Skip to content

v0.2.5

Choose a tag to compare

@fioenix fioenix released this 13 Nov 04:05
· 730 commits to main since this release
08cb2a6

Added

  • ZCode, Z.ai's coding agent (Apache-2.0), as a fourth delegate family (contributed by nta-jeremy): it can be recorded in the host config's inventory, the environment allowlist and the assignment matrix, and the coordinator dispatches it for code tasks in --mode edit (no shell; the coordinator runs the tests) and for advice in --mode plan, never yolo, guarded or build, with facts taken from ZCode's own CLI documentation on 12/11/2026 and a permission check on a machine's first dispatch (delegate references/agents.md "ZCode"). Review routing skips it, the away lease denies it, and a host config without a ZCode row still passes setup check.

  • A repository can set its own away choices (owner decisions, 2026-11-11 and 2026-11-12): /setup add afk, inside a repository, asks Task choice, Push and pull requests, Merge, Trusted authors and a new Roots list (top-level code folders such as packages), each with "Use my user setting" as the default, previews the object and, after a yes, saves it as autopilot in the repository's approved settings.json (project_progress.py settings --set autopilot=<JSON>, --dry-run to preview, --clear autopilot to undo). A set project value wins over ## afk in host.md for that repository; hosts, lease hours, limits, the quota floor and the categories stay user scope only. An unapproved or branch-supplied file sets nothing, and an approved file that is now broken turns those five keys off, never wider. afk.py check re-reads them on every call, and the lease hash covers the user section only, so a project change does not end a live lease.

  • afk.py check accepts a path under a project root with the built-in roots' rules, a top-level README.md, CHANGELOG.md or CONTRIBUTING.md, and git merge [--no-ff] [--no-edit] of an existing local branch (refs/heads/<branch>, or an unambiguous bare name) into a work branch when Merge is yes for the repository; into the default branch it stays Tier 3. Delegated claude -p and codex exec runs outside the read-only delegate_env.py advisor form, including write-capable dispatches and shell-function commands, stay Tier 3.

  • afk.py check binds away permissions to one repository identity, one approval and one target. A single resolver (project_store.identity, on the reciprocal-worktree rules of git_metadata) feeds the settings, branch and remote reads; a repository it cannot identify, including a linked worktree with a missing, mismatched or unsafe back-pointer, gets the strict values (Task choice listed, push and merge no, no extra roots), and so does an approved object whose recorded origin no longer matches. Both override directions stay. Breaking for away commands: gh pr merge, create and edit are Tier 1 only with one explicit --repo github.com/OWNER/NAME equal to that origin, creation also needs --head <branch>, and a push qualifies only when origin's push destination (pushurl included) is that repository. afk.py effective prints the values check applies, read-only. A current branch that is a symbolic ref to the default branch counts as the default branch, and a bare merge name of 4 to 64 hexadecimal digits in any case is ambiguous. Each settings approval record now also holds the full SHA-256 of its main checkout root and is checked against the resolved identity, so a checkout whose eight-character code collides with another's never uses that checkout's approval (it reads as broken), and an autopilot object approved before this binding is strict until it is approved again. Path operands are read from the working directory and judged by their path relative to the repository root, so from a subfolder a root never widens to the same name below it and a protected folder above it still counts; without an identified repository no path is a known path.

  • /setup add afk in a repository now replaces the whole autopilot object, or clears it when every answer is "Use my user setting" (other project settings stay). Its preview shows the repository, the action, the effective values and a token; project_progress.py settings --set autopilot=<JSON> and --clear autopilot write only with --expect <token>, and refuse when the repository, its settings or its origin changed since the preview.

