Skip to content

Releases

v0.3.2

Choose a tag to compare

@fioenix fioenix released this 17 Nov 07:44
8ccfeb3

Added

  • afk check hook on Claude Code (tenth scoped exception, owner decisions 17/11/2026; docs/specs/afk-check-hook.md): while an away lease is live, a PreToolUse Bash hook judges each command as afk.py check does and denies every Tier 3 answer, in both modes, and every form outside the syntax it reads in full (a syntax allowlist: variables other than a plain $HOME, command substitution, xargs, find -exec, directory changes with a relative glob, glued operators and more; false denials are expected); once a lease is live any error denies as unsafe, and it never answers allow and stores no command text. The agent no longer has to call check before a Bash command on Claude Code. FNL_AFK_HOOK=off disables it.

  • tidy: when the agent asks the user to compact, it hands over the command ready to copy in its own block, on Claude Code /compact <focus> with a one-line focus naming what the summary must keep (goal, current task, open runs and pull requests, next action, handover saved), a bare /compact on Codex, and /new or /fork with the handover on Antigravity, whose /compact may not exist.

  • /setup add delegation asks which AI helpers vibecode uses as one native multiple-choice question built from the agent scan (AskUserQuestion, request_user_input on Codex, ask_question on Antigravity; numbered text choices where a host has none), then one single-choice question per role (code-review, security-review, navigator, advisor) among the chosen helpers plus "None", recommending a family other than the coordinator's. A helper left unchosen gets no matrix line, and the levels, types and model questions are asked only for the chosen ones (Quick mode gives them levels 1-5). The new read-only matrix_setup.py choices prints every helper name the matrix accepts, from ## Agents on this machine and the ~/.claude-<name> folders, with its family, presence and the coordinator mark, never a path, and says when the inventory is missing or stale and /setup scan should run first.

  • README: a warning under "Supported agents" that delegation, cross-agent reviews, advisor and the quota and context probes run other agents' CLIs headless under your own accounts (afk included), that a consumer plan may forbid that use, and that which agents run is your decision and responsibility.

  • The afk overlay (docs/specs/afk-overlay.md, owner decisions 16/11/2026): /setup add afk lists the lines of AGENTS.md, CLAUDE.md and the Spec Kit constitution that still require your approval for an action your away rules allow, and offers a local AGENTS.afk.md holding only override entries for those rules, each naming the rule it replaces. instructions_write.py writes it after its own question, keeps it out of Git, keeps earlier versions in the private state and records its SHA-256; while a lease is live the session-start hook (Claude Code, Codex) and /afk start print it only when its bytes still match and hold no control or format character other than tab and newline, else one fixed "not in force" line (the writer refuses such text too), and /afk stop says it no longer applies. In strict, afk.py check protects AGENTS.afk.md* and .specify/memory/constitution.md and the writer refuses during a lease; in open neither applies. A constitution rule is never replaced: setup only proposes the amendment. FNL_AFK_OVERLAY=off stops the printing; /setup check reports the overlay's state.

  • A recorded rework limit: - Rework limit: <2-4> under ## Reviews of host.md (default 2 when absent or invalid), asked with the review roles in /setup add delegation. review-gate stops at that limit instead of always after two rounds (advisory), and ledger.py finish refuses more rework rounds than the limit without a --reason.

  • The proof nudge (docs/specs/proof-nudge.md, owner decision 17/11/2026): on Claude Code a Stop hook adds one fixed line to the agent's context, asking it to run the targeted tests and read the output (or run checks) before reporting done, when this session committed the current head and no checks or verify_claim.py receipt names it; once per head per session, never blocking. A SessionStart hook records the session's start. FNL_PROOF_NUDGE=off turns it off.

  • compact_setup.py preview --host claude --project <repository root> writes the auto-compact window to that repository's untracked .claude/settings.local.json (this machine only), with the same preview, digest, 60-minute expiry, backup and atomic write as user scope; it refuses a path that is not the checkout root or a file in Git's index, and creates .claude only when applying.

Changed

  • Session titles follow a coordinator's work (owner request, 17/11/2026): the session-start rule and tidy "Titles" rename the session when a spec, story or large pull request it ran is done, and the merge notice after gh pr merge reminds the agent to; the help pages for tidy, setup, review-gate, troubleshooting and /setup check describe the 0.3.2 changes; CLAUDE.md "Versions" lists the docs every user-facing pull request updates.

  • The README diagrams are re-rendered from their 0.3.2 sources (task lifecycle, session health, skills architecture, delegation sequence, afk mechanism, request to merge), each as the diagram alone at 2x like the others; the afk mechanism names the afk check hook on Claude Code, and request to merge says an open lease merges on its own as a strict lease with Merge yes does.

  • /setup add compaction proposes an auto-compact level instead of prescribing none (owner decision, 17/11/2026): before the token question it shows an advisory value by session kind with its reason, never preselected (about 300000 to 500000 for a coordinator on a 1M-window model, later or a fresh session per story for single-task work, about 70% to 80% of a smaller window), then asks where it applies, user scope or, on Claude Code, this repository only. The beginner path and Quick setup now ask the threshold with its own question and say that without one Claude Code's tidy nudge stays silent.

  • Breaking: afk gains two modes (owner decisions, 16/11/2026; docs/specs/afk-autonomy.md). open, the default, lets every command, path, category and merge run under a live lease except the credential floor: afk.py check answers floor (Tier 3, never approvable through Decide) for a request that changes a credential or reads, prints, copies or sends a secret or key, after splitting operators, dropping wrapper prefixes and reading inline code. Two guards are on by default in open: instructions from third-party content never grant actions (advisory), and check answers guarded for a commit on the default branch, a push to it, and every force, delete and mirror push. - Default branch: open and - Third-party instructions: open switch them off; a repository's approved mode, default-branch and third-party-instructions keys win there. - Mode: strict keeps every earlier list, category, limit and gate. A lease set up before this change, with no Mode line, is now open; add - Mode: strict to ## afk to keep the earlier rules. In open the Merge answer is ignored and optional, and /setup add afk (afk_setup.py preview --mode) writes the mode first.

  • merge_check.py --independent --fallback-family <family> under a lease whose effective mode is open (the ## afk Mode line, or the repository's approved mode; any doubt counts as strict) accepts same-family markers that carry the fallback reviewer line without quota evidence, a router-trigger limit or two models for an enforcer change, and counts a marker without that line as another family's when the configured review role for its kind names one (docs/specs/afk-autonomy.md D). A strict lease and runs outside a lease keep the earlier conditions.

  • /setup add afk asks the mode first, open recommended with one plain paragraph on what agents may do, then the two guards of open, each defaulting to guard, and for open only the apps and the lease hours; strict keeps every earlier question. A repository's own away rules ask its mode and guards first, and the repository settings preview shows Mode, Third-party instructions and Default branch first.

