✨ feat: template: finish proves the first failing run and one state of open work, and release runs finish again by pascalandy · Pull Request #421 · pascalandy/nicely

pascalandy · 2026-11-01T00:46:09Z

Outside the plan: Pascal assigned #420, which holds #419 in full as its point 1. This pull request stacks on feat/417-round-ledger of #418, which is open and not merged, at 39bff54. It closes #419 and #420.

In round 4 of the day-1 trials, the loop held better: all three agents kept every finding and stopped at 5 rounds. But no build reached "Done when". The judge found the same gaps in 2 or 3 of the 3 extensions:

intakes that cite sources nobody read
no proof that the first scenario failed
generic cases with a name but no test
broken promises graded advise at first
## Open lines that said None. beside open findings

#419 added two faults of #418. just finish read "block" in a title as a severity. just release trusted a receipt that a hand can write.

What changes

Severity words (template: finish reads common words as severities, and release trusts a hand-written finish in the receipt #419.1). just finish reads a severity only in the line - **Severity**: of a finding. "Large inputs block the terminal" or "the code block in README.md" now pass. Forms that grade still fail, so no grade hides:
- a severity in code, such as `warn`
- a cell of a table that holds only a severity, also with marks such as **Block.**
- "Severity: block" in prose, in any part
- a finding heading, or an ID of no finding, outside ## Findings
Release (template: finish reads common words as severities, and release trusts a hand-written finish in the receipt #419.2). just release runs just finish again on the clean, signed HEAD. It runs before acceptance and before the tag. A forged receipt on a repository that finish refuses pushes no tag.
First failing run (template: prove the intake and the first failing run, test each generic case, and deliver #419 #420.3). readiness.md holds ## First failing run with four lines:
- the Scenario commit of ## Order of round 1
- the command that ran the scenario
- its exit status, which is not 0
- the assertion that failed
just finish fails when the record is missing, belongs to another commit, exits 0, or names no command or no assertion. The reviewer refuses a failure that is only a compile error.
One state for open work (template: prove the intake and the first failing run, test each generic case, and deliver #419 #420.6). A finding with the state open needs a line under ## Open lines of spec.md that names its ID.
Prose (template: prove the intake and the first failing run, test each generic case, and deliver #419 #420.2, 4, 5, 7, 8).
- Intake. The intake records the source, its version, and the part that the agent read, with each experiment apart (intake.md, vision.md).
- Generic cases. Each case of the Output group that applies gets a test file that the test makes (generic-cases.md, test-kit.md, acceptance.sh). This covers the two cases of template: prove the intake and the first failing run, test each generic case, and deliver #419 #420.4.
- Severity. A broken promise is at least warn at its first review, with three calibration examples (readiness-review.md).
- Rounds. Round 1 tests each generic case, and the first review tests families of cases. A root cause is fixed before the next review, and the last rounds verify.
- Test files. The tests never copy a file of the user.
Each new rule has a row in AGENTS/rule-index.md, with its check or "Prose only".
Moves. Two moves keep files under just limits, with no change: the proof of the release URL (template-release-url-test.sh), and the intake of the skill (intake.md).

Decisions made without Pascal

template: prove the intake and the first failing run, test each generic case, and deliver #419 #420.6, rule 1. ## Open lines lists each open finding by its ID. It holds the other rule too, because None. names no ID. An author lists the open findings of the review before in the next round, because the reviewer commits only readiness.md. At the stop of a build, spec.md can still say None. beside a finding that the last review marked open. just finish then fails and names that finding.
template: finish reads common words as severities, and release trusts a hand-written finish in the receipt #419.1 reverses a choice of ✨ feat: template: a build stops at 5 rounds, and a ledger keeps every finding #418. ✨ feat: template: a build stops at 5 rounds, and a ledger keeps every finding #418 failed the bare words block, warn, and advise anywhere. template: finish reads common words as severities, and release trusts a hand-written finish in the receipt #419 asks for them to pass, so only the grading forms above fail. The word severity alone still fails only under ## Findings, as template: release trusts a signoff receipt without finish, and finish misreads a few readiness forms #416 chose. Reviews 1 to 3 found three ways to hide a grade after the change, and each one now fails.
Release runs finish in its checkout, before acceptance. It runs on a clean HEAD that equals the signed and pushed SHA, so it needs no new clone. Running it before acceptance stops a release before slow work.
The run record is in readiness.md, not in a commit message. The reviewer writes it with the ledger. Finish does not reread it for earlier builds, so reviews before template: prove the intake and the first failing run, test each generic case, and deliver #419 #420 still end their builds. After just update, the agent runs the scenario again at its commit and records that run.
What finish cannot check stays prose. Finish cannot see a read, a compile error, a test file of each case, or a first grade. The reviewer of round 1 checks them, as readiness-review.md says.
The release test makes a repository that passes finish. It commits filled forms and a passing review before each release that must succeed. A later round starts a new build with its own steps.
Stack. I ran just signoff on this head only. just signoff-layers would also sign the head of ✨ feat: template: a build stops at 5 rounds, and a ledger keeps every finding #418, which this task does not touch.