  • commit-messages' worktree_sweep.py list (copy in review-gate) helps clean up worktrees after a pull request merges, so they no longer use up storage (owner decision, 2026-11-12). It is read-only and deletes nothing: git does the removal with its own checks. For each linked worktree it prints the path as git lists it, branch, size on disk, clean or dirty, unpushed commits, merged (an ancestor of the local origin/<default>, or, when gh is present, a merged pull request whose head commit is the branch's current head; otherwise unknown) and a candidate flag, set only when the worktree is merged, clean, has nothing unpushed, is not locked or detached, is not the working folder or a parent of it, and neither holds nor lies inside another registered worktree (compared as written, resolved and case-folded). Under each candidate it prints the two commands to run: git worktree remove <path>, never --force, and git branch -d <branch>. commit-messages "Branches", the review-gate merge step and the afk merge gate now run it right after a merge and run those two commands only for the worktree of the pull request just merged, when it is a candidate this agent or its runs created; git refuses a dirty worktree and -d an unmerged branch, and git worktree remove also deletes the files git ignores inside that worktree. Under afk both commands go through check and run only on Tier 1. The rest of the list goes to the user with sizes (advisory).

  • On Claude Code, a hook prints one fixed line after a successful gh pr merge that is the whole command or the last command of a line joined only by ; or &&: "fnl: a pull request merged; run worktree_sweep.py list and, for its worktree when a candidate, the two git commands it prints (advisory)." Other compound commands print nothing, because the payload's outcome is not the merge's own; nor does a quoted excluded flag such as '--auto', which the shell passes on as --auto, or a quote glued to other text. It stores, logs and sends nothing, never blocks and always exits 0. Codex and Antigravity get no hook, because their hook output is not verified here.

  • context_snapshot.py probe --host claude|codex --session <id> reads the current session's context size on request, as the context probe exception in CLAUDE.md allows: Claude Code's /context on a forked, unsaved copy of the session, or Codex's app-server resuming an idle or unloaded thread with no goal. It runs only the program confirmed in /setup add quota, with that probe's environment, private working folder, 30-second limit and output cap, stores the existing snapshot fields (Codex as host codex) and prints the read JSON; any failure exits 3. FNL_CONTEXT_PROBE=off turns it off; no hook runs it. tidy uses it when no fresh status-line snapshot exists. Live delivery on a real host is unverified, including whether Claude Code resumes a session from the probe's neutral working folder.

  • shape, a model-invoked interview for brainstorming, discussing, ideating on or stress-testing an idea, plan or design, or for unclear requirements before building, in any language: it maps the decisions as a tree, asks in rounds only the questions whose prerequisites are settled, each as one question of the host's native question tool with two to four options and the recommended one first, never as a text list (rounds split at four questions per call on Claude Code), looks facts up itself, records settled decisions with their date in the work record or plan, keeps undecided ones open with an owner, never invents answers in a delegated or headless run, and acts on nothing until the user confirms the result. shape runs first and installed grilling (/grill-me, /grill-with-docs) runs when shape is missing or the user asks for it; superpowers brainstorming only for many divergent design ideas. Learned from Matt Pocock's grilling (MIT); no text copied. It has no fnl- alias. New trigger evals trigger-shape and trigger-shape-vietnamese.

  • The session-start reminder adds one fixed line: the user-only skills (help, ask, setup, status, afk, brb, pause, gtg, update, restore-point, session-hygiene, verification and every fnl- alias) are for the user to type and never called with the Skill tool; questions about the plugin are answered from help's references, coding work goes through coding-router (left out when Routing is off) and second opinions through advisor, and when a user-only skill is needed the agent tells the user the command to type. A test keeps the list equal to the skills marked disable-model-invocation: true.

  • The compaction nudge (eighth scoped exception in CLAUDE.md, docs/specs/compaction-nudge.md): on Claude Code a UserPromptSubmit hook in its own group, on by default, reads only the integer usage counts of the latest assistant entry in the last 256 KiB of the session's transcript (a regular file, not a symlink, inside the config folder's projects) and the compaction settings, and at 80% and again at 90% of the threshold prints one fixed line asking the agent to run tidy now: save the handover, write back decisions and lessons, note background runs and open pull requests, then offer to compact or start a fresh session. It stores only the last band per session under a hash of the session id, prints again after a compaction lowers the band, and prints nothing when auto-compaction is off, no window is set (the host default is not recorded yet), the mark cannot be written or anything fails. FNL_COMPACT_NUDGE=off turns it off. Codex and Antigravity are not covered. Live delivery of its line into the model context is unverified.

  • The session-start hook prints one Session id: <id> line when the host's id is a lowercase UUID, on every source, so the agent can probe its own session.

  • Setup and update onboarding fixes (owner decision, 2026-11-12):

    • /update upgrade notes now cover 0.2.0 to 0.2.3 (Codex hook trust, the ## Limits parallel lines, away setup and decisions, agent levels, the matrix migration, fnl-ask to help, the question-gate hook) and the missing 0.2.5 steps (Codex hook trust, quota re-pin after a Claude Code or Codex update, the context probe's need for /setup add quota, and information on worktree cleanup, /afk start's goal and loop, the main checkout and the full /finolabs-vibecode:help and :status names). Items show [done] from read-only checks in the new setup script upgrade_state.py, which reads host.md (and, for one item, the current repository's approved check list); a note can be info, and "where": "repository" runs a check in the user's current folder.
    • update_check.py onboard first marks the installed release seen only on a fresh data root (no host.md, no onboard.json); setup runs it before its first write, so a new user is no longer sent through every earlier version's notes.
    • /setup check lists every old fnl-<skill> name in instruction files with its new name (row old skill names and I14), reports a host config with no delegation section as delegation not set up instead of failing, adds one legacy host settings row for ## Checks, ## Project settings, ## Project labels, the global Ledger project:/Task file: lines and the ## Limits parallel lines, names each quota family whose recorded program is gone, moved or unconfirmed (path checks only, nothing runs), and adds a compaction settings row when a window is set while auto-compaction is off in user, project, local or managed Claude Code settings or DISABLE_AUTO_COMPACT or CLAUDE_CODE_AUTO_COMPACT_WINDOW is set (key, scope and variable names only). The compaction preview warns about the same settings.
  • Delegated runs record that they ended, and the coordinator is reminded of runs with no outcome (owner decision, 2026-11-12; docs/specs/ledger-closure.md). After its child exits, delegate_env.py appends one ended ledger line (dispatch id from FNL_DISPATCH_ID, time, exit code, negative for a signal) when the id is valid and the ledger exists; it never records command text and a failed write never changes its exit status. open shows ended exit=<n>, stale treats an ended record as due after 24 hours without waiting for twice its budget and proposes error for a nonzero exit or cancelled for exit 0, and summary ignores the line. On Claude Code a new UserPromptSubmit hook, in its own group, prints one fixed line at most every 30 minutes per session with the count and up to five ids of runs that session started that have no finish line and ended or passed their budget; it only prints, keeps one integer mark per session, records no metrics event, sends nothing and is off with FNL_DISPATCH_NUDGE=off. Codex and Antigravity get no hook.

Changed

  • Session titles start with one emoji for the kind of work (owner, 13/11/2026): <emoji> <project> <scope>: <activity>, from a fixed set of 25 in tidy references/titles.md (🐛 fix, ✨ feature, 🔍 review, 🔒 security, 🚀 release, 🎨 design, ⚡ performance and others), changed when the activity changes kind; ledger labels and log file names stay plain text.

  • Two skills are renamed (owner decisions, 12/11/2026 and 13/11/2026). session-hygiene is now tidy; session-hygiene and its fnl- name stay as user-only aliases that load tidy, and the hooks' printed text names tidy. autopilot is now afk, which holds the away skill itself: the autopilot name and its fnl- alias are dropped with no alias, brb and the fnl- names of afk and brb load afk, autopilot.py becomes afk.py, autopilot_setup.py becomes afk_setup.py, and setup add autopilot is gone (use /setup add afk). Breaking: the away settings now live in ## afk of host.md and an old ## Autopilot section is no longer read, so a lease does not start until you run /setup add afk again; then delete ## Autopilot by hand (the /update note 0.2.5/afk-section shows [done] once only ## afk is left). A repository's autopilot settings object and the lease's private state keep their names, so stored data still reads. /setup check lists session-hygiene -> tidy and autopilot -> afk in instruction files, and the usage metrics count events stored under the old names as tidy and afk.

