Settle the check judge contract and divergence corpus before porting it

Problem

The merge scripts judge a saved check rollup with the shell function mp_judge_checks in packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh, using jq. internal/checks holds a Go evaluator, which the unattended merge path calls through agentrepo checks evaluate (ma_gates in merge-agent.sh). Moving the attended path to the Go evaluator is feature 3 of #588. It is unsafe to do before the two judges' differences are written down, because a difference can turn a failure green or a green into a refusal. The Feature that does the move is #723, and it is blocked by this Spike.

The two judges differ on the same data (read on main at d41369e7):

Malformed input: the shell exits 3, Go exits 5. A missing or null statusCheckRollup: the shell reads it as an empty rollup (exit 4), Go as unreadable input.
Unknown states: the shell reads an unknown conclusion, status or status-context state as ok. Go returns the failing outcome (exit 1) for them, not the unreadable outcome (internal/checks/checks.go:180, internal/checks/checks_test.go:82). A missing status reads as completed in the shell and as failing in Go. The shell upper-cases states before comparing, so a lowercase spelling reads like its upper-case form; Go compares exactly and treats a lowercase spelling as failing.
Superseded entries: both judges select the latest run of each check before reading states (internal/checks/checks.go:148, pull-request.sh:88), so an unknown state on an earlier run that a later success replaced is never read. The contract has to say whether a refusal for unknown states covers such runs.
Entries without a name or with an unknown __typename: the shell judges them, Go rejects them as unreadable.
Start times: the shell compares startedAt as strings (pull-request.sh:71, pull-request.sh:88), Go parses instants (internal/checks/checks.go:110, internal/checks/checks.go:269). A failing run at 2026-10-01T02:00:00+02:00 and a passing run at 2026-10-01T01:00:00Z of the same check make the shell choose the failure and Go choose the success. An unparseable timestamp is unreadable input in Go and a string in the shell. Tied start times: the shell takes whichever max_by returns, Go takes the worse verdict.
Repeated status contexts: the shell judges every entry as it stands, so an older failing entry fails the rollup. Go keeps only the newest entry per context, so a newer success hides it.

The callers also depend on the shell judge's behaviour. preflight.sh calls mp_require_reviews first and exits 75 when a review is required (preflight.sh:94; the regression at tests/skills/merge-pr/scripts/merge-workflow.bats:385 requires it), runs the unattended gates, and only then reports each check's state and name, sorted by name, without failing or still-running checks making that section refuse. squash-merge.sh prints failing or still-running checks as <state> <name> lines on standard error and refuses with a message that depends on the outcome. Both callers parse the saved response before they judge it (preflight.sh:78, squash-merge.sh:116, pull-request.sh:39), so a truncated live response can stop them with a parser error before any judge message. The Go command prints a JSON summary of the non-successful names only, with exit codes 0 to 6 (internal/agentrepo/command.go:55). The shared loader accepts agentrepo from v0.16.0 (packages/agents/source/lib/lib.sh:140), and docs/agent-skills-migration.md:9 lets a consumer run with none.