Checklist

1. Spec: finish-gate.md, extensions.md, generic-cases.md, and test-kit.md cover each check and each rule
2. Scenario: each new case failed first, see "Proved by"
3. Code: those cases pass in just template
4. Text: n/a. The messages of just finish and just release stay in code, as every message of sdk/extensionrepo does
5. Checks: just check passes
6. Verify: an agent that did not write the change ran verify-nicely on template. Every step passed, see "Proved by"
7. Docs: n/a for the backlog, because the card is outside the plan. README.md shows no change for a user. The skills of the template teach each new rule
8. Sign off: just signoff passes on this head

Proved by

Each new case failed first on the code before it, then passed.

At the base 39bff54, the new tests of template: finish reads common words as severities, and release trusts a hand-written finish in the receipt #419 and template: prove the intake and the first failing run, test each generic case, and deliver #419 #420 ran with a fct_fail that does not stop:
- Open lines: want present "spec.md: ## Open lines does not name the open finding R2-F1 of readiness.md"
- First failing run: want present "readiness.md: it holds no part ## First failing run", "the first run of the scenario exited 0", "names no assertion that failed", and "names no command". Each time just finish exited 0, want 1
- Plain words: just finish exited 1, want 0
- Forged receipt: just release v0.1.0, exit 0, then RED: release trusts a receipt that names just finish, on a repository that just finish refuses and RED: a release that just finish refuses pushed a tag
Review fixes. Each reviewer judged whether the new tests of the round before fail on the head before their fix. The last reviewer mutated the code back. Without || section, the R3-F7 case fails. Without fence == "", the R3-F5 case fails.
Checks at the head 2f813aa. On Linux (om1), just signoff ran just check in a scratch worktree: exit 0 in 11 min 7 s, then Signed off on 2f813aa8fb91df9094ebd786a8d510269dea882e. On macOS (the mini), check_remote.py ran uv run --no-project --python 3.12 just check --keep-going: all 11 checks exit 0, template included, so the test scripts run under the bash 3.2 of macOS. The first two runs on the mini stopped before any check: just check refuses --verbose, and the login shell found the python3 3.9.6 of the system
verify-nicely, template at 1fa2f51, by the reviewer of round 3, with no keyring reachable:
- Every step passed: Launch, Doctor, commands, Render and install, Release text, Loop text, Finish gate, Discovery, Help, Inspect, Dry run, Missing file, Refusal, Describe, Skill, Doctor probe, and Cleanup
- "Loop text" lists ## Intake, ## Acceptance lines, ## Test files, ## Rounds, and ## Readiness review in SKILL.md, and ## Intake of a script, ## Intake of a program, and ## Evidence of the intake in intake.md
- "Finish gate" exits 1, names the 20 Fill: lines, and says readiness.md: no readiness verdict is recorded
- AGENTS.md:104 reads "It runs just finish again on that commit, then just acceptance, before it makes the tag"
Reviews. Three rounds, one reviewer for each, all Opus 5.5 high, none of which wrote the code. Round 1: 3 warn and 5 advise. Round 2: 1 warn and 7 advise. Round 3: APPROVE with 4 advise. Each finding was fixed. A fourth reviewer reread the fixes of round 3 and approved, then approved its own two advise after their fix.

Remaining risks

The run record and the intake are text that the agent writes. Finish checks their form, not that the run or the read happened. The reviewer must run the command again when in doubt
A complete forged review can still end a build, as ✨ feat: template: a build stops at 5 rounds, and a ledger keeps every finding #418 states
An earlier committed review that names, outside ## Findings, the ID of no finding now fails its "round N says pass" check. Reviews 3 and 4 found no honest ledger that does this

Sol review of the stack. Round 1 found that an open finding could not end a build as an accepted limit: ## Open lines must name it, and moving its row to Limits. changed spec.md outside Limits., so finish asked for a review of a new round. e114cd1 lets a row of ## Open lines leave when a row of Limits. takes its ID; any other change still needs a review. The new case of scripts/template-ledger-run-test.sh failed before the fix. Two other findings were declined: a fenced bare severity word is inside the documented parser limit, and a receipt that names echo finish affects provenance only, since release reruns finish.

Closes #419
Closes #420

Reviewed with codex-loop: 2 rounds with gpt-6.1-sol over the whole stack. Sol verdict: approved

by Opus-5.5 via Claude Code
Stack review and fix by Claude Opus 5.5 in Claude Code.

readiness.go nears the size limit. The verdict check moves to readiness_verdict.go, with no change of behavior. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… to its helpers template-release-test.sh nears the size limit. Its loop over the unsuccessful receipts moves to template-release-helpers.sh, with no change of behavior. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

A receipt with only just check, from before #415 or written by hand, let just release tag a repository whose loop was not done. Release now needs a just finish check with exit 0 in the receipt. Part of #416 and #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

extension-repository.md nears the size limit. The list of what just finish checks moves to finish-gate.md, with no change of rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…diness.md readiness.md now holds the ledger of a build. ## Rounds records each round, its head, its reviewers, the commits that each review covered, and the result of just finish. ## Order of round 1 records the commit of each step. Each finding keeps a stable ID and gets a state. just finish fails on more than 5 rounds, with a message to stop and to report to the user. It fails on a finding of an earlier round without a state, on a warn or block that is declined, on an open finding, on a commit outside HEAD, and on a last round whose head is not Commit:. It compares the findings with each committed readiness.md, so a finding never leaves the ledger and its severity never goes down. Finish also names the cause of a severity outside its line: the word, the table, or the fenced code. It reads a severity in each part after the header, ignores a byte order mark, and says that it reads the header in the first three lines. Part of #416 and #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…own file The skill nicely-extension nears the size limit. Its readiness review moves to readiness-review.md beside it, with no change of rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…unds, and each case has a safe outcome Round 1 of the skill records the intake facts in vision.md, writes spec.md, commits a failing scenario at the command boundary, and only then changes the handler. AGENTS.md no longer gives another order. The skill names three acts: a pull request for review, a stop report at the limit of 5 rounds, and the delivery after just finish passes. An answer that lets the agent choose starts no round. The readiness review teaches finding IDs, states, and the ledger. A command that replaces a user file proves its invariants with the test kit, and create-formal-verification-skill stays for paid work. generic-cases.md gives the safe outcome of each case, with new cases for a large expansion, an input with several parts, a temporary file, data that a reader needs, and data that changes the meaning. acceptance.sh compares the primary output, from a reader that is not the wrapped program, with an independent reference. Part of #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The history of readiness.md went back before the build. So the build of a second command had to keep the findings of the first, and a new finding could reuse an old ID. Finish now reads only the reviews committed after the Intake commit of the build. Part of #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The check moves without a change of behavior, so that wrap.go keeps room under the size limit for the next fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… to wrap_stop.txtar The timeout, Ctrl-C, and version probe cases move without a change of behavior, so that wrap.txtar keeps room under the size limit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Program.Input holds the item or path that the call works on. When it is set, each RUNTIME message that Run and Ended make names it. A signal adds an INTERRUPTED or TERMINATED error that names it, which Finish puts first. An empty Input keeps the messages of today. A version probe that uses up the deadline of Timeout now gives the timeout, RUNTIME with exit 1, instead of PREREQ_MISSING. A probe that reaches ProbeLimit inside a longer timeout stays PREREQ_MISSING. hello count sets its file as the input, and hello scan sets none, so the wrap scenarios prove both paths. Refs #407 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The inspect example sets its path as the input of sdk/wrap. Its scenario checks that the answers to a failed exit, the timeout, Ctrl-C, and SIGTERM name the file. A version probe that uses up the timeout now expects the timeout, RUNTIME with exit 1. A probe that fails, or a python3 that is too old, stays PREREQ_MISSING. spec.md moves the probe that uses up the timeout to worker-fails. Its no-python line drops the slow probe, which no scenario proves. Refs #407 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… own file test-kit.md is near the limit of 2,500 tokens. The section moves without a change of text, so later rules about the kit commands fit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

An interrupted run must keep its input, leave an absent output absent, and keep an old output. A protect selection now applies to the next signal call, which ends it for signal. refuses keeps the selection until the next protect call, as before. The comparison skips the marker. signal and the coming appears command share the foreground run, the wait for the marker, and the bounded wait for the exit. A kit test also proves that protect and refuses catch a run that writes through a link. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… new file appears <marker> <path> <command> <arg>... runs a registered call as signal does. At the new marker, the kit creates the path with an exclusive create and its own bytes. The run must then fail with a valid answer, and the path must stay the same regular file with the same bytes. The command refuses --force, and the next protect selection applies to it. Scenario commands lists which kit calls prove each invariant of a command that replaces a user file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… reader A check that reads only a header or the file type cannot see a lost content. The check now compares what a reader other than the wrapped program gives with an independent reference. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…nd that replaces a file Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…lace Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Review round 1 found that a readiness.md could move its Intake to an earlier review, drop a warn, and restart the rounds at 1. A label line under ## Rounds could also hide rows. The history of a build now starts only after a committed review with a passing verdict, so a review that needs work never ends a build. Each round of an earlier committed ledger must stay unchanged. Each line under ## Rounds and ## Order of round 1 must be a row. Each of the first four steps of round 1 needs its own commit. The rule index lists appears among the call forms. The no-python line of the template spec names the probe that does not finish within 10 s again, and a hello scenario proves it. Part of #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…nding without an ID never leaves the ledger Review round 2 found four ways past the ledger, and no way out for an honest build at the limit. A review ends a build only when its verdict follows from its findings and from the spec.md of its commit, so a changed verdict line ends nothing. Each committed review of the build names the head of a round, and each round before the last has a committed review of its head. An earlier finding without an ID keeps its title in later reviews. The user can now accept a limit at any round, also at the limit of 5: after a review, only the part **Limits.** of spec.md may change, and the reviewer gives the finding the state limit in a review of the same commit. Part of #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… pass Review round 3 found that a limit accepted after the review blocked every later build, that a review of the form of #409 never ended a build, and that a two-line edit of a review that needs work ended one. A review ends a build only when its verdict, its states, and the history of its own build pass, as just finish checks them. Its limits come from the spec.md of the next intake when only **Limits.** differs. A line - **Limit**: of the form of #409 reads as the state limit, and a review without a ledger needs a passing verdict only. The readiness review says that only readiness.md and **Limits.** of spec.md may differ from the reviewed commit, and that the limit is committed before the review. Part of #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… Finish pass ends a build Review round 4 found that the inferred end of a build failed at the 10th build, grew three times slower with each build, broke again when spec.md changed before the next intake, and let a 7-round build end. Three fixes of that inferred boundary failed in a row, so the ledger drops it. History is every committed readiness.md of HEAD. The round numbers go on from build to build, and the ledger keeps each finding. A round before the last whose Finish is pass ends a build, when its newest committed review passes: a verdict that follows from its findings, valid states with the current limits, and each finding of the reviews before it. Each build, also one that ended, has at most 5 rounds, and the intake of a build follows the end of the build before. A Limit line no longer replaces a Spec line. Part of #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…t finish passed then Review round 5 found that finish checked a review that ended a build with the current spec.md, so a limit or a renamed spec ID of a later build broke an honest repository for good. A Finish pass on a Ready review whose finish failed, such as on an open line, also reset the 5 rounds. Finish now checks such a review as the repository stood when it was committed: no Fill: line, no open line, only readiness.md and **Limits.** of spec.md changed since its head, and a verdict and states that follow from its findings. A spec.md before the next round stands in when only **Limits.** differs. Each review is parsed once and each commit resolved once, so 80 builds take 0.76 s in place of 1.48 s. Part of #417. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…xture Each case of appears and protect ran its own go test of the standalone fixture, 29 in all. On the mini, that load made the 2 s budget of completion values and the 1 s timeout of a hello scenario fail in the full just check. The cases now run as scenarios of one go test per test function, and each case reads its own result and log. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