  • open-source-review keeps its steps in the skill and moves the repository settings, community files and plugin package checklists into references/, so the skill fits the size limits; every item there is still marked. plugin-packaging moves the Anthropic directory facts into references/directory-anthropic.md. A test now holds both skills and these references to the size limits, and the prepare-only eval fails an answer that asks for approval and then goes on with a portal step anyway.

  • Record the owner decision for a compaction nudge (eighth scoped exception in CLAUDE.md, docs/specs/compaction-nudge.md, PRIVACY.md): on Claude Code a UserPromptSubmit hook will read only the token counts of the session's latest turn and the compaction settings, and at 80% and 90% of the threshold tell the agent to save a handover and write back decisions before the host compacts; the code is under Added.

  • Record the owner decision for a context probe (seventh scoped exception in CLAUDE.md, docs/specs/context-probe.md, PRIVACY.md): on request it reads the current session's context size from Claude Code's local /context or Codex's app-server, so tidy gets real numbers where no status line runs; the code is under Added.

  • fnl-checks' checks.py run --dispatch <ledger id> writes the verification receipt that fnl-delegate's ledger.py finish --outcome accepted requires, for the task checkout and its head, after a whole-list pass on a clean tree whose head and list did not move, whatever the local-receipt mode; each such run first removes the dispatch's earlier receipt. Until now checks.py run wrote only the merge gate's local receipt, which the ledger never reads, so most acceptances used --accept-unverified. fnl-delegate's accept steps now use it where a check list is approved, verify_claim.py over the brief's tests where none is, and keep --accept-unverified for work with nothing to check.

  • The main checkout stays on the default branch: commit-messages "Branches" now cuts every branch, a coordinator's own small fix included, in a worktree of its own (git worktree add -b <name> <path> origin/<base>) and never runs git switch or git checkout <branch> in the main checkout (advisory), and coding-router repeats it. The session start prints one fixed line when the repository's main checkout is not on its default branch (from refs/remotes/origin/HEAD; silent when that is unknown or on any error), read from local Git metadata without running Git and without printing a branch name or path.

  • coding-router fires when the user wants to brainstorm, discuss, ideate or shape an idea or design for code, and the none workflow's step 1 now has two paths: missing criteria keep the bounded three questions; an open intent or design, or a request to brainstorm or discuss, runs an interview through the native question tool, shape first, installed grilling (/grill-me, /grill-with-docs) when shape is missing or the user asks for it, and superpowers brainstorming only for many divergent design ideas. The mattpocock and superpowers Clarify rows name shape as the fallback, and in Spec Kit shape works inside a phase and yields to Spec Kit's own clarify step. New trigger eval trigger-coding-router-brainstorm.

  • ledger.py stale --older-than-hours <n> (1 to 8760) sets the stale threshold for one preview; a record inside its budget is still never listed. fnl-status now says when to preview and close stale records.

  • Documentation matches 0.2.5: the README describes the release, the context probe, the compaction nudge, the merge notice and the dispatch nudge, embeds the new diagrams and lists ZCode under "Supported agents"; docs/how-it-compares.md, the help map and the away and session diagrams say that /afk start prints a /goal line and starts /loop; the worktree sweep's docs say it escapes paths and quotes its commands; and /update gains 0.2.5 notes for the repository's own afk choices, the compaction nudge's window, shape, ZCode, the ledger closure and the dispatch nudge.

  • heavy_run.py now also waits on macOS until the kernel's memory pressure has read normal at every poll for a whole window (default 60 seconds) and swap-outs over that window average at most a limit (default 5 MB/s); an unreadable pressure or swap-out reading waits and never counts as a pass, and other platforms skip these two gates. The memory minimum default rises from 3 to 4 GB. New ## Heavy jobs keys: Memory window seconds (1 to 900) and Swap-out limit MB per second; readings are stamped when they complete, one sample is never a rate, and a swap-out count that goes backwards restarts the window. status prints a pressure: line. Light targeted tests skip the lock and the load gate but still wait while pressure is critical.

  • plugin-packaging covers directory submissions: a new references/directories.md holds what Anthropic's and OpenAI's plugin directories ask (who submits, the tracked branch or tag, field limits, the desktop-only path for plugins that run a local program) and the portal holds no local validator shows, and a new "Directory editions" part describes building a per-directory edition from the same sources with a test that fails while a committed edition is stale. Every portal step (upload, draft, repository access, portal validation, submitting) waits for the owner's approval, a new prepare-only eval checks that the skill stops before it, and the form's answers and attestations stay the owner's. It also records, as observed and reported rather than published, that Anthropic's directory refuses after submitting local code (command hooks, local MCP servers, mods, scripts, broad allowed-tools) from organizations that are not Verified, keeps the design-only exception to the image-generation restriction, and that hooks shared with Codex stay in shell form because Codex hooks take no args.

  • open-source-review scans with gitleaks git and gitleaks dir, runs the plugin validators on a fresh clone so untracked files such as .DS_Store do not show, and walks the target directory's own pre-submission list against the submitted folder or ZIP.

  • /afk start now keeps the session working while you are away: it prints a goal you can type in the host's /goal command before you leave (the lease only bounds actions and never starts a turn), and starts a self-paced /loop on Claude Code when work waits on CI, reviews or a delegated run, stopping it when the lease ends. Neither widens what the lease allows.

  • ledger.py stale is more conservative by default: an open record is stale only past twice its budget and at least 24 hours (before: past its budget), and a record with no usable budget becomes stale after 24 hours (before: never). Each preview line now shows the record's label, and stale --close rechecks each record against the threshold its preview used.

  • A second /afk start while a lease is live now reports when that lease ends and tells the session to work under it, instead of a generic refusal that led a second session to work outside the lease. The lease covers every session on the machine; nothing is extended or reset.

  • The assignment matrix is the one setup path for levels, roles and the judge: add roles defers to it when it exists, and its Jev choice asks the same separate TypeSafe consent question as add roles. The Agent skills block now names help as /finolabs-vibecode:help (Codex $help).

  • The away-session setup is now /setup add afk; its choices move to ## afk in host.md (see the rename of autopilot to afk above).

  • Breaking: skill names drop the fnl- prefix so they are short and easy to remember (owner decision, 2026-11-11): fnl-setup is now setup, fnl-coding-router is coding-router, and so on for every skill, the short aliases included (afk, brb, ask, verification, gtg). Every old name stays as a user-only alias (skills/fnl-<name>/) that loads the renamed skill with the same arguments, so /fnl-setup typed by you keeps working; the model does not see the aliases, so an instruction file that tells it to invoke an old name must use the new one (/setup check lists them).
    On Claude Code, type /finolabs-vibecode:help and /finolabs-vibecode:status in full, because /help and /status are the host's own commands. Usage metrics record the current name and count an old name, typed or stored, as the same skill. The agents (fnl-reviewer, fnl-security-reviewer, fnl-navigator) and the fnl-review and fnl-router markers keep their names, and a repository whose ## Agent skills block names fnl-review-gate still counts as adopting the review gate.

  • A repository's settings (ledger, tasks, parallel, parallel-caps) and its dashboard label now live in its own folder, <main checkout>/.finolabs/vibecode/settings.json (linked worktrees share it), instead of ## Project settings and ## Project labels in the shared host.md (owner decision, 2026-11-11). Every setter records the SHA-256 of the file in <data root>/state/settings-approved/; a file with no approval record (such as a branch-supplied one) reads as unset, and an approved file that is now missing, unreadable, unsafe, invalid or edited makes capacity and the ledger join unknown, turns the task file and label off and says the settings need setup again, so a repository cannot raise parallel, relax an approved limit or redirect tasks.
    The setters keep their consent rule, lock and validation, add .finolabs/ to .git/info/exclude when missing (never to .gitignore) and leave host.md byte-identical. While the settings are broken every setter refuses with a message naming project_progress.py settings --reset, which discards every project setting (the file and its approval record) after the user's yes and returns the repository to never configured; Check reports broken settings as needing setup again. The dashboard shows a label only for the repository it runs in; other repositories show their registered folder name.