Acceptance criteria

AC1: The findings comment holds a contract table with one row for each of these input classes: malformed JSON, missing or null rollup, empty rollup, unknown state, lowercase state, missing status, empty check name, unknown entry type, tied start times, a rerun that has not started, repeated status contexts, start times in different UTC offsets, an unparseable start time, and an unknown state on a superseded run. Each row states the shell judge's outcome and exit status, the chosen preflight and squash outcomes separately, and whether the change is deliberate. For every case, each caller's chosen result records its report/continuation or refusal, numeric process exit status, and exact standard output and standard error (including the diagnostic message and explicitly empty streams). These are caller process statuses, separate from the judge's status; otherwise passing prerequisites and later gates isolate the case, with merge writes stubbed. Invocation modes and earlier stops are identified as in AC7. Empty, missing and null rollups have separate cases. These fields specify findings for Replace the shell check judge in the merge scripts with the Go evaluator #723 to consume; they do not require implementing the chosen behavior in this Spike.
AC2: Every row of the table cites at least one saved rollup, with the output and exit status the shell judge produced for it and the result the current Go evaluator gives for it, both recorded by running them on the head named in the comment.
AC3: The comment states the sweep's domains, generation rule and expected total before reporting results. The sweep uses homogeneous ordered pairs of two entries for one check, separately for CheckRun (ordered by startedAt) and StatusContext (ordered by createdAt); it does not pair different entry types. For each CheckRun, the status domain is COMPLETED, IN_PROGRESS, PENDING, QUEUED, REQUESTED, WAITING, all six exact lowercase spellings, the literal UNKNOWN_STATE, and omission of status (14 alternatives). Its independent conclusion domain is ACTION_REQUIRED, CANCELLED, FAILURE, NEUTRAL, SKIPPED, STALE, STARTUP_FAILURE, SUCCESS, TIMED_OUT, all nine exact lowercase spellings, the literal UNKNOWN_STATE, and omission of conclusion (20 alternatives). Cross the full status × conclusion product on each entry, including combinations in which status causes a judge to ignore conclusion; do not vary one field at a time. For each StatusContext, the state domain is ERROR, EXPECTED, FAILURE, PENDING, SUCCESS, all five exact lowercase spellings, the literal UNKNOWN_STATE, and omission of state (12 alternatives). Omission means the property is absent, not null or an empty string; neither null nor empty-string state fields nor a lowercase unknown sentinel are additional sweep alternatives. Each ordered state pair takes all seven concrete timestamp pairs defined in Approach. Other fields are fixed as stated there. The expected total is 7 × ((14 × 20)^2 + 12^2) = 549808 cases: 548800 CheckRun pairs and 1008 StatusContext pairs, before any deduplication of equivalent judge results. The recorded cases cover exactly that product, and zero combinations that the shell judge reports as failing or still running are green under the table's outcomes.
AC4: The comment states one general decision rule for any pair of runs of one check in which the string order and the instant order of the start times disagree. The rule is never green when either ordering would select a failing or unfinished run, and the table's row for the failing run at 2026-10-01T02:00:00+02:00 against the passing run at 2026-10-01T01:00:00Z applies it.
AC5: For an unknown status, conclusion or status-context state, on the latest run and on a superseded run, the table records the chosen preflight and squash outcomes separately, each with its numeric process exit status and exact standard output and standard error, including its message or explicitly empty streams, using AC1's caller fields and boundary. It states whether validation covers every entry or only selected entries.
AC6: A second table covers every exit status of the evaluator command (0 to 6), any other exit status, an absent command, malformed or non-result output, valid-looking success output with a nonzero exit status (including mapped statuses 1 and 2), and a zero exit status with missing result fields. For every case it records the evaluator's process status and output (or absence), the chosen evaluation category (green, failing, still running, empty or unreadable), and separate chosen preflight and squash results using all AC1 caller fields: report/continuation or refusal, numeric process exit status, exact standard output and standard error with the message or explicitly empty streams. Only a zero evaluator exit status with a complete success result is green; preflight continuation is not evidence of green checks. The table identifies whether an earlier stop or the check section supplies each result.
AC7: A third table lists each refusal and parser stop before the shell check-reporting and judging section, including those raised by the shared loader (for example its version refusal) and by the unattended helpers it invokes (command usability and the unattended gates). For every case and relevant invocation mode it records preflight and squash results separately using AC1's caller fields: report/continuation or refusal, numeric process exit status, exact standard output and standard error with the message or explicitly empty streams, the first stop reached and whether the check section was reached. It states whether each result stays unchanged. Record the other caller's result even when that caller does not take the same stop. It states that preflight's check section reports failing and still-running checks without refusing, and that a required review still exits 75.
AC8: The comment names the unattended path's existing fixtures that the table changes and the outcome, exit status, messages and skipped-check hint each gets under the table, or states that none change.
AC9: The comment lists the evaluator outputs the table promises that the current evaluator does not yet give, and either names the first devtools release that gives all of them or states that no compatible release exists yet and names the issue that implements and releases it. In both cases it states how the callers behave with an older release or with none, and names the issue that installs a compatible release in consumer repositories.
AC10: Non-functional: the comment records the median time of 20 runs of the shell judge and of the Go evaluator over a generated 500-check rollup, and proposes the time limit for the merge path.
AC11: Each AC of this issue is answered in the findings comment, and the comment links the Feature issue for the caller migration and any further issue filed.
AC12: Non-functional: before posting, the complete final findings comment and every complete final follow-up issue body, including embedded corpus and command output, each pass gitleaks, the repository's check-leaks.sh with its private-strings list, and a separate leak review of each post's full text. The findings comment records a clean result for each of the two scans and the leak review for each post. Any text changed after a check is checked again before posting.

Scope and non-goals

In scope: reading and running the two judges, recording the three tables, the corpus and the measurements, and filing follow-up issues.
Non-goal: changing the shell judge, the Go evaluator, the merge scripts or the generated skills; the findings are a comment, and any prototype stays on a throwaway branch that is deleted.
Non-goal: review-verdict parsing (#589), approval selection (#590), policy inspection and merge execution (later features of #588).
Out of scope: distributing the Go binary (#333) and installing it in consumer repositories (#544).

Approach

Build the evaluator command from main, then run the shell judge and the Go evaluator over hand-written rollups for each class of AC1 and over the generated sweep of AC3. Record every rollup with both outputs, so the Feature can commit them as fixtures. Choose each row's outcome by one rule, with no exceptions: the merge path is never greener than the shell judge for a failing or unfinished check, and a rollup that cannot be read is refused. Before running the sweep, write down its domains, generation rule and expected total (AC3). The seven ordered timestamp pairs (first, second) are: equal (2026-10-01T01:00:00Z, 2026-10-01T01:00:00Z); earlier (2026-10-01T00:00:00Z, 2026-10-01T01:00:00Z); later (the preceding pair reversed); offset-reversed (2026-10-01T02:00:00+02:00, 2026-10-01T01:00:00Z); unparseable first (INVALID_TIME, 2026-10-01T01:00:00Z); unparseable second (the preceding pair reversed); and both unparseable (INVALID_TIME, INVALID_TIME). Apply these to startedAt for CheckRuns and createdAt for StatusContexts. Keep JSON array order first then second, __typename fixed to that sub-sweep's type, one fixed nonempty check name and workflow for CheckRuns and one fixed nonempty context for StatusContexts; omit fields belonging only to the other type. Fix the head and use no required-check names or expected-head constraint for this comparative sweep; AC8 separately records unattended requirements. No other axis is multiplied into AC3; AC1's additional cases remain separate.

Trace the shared loader and the unattended helpers invoked before the check section, as well as the three callers, to fill the third table, and read the unattended fixtures under tests/skills/agent-merge-pr/ for AC8. Post the findings as one comment on this issue.

Dependencies

Native parent: #588. No blocker. The Feature that migrates the callers, #723, is blocked by this issue, by #333 and by #544.

Size

S. One findings comment and, at most, a few follow-up issues.

Verify plan

AC1: the table has a row for every class named in AC1, with distinct empty, missing and null rollup cases. Inspect every row's separate chosen preflight and squash report/continuation or refusal, numeric process exit status, exact standard output and standard error (message or explicitly empty streams), invocation mode and boundary, rather than accepting only a merge-path category or evaluator exit code.
AC2: each row cites a rollup and both recorded outputs; a reader can rerun the two commands on the cited rollup and get the recorded result.
AC3: independently reconstruct the specified field-level products: check all 14 status alternatives, all 20 conclusion alternatives and all 280 crossed CheckRun entries, including simultaneous lowercase/missing fields and combinations with non-completed status; check all 12 context-state alternatives. Compare the generated case keys with the complete ordered-pair × seven-timestamp set, requiring no missing or duplicate keys and exactly 548800 CheckRun and 1008 StatusContext cases (549808 total). A match between a researcher-selected total and a recorded total alone does not pass. Inspect the domain manifest and the seven timestamp pairs, then confirm zero shell-failing or still-running combinations become green under the chosen outcomes.
AC4: the stated general rule, the row for the two runs above, and the sweep's offset-reversed combinations (AC3), none of which shows a green outcome when either ordering selects a failing or unfinished run.
AC5: inspect unknown status, conclusion and context-state rows on both latest and superseded entries; require the validation-scope decision and separate chosen preflight and squash results, each with numeric process exit status and exact standard output and standard error, including its message or explicitly empty streams.
AC6: inspect every named case in the second table, including malformed evaluator output, success-looking output followed by exits 1 and 2, and exit 0 with missing result fields. Check the recorded evaluator status/output separately from both callers' chosen report/continuation or refusal, numeric process exit status and exact standard output and standard error (message or explicitly empty streams); confirm only a complete success result with evaluator exit 0 is green. Confirm the current evaluator's statuses against internal/agentrepo/command.go and internal/checks/checks.go on the cited head, and record the chosen caller contract as research findings.
AC7: confirm the third table against preflight.sh, squash-merge.sh, pull-request.sh, the shared loader and merge-agent.sh on the cited head, tracing the helpers invoked before the check section for each invocation mode, and the regression at tests/skills/merge-pr/scripts/merge-workflow.bats:385. Inspect both callers' numeric process exit statuses, exact standard output and standard error (including messages or explicitly empty streams), first stop, reachability of the check section and unchanged/change decision for every case; do not substitute a judge exit status for a caller status.
AC8: the list of unattended fixtures, each run on the cited head under the proposed rules.
AC9: the list of promised outputs the current evaluator lacks, then either the named release with evidence that it produces them or the "no compatible release yet" statement with the implementing issue, and the issue number that installs it.
AC10: the recorded medians and the commands that produced them.
AC11: the comment's answer to each AC and the links.
AC12: before posting, run gitleaks and the repository's check-leaks.sh with the private-strings list over the complete final findings comment and every complete final follow-up body, including embedded corpus and command output. Also obtain a separate leak review of the complete final text of each post before it is posted. Inspect the findings comment's recorded clean results for both scans and that post's leak review, covering every post; recheck any changed text. This verifies D4's three requirements.

Risks

A wrong row would let the Feature turn a failure green; AC3 and AC4 exist to catch that, and the Feature's Verify plan replays every row. The sweep might miss a class; AC1 names the known ones and the Feature adds a case for any a reviewer finds. The undo is deleting the throwaway branch; nothing else changes.

Operator authority

None

Refinement: NOT READY — clarify observable behavior, integration verification, and rollout prerequisites

Checked

Problem: Clear and supported by code. The shell judge accepts unrecognized states as green; Go rejects them. Their handling of repeated contexts and tied runs also differs (packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:69, internal/checks/checks.go:180, internal/checks/checks.go:257).
Acceptance criteria: Six correctly numbered AC<n>: entries, within the limit, including a non-functional criterion. Implementation detail, compound outcomes, and verification gaps remain below.
Scope and non-goals: Explicit and bounded. The change matches Feature 3 of Replace complex shell with tested Go, starting with gates and merging #588. No overlap identified with Port the gate runner and pass cache from Python to Go #713–Replace the commit and pull request workflow guards with Go hooks #719 from their current titles; their bodies are empