Repository navigation
Releases
Pendulum 0.10.7
Pendulum 0.10.7
Upgrading from v0.10.6
- Re-pin to
clockgrove/pendulum@v0.10.7. The workflow template is unchanged.
Fixed
- #631: a fixed review finding could Block. The finding check shows its judge the finding's own file and then the PR's other changed files, cut at 8,000 characters in git's path order. When an answer fixed a finding in a different file from the one the finding was anchored on, for example a doc file for a finding on a code line, that fix could fall past the cut. The judge then answered "unresolved" and the stage Blocked with the fix already on the head. Files the reply names now come first, then those the finding names. A cut list ends with the paths it left out.
Release check
Claude passed end to end on Actions, on pendulum-scratch:
- one finish card;
- every stage records the candidate build;
- the progress comment's total matches its headline.
Pendulum 0.10.6
Pendulum 0.10.6
- Test razor (group 2): 626 tests to 270 across 34 files (#629)
- Test razor (group 4): 34 test files cut to must-have cases, 552 tests to 248 (#627)
- Test razor (group 1): 696 tests to 256 across 36 files (#625)
- Test razor (group 3): cut non-must-have tests in 35 plugin test files (#623)
- withClaim ends its heartbeat with its stage: no claim written back after the drop (#626) (#628)
- Declared generated files: take the default branch's side and regenerate, checked exactly (#619) (#624)
- Card: decision records counted apart, and the card and progress figures agree (#617) (#620)
- Tests: no file outlives its tests; unclaimed and stuck-checks cut to must-haves, deterministic (#622)
- Finish card: working and waiting time apart, each wait with its cause (#614) (#618)
Upgrading from v0.10.5
- Re-pin to
clockgrove/pendulum@v0.10.6. The workflow template gains one line,PENDULUM_REGENERATE: "". Re-render with setup to have the pin and the new line written for you; leaving it out means no regeneration. - Generated files in a conflict (#619). A repository can now declare generated files (goldens, snapshots) with the
pendulum-generatedattribute in.gitattributeson the default branch, and set theregeneratecommand (PENDULUM_REGENERATE, workflow-only). In a send-back that conflicts, Pendulum then takes the default branch's version of those files and has the command write them again, instead of judging them hunk by hunk. In every stage, a head whose change touches a declared path is checked: the committed files must be exactly what the command writes. See docs/INSTALL.md, "Generated files". Repositories that set nothing see no change, except that a conflict over the 64 KiB limit now names this step as the fix. - Finish card summary.
passes.totalno longer counts decision records. They are counted apart, in a newdecisionRecordsfield, as the report counts them (#617). An older card is still read correctly.
Changed
- The finish card shows working and waiting time apart (#614). A wait on a blocked-by issue or a pause is its own row with its cause, for example "On hold · blocked by #42".
- Card, ledger and report agree to the cent (#617). Each pass's dollars are rounded once, the same way everywhere. When a row's cents are shared out to add up to the total, the card says so.
- The live progress card (#617) shows one row per CI wait, and a rebase stage is named as one from its first write.
Fixed
- #626: in local mode, a claim heartbeat still in flight when its stage ended could write the claim back, or hang in an outage wait. The stage now ends its heartbeat first.
- #621: a test could hang the suite under load. CI now bounds every test at 120 s and forces each test file to exit.
Tests
The suite was cut to must-have tests: money, trust and security, state safety, machine-read formats, and one regression per real defect. It went from about 2,400 tests to about 1,130.
Release check
Claude and Copilot passed end to end on Actions, on pendulum-scratch:
- the live progress card held across 28 written states;
- one finish card per Objective, with working and waiting time apart on the blocked one;
- the card and
pendulum reportagree on passes, decision records and dollars; - the repo owner is trusted;
- every stage records the candidate build.
Pendulum 0.10.5
Pendulum 0.10.5
- Re-check a usage-limit hold whose reset is far off, and release it once the account answers (#613) (#615)
- Report: a --self login is trusted only on a positive answer (#607) (#612)
- Count the card's and the report's records only from verified writers (#607) (#610)
- Progress card while running: a run's cost at its end, no ended Now row, rows in time order, reruns stay live (#611)
Upgrading from v0.10.4
- Re-pin to
clockgrove/pendulum@v0.10.5. The workflow template is unchanged. - Finish cards and
pendulum reportcount stage, spend and decision records only from authors whose write access is verified (#607). A read-only member's markers count nothing. In the report, an author whose permission the call budget didn't reach, or GitHub refused, givesauthor-uncheckedorauthor-unverifiableinsource.missing, and their records are unknown cost, never dollars. pendulum report --self: a login is listed insource.trusted.loginsonly once it is decided a writer. Otherwise it is listed insource.trusted.refusedwith a reason:cannot-push,unverifiable,uncheckedorno-record.- Local mode: while a usage-limit hold's reset is more than an hour away,
servere-probes the held account about every 30 minutes with one tiny call, and releases the hold when a model the limit applies to answers (#613). The probe's cost is in serve's log.
Fixed
- #608, the live progress card:
- a run's cost shows in the headline and its row as soon as the run ends;
- a step that has ended is never the Now row, and there is none between steps;
- rows are in time order with Now last;
- a rerun pass shows its own Now row;
- every write goes through one queue, so the exit write is always last.
Release check
Claude and Copilot passed end to end on Actions, on pendulum-scratch:
- the live progress card was read through 33 written states, and all checks held;
- the finish card and the report agree;
- the repo owner is trusted;
- every stage records the candidate build.
Known limits
- A card's
passes.totalcounts decision records, while the report counts them apart (decisionRecords). A stage with AI judge calls shows one more pass on the card.
Pendulum 0.10.4
Pendulum 0.10.4
- Trust a MEMBER or COLLABORATOR author only on a verified write permission (#605) (#606)
- Copilot stages: cards name models by the report's table; ran names Copilot's own models (#600) (#604)
- Unlisted reviewers' threads reach the answer pass under a resolved-conversations rule (#595) (#597)
- Pendulum's own git goes through the stage's proxy (#599) (#602)
- Local mode: an unread review list counts no line comment as Pendulum's reviewer's (#601) (#603)
- A proxy's 5xx to CONNECT reads as this machine not reaching the host (#594) (#598)
Upgrading from v0.10.3
- Re-pin to
clockgrove/pendulum@v0.10.4. The workflow template has not changed, so re-render only if you want the pin written for you. - Trust is stricter. An org member or collaborator counts as a writer only when Pendulum reads their real permission as write, maintain or admin (#605). Writers listed in the repository's collaborators, the owner and pinned maintainers are unaffected. Other members cost one permission read per 10-minute window. Read-only members' comments, replies and markers no longer count as a writer's.
- Review Apps you didn't list (GitHub's Copilot reviewer, the Codex connector, and so on) now reach the answer pass when your rules require resolved conversations, instead of Blocking at merge (#595). Their text is always untrusted data. Threads from strangers and from
github-actions[bot]still Block.
Fixed
- #594: a proxy that can't reach GitHub (5xx to CONNECT) reads as an outage, so stages wait it out.
- #599: Pendulum's own git uses the same proxy as its API calls (local mode).
- #601: local mode: a failed reviews read no longer lets other comments pass as Pendulum's reviewer's.
- #600: Copilot stages show display model names. A bring-your-own-key model id, which can carry an AWS account number, is never published, and any model Copilot didn't confirm as non-BYOK is withheld.
Release check
Claude and Copilot passed end to end on Actions, on pendulum-scratch. That covers the cards, the recorded Pendulum version on every stage, Copilot model names with no raw ids, Pendulum's own finding answered and resolved, and the report reading cards and stages identically.