  • afk.py rules prints the afk rule table, generated from the data check reads (modes and guards, the credential floor, the default-branch guard, the strict settings, paths and rule rows, the log kinds); skills/afk/references/rules.md is that output and tests/test_afk_rules.py keeps them equal. The afk skill, its lease rules, the decision and merge-gate references point to it instead of restating rules; the skill and its lease rules shrink from 16.5 KB to 7.4 KB.

  • afk: the bare export, set and env exception of the credential floor applies only to a word cut out of a search pattern after a lone | or &; find . -exec sh -c 'a;export' ';' now answers floor. merge_check.py reads a lease as strict when its ## afk section has a key afk.py does not read or an invalid guard value.

  • The afk mechanism and away decision diagrams say they show strict and that open, the default, runs every action except the credential floor and the default-branch guard (git commit and git push only), with an advisory third-party-instructions guard and no enforced review or merge_check.py; their PNG renderings and the README alt text are updated. The README, help map and pages, and the comparison page describe the modes, the guards, the overlay, the setup mode question and the helper picker.

  • Long sections load only when needed: commit-messages "Branches" keeps a short summary and moves the full branch flow to references/branches.md; checks "Accepting someone else's work" moves to references/accepting.md; review-gate's merge step points to the worktree sweep in commit-messages "Branches" instead of repeating it.

  • delegate `references...

Read more

v0.3.1

Choose a tag to compare

@fioenix fioenix released this 15 Nov 13:33
e6fc1dd

Added

  • The compaction nudge runs on Codex too: one UserPromptSubmit hook (compact_nudge.py --host codex) reads only last_token_usage.total_tokens and model_context_window of the latest token_count event in the session's transcript under the Codex home's sessions folder, and the top-level model_auto_compact_token_limit of its config.toml; the threshold is the smaller of that limit and 90% of the window, as Codex resolves it. Same bands, message, mark and FNL_COMPACT_NUDGE=off; it never blocks. Codex asks to trust the plugin's hooks again after this update.
  • merge_check.py --ci-context <name> runs the merge check as a hosted commit status: it leaves its own status out of the checks, accepts merge state BLOCKED, consults no local receipt and exits 3 while another hosted check is still running. This repository's new Merge check workflow posts it as merge-check from the default branch's code only (never checking out the pull request), on pull request changes, writers' comments, a finished Validate run and by hand (owner decision, 15/11/2026).
  • /setup instructions covers the away mode and the assignment matrix: the contract template points to afk for decisions while the user is away and keeps the lease limits in the host config, says model routing lives in the assignment matrix and machine limits in ## Machine rules and ## Heavy jobs, and the explore step offers to replace hand-written lines that restate those features with a pointer.
  • An away lease decides unknown requests instead of parking them (owner decision, 15/11/2026). When afk.py check answers unknown, the agent asks the away ladder "run this exact request?"; a rule, advisor or host attempt that settles on option 1 adds --approve-argv-json '<argv>' or --approve-paths <paths> to its log --kind decision call, and the next check of that exact request in that folder answers decided (Tier 1) once. The approval is refused unless the request is unknown, runs in a found repository with user or project settings, keeps every path inside the checkout and outside its protected paths, and has one of the allowlisted shapes: a script file of this checkout run directly or by python or node; make with plain targets and VAR=value of a plain value, never a variable naming a program, flags or a path; npm, pnpm, yarn or bun with test, run <script> or an install with no package named; python -m with a listed test or lint module (unittest, pytest, ruff, mypy and the like, never one that runs other code such as coverage run or pylint --init-hook), or pip install only of -r, -c and -e operands inside the checkout; cargo build, test and lint verbs without --config; go build, test and vet verbs with flags from a closed list and local package paths only; go mod download, tidy or verify; uv sync, poetry install, bundle install; git with a listed verb (fetch from origin by branch name, cherry-pick, revert, add, mv and the read verbs) and listed flags only; gh with a listed group and verb (pr view, checks, diff; issue create, comment, edit, close; run view, watch, rerun; workflow run; read verbs) and no other repository than the verified one in any -R or --repo form; and this install's delegated run through delegate_env.py with an approved program of a lease family. A repository's own CI script, make target, module test run or app launch no longer parks the lease.
  • /setup instructions offers a dated standing relay line for a repository whose reviewers cannot post, including security reviews while an away lease with Merge yes is live, and /setup add afk points to it after a Merge yes; this repository's CLAUDE.md "Reviews" records that clause (owner decision, 15/11/2026). Before, an away lease could not merge a pull request whose security reviewer could not post.
  • afk: a third rework round goes through the away ladder like every other stop after two rounds, instead of waiting for the user.
  • help has one short page per skill (what it does, when to use it, the command on each host and a few tips), listed under "Skill pages" in its contents reference, plus a tips page and a troubleshooting page (native Windows needs WSL, Codex hooks to trust again after an update, a stale update cache and more), each routed from help.

Changed

  • Breaking: ## Agent levels in the host config is no longer read (owner decision, 15/11/2026). The assignment matrix is the only source of levels; without a valid matrix no levels are set and routing falls to the strength table. levels_setup.py and the setup reference levels.md are removed; matrix_setup.py migrate is the only reader left, so /setup add delegation still moves the levels into the matrix. /setup check reports a leftover section as no longer read with that fix, and the update checklist gains the 0.3.1/agent-levels-retired note.
  • Breaking: the verification aliases (verification and its older fnl- name) are removed; type checks (owner decision, 15/11/2026).
  • The repository moved to finolabs/vibecode (owner decision, 15/11/2026); the plugin keeps its name, finolabs-vibecode, so commands, the catalog entry and installs are unchanged. Links, install commands and the update check's release endpoint use the new path. An install of 0.3.0 or earlier stops seeing update notices, because its update check follows no redirect: update it once by hand with the host's update command.
  • versioning ends every release with tidy: once the tag, the forge release, the -dev bump and any approved catalog pull request are done, the agent saves the handover, writes back the release's owner decisions and lessons, and renames or retires the session; tidy counts a finished release as a natural boundary (owner decision, 15/11/2026; advisory).
  • afk: git commit under a lease is Tier 1 only on a work branch; on the default branch it answers protected (owner decision, 15/11/2026).
  • ledger.py finish prints a warning when it accepts a run with --accept-unverified, or a review without --review-url; the acceptance is still recorded.
  • The afk mechanism and away decision diagrams show the third answer of afk.py check: an unknown request of an allowlisted shape goes up the decide ladder as "run this exact request?", and an approval recorded with the decision makes the next check of that exact request answer decided (Tier 1) once; the afk mechanism also shows that a merge needs a recorded MERGE OK under 15 minutes old. The PNG renderings and the README alt text are updated.
  • The request-to-merge diagram shows shape interviewing you before the router when the idea is unclear, and the implementer working in one worktree per branch while the main checkout stays on the default branch.
  • The eighth scoped exception in CLAUDE.md, PRIVACY.md ("Compaction nudge") and docs/specs/compaction-nudge.md gain a Codex clause (owner decision, 15/11/2026): the same nudge as one Codex UserPromptSubmit hook that reads only two integers of the latest token_count event in the session's own transcript under the Codex home's sessions folder and only model_auto_compact_token_limit from its config.toml. The code is compact_nudge.py --host codex, under Added.
  • afk: checks.py run --dispatch <dispatch id> is Tier 1 under a lease, so an away coordinator can write the receipt that accepting a delegated run needs; checks.py still writes it only after the whole approved list passes on a clean tree at an unmoved head (owner decision, 15/11/2026).
  • afk consumers follow the unknown-request ladder: afk names the own-worktree branch form and checks.py as test runners, the delegate route rule for an absent user defers to afk "Decide" during a lease, review-gate sends the stop after two rework rounds through Decide during a lease, the /setup add afk test-command question says other allowlisted shapes may still be decided, and the spec, README and index wording match.
  • The review-gate low-risk lane also takes a post-release record: merge_check.py --lane qualifies a pull request that changes only the Status: line of a docs/specs/release-X.Y.Z.md record from its prepared form to released; tagged vX.Y.Z on YYYY-MM-DD; ... and that release's row in docs/specs/index.md the same way, beside the changelog and the manifest version bumps. A second changed line in either file, a record without its row, a wrong version or date form, or any other spec still takes both reviewers.

