Repository navigation
Releases
Releases · sattva2020/keystone
Keystone 0.9.40
Changed
- One sweep beside the judge on L. The first
/keystone:verifyrun of an L plan whose checks are green runs review
(an Opus finder plus the review validator) and/keystone:securityonce, in the same message as the verifier; all
findings land in oneverify.md(## Sweep) and one/keystone:fix verifypass. Only critical review items block;
suggestions areAdvisory. The sweep repeats only on--fullor for tasks added after it; a failed build, test or
agent leaves it due for the next run. The judge and the sweep never see each other's output. - No second review after the sweep. A plan that had its sweep gets no
Review the diff/Security checkoptions,
no/keystone:reviewinNext:and no built-in/code-review; running them by hand still works. - Late non-critical findings go to Advisory. The judge adds
Impact: critical | non-criticalto everywrongand
missingline. After the first PASS / PASS-UNPROVEN a late non-critical one isLate: <id> (advisory), outside the
verdict; the chat shows onlyLate advisory: <k>. - BLOCKED waits for you without a judge. A
BLOCKEDrow whose evidence cell, task tail and files are unchanged is
carried (carried: … — BLOCKED — waiting for you …); with nothing else to judge the verifier is not called. - A ceiling of three judged FAILs. From the third judged FAIL run (failed builds and accept commits do not count)
the verifier is not called until every open finding is decided — accept, debt or back to the plan
(Accepted finding by …,## Debt … deferred). A plan edit or--fulllifts it for one run. It replaces the trend menu. --fullbrings back the old behaviour for one run: full judge, sweep due, late rule off, no ceiling, no carried
BLOCKED.- Sweep findings stay in their run. A critical review item and a
Security gate: failof the sweep belong to the
verify.mdwhose## Sweephas them;/keystone:fix verifytakes them from there, and a later run does not carry
them. A sweep that failed on unchanged code runs again on--full. Thenothing to judgestop prints
open findings:(findings of the last judged run that are not rows, failed gates — its## Sweepincluded); when
there are any,Next:is/keystone:fix verify, notaccept. - Parked rows. A late row routed to
Late: … (advisory)and a row accepted at the ceiling are carried without the
judge (carried: … — advisory | accepted) until their files, their plan text or a cited row change. An accepted row
id covers only the runs up to its accept: judged again, it is a finding again.
Keystone 0.9.39
Added
- Feedback summary: coverage.
/keystone:feedbacknow has aCoveragesection: per chain skill, in how many sessions that started with feedback on the skill ran, and in how many of those its self-report was written — a skipped report no longer looks like "the skill did not run". Read from the transcripts on this machine (<config>/projects), nothing is stored. The instruction of theSessionStarthook carries aplugin: <version>line, so--since X.Y.Zcounts only the sessions of that version and later; sessions of 0.9.37 and 0.9.38 have no such line and drop out under any--since.
Keystone 0.9.38
Fixed
- The band no longer disappears mid-task (mod). A session's binding to its plan is also kept in the mod's store by session id, so an app restart or update no longer leaves the session without status, buttons and cost. The band (plan and context bars, the next-command button) now stays after every turn while the session has a plan — also after a plain prompt or an answer to the agent's question. A session that starts in a repo with a plan in view it has not worked with gets one dim row naming that plan, with its next command as a button; the row leaves with the first prompt.
v0.9.37
Added
- Agent feedback (off by default). The plugin option
feedback(orKEYSTONE_FEEDBACK=on;KEYSTONE_FEEDBACK=off
switches it off for good, the benchmarks set it) makes aSessionStarthook put a short instruction into the context.
The chain skills — start, intent, plan, check, implement, verify, fix, review — then end with a self-report of at most
10 lines about the skill text (unclear / deviated / guessed / extra, eachskill § section), in the language of the
conversation; the chat gets one line. Records are appended to~/.claude/keystone-feedback/<project>/feedback.jsonl
(all worktrees of a repo share one folder;KEYSTONE_FEEDBACK_DIRmoves it) — only on this machine, never sent
anywhere, by rule no project code, file contents or secrets (known secret patterns are masked — best effort), no automatic cleanup. /keystone:feedback.summary [--since X.Y.Z]folds the records of all projects into repeated problems with a
count;judge <plan path>/judgeruns the new read-onlykeystone:feedback-judgeagent on the signals of the work
(tool errors, repeats, interrupts, verify rounds, user corrections — counted byfeedback.mjs, no model): it confirms
or doubts each claim of the reports and adds what was left unsaid, with cause and class;auto <plan path>is called by
verify on M/L — once per task, only when the signals call for it. Feedback never blocks work, is not a gate and does
not touch the verdict.