  • Breaking: ## Project settings, ## Project labels and the global Ledger project: and Task file: lines of host.md are no longer read, with no fallback and no migration command (owner decision, 2026-11-11). Unset means the defaults: parallel 3 with no caps, the folder name as ledger name, no task file, the registered folder name as label. Set them again per repository with /setup add project or /setup add delegation, then delete the old lines by hand. Check no longer reports invalid or deprecated host settings.

  • Repository identity is validated: a .git file counts as a linked worktree only in the canonical layout <main>/.git/worktrees/<name>, with its commondir resolving to that .git, its gitdir pointing back and every metadata file and folder owned and not group- or world-writable, as the review gate already required; anything else resolves to the checkout's own root, so a checkout cannot borrow another repository's code, check list, settings or exclude file. Codes of real repositories and real linked worktrees are unchanged.

  • Sessions get named more often: the session-start rules now tell the agent to invoke tidy at the first task, on a task change and after every compaction, and name the rename mechanism per host; after a compaction the hook adds one fixed line to check the title. The skill's description leads with naming and lists plain requests such as rename, wrap up, save progress and hand over. Local usage data from November 2026 showed no model-initiated call of the skill in 691 transcripts and a rename in 42 of 312 long sessions.

  • The router and verification fire more often. coding-router step 1 now tells the agent to invoke tidy to name the session once the task is known and again at each phase change (planning, implementation, review); its gate line and the session-start rules tell the agent to invoke checks before reporting work as done, fixed or passing and before accepting a delegated run. The checks description leads with the questions users type ("is it done?", "did the tests pass?", "is it ready to merge?"), with a new trigger eval. Local usage data from November 2026 showed no call of the verification alias, which hosts never offer to the model. The first-edit router reminder stays on Claude Code only: Codex's PostToolUse output contract and Antigravity's hook output are not verified in this repository.

  • A repository's check list now lives in its own folder, <main checkout>/.finolabs/vibecode/checks.json (linked worktrees share it), instead of ## Checks in the shared host.md (owner decision, 2026-11-11). Apply records the SHA-256 of the approved file in <data root>/state/checks-approved/; a file edited, committed or checked out on a branch, or without approval, runs nothing and turns the local receipt off until /setup add checks runs again.
    checks_setup.py adds .finolabs/ to the repository's .git/info/exclude once (never to .gitignore), keeps a replaced file as checks.json.bak-<date>, and never opens host.md. Undo is checks_setup.py remove for this repository; a stale remove preview changes nothing.

  • afk.py check splits its protected set so an away session can change the plugin's own source (owner decision, 13/11/2026). On a non-default work branch, with Push and pull requests or Merge yes for the repository, the plugin's own source (skills/** including SKILL.md and the skill scripts, scripts/**, top-level CLAUDE.md and AGENTS.md, .<host>-plugin/*.json hook and plugin manifests) is a known path for edits and for the merge gate's changed-path check, run in the pull request's worktree. Everything that changes live behavior or reaches outside the repository stays Tier 3: paths outside the repository (user and host config, the data root, host.md, lease and setup state), .git/ and git hooks, .github/, the repository's .claude/ settings and hooks, secret-looking names, every source path on the default branch, and the running install's own checkout on every branch (matched by its validated checkout, so its linked worktrees under .claude/worktrees stay editable; this also closes an older gap where the install's top-level scripts/*.py were Tier 1). Pull request and commit text carrying a review gate marker (<!--, optional whitespace, fnl-) is forbidden, so reviews post only through post_review.py. Breaking for scripts that rely on exit 0: check now exits 3 on Tier 3 (see Fixed).

  • Breaking for lists saved with 0.2.4: the ## Checks section of host.md is no longer read, with no fallback and no migration command (owner decision, 2026-11-11). Until you run /setup add checks again in each repository, it has no checks and its local receipt is off; then delete the old section from host.md by hand. The 0.2.5 upgrade note in /update says so.