Fixed

  • /setup check and /setup scan on native Windows print the one WSL line and exit 1 instead of a traceback (issue 193).
  • checks names the host config's Load limit (12 by default) for manual heavy jobs instead of a fixed 12; the setup beginner line says where the skipped steps come back, and the matrix migration's TypeSafe No branch says its plan removes no old section.
  • An approved check list (checks-approved) is bound to the full identity digest of its main checkout, so another repository whose eight-character code collides never inherits it; a record written before this release still counts until /setup add checks writes it again. The update checklist gains the missing 0.2.5 note for the compaction setting check.
  • readme_check.py hardening: a local link that resolves outside the repository fails K1 even when the file exists; file:, loopback, private and link-local URLs fail K1 (I1 for images) instead of being listed as EXTERNAL; heading and comment scanning run in linear time; override heading patterns over 200 characters are bad input; a closing fence must be at least as long as its opening one; anchors, table rows (C2) and the License line (L1) ignore HTML comments and code; a non-string manifest description fails D1 instead of crashing; and a second, unclosed <!-- after a --> on one line keeps the comment open.
  • afk: a merge pin and the merged...
Read more

v0.3.0

Choose a tag to compare

@fioenix fioenix released this 14 Nov 18:50
9cb1b55

Added

  • Models per kind of work in the assignment matrix: an agent line may carry models <rule> | <rule> ..., each <model>[ then <fallback>] for <kind, type or levels>, read in order with the first match winning (a rule naming no kind applies only to implementation). route_pick.py pick gains --kind (implement, code-review, security-review, navigator, advisor), makes --type optional and prints the model chosen for each dispatch, with quota-fallback when an Antigravity rule's then replaced a model whose quota pool (argon, gemini, claude-gpt, other) is out; an agent whose chosen model is out is dropped. review_route.py judges Antigravity's quota per review kind on the model its rules pick (--level for level-bound rules), matrix_setup.py show lists the rules, and /setup add delegation asks for them. Any malformed rule makes the whole matrix invalid.
  • merge_check.py --independent --fallback-family <family> lets an away lease merge on reviews by the author's own family when no other family is reachable: it needs a live lease, the author's family, positive out-of-quota evidence (a fresh reading with every pool below the floor, or a live rate-limit marker) for every other present family, a router trigger other than security, permission or privacy, no changed enforcer script or instruction file, and the exact fallback reviewer line on each counted marker. afk.py log --kind review-fallback --pr <n> --head <sha> records it for the user's return.
  • ledger.py finish --retro marks a record written after the fact.
  • /setup add afk inside a repository can list its own test-commands (exact commands, never starting with git, gh or rm) and top-level-files (exact top-level file names such as TASKS.md). During a lease, afk.py check answers Tier 1 for a listed test command, bare or through heavy_run.py run --, and treats a listed file as a known path, in that repository only; the strict and no-repository states list none, and a repository without its own lists uses the account defaults in ## afk (see the account-default lists below). Before, a lease could not run a repository's tests, which its own merge gate needs, or tick its task list.
  • afk.py start adds a This folder line with the rules of the repository it ran in (limits, decided categories, deploy targets, test commands, top-level files), and afk.py effective adds them as rules; before, both showed only the account values, so a user saw rules that did not apply in the repository. Re-running /setup add afk now offers the saved limits and table answers instead of the defaults.
  • afk.py check allows creating a new own work branch during a lease: git switch -c <branch> or git checkout -b <branch> (no start point) is Tier 1 in a found repository with Push and pull requests or Merge yes, never with a default-branch name; before, no branch could be created and the own-branch flow parked at its first step.
  • afk release targets: a tag, GitHub release or publish (git tag, gh release, npm publish, pnpm publish) runs under a lease only when the exact command is listed (Release targets in ## afk, or a repository's own release-targets), in addition to Tags and releases or Publishing at decide; /setup add afk asks for the list in its own question with the commands read back in the preview. Reading tags and releases (git tag -l, gh release view or list) needs no entry.