template-release-test.sh is near the token limit of a file, and the next commit adds the proof of a forged receipt to it. The proof of the repository URL of a module with a major version suffix is its own responsibility. It moves, unchanged, to template-release-url-test.sh. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…finish again Finish read the words block, warn, and advise anywhere after the header as a severity. A title such as "Large inputs block the terminal", or prose such as "the CLI does not warn", failed it. Finish now fails only on a severity in code, such as `warn`, on a severity alone in a cell of a table, and on the word severity under ## Findings. The plain words pass. Release trusted a signoff receipt that named just finish, and a hand can write such a receipt. It now runs just finish again on the clean, signed HEAD, after acceptance and before the tag. The release test proves that a forged receipt on a repository that finish refuses pushes no tag. Its releases that must pass now commit filled forms and a readiness review first. Closes #419. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…lines names each open finding In round 4 of the day-1 trials, two of three agents committed the scenario before the handler but never ran it. Each spec.md also said None. under ## Open lines while the ledger recorded open findings. readiness.md now holds ## First failing run, with the command that ran the first scenario at its commit, its exit status, and the assertion that failed. Finish fails when the part, the command, or the assertion is missing, or when the exit status is 0 or no number. A finding with the state open now needs a line under ## Open lines of spec.md that names its ID, so spec.md and the ledger agree on the open work. Part of #420. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The skill of the template is near the token limit of a file, and the next commit adds the evidence of the intake. The parts "Intake of a script" and "Intake of a program" move, unchanged, to intake.md, which SKILL.md links. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ets a test file, and a broken promise is a warn Round 4 of the day-1 trials found gaps in 2 or 3 of 3 extensions. This commit writes the rules of #420 that no check enforces, and the docs of the checks of the commits before it: - intake.md: the intake records the source, its version, and the part that the agent read, with each experiment apart. A link alone is not evidence of a read - generic-cases.md, test-kit.md, and acceptance.sh: each case of the Output group that applies gets a test file that the test makes. A block that a reader needs keeps its content and loses its private text. A declaration that changes the meaning gets a value that is not the default, and the reference applies it - readiness-review.md: a broken promise is at least warn at its first review, with three calibration examples. The reviewer of round 1 checks the intake and the first failing run, and refuses a compile error - SKILL.md: round 1 records the first failing run, tests make their own files, and four rules help a build reach "Done when" in 5 rounds - finish-gate.md, extensions.md, and rule-index.md: the checks of #419 and #420, and "Prose only" for each other new rule Closes #420. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… no grade hides outside a finding An independent review of the commits before found three gaps: - A record of ## First failing run from an earlier build still passed. The record now names its commit, which must be the Scenario commit of ## Order of round 1. A record in fenced code, or in the form **Name:**, now reads as finish expects - After #419, "Severity: block" in a part after ## Findings, a finding heading under another part, and a cell such as **Block.** passed. Finish now fails on the word severity before a severity in any part, on a finding heading with an ID outside ## Findings, and on a cell that holds only a severity with marks around it. "Each finding names its severity" still passes - Release ran just finish after acceptance, which can be slow. It now runs just finish first Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… grade, and the order of release The docs follow the fix before. finish-gate.md keeps its list of failures in one piece, and names the Commit line of ## First failing run and the forms that grade. readiness-review.md lists those forms, says what to do after just update, and keeps a sentence out of the last calibration example. The rule index gets the three rules it missed and links ledger_run.go. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… record reads like a finding A second independent review found these gaps: - A line outside ## Findings, such as "- R3-F8 Long lines are cut: block", hid a graded finding. Finish now fails on each finding ID outside ## Findings that no finding under it has - The record of the first failing run closed a fence of four marks at a line of three, and refused a value that ends with a period. It now reads fences and values as readinessFindings does - The proof of a forged receipt now checks that release ran no acceptance before just finish refused, and the proof of the hidden forms checks each message again Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ely reads intake.md finish-gate.md and readiness-review.md name the check of a finding ID outside ## Findings, and finish-gate.md splits a sentence of 30 words. The step "Loop text" of verify-nicely reads the parts of the skill after the move of the intake to intake.md. intake.md limits experiments to commands that neither write nor bill. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… ID either The third independent review found that a heading, such as "## R3-F7 Long lines are cut" or "#### R3-F6", still carried a finding ID outside ## Findings with no grade. Such a heading now counts as a mention. A heading of level 3 with an ID keeps its own message. A test proves that an honest summary of existing IDs still passes. The step "Loop text" of verify-nicely reads 12 lines after "## Done when", so it reaches the line of a "full stop" request, and a sentence of finish-gate.md is split. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… is a stray ID too The review of the last fixes found that a heading such as "### R3-F9 x" in fenced code outside ## Findings got no message: the mention check skipped it as a finding heading, and the fence skipped the check of a heading outside ## Findings. It now counts as a mention. The proof puts "## R3-F7" right after the findings, so it covers the heading that closes ## Findings too. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

coderabbitai · 2026-11-01T00:46:18Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

Source mode: Auto
Repository: Unavailable
Branch: Unavailable
File: No configuration file from this repository applied
Commit: Unavailable
Layers: Repository UI
Review profile: ASSERTIVE
Plan: Advanced
Run ID: 5b717168-11e0-4fe2-9202-e417ff4da75f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

Trigger review

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

…rounds A Sol review of the stack found that an open finding could not end a build as an accepted limit. Its row under ## Open lines must name it, and any row there fails finish. Moving the row to **Limits.** changed spec.md outside **Limits.**, so finish asked for a new review of a new round, after the last one. Finish now also lets a row of ## Open lines leave when a row of **Limits.** takes its ID. Any other change of spec.md still needs a review of its commit. Refs #420 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The Sol fix of #421 added a sentence to finish-gate.md, and with the case matrix of #424 the file went over the 2500 tokens of just limits. The part "The limit of the parser" now has its own file, with no change of its text. finish-gate.md and readiness-review.md link it. Refs #423 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

pascalandy and others added 30 commits October 31, 2026 14:27

♻️ refactor: template: move the verdict of readiness.md to its own file

0892931

readiness.go nears the size limit. The verdict check moves to readiness_verdict.go, with no change of behavior. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

♻️ refactor: contract: move the checks of just finish to their own part

6192786

extension-repository.md nears the size limit. The list of what just finish checks moves to finish-gate.md, with no change of rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

♻️ refactor: template: move the readiness review of the skill to its …

b22166d

…own file The skill nicely-extension nears the size limit. Its readiness review moves to readiness-review.md beside it, with no change of rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

♻️ refactor: sdk/wrap: move the program check into check.go

a1c2fdb

The check moves without a change of behavior, so that wrap.go keeps room under the size limit for the next fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

📝 docs: contract: move the scenario commands of the test kit to their…

95e94df

… own file test-kit.md is near the limit of 2,500 tokens. The section moves without a change of text, so later rules about the kit commands fit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

📝 docs: template: the skill names the kit commands that prove a comma…

fd00b34

…nd that replaces a file Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

✅ test: template: the proof of finish names an author in place of a p…

5859d69

…lace Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

📝 docs: template: the rule index names the checks of the ledger history

b3dc8b4

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

pascalandy and others added 6 commits October 31, 2026 20:12

pascalandy changed the base branch from feat/417-round-ledger to main November 2, 2026 01:59

pascalandy merged commit e114cd1 into main Nov 2, 2026
3 checks passed

pascalandy mentioned this pull request Nov 2, 2026

🚑 fix: transcript: a dry run under a busy run ID says busy, as the call does #426

Merged

8 tasks done

Navigation Menu

Repository navigation

✨ feat: template: finish proves the first failing run and one state of open work, and release runs finish again - #421

✨ feat: template: finish proves the first failing run and one state of open work, and release runs finish again#421
pascalandy merged 37 commits into
mainfrom
feat/420-intake-proofs

pascalandy commented Nov 1, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Nov 1, 2026 •

edited

Loading

Review skipped

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

pascalandy commented Nov 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes

Decisions made without Pascal

Checklist

Proved by

Remaining risks

Uh oh!

coderabbitai Bot commented Nov 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

pascalandy commented Nov 1, 2026 •

edited

Loading

coderabbitai Bot commented Nov 1, 2026 •

edited

Loading