Known limits
Pendulum 0.10.3
Pendulum 0.10.3
- CI quality: plugin tests in four slices, one quality check; the slowest tests off the wall clock (#591) (#593)
- Progress comment and finish card read at a glance, in one design (#589) (#592)
- Report clocks a renamed Objective as renamed in every scope (#587) (#590)
- Visible times in a chosen time zone: PENDULUM_TIME_ZONE and the Objective's own (#586) (#588)
- Finding check judges a PR-body answer on the body's change (#585)
- Docs: list a test runner as a validation command so answer rounds can settle a finding by test (#583)
- run --wait: honour a failed start's backoff between resumes; heartbeat on a due poll that resumed nothing (#575) (#580)
- Report answers the owner's first questions: churn, per-merged cost, route, day, Blocks (#570) (#581)
- Report fields measure what their names say (#569) (#578)
- Report books model dollars by the model each pass ran on, with Claude sub-agent rows (#568) (#576)
- Report names the stages recorded before the rename it does not read: legacy-records (#567) (#579)
- run --wait waits out a network pause in place; exit 75 when left paused (#575) (#577)
- Report periods read each Objective's whole thread; planned calls equal calls made (#563) (#572)
- Recording fixes from the report validation: priced gateway decisions, finish cards, spend lines (#565) (#574)
- Report correctness batch from the validation (#566) (#573)
- Report learns Pendulum's App id exactly when the token cannot read variables (#564) (#571)
Upgrading from v0.10.2
- Re-render the workflow with this release's setup:
/pendulum:pendulum-setup, orsetup.mjs --apply workflow --action clockgrove/pendulum@v0.10.3. The template gains one setting,PENDULUM_TIME_ZONE. - Time zone (optional): set the repository variable
PENDULUM_TIME_ZONEto an IANA zone (e.g.America/Los_Angeles).- Visible times in progress comments, finish cards and notices then read like "9:00 PM PST".
- An Objective filed through the pendulum-file skill records its filer's zone on its first line, which wins over the repo setting.
- Local mode defaults to the machine's zone, Actions and the cloud sandbox to UTC.
- Hidden markers and machine-read text stay UTC.
- Validation commands: list a test runner that can take one test file as a validation command (e.g.
pnpm check, pnpm test), so an answer round can settle a finding by running its test. pendulum reportnow writes schemapendulum-report/2. Fields were renamed or redefined to say exactly what they measure; see docs/REPORT.md. Anything that reads/1needs updating.- Finish cards:
- The card has a new design with no images.
- Blocked, Question and closed-unmerged Objectives now get a card too, one per Objective, updated in place.
- A non-merged card never counts as a merge.
New
pendulum report, validated against an independent oracle: zero differences on synthetic and real data.- Dollars are booked by the model each step ran on, with Claude sub-agent rows.
- Answers cover churn, cost per merged Objective, route, day, Blocked stages and Factory-era records (
legacy-records). - Planned calls equal calls made.
- Progress comment redesign: a headline, a table of steps with time and cost, a timeline chart, and the full log in Run details.
run --wait/review --waitwait out a network loss in place, keep their claim alive and back off between retries. They exit 75 ("paused, resumable") at the limit, and exit 0 only when the PR merged.- A finding fixed in the PR description is judged on the description's change (
"pr_body": truein the reply).
Known limits
- New layouts on adopters: the new progress comment and finish card are validated on synthetic comments and on the release check's two Objectives. Adopters' first runs on v0.10.3 are their first real use.
- Network-drop test: run against a real
run --waitis not yet done. The test is ready, but this machine has no local token for the scratch repo. The fix is covered by unit tests. - Copilot stages show the raw model id on the card (e.g.
claude-sonnet-5.5). - Codex dollars are at OpenAI's Standard rates; the Codex release-check leg is still blocked by the test organisation's OpenAI account.
- Release check: Claude and Copilot passed end to end on Actions, on pendulum-scratch. That covered the new cards, time zones (repo zone and a per-Objective zone), the recorded Pendulum version on every stage, the report reading both cards and stages identically, and the App reading PR description edits.