Changed

  • The README's afk mechanism diagram shows the whole away flow on one page, and the away decision diagram shows the evaluation table instead of Jev at 0.8.
  • Breaking: afk tags, releases and publishing that were Tier 1 with their category at decide alone now park until the exact command is listed as a release target (owner decision, 15/11/2026).
  • afk command and file lists load like environment variables (owner decision, 15/11/2026): Production targets, Test commands, Top-level files and Release targets in ## afk (afk_setup.py preview --production-target, --test-command, --top-level-file, --release-target) are the default for every repository, and a repository's own list replaces the whole list there. The strict and no-repository states use none, and an invalid ## afk list line fails start. Before, the three older lists were read only from a repository's settings. Production stays decidable only in a repository's own decide list.
  • The same-family review fallback (merge_check.py --fallback-family) may now merge a pull request that changes an enforcer script or instruction file when each kind of review has approvals at the head from two different models of the author's family, each marker naming its model as model=<name> right after the fallback reviewer line; one model, the same model twice, a REWORK from either, or a missing model name still refuses (owner decision, 15/11/2026). Every other fallback condition stays.
  • afk limits are counted per repository: each repository under one lease has its own runs, runtime minutes (from its first allowed action) and actions an hour, and a repository may set its own max-runs, max-runtime-minutes and max-actions-per-hour in /setup add afk (wider or narrower than the user value). Breaking for away loops: a spent repository answers spent with no end line, so the loop goes on in other repositories, exhausted is no longer returned, and a changed setup now adds the end line like an ended lease; a lease started before the update reads as invalid, so start it again.
  • afk categories and deploy targets can differ per repository: /setup add afk inside a repository may set its own decide list (an explicit list replaces your account's table there; every unlisted category stops), its own non-production-targets, and, only there, production-targets: exact production deploy commands that run while you are away only when production is in that repository's decide. Breaking for away flows: outside a repository and in the strict state every category stops and no deploy target applies, and gh release create now needs --repo github.com/OWNER/NAME equal to the verified target, like a merge.
  • afk decisions take an evaluation table before the deciders: judge.py ask --site away-evaluation prints the probability of each option ({"source", "of", "probabilities"}), and without Jev the host's small fast model writes the same table, checked by the new judge.py check-table --file <brief> --of <k>. The table goes to the advisor and the host as data and never settles a decision. Breaking for away flows: the away-decision site and its 0.8 threshold are removed, afk.py log --kind decision replaces rung jev with table (--settled yes --by jev|host-small --of <k> --probability <top p>, or --settled unavailable --by jev|host-small), and the ledger's away summary and the dashboard show the table's top probability instead of a Jev spread.
  • ledger.py summary leaves cancelled records closed late (by ledger stale or the coordinator) out of the counts, acceptance rates and medians and prints them as a separate total, and leaves finish --retro records out of the medians (a group with only such records shows -). The text table gains the quota, error and cancelled columns that --json has, and below it flags accepted runs whose log scan had findings and review acceptances recorded without --review-url. --json rows gain cancelled, closed_late, retro, accepted_findings and review_unposted; median_min may be null. The status dashboard uses the same rows.
  • ledger.py finish refuses a negative --rework-rounds, more than 2 rework rounds without a --reason, and accepted with --tests failed; start warns, and still records, when the model is not in the assignment matrix or the project label differs from the repository's ledger project setting.

Fixed

  • ledger.py finish --log-path accepts a log directly in .finolabs/logs/ of the --checkout checkout or of the main checkout, not only of the current one, so a coordinator in another worktree can close a run; a refused log gets one fixed reason line instead of a bare ledger: log refused. An empty log is now not_scanned, not no_findings. delegate "Where files go" names one place for logs.
  • Two ledger.py start calls in the same second can no longer share a dispatch id: the id is checked against the ledger under the finish lock.
  • afk no longer parks review posting: afk.py check allows post_review.py <number> and post_review.py <number> --body-file <run body> (either order) when the repository's target is verified and Repository comments is decide; the body must be post_review.py's own private run body, its first line one review marker and the rest free of secrets, machine paths and other gate markers.
  • The afk advisor rung follows the model table: afk.py check accepts an advisor run only on a model that delegate references/models.md lists as that family's frontier or strong tier (an unknown or fast-tier model such as Flash is denied), and for a Claude coordinator denies a Claude model served through Antigravity, so the advisor is frontier else strong as advisor choose mode says. Breaking for away flows: a new model name is denied until the table and the check list it.
  • afk.py log and review refusals of a schema error now name the flag and why (for example the advisor no row takes no --of, --option) instead of the generic setup message, which stays for missing, invalid or unsafe private state; no passed value is echoed.
  • advisor choose mode says to check the advise dispatch's ledger.py start argv with afk.py check, the only check that counts a run against the lease's run limit, and that an advisor family whose program moved after an update stays denied until /setup add quota records it again (the exact program record is kept on purpose). afk decide names delegate's judge.py as the copy check allow...
Read more

v0.2.6

Choose a tag to compare

@fioenix fioenix released this 13 Nov 17:56
66c597a

Added

  • plugin-packaging ships a README standard and checklist (references/readme.md): the sections in order, required and conditional, the facts a README takes from the repository, and checkable items for structure, badges, counts, hosts and install commands, descriptions, license, images, links and placeholders. readme_check.py checks the mechanical items offline and lists external links to open; it is advisory. A repository may adapt headings, aliases and a house style in .github/readme-standard.json. versioning and the plugin-packaging release checks now run the checklist before a release.

Changed

  • The README gains a short Privacy section linking PRIVACY.md, and the repository's own marketplace entry now carries the manifests' description.
  • Breaking: ## Agent roles in the host config is no longer read (owner decision, 13/11/2026). The assignment matrix is the only source of review roles (code review, security review, navigator, advisor) and the judge: a user with only ## Agent roles loses those roles and the Jev judge (reviews follow the capability order, the judge is the host model) until running /setup add delegation, which converts the section with matrix_setup.py migrate and asks consent again. /setup check reports that case as roles: missing, names an ## Agent roles left beside a matrix in legacy host settings, /setup add roles records roles and the judge in the matrix, and /update lists the step. The judge-call rule in CLAUDE.md and the judge section of PRIVACY.md now name only the matrix. ## Review posting and ## Agent levels are unchanged.
  • The README opens as a landing page: the icon, the tagline and subhead, badges for the release, licence and the three hosts, the banner, a "What you get" block, the install commands for each host and the request-to-merge diagram. "For engineers" (release notes, compaction steward, diagrams, Bricks) moves below the reference sections; nothing is removed. The banner (assets/banner.png) is also the new GitHub social preview, generated from .github/social-preview.json.
  • The setup beginner path asks only the essentials: the Delegation mode and Routing questions in user scope, and the workflow and Agent skills block questions in project scope. Moving the data root, the agent scan, the notes read-back, hiding local files, companions, extras and the instructions offer wait for a later run or add <feature>, each still with its own question. User scope ends with one line on how to adopt the review gate in a repository. A feature that needs the agent scan reuses the answer given earlier in the same run instead of asking it again.
  • afk.py check answers a missing, expired, stopped or exhausted lease with one fixed end line in its JSON telling the agent to stop any loop or wakeup it started for the lease and report to the user, so a host /loop started under /afk is told to end at its first check after the lease ends, without the user typing /afk stop (advisory: the agent stops the loop; no script kills it). A lease whose runs or runtime minutes are spent now answers exhausted instead of limit; limit is left for the rolling-hour action cap, which frees up. Breaking for scripts that match limit; match exhausted instead.
  • Breaking for away flows that log a merge: run the allowed gh pr merge <number> --match-head-commit <sha> check first. afk.py log --kind merge --pr <number> --head <sha> is refused unless an allowed gh pr merge <number> --match-head-commit <sha> check of the live lease pinned that number and head, so the head that lets the merge gate's cleanup remove a worktree is bound to an allowed merge check (not to proof that GitHub merged it); an allowed merge check records the pair in the private afk log. A check whose repository identity cannot be validated keeps its own folder, so the work-tree gate never falls back to the process's working directory.