Keystone 0.9.36
Fixed
- The push gate says what it closes. Its deny text read "push/PR is closed" for every case, while in a cloud
session a push of the work branch passes — agents in two live cloud runs reported the message as wrong. It now
names the case: "this push carries commits of<task folder>and is closed while …" or "a PR carrying commits of
<task folder>is closed while …" (gh pr createand the GitHub MCP tools alike), and in a cloud session adds that a
push of the work branch itself stays open.
Keystone 0.9.35
Fixed
- Push gate covers the GitHub MCP tools. The gate watched only Bash and PowerShell, so in a cloud session a PR made
through the GitHub MCP tools skipped it while the plan was a draft (found live, 0.9.33). A secondPreToolUsegroup
now sendscreate_pull_request,merge_pull_request,push_files,create_or_update_fileanddelete_filefrom any
MCP server to the gate: a PR or a merge is judged likegh pr create(its head against its base), a write straight to
the default branch is closed while a plan is active, a write to another branch and every read tool pass. The first
group's matcher stays a list of exact names, soTodoWriteand other tools do not wake the gate.
v0.9.34
Added
- One session per plan.
/keystone:implementand/keystone:fixtake the plan before their first edit, so a
second session (another worktree, another machine, the cloud) cannot run the same plan at the same time. The owner
record — place (worktree folder or cloud session), branch, stage, time; no machine name, no path — lives on the
service branchkeystone/ownersand is pushed to origin byskills/implement/scripts/owner.mjs(plumbing only: the
work tree, HEAD,mainand the plan folder are untouched, the push gate is unchanged). A second session stops with
who runs the plan and asksСтоп/Забрать план; without AskUserQuestion it stops, and--forceis only a human's
pick. The owner beats before each task and commit; no beat for 2 h = stale, and taking over becomes the recommended
answer. A taken-over session stops at its next beat. The end of a run,verifiedandabandoned(verify, accept,
start, fix) drop the record. Of two takes on one tip exactly one wins (push without force + one retry; on the local
branch a CAS on the decision's snapshot). Offline or without origin the mark stays local with one warning.
/keystone:sessionsdoes not wake a child whose plan another live session runs.KEYSTONE_GATES=offturns it off.
Keystone 0.9.33
Fixed
- Push gate: a pushed work branch no longer hides the plan's commits. The gate counted a push as carrying the plan
only when it shipped commits not yet on any remote branch. In a cloud session, where the work branch may be pushed
before verify (0.9.30), that made a latergit push origin HEAD:mainandgh pr createlook empty, and both passed
while the plan was still a draft — found live on a sandbox repository with 0.9.31. What a push ships is now measured
against the branch it lands on (origin/<dst>; for a PR, the default branch); a new branch,--allor a repository
withoutoriginkeep the old "not on any remote" rule. Commits already on the default branch count as landed, and
merge commits are still skipped, so main merged into a work branch keeps its push open (0.9.32).
Keystone 0.9.32
Fixed
- Push gate: a merge that only combines plan-folder commits already on a remote (main merged into a task branch) no
longer closesgit push; the gate counts non-merge commits not on any remote (git log --no-merges). A plan's own
unpushed commit still closes it. - Verify: merging main into the task branch no longer forces
delta: full. The base is found on the first-parent line
(an older merged-in commit no longer moves it),merge since the last verifyis gone, and files only a merge brought
are not counted —delta: since <sha> — <k> files changed …; <m> more came in by merge (not counted). Rows whose
evidence names such a file are still judged. - Verify: paths written to
verify.md(thefiles:list,rules scope:, Architecture gate,accepted violations:
and theAccepted violation by …line, changed files in the build / lint command and the check output, the
verifier'sfile:line) are in backticks, so Prettier via lint-staged no longer turns__tests__into**tests**;
findings and accepted violations read the same with or without backticks./keystone:setupproposesdocs/keystone/**in
.prettierignorewhen the project runs Prettier. - Verify: a check over the time limit says how to raise it —
not run (timeout after 600 s — raise KEYSTONE_CHECK_TIMEOUT_MS).
Keystone 0.9.31
Fixed
- The plugin passes claude.ai validation again. claude.ai now rejects a skill whose
descriptionholds
<placeholder>text ("SKILL.md description cannot contain XML tags"), and one such skill blocks the whole plugin: the
account marketplace kept serving 0.9.29 to cloud sessions, so 0.9.30's cloud push-gate fix never reached them. Nine
descriptions (docs, evolve, fix, plan, qa, qa-check, reference, roadmap, security) now write placeholders as
[placeholder]; a test keeps every skill and agent description free of<…>.