Fixed

  • afk.py check no longer stops the own-branch release work an away mandate authorizes (owner report, 13/11/2026). A path inside a linked worktree of the repository (validated by project_store.identity, never a lexical .claude/worktrees allow) is judged by its path relative to that worktree, as from inside it, and git -C <worktree> as the same command run there; any other folder under .claude/ stays Tier 3. gh pr create and gh pr edit accept --body-file, and git commit accepts -F <file>: a regular file of at most 64 KiB inside the check's own checkout (for example its git-ignored .finolabs/), outside the data root, with no link from that checkout root down, whose text is screened for controls, secrets and machine paths (else unsafe), so Markdown and shell examples go there. Inline title, body and message values may hold newlines, tabs and a leading -, but a backtick, $( or $ before a letter, _ or { is unknown. gh pr comment <number> --repo github.com/OWNER/NAME follows the Repository comments category. Clarified: as before, only a Tier 1 answer reserves an action or a run. Breaking: check exits 3 on Tier 3 (0 only on Tier 1, same JSON), so check ... && command no longer runs a denied command. With Merge yes, the merge gate's cleanup is Tier 1: git worktree remove [--] <path> of a linked worktree whose HEAD is the full SHA a log --kind merge --head <sha> entry of the current lease recorded (new optional --head on the merge kind; never forced, the main or the current checkout, or a worktree holding this install) and git branch -d [--] <branch> (never -D, the current or a default branch); git rm [-r] [-q] [--] <paths> on a work branch classifies each path like a change. A git commit -m message is data the same way (-q allowed), so a message that names hooks, settings or SKILL.md, or spans lines, no longer makes the commit Tier 3. A git command that changes a work tree (merge, commit, add, pull and the like) stays Tier 3 when it would run in the running install's own checkout, directly or through git -C, and log --kind merge --head needs its --pr.
  • Delegate ledger records no longer stay open or get closed from the wrong checkout (owner decision, 2026-11-12; docs/specs/ledger-closure.md). ledger.py stale (preview and --close) now considers only records whose project is this repository's ledger project, resolved as status resolves it (the approved ledger setting, else the checkout folder name), and refuses when none resolves; --all-projects widens the preview with a warning, and the preview's digest binds that scope, so a close must come from the same project or repeat --all-projects. A review dispatch (--kind review) is accepted on its posted review: finish --outcome accepted needs no verify receipt and no --accept-unverified, takes --head (recorded as reviewed_head, which never binds authorship) and an optional --review-url to a pull request comment, checked by pattern only and recorded as review_posted. --reason may go with any outcome, kept only as a digest. Other kinds keep the receipt rule.
  • coding-router's Spec Kit adapter follows a core speckit-* skill marked disable-model-invocation: true (Spec Kit 1.4's --no-model-invocation integration option) by reading its SKILL.md, instead of calling the skill tool that refuses it or reporting an integration gap.
  • afk no longer treats the plugin's own wrapper commands (ledger.py, record.py, heavy_run.py, judge.py, delegate_env.py, merge_check.py) as protected when the plugin is installed under a .claude or .codex folder, as Claude Code and Codex install it. Only that one exact script path is spared the protected-name check; every other token, including a nested command's, is still checked.
  • An alias whose skill is user-only (fnl-help, ask, brb, gtg and the older names of afk, pause, restore-point, setup, status and update) no longer tells the agent to load that skill, which made it call the Skill tool and get refused: it now says the skill is user-only and to read its SKILL.md in the plugin's skills folder and follow it, never through the Skill tool. Older names of short names load the final skill directly (fnl-ask → help, fnl-verification → checks, fnl-afk and fnl-brb → afk, fnl-gtg → pause); a test and check-references.py hold both rules for every alias.
  • precompact.py no longer asks for a handover it cannot get: PreCompact runs after compaction has started and Claude Code uses its output as summary instructions, so it now tells the summarizer what the summary must keep (goal and acceptance criteria, settled decisions with dates, work in progress, open background runs and pull requests, the next action, and whether a handover was saved). The handover is now prompted earlier by the compaction nudge.
  • After a Claude Code or Codex update moves the program confirmed in /setup add quota (a new versioned folder, for example), quota_probe.py read, context_snapshot.py probe and /setup check now say ": program changed since setup", with the old and new version folder names when both paths differ only there and never a path, and name the fix /setup add quota to re-confirm the new path. Before, the quota probe recorded an execution error and the context probe gave only its generic refusal. The changed program still never runs and the new path is never accepted on its own.