Fixed

  • readme_check.py reads a hosts list in .github/readme-standard.json, so a plugin for Claude Cowork only, whose manifest is .claude-plugin, is no longer asked for a Claude Code badge, header mention and install section. A code fence that opens inside an unclosed HTML comment no longer hides the rest of the README.
  • While an away lease is live, agents no longer stop to wait for the user: afk "Decide" now names every hand-off it covers (a question tool or text question, a wrap-up question, an open decision in a plan, a skill step where the owner decides such as review-gate's stop after two rework rounds) and says never to end a turn waiting for an answer. The away question gate's denial and the session-start away line say the same: never ask or wait for the user; only stop work parks.
  • afk.py check lets one exact <!-- fnl-router depth=... trigger=... --> line through in pull request and commit text, so an away flow can open the pull requests that merge_check.py accepts; every other <!-- fnl- marker, a second router line and a malformed one stay forbidden.
  • /setup add afk asks fewer, clearer questions. Inside a repository it first asks whether the rules are for your account (every repository) or only this repository. After the account rules it asks once whether this repository should differ and then only the ticked keys, instead of asking the five repository keys again; the repository's stored object is shown first, a No previews its removal behind the save question, and project_progress.py settings ... --dry-run now returns current (the stored object) and changed (the effective values the change would move). The five limits are one question with the suggested values, asked as numbers in one round only on No. The table of what waits for you starts with "Keep the table as shown?", and only on No lists the seven changeable rows as two multi-selects (ticked kinds the agents may decide; unticked or unanswered ones wait for you), then confirms each row moved to decide. Each permission keeps its own question. The repository part lives in setup references/afk-project.md.
  • readme_check.py no longer hides README text between a <!-- and a --> written inside code spans or blocks: comments are blanked only outside code. The README reference states which override errors exit 2.
  • afk.py check lets an install that is the main checkout (a -dev install read from it) remove a merged linked worktree, as afk "Merge gate" says: git worktree no longer counts as a command that changes the install's own work tree, since its only Tier 1 form removes another worktree and still refuses the main checkout, a forced removal and any worktree holding the install (owner, 13/11/2026). The worktree rule docs name log --kind merge --pr <number> --head.
  • afk.py check: a merge check's number takes the ASCII grammar of log --pr.

v0.2.5

Choose a tag to compare

@fioenix fioenix released this 13 Nov 04:05
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 un...
Read more

v0.2.4

Choose a tag to compare

@fioenix fioenix released this 11 Nov 04:23
1600861

v0.2.4 - 2026-11-11

Highlights

This release helps you check your work when cloud CI is expensive or out of minutes, and makes updating the plugin walk you through what is new.

  • Repository checks. /fnl-setup add checks finds a repository's checks and saves the list you approve. checks.py run runs it locally, with an optional pre-push hook.
  • A local receipt when cloud CI cannot run. Opt in per repository, and a passing full run can stand in for missing or failing cloud checks at the merge gate. It is self-reported, so it suits a repository one person works on.
  • Cheaper CI. Setup can write a GitHub Actions workflow that skips document-only changes and cancels superseded runs, and checks.py usage shows what your CI minutes cost.
  • Guided updates. /fnl-update now goes on to a checklist of what the new version adds; /fnl-update whats-new only shows it. You can also pick or pin a version.
  • Renamed. fnl-verification is now fnl-checks; the old name keeps working through 0.2.x.

Not verified on a live host: a real git push through the installed hook, and a run of the generated workflow on GitHub Actions.

Added

  • /fnl-setup add checks reads a repository's package.json scripts, declared Python test runner and Makefile targets without running them, lists run: lines from its CI workflows as unverified candidates it never converts, proposes an ordered check list (a name and a JSON argument array each, never run through a shell by the plugin; known shells, launchers and inline-code forms are refused as a best-effort guard, not a sandbox; skipped items named) and, after a digest-bound yes, checks_setup.py saves it as that repository's entry in the host config's own ## Checks section (at most 20 checks of 1 KB each); ## Project settings is unchanged.
  • fnl-checks' checks.py run [--changed] [--base <ref>] runs a repository's stored checks in order through heavy_run.py and stops at the first failure (--changed narrows python -m unittest discover -s tests over a flat tests folder to one discovery per module select_tests.py names, and runs every other form and layout whole); an opt-in per-repository local-receipt mode lets a whole passing run on a clean, unmoved head stand in for missing or failing hosted checks in merge_check.py (only when the checkout's origin is the --repo repository, and the latest full run decides, one full run per repository at a time), which names it checks: local receipt (<head>); and checks_setup.py hook install|remove adds or removes an optional git pre-push hook that runs checks.py run --changed, refusing when core.hooksPath is set and never overwriting a different hook without a second flag.
  • checks_setup.py ci preview|apply writes, after a digest-bound yes, one new GitHub Actions workflow from a repository's stored checks: pull requests, pushes to the default branch, a weekly schedule, superseded runs cancelled, document-only changes skipped, caching for a known ecosystem. It refuses a link on the path and an Actions expression, keeps a backup, never overwrites without --replace-existing, and never commits or pushes; fnl-checks' references/ci.md describes it.
  • fnl-checks' checks.py usage --from <file> [--files-from <file>] [--limit-minutes N] summarises hosted CI runs from a gh run list --json file: run count, minutes rounded up per run, minutes per workflow and per event, the share spent on document-only changes and days left at the observed rate. It makes no network call and starts no process.
  • /fnl-update now goes on, after the host update, to a checklist from the new references/upgrades.json (one closed-schema entry per release, starting with 0.2.4) and offers each open setup step through fnl-setup with its own consent; /fnl-update whats-new only shows the checklist, and the session-start hook adds one read-only line naming /fnl-update while the installed version's notes are unseen.
  • update_check.py gains whats-new, onboard seen <version> and onboard skip <id> (the only writers of onboard.json: last_seen and skipped ids), target [<version>] [--evidence <file>] (latest by default; a chosen version only with matching gh release view evidence, an older one reported as a rollback with what is lost) and pin <version> / unpin, which silences the newer-release notice; the network check is unchanged. The Codex refresh procedure is now marked verified, and fnl-versioning asks the release pull request to write the upgrade-notes entry, which a test enforces.

Changed

  • fnl-verification is now fnl-checks, which also covers where a repository's checks run (locally or in CI); /fnl-verification keeps working as a user-only alias through 0.2.x.

v0.2.3

Choose a tag to compare

@fioenix fioenix released this 10 Nov 06:11
30d6616

v0.2.3 - 2026-11-10

Highlights

This release lets your agents keep deciding while you are away, gives you one table for who does what, and adds a safe way to stop before restarting the machine.