Pendulum 0.10.2
Pendulum 0.10.2
- Report stage path: a stage's model from its own records alone; review stages mark no sub-agents (#559) (#560)
- Model-usage proposals in the report skill; pendulum report --rates (#547) (#558)
- Finish card: a machine-readable Objective summary marker; the report reads merged Objectives from it (#544) (#552)
- Cross-OS CI on the cloud-sandbox route: the launcher dispatches the run (#538) (#554)
- Setup adds and removes harness accounts; no hand edit of the workflow (#555)
- Record each review finding's outcome as data; pendulum report counts them (#546) (#550)
- Workflow write-grant check reads the permissions block exactly (#556) (#557)
- pendulum:experiment: labels and pendulum report --compare (#545) (#549)
- Each stage's per-pass ledger rides on its progress comment (#543) (#553)
- README: Coding agent accounts: one account per harness and route, and several (#548)
- README: an Objective is a feature or an epic; what Pendulum records, led by the finish card (#542)
Upgrading from v0.10.1
- Re-render the workflow with this release's setup:
/pendulum:pendulum-setup, orsetup.mjs --apply workflow --action clockgrove/pendulum@v0.10.2. Nothing is required beyond the pin; the re-render only refreshes comments. - Check your CI workflows' permissions. Doctor's
ci tokenrow, and the check on a cross-OS workflow, now count anywritescope orwrite-allas a write grant (#556). Before, they counted only four scopes. A pull-request or merge-queue workflow that grantsid-token: write,packages: writeand the like now FAILs doctor. Move that job to a workflow that doesn't run branch code, or drop the grant. - Progress comments now carry two hidden lines: a per-pass ledger (#543) and, on review stages, a findings record (#546). They hold numbers and labels only. A per-pass model is written only when the harness line already names it, or is the stage's own model when the harness line names none (Copilot at its default).
- Finish cards now carry a hidden summary line (#544), so
pendulum reportreads a merged Objective from one comment.
New
pendulum report:--compare A,Bforpendulum:experiment:<name>arms (#545);- review findings and their outcomes (#546);
--rates MODEL,..., a ratio of published rates between models of one vendor, never dollars (#547);- an on-request proposals page in
/pendulum:pendulum-report, each proposal an observation, hypothesis and experiment (#547).
- Setup:
--add-account,--remove-account,--accountsand--auto-accountmanage harness accounts. The output names the exact secret to save (#551). - Cross-OS CI on the cloud-sandbox route. The launcher job starts the run; no new App permission is needed (#538).
Known limits
- Codex dollars are at OpenAI's Standard rates. A project set to Fast bills more, and Codex can't report it.
- Rate ratios exist only for OpenAI models: Claude Code reports its own dollars, so there is no Claude price table.
- The one-click App creation flow (#486) has still not been hand-tested in a browser.
- Release check: Claude and Copilot passed end to end on Actions, on pendulum-scratch, including an experiment comparison that matched the finish cards. The Codex leg is blocked by the test organisation's OpenAI account setting.
Pendulum 0.10.1
Pendulum 0.10.1
- A stage with a spend line takes its model from its usage record when its harness line names none (#539) (#540)
- A cancelled Actions stage stops its agent's harness after its record (#532) (#537)
- Cross-OS CI in the work stage: opt-in macOS and Windows legs as validation (#518) (#526)
- pendulum report: read the per-pass ledger for Actions stages that have a spend line (#533) (#536)
- A cancelled Actions stage writes its own result.json before it exits (#529) (#531)
- Finish card: where an Objective's dollars went, churn apart; max_objective_usd caps one Objective (#517) (#527)
- Rate limit: read the count's reset through its Date against the time it arrived (#534) (#535)
- Finding check: run the test a "fixed" reply names before a round is spent (#516) (#523)
- Judge a Running stint whose work job ended inside its grace; wake at the grace's end (#530)
- pendulum report: the cost and token data layer and the pendulum-report skill (#512) (#525)
- Triage: a CI job cancelled at its time limit is said so, never a flake (#519) (#524)
