed when it runs over (ETA overdue (still running, not killed)); a note is freed. Projects: nothing to do.
2026-10-01 wf res status/wait: warn when a running job is stuck at its own memory limit (≥ 85 % of --mem and ≥ 20 % memory stall) — r-N throttled at its memory limit …: likely too small. Projects: nothing to do.
2026-10-01 wf res: a job killed with its wrapper (OOM, signal, timeout) now says why in status/wait, e.g. done rc=? peak 3.9 GB … (killed: oom-kill by systemd-oomd, limit 4.0 GB; raise --mem), read from the journal; was rc=? peak ?. Projects: nothing to do.
2026-10-01 wf projects skips a wf checkout (its templates/workflow.toml is no project). Projects: nothing to do.
2026-10-01 Repo split: the tool repo holds only the tool (fresh history); the workflow's own tasks and history moved to a separate private wf project. Reference docs: docs/design.md, docs/resource-ledger.md. wf report / inbox unchanged. Projects: nothing to do.
2026-10-01 wf res: background Claude sessions (process named by version, e.g. 2.1.283) are now recognised: adopt/timer move them into agents.slice, and their notes are owned by the session, not its shell. Projects: nothing to do.
2026-10-01 wf res: shared memory/CPU ledger for agent jobs — run (exit 3 = busy, with ETA; --queue), status, wait, release, note, game on|off (gaming reserve, auto-off 4h), clean (stale /tmp/claude-* scratch auto-deleted after 2h), adopt (running Claude sessions → agents.slice, no restart), timer on|off, shell-init. Rule in shared CLAUDE.md: big/long jobs go through wf res run. Projects: nothing to change; move ad-hoc systemd-run jobs to wf res run.
2026-10-01 wf check errors on bare ids in After: / Slices: (were silently ignored, so next could pick a task before its dependency). Project: write them as [[id]].
2026-09-30 Inbox fixes: wf rename OLD NEW (id + all links); add takes a leading <id>: as the id; --parent --id slices chain After: via Slices: and count for done/next; next skips parents with open slices; set --title "Title. Goal." / question replaces the text. Projects: nothing to do.
2026-09-29 First release: wf (next list show ctx search log projects check add done prio move status set note body tick report init migrate), shared CLAUDE.md, TASKS format 1. Projects: wf init or wf migrate --write, pointer line in CLAUDE.md.