  • Away decisions. During an /fnl-afk lease, a question the agent would ask you goes up a ladder instead: a recorded rule, Jev, an advisor of another model family, then the agent's own choice by the project's goal. Production, live environments, real users and customer data always wait for you, and setup lets you choose which other categories stop or are decided. Every attempt is recorded in the ledger, and on return you mark each decision kept or reversed. On Claude Code a hook stops the question tool during a lease.
  • Assignment matrix. /fnl-setup add delegation writes one ## Assignment matrix with each agent's levels, kinds of work, model and roles. route_pick.py answers which agent should take a task, and the ledger records when a dispatch goes off that order.
  • Pause. /fnl-pause (short name /fnl-gtg) drains running work before a restart or shutdown, writes handovers and reports what recorded work remains. heavy_run.py refuses new jobs while paused.
  • Names and help. /fnl-afk is the main short name for the away mandate (/fnl-brb still works), /fnl-help replaces /fnl-ask (kept as an alias), and every background run gets a name from ledger.py start.
  • Fixed: judge.py now accepts the judge service's real reply; before this, calls fell back to the host model.

Two sequence diagrams show one delegated task and one away decision step by step.

Not yet proven on an installed host: the question-tool hook and the typed-skill event. The attempt to cut log-scan false positives was dropped after security review; only a clearer folder message ships.

Added

  • With metrics consent on Claude Code, a skill typed as /fnl-name records the existing skill event (name only, never prompt text), and the first file edit of a session with no router receipt prints one advisory reminder to invoke fnl-coding-router; Codex and Antigravity stay unmeasured and advisory.
  • select_tests.py [--base <ref>] in fnl-verification lists the tests.test_* modules that name a file changed since the base (default origin/main, copies of a shared file included), reports changed files no module names, and groups the modules two per command; it runs only git diff --name-only and no test.
  • fnl-afk (away from keyboard) is the main short name for the away mandate; fnl-brb remains an older alias, and both load fnl-autopilot.
  • User-invoked fnl-pause (short name fnl-gtg) drains work before a restart or shutdown: pause.py writes a pause marker, reports the recorded work that remains and resumes afterwards, and heavy_run.py run refuses new commands (exit 77) while a valid marker exists.
  • matrix_setup.py (validate, preview, apply, show, migrate, candidates) and the shared assignments.py parser write and read the ## Assignment matrix section (levels, types of work, model and roles per helper, plus the judge) with a digest-checked, backed-up, section-only write; the readers still use the old ## Agent roles and ## Agent levels sections until a later release.
  • route_pick.py pick --type <type> --level <1-5> [--author <family>] answers the routing query from ## Assignment matrix and local state in at most five lines, and ledger.py start takes --install, --work-type and --level, records pick_rank and warns when a dispatch is off the preferred order, with the shares per agent in ledger.py summary; a preference never refuses a dispatch.
  • Away decisions, scripts part: autopilot.py log --kind decision records each ladder attempt against a closed schema in the ledger (no question or option text) and autopilot.py review marks it kept or reversed; ledger.py summary adds an away section per lease; autopilot.py check reads the stop-or-decide categories and Non-production targets, opening tags, releases, publishing and listed deploys only when chosen, and knows the judge and advisor calls of the ladder (the advisor only through the family's executable recorded for the quota probe, with a literal -- or -p before a screened brief); judge.py gains the away-decision choice site; autopilot.py packet-check caps and screens the precedent packet; on Claude Code the opt-in question gate (away_gate_hook.py, FNL_AWAY_GATE=off) denies the question tool during a live lease, and session start prints a fixed line while one is active.
  • Away decisions, skills part: fnl-autopilot "Decide" (references/decide.md) runs the ladder (a recorded rule, Jev, the advisor, the host) with two checks against "Where it must stop", the precedent packet and one record per attempt, and on stop lists the lease's decisions for kept or reversed; fnl-advisor gains a choose mode that returns exactly one option or none during a lease; /fnl-setup add autopilot asks the stop-or-decide categories (two fixed at stop), optional Non-production targets and a separate excerpt consent, off by default; fnl-status shows the away decisions per lease.
  • ledger.py stale previews the open ledger records past their budget with what their run log shows (exists, last-write age, exit line), and ledger.py stale --close --digest <prefix> closes exactly those as stalled, error or cancelled with a closed-late note within 60 minutes of the preview; it never marks a record accepted.
  • Two sequence diagrams in docs/diagrams/ (one delegated task from dispatch to merge, and one decision during an away lease) and an updated architecture diagram; fnl-versioning now asks for diagrams, README, docs and in-product help to be reviewed against what ships before a release is proposed, and fnl-open-source-review checks that they match (both advisory).

Changed

  • Review routing, the judge setting, checkup.py and the session-start summary read roles and judge from ## Assignment matrix when it exists: a valid matrix with its own yes consent alone decides (a role no line names is unassigned), an invalid matrix or one without consent leaves roles and judge unset, and without a matrix the old ## Agent roles section applies until 0.3.0; the matrix gains the optional role advisor, and session start adds Assignments: <n> agents set when the matrix is valid.
  • The plugin help skill is now fnl-help (fnl-ask stays as a thin alias); its map lists every shipped script and setup feature and answers which agent takes which work, a test fails when either drifts, and one test computes the skill count from the folders, which only the manifests and README now state.
  • ledger.py start prints a ready-made run name on stderr (label: codex sol: demo PR 7 code review) for the host's name field, because only 8% of 1,555 background runs measured on one machine used the documented form; --task now holds only scope and activity.
  • fnl-review-gate fits the 10,000-byte skill limit (17,999 to 9,925 bytes): engines, packet contents and the impact set move to references/reviewers.md, the manual merge check, batching and the low-risk lane to references/merge-check.md, relay and posting setup to references/relay-and-posting-setup.md; no rule changes meaning and every cited section keeps its name.

Fixed

  • judge.py accepts the reply the judge service actually documents (model and usage beside answers, type beside the answer); it refused every real reply and fell back to the host model.
  • A too-open run-log folder is refused with one fixed message that says to run chmod 700.

Full changelog: v0.2.2...v0.2.3

v0.2.2

Choose a tag to compare

@fioenix fioenix released this 09 Nov 09:21
5a282ef

v0.2.2 - 2026-11-09

Highlights

This release lets you tell the plugin which task levels each of your agent commands can take, finds your extra Claude Code commands for you, and fixes two defects in the heavy-job wrapper and the quota probe that only showed up on a real machine.

  • /fnl-setup add delegation asks which of five task levels (1 easy to 5 ultra hard) each helper takes, including extra Claude Code commands such as a claudework shell function; fnl-delegate routes by those levels (advisory) and prefers a declared command inside its range while it has quota.
  • levels_setup.py candidates lists, as names only, the claude* programs on PATH and the ~/.claude-<label> folders, so setup can offer them. It runs nothing and never lists shell functions.
  • heavy_run.py reads free memory from the kernel's own figure instead of a figure that refused a review on a 16 GB machine, and status prints waiting: about <n>. A live check found and fixed two counting defects (a 450 KB ps listing and Homebrew Python shown as Python).
  • The quota probe reads a Claude Code reset time printed on the hour (6pm), so Claude installs are no longer all marked unverified.
  • Native Windows is documented as unsupported: scripts that need pwd or fcntl stop with one clear message and the README points Windows users to WSL (issue 193).

Not verified on a live host: /fnl-brb, Antigravity review posting, pgrep/ps -p on Linux and native Windows. See the release record and the changelog.

v0.2.1

Choose a tag to compare

@fioenix fioenix released this 08 Nov 21:55
00dee56

v0.2.1 - 2026-11-09

Highlights

This release adds a wrapper that serializes heavy local jobs run through it, lets the plugin run a time-boxed away mandate for your host agents (/fnl-brb), and tightens setup, delegation and review checks.

  • heavy_run.py runs one heavy command at a time per machine, with load and memory gates, a wait cap and a maximum run time.
  • fnl-autopilot (alias fnl-brb) starts and stops a time-boxed standing mandate; each action is classified by code and merges need your recorded Merge answer, reviews and green checks.
  • The parallel run limit is per repository (default 3); setup offers the Agent skills block in AGENTS.md and CLAUDE.md; Check reports plugin enablement per Claude home.

Not in this release: the opt-in heavy-job gate hook described in docs/specs/heavy-jobs.md (the heavy_run.py wrapper is in), the review residual notes and the smaller setup items; see the release record.

Added

  • /fnl-setup add autopilot asks each away-session setting separately, with an explicit Merge decision as standing authority. Its writer previews the canonical ## Autopilot section and applies only a confirmed, unexpired SHA-256 digest, preserving other host bytes with a backup and refusing links or changed files.
  • fnl-delegate/scripts/autopilot.py (first slice of autopilot, part 1): start, stop, check and log for a time-boxed lease read from ## Autopilot in the host config. check classifies a command's argv or a change's paths into Tier 1 or Tier 3 by code (unknown input is Tier 3), enforces lease expiry, the stop marker, setup-hash changes and run, runtime and hourly action limits, and denies merge unless setup says Merge: yes and the argv pins one full head SHA (that the SHA is the reviewed current head stays the caller's duty). State is private and local; nothing is executed or sent. merge_check.py --independent additionally refuses the merge when the families routing would pick for the reviews include the pull request author's own; which family posted each counted marker stays the caller's duty. The skill, the /fnl-brb alias and the setup section are available.
  • User-invoked fnl-autopilot starts and stops a time-boxed away mandate with action classification and an independent merge gate; fnl-brb (be right back) is its thin alias. Recorded setup is required; use /fnl-setup add autopilot to record it. The live loop, dashboard tab, scheduler, notice hook and host backstop remain deferred.
  • fnl-delegate/scripts/heavy_run.py runs one heavy command at a time on the machine: it holds a kernel lock in the fixed anchor folder, waits for the one-minute load and free memory, exits 75 after a wait cap, refuses detached Docker launches, starts the command with the allowlisted environment records who started the job (label, repository, dispatch id, program name, pids, times) so status can show it without guessing the holder, and stops a job past its maximum time (default 90 minutes). It replaces the mkdir recipe.
  • Record an optional coordinator session hash on dispatch starts from CLAUDE_CODE_SESSION_ID (never from a command-line option); keep only a short SHA-256 hash, or null when unknown.

Changed

  • The user-contract template in fnl-setup instructions now has a delegated-acceptance line: all tests of the package or test module holding the changed code and its traced consumers, chosen from the diff, never a repository-wide suite without the owner's request, unless a repository's owner instructions scope task-level proof more narrowly.
  • The parallel run limit and per-family caps are now set per project (repository) in ## Project settings (parallel, parallel-caps), default 3, and count only that repository's open dispatches; there is no machine-wide total, and family quota pools stay shared. fnl-setup add delegation asks for the limit inside the repository.
  • fnl-delegate's description, the fnl-ask map and the README now say plainly that it hands work to other agents and runs it; the skill name is unchanged.
  • Delegated runs must start in the host's own background mode, one runner per call. The dispatch step and the background rule now name ( … ) &, nohup and disown as the forms that hide a run from the host's task list; the handoff step and the after-compaction reminder name the first two. The hook script changed, so Codex asks to trust the hooks again.
  • After a compaction, the session-start hook prints one fixed delegation reminder (unless Delegation mode is manual), and the handover template carries a Delegation state line, because delegation to other agent families fell after compaction. The hook script changed, so Codex asks to trust the hooks again.
  • fnl-setup now offers the ## Agent skills block in every project context file the hosts read, AGENTS.md and CLAUDE.md, each with its own preview and yes, and skips a CLAUDE.md that only imports AGENTS.md.
  • /fnl-setup check reports, per Claude Code home, whether this plugin is enabled, disabled, absent or unverified in enabledPlugins, showing no other settings content. fnl-delegate and fnl-verification now document the Codex review and posting run without a sandbox, record a run stopped by a permission request as needs-permission rather than quota, and explain that ledger.py finish refuses --head and an unverified accepted for a review dispatch, which is accepted with --accept-unverified after reading the posted marker.

Deprecated

  • The ## Limits parallel lines in host.md are no longer read; check-config reports them as legacy_limits.

Fixed

  • Check shows the version a Claude Code folder marketplace declares beside the one the installation registry records, and stays unverified when they differ.
  • Check reads the Claude home and instruction file selected by an absolute CLAUDE_CONFIG_DIR, and lists other ~/.claude* homes that hold plugins with their recorded plugin version.

v0.2.0

Choose a tag to compare

@fioenix fioenix released this 08 Nov 04:20
8dba36b

v0.2.0 - 2026-11-08

Highlights

This release makes delegated work easier to direct, review and check, with a shorter setup and clearer evidence of what each host can actually do.

  • Assign code and security review roles, prefer a family other than the author’s and post through one guarded path.
  • Check project direction with the navigator and use one judge interface, with a small host model by default and optional Jev.
  • Require verification receipts for delegated ledger acceptance, bound to the checkout and current commit.
  • Start headless runs with environment allowlists and inspect finished logs through local scanning and redaction.
  • Track repository outcomes, one-file work records and delivery progress alongside Spec Kit.
  • Choose Quick or Advanced setup, with plain-language approvals and optional features loaded when needed.
  • Record opt-in context snapshots from Claude Code and Antigravity status lines for compaction decisions.
  • Run parallel agent pools within advisory total/family limits and see capacity, quota pools and readable times on the status board.
  • Read corrected host contracts that distinguish documented support, local checks and unverified live delivery.

Upgrade notes

  • Run /fnl-setup check after updating ($fnl-setup check on Codex); it reports configured, missing and unverified features without changing them.
  • Use /fnl-setup add delegation to review the new ## Limits lines: - Parallel dispatches: at most 6 at one time. and optional - Parallel per family: claude=2, codex=3 are examples, not prescribed values; missing or invalid totals use 5, and older parenthetical wording needs replacement.
  • Use /fnl-setup add roles to choose review families and the judge; assignments grant no posting rights, and each posting grant needs its own approval and observed check.
  • Claude Code may ask you to trust the changed plugin hooks after updating (a new StopFailure entry records rate-limit markers only when the quota probe is on).
  • On Codex, open /hooks and check trust again: changed bundle contents can require renewed trust even when hook declarations are unchanged.
  • Context recording needs a status-line command in your own host settings; setup previews a new command or gives a recipe to compose with an existing one, which it leaves untouched.
  • Follow the replacements under Removed; data-root location and existing consent remain unchanged, and new opt-ins still need separate approval.

Added

  • Review gate: User-assigned review roles and deterministic routing prefer a family other than the author’s and disclose same-family fallbacks.
  • Review posting: A guarded wrapper binds review comments to the checkout repository and checked drafts, while posting status needs an observed author and marker.
  • Navigator: Bounded course checks compare confirmed outcome anchors with repository evidence and suggest up to three ways to address drift.
  • Verification: Current-head receipts record targeted tests and criterion judgments before delegated ledger work can be accepted, with a reason-digested exception counted separately.
  • Judge: One interface supports a small host model by default and optional Jev, with explicit fallback and judgments that grant neither authority nor acceptance.
  • Context: Opt-in Claude Code and Antigravity status-line recorders supply fresh token counts to the compaction steward without reading conversation content.
  • Delegation: A task-routing helper applies the grade matrix, and a read-only Claude Code/Codex startup summary shows roles, posting state, cached quota and open dispatch counts.
  • Quota: An advisory Claude Code rate-limit marker (family and time only, recorded from the StopFailure event while the quota probe is on) steers routing away from a limited family for at most five hours; a newer reading or your own statement overrides it.
  • Capacity: Advisory parallel-slot checks expose a user-selected total from 1 to 32 and optional per-family caps in setup, session start and the dashboard.
  • Work tracking: Per-repository settings, dated outcome anchors and bounded one-file work records complement Spec Kit artifacts.

Changed

  • Setup: A short core uses user/project scopes, Quick/Advanced modes and on-demand features, with plain-language approvals explaining the effects of yes, no and undo.
  • Setup: Companion questions show checked install commands without installing anything, and explicitly declined features stay declined until requested again.
  • Delegation: Agent families act as pools of concurrent runs, with family/model labels and the coordinator retaining model selection, quota and permission decisions.
  • Quota: Opt-in readings happen on demand before routing, at least 15 minutes apart, replacing hook-started periodic workers and adding advisory Claude Code rate-limit markers.
  • Review gate: Adopted repositories require exactly one valid router marker in the pull request body at merge time, in addition to current-head code and security review comments.
  • Session hygiene: Compaction decisions and the safe-boundary procedure move into the session-hygiene skill, informed by the configured judge when needed.
  • Dashboard: Repository pipelines, task/package states, ledger mismatches, quota pool models and local readable times appear on a page that scrolls with the browser.
  • Dashboard: Repository-scoped status replies retain cross-repository detail on the board, with bounded review chains, detailed tables and five retained repository snapshots.
  • Docs: Host contracts and architecture/lifecycle illustrations distinguish advisory instructions from enforcement and documented contracts from live proof.

Deprecated

  • Work tracking: Global Ledger project: and Task file: lines remain readable in 0.2.0 and 0.2.1; move them to per-repository settings before their planned removal in 0.3.0.

Removed

  • Verification: The fnl-verifier agent is replaced by fnl-verification; run its verify_claim.py, read the evidence and decide acceptance as coordinator.
  • Session hygiene: The fnl-steward agent is replaced by the Context pressure procedure and threshold script in fnl-session-hygiene.
  • Setup: The fnl-scan-agents skill is replaced by /fnl-setup scan for inventory refresh.
  • Setup: The fnl-set-mode skill is replaced by /fnl-setup mode for Delegation mode and Routing.
  • Setup: The fnl-setup-instructions skill is replaced by /fnl-setup instructions in the same setup session.

Fixed

  • Review gate: The posting wrapper accepts standard macOS Homebrew gh while retaining narrow prefix, ownership and nested-repository checks.
  • Setup: Correct starter commands and retired-pointer checks support custom Codex instruction folders and Antigravity global rules through the guarded writer.
  • Run hygiene: Benign log pairs no longer consume the secret-hit cap, and environment-dump detection requires ten distinct environment-style names.

Security

  • Judge: Canonical opt-in, capped closed payloads, secret/path screening, verified transport and a 24-attempt process cap bound external judgments, with explicit host fallback.
  • Run hygiene: Brief lint, child-environment controls, bounded raw/normalized log scans and digest-bound redaction reduce exposure while checking file identity and recent writes.
  • Setup: Malformed declined-feature records, nonregular instruction targets and environment-selected custom Codex targets are refused before writing.

Known limits

  • Local fixtures and vendor documentation do not prove live hook, setup, review-posting or status-line delivery on each host; navigator/judge validation remains limited and not independently established.
  • Codex has no external status-line command for this recorder; Claude desktop Code-tab delivery and composition with user-owned status lines remain unverified, and latest-snapshot selection can be ambiguous across parallel sessions.
  • Antigravity invocationNum == 0, interactive stop delivery and /compact availability remain unverified; the fixed stop message still names /compact, with /new or /fork documented as alternatives.
  • Relay receipts and dashboard relay counters are not built; temporary relays are uncounted and still subject to existing approval rules.
  • Run hygiene covers allowlisted headless command forms; reachable HOME credential files, interactive/app sessions and in-session subagents remain outside that protection, and a private HOME is not implemented.
  • Parallel-slot checks are snapshots, not reservations or ledger enforcement; live scheduling, provider quota reads and the Claude Code StopFailure payload remain unverified.
  • The update check remains on by default, independently of metrics consent, and contacts the fixed public GitHub endpoint; FNL_UPDATE_CHECK=off disables it, while Codex check/notice delivery and Codex/Antigravity update procedures remain unverified.
  • Optional Jev calls send bounded English evidence to TypeSafe only when invoked with opt-in and authorized key/budget; live network behavior is not established by this release's local checks, and unavailable calls require host fallback.
  • Autopilot, the merge permission guard and the goal skill are deferred to 0.2.1; multi-account agent routing, relay accounting, private HOME and the Antigravity stop-message/Codex PreCompact decisions also remain deferred.

Full changelog: v0.1.10...v0.2.0