- Judge a diff over 32 KiB in parts, not skip it: answer gate and rebase check (#515) (#522)
- Use the short name Pendulum; make the workflow's auto-merge comment true for both values (#521)
- doctor: install: none passes the lockfile row (#514) (#520)
- README: a complete first run, with the merge and spending limits beside the claims (#513)
- Finish card: a summary of time, cost and effort on the issue when an Objective merges (#498) (#511)
- Count gateway-route decision calls; stop drift at an exhausted cap (#479) (#510)
- Scorecard: price every stage on every route once; read Codex totals at the last event (#480) (#506)
- Cost: say what each dollar figure is, billed or API-equivalent (#481) (#508)
- Codex prices: OpenAI's 2026-11-16 rates, the long-context tier per request, the gpt-5.6 alias, lapsing promotional rates (#482) (#509)
- Cost: record unreported spend as unknown; remove every estimate and bound (#483, #478) (#504)
- Sealed sessions bind the whole usage record; published sessions resume only from Pendulum's runs (#505) (#507)
- Seal the agent's event stream and sessions in a public repository's usage artifacts (#490) (#503)
- Reach 75 on the Codex catalog scanner: SHA-pinned actions, plugin LICENSE and SECURITY.md, smaller icon, local Codex source (#501)
- Turn off the coding-agent CLIs' vendor telemetry in every stage (#502)
- Safer first setup: three questions, a first-run preset, a read-only default workflow token (#493)
- doctor: check the review job's pendulum:review label guard in the parsed jobs (#488) (#500)
- Plugin folder README for the Claude plugin directory listing (#499)
- Brand README: drop the trademark sentence (the trademark question is tabled) (#495)
- Pendulum logo: Kirk's updated mark (pendulum hanging straight), regenerated package, listing icon and social preview (#494)
- README first screen: logo, a real run, quick start, sourced comparison (#492)
- Setup creates the GitHub App from a pre-filled definition (#486)
- Pendulum logo: the painted grandfather-clock mark, its brand package, listing icon and social preview (#491)
- Skill set for newcomers: twelve user skills, one verb each, and a status skill (#487)
- Trust pages: data and privacy, release cadence, security targets, contributing, issue forms, code of conduct (#485)
- Listing metadata: icon, social preview, manifest fields, Action branding (#466) (#484)
Upgrading from v0.10.0
- Re-render the workflow with this release's setup (
/pendulum:pendulum-setup, orsetup.mjs --apply workflow --action clockgrove/pendulum@v0.10.1). This addsPENDULUM_MAX_OBJECTIVE_USD, updates the wake job's comment, and fixes the auto-merge comment. The Action's work and review steps nowexecthe stage, so a cancelled job writes its own usage record. - Set the
PENDULUM_APP_IDrepository variable to your App's numeric id (the same number as the secret; it is not a secret):gh variable set PENDULUM_APP_ID --repo OWNER/NAME --body <id>.pendulum reportuses it to recognise the App's stages, and doctor checks it againstapp_id. New installs get it from setup. - Every stage now writes a spend line on its progress comment, on the Actions route as well. The finish card,
daily_usdand the new Objective-wide cap read the line, so they no longer download artifacts.
Known limits and trade-offs
- Actions are pinned to SHAs in the workflows,
action.ymland the workflow template. Dependabot keeps them current. - Codex dollars are at OpenAI's Standard rates. Codex can't report or pin the service tier, so a project set to Fast bills 2× (2.5× on gpt-5.5) while Pendulum shows Standard. A doctor row says so. The rate for
gpt-5.6-solreads as unknown after its promotional price ends on 2026-11-30. - Drift checks hold back a 10% reserve of
max_usd(at least $0.30). Passes get what's left. - Opt-in cross-OS CI (
cross_os_workflow,cross_os_checks) works on the Actions route and in local mode. On the cloud-sandbox route it Blocks as a setup gap for now (#538). - The one-click App creation flow (#486) has not been hand-tested in a browser yet.
- Release check: Claude and Copilot passed on Actions, including a mid-run cancel. That cancel left a readable
result.json, Blocked on the next tick and reaped the harness. The Codex leg was blocked by the test organization's OpenAI account setting, not by Pendulum.
Pendulum 0.10.0
Pendulum 0.10.0
- Rename Clockgrove Factory to Clockgrove Pendulum (#476)
- The finding check reads a finding settled by a maintainer's recorded decision as answered (#475)
- factory CLI first run: doctor, --version, plain errors, token set (#457) (#464)
- Skills: consent before writes, hidden internal stage skills, harness parity, accuracy (#456) (#473)
- Local mode takes every Objective when the default branch has no Factory workflow; serve says what it leaves to the Action (#455) (#460)
- Docs: local mode leads with setup and running; README local quick start; glossary (#458) (#463)
- Skills name only what an installed plugin holds; setup walks local mode from its own local.md (#454) (#462)
- Claude on a subscription on the Actions route: INSTALL section, setup offer, doctor's plan row (#459) (#461)
- The rebase check asks per hunk of the branch's change against the resolved file; local mode keeps a stopped conflict pass (#452) (#453)
Upgrade notes
Clockgrove Factory is now Clockgrove Pendulum. This release is a clean break: nothing of the old name is read.
- Repository and Action:
clockgrove/pendulum. Pinuses: clockgrove/pendulum@v0.10.0. GitHub does not redirect Actionsuses:across a repository rename, so workflows pinned toclockgrove/factory@…stop working. doctor FAILs any pin before v0.10.0. - Everything renamed:
- labels
pendulum:*, branchespendulum/<n>, settings and repository variablesPENDULUM_*; - secrets
PENDULUM_APP_ID,PENDULUM_APP_PRIVATE_KEY,PENDULUM_CODEX_GATEWAY_KEY; - workflow
.github/workflows/pendulum.yml, rulesetpendulum, boardPendulum: <repo>; - CLI
pendulum, pluginpendulum@clockgrove, skillspendulum-*; - local paths
~/.config/pendulum,~/.cache/pendulum,.pendulum/.
- labels
- Before you migrate:
- Let Running and In review Objectives finish.
- Objectives left Blocked or in Question are not carried over; restart them under
pendulum:objectiveafter the migration. - Stop local loops.
- Open PRs on
factory/<n>branches: do not rename those branches in GitHub's UI. A branch rename closes an open PR whose head it is. Run~/work/ops/rename-pendulum/carry-prs.sh OWNER/REPOinstead (--dry-runfirst). It createspendulum/<n>at the PR's head, opens a new PR with the old title and body, closes the old one with a link, and deletes the old branch. Then re-applypendulum:objective, and the worker resumes on the branch and the new PR. Review threads stay on the old PR. - Migrate:
- rename the labels in place, move every
FACTORY_*repository variable and workflow env key toPENDULUM_*, and create thePENDULUM_*secrets; - re-render the workflow with
setup --apply workflowintopendulum.ymland delete the old file; - move (not copy)
~/.config/factoryto~/.config/pendulum.pendulum serve,runanddoctorrefuse to start while the old folder exists, because the sandboxes hide only the new path.
- rename the labels in place, move every
- Plugins: Claude Code:
claude plugin marketplace update clockgrove, then/plugin install pendulum@clockgrove. Codex and Copilot: remove the old plugin and addpendulumfrom theclockgrovemarketplace. - 1.0 streak: the walk-away count before the rename is recorded on #465; the streak now counts from v0.10.0.
Factory 0.9.37
Factory 0.9.37
- A re-applied factory:objective counts only from a person who can push; a ready mark beside one starts no review on Actions (#448, #449) (#450)
- A Question re-labelled since it was asked gets no ready-mark review in local mode: the label reruns the worker (#446) (#447)
- A session that ends abnormally while its own background job runs is resumed and told so; instructions match how -p waits (#440) (#445)
- A hand-off does not wait for a sibling a person holds out of the merge queue (#439) (#444)
- run --wait follows a merge on the repository's own bound, not a fixed 120 minutes (#438) (#441)
- A checks_hung queue removal counts toward QUEUE_TRIES only when the step really hung (#437) (#442)
- A sandbox record that cannot be read is no absent record: the dispatcher skips that Objective for the tick (#430) (#443)
- Limits a person owns become settings: relay, usage-limit waits, claim pause, local waits, stalls, sibling wait, part cap (#423) (#436)
- Launch loss, vendor holds, transcript lookup, queue conflicts, the board and key shapes decide on structure (#422) (#435)
- Local mode answers its stand-ins exactly or by decision call (#421) (#434)
- Log windows, rate limits, git transport and wide reds decide on structure (#420) (#433)
- Install runs every ecosystem's lockfile; the sandbox image rebuilds only when what it is made from changed (#419) (#431)
- An unread writer comment holds the merge only when it asks for something (#416) (#432)
- The fix-pass fence asks a decision call before refusing a toolchain file; name guesses become toolchain_paths (#414) (#428)
- Doctor reads token permissions and the CI workflow from structure; Claude model ids are the provider's to refuse (#418) (#427)
- A setup or install that fails while the network is down waits for it and runs again (#415) (#429)
- Acceptance entries and blockers come from exact structure, not heading or line regexes (#417) (#426)
- The same failure is a decision call where signatures cannot tell; the confidence floor is a setting beside the triage model (#413) (#424)
- An Objective's title is vouched for before any stage uses it (#410) (#425)
- An edit after the plan is judged once per new text: wording only keeps the plan (#400) (#409)
- The fix loop goes on while each fix pass makes progress, not for a fixed two passes (#399) (#408)
- Probe and serve tests decide on their own temp directory and the loop's lines, not the host's /tmp or a clock (#411) (#412)
- workflowEnv reads the parsed env block; the setup fence resolves the command by bash's grammar and fails closed (#387) (#397)
- A Question's answer is a decision call, not the first writer comment (#398) (#407)
- Network outages: GitHub's status page decides a persistent 5xx; a pass is cut off only on the harness's own error (#389) (#395)
- Decide a silent harness is hung from its own state, not 30 minutes of silence (#388) (#393)
- Required checks are waited for while they progress, not for a hidden 60 minutes (#401) (#403)
- Every stage call carries the current token; a 401 on an App token is refreshed once (#405) (#406)
- Owner-held limits become settings; triage's rerun waits within the checks' wait (#402) (#404)
- Provider busy, request-only refusals and set-up job failures decided on structure (#390) (#396)
- Automatic account choice prefers the account with the most quota left (#391) (#394)
- Decision calls fall back to the stage's own harness on its subscription (#379) (#385)
- Sibling hold and send-backs decide by trial merge, not file counts (#365, #366) (#384)
- Doctor reads CI triggers from parsed YAML and parses setup and validation with bash -n (#386) (#392)
- The rebase check decides on content, not commit ids (#362) (#375)
- The answer gate asks the decision model, not a 150-line threshold (#367) (#378)
- A reviewer finding counts as answered only when the decision model reads the answer as resolving it (#368) (#382)
- factory hold list / release: lift an account's usage-limit hold without deleting files (#376) (#383)
- Local sandboxes see what the repository's commands run, not a fixed tool list (#371) (#381)
- serve holds new starts on a machine fault instead of Blocking each Objective (#370) (#380)
- triage: only an exact changed path vetoes a rerun; name matches go to the model (#369) (#377)
- setup names the App permissions a personal account needs, without projects (#360 follow-up) (#364)
- Doctor's app permissions row names projects only where an App needs it (#360) (#361)
- A doctor Block names the failing check once and fixes with doctor's step (#357) (#358)
Upgrade notes
-
Who starts an Objective is checked now. Applying
factory:objective, removingfactory:draftor applyingfactory:restartcounts only when it comes from a person who can push, or from a pinned maintainer. An app's own token never counts. A start by anyone else isn't taken: the Objective is held Queued, and one comment says so. Queued Objectives whose latest start came from someone without push access (including a writer who has since lost it) are held at their next pick. A writer re-applyingfactory:objectivereleases them. -
The
*toolchain*file-name andtoolchain/directory guesses are gone from the fix-pass fence. An adopter who relied on them should list those paths in the newtoolchain_pathssetting. -
New settings, all with today's values as defaults:
decision_routetriage_min_confidencerebase_limitthird_party_wait_minutesguess_hourstoolchain_paths- machine.json
pressureandguess_hours - the local wait and outage bounds
INSTALL.md and LOCAL.md list them all.
-
Decision calls work without an AI Gateway key. They run on the stage's own harness subscription. Set
decision_route: offto turn them off; ADAPTERS.md lists what each call costs. -
Re-render the workflow (
factory setup --apply workflow) when you repin. This picks up the new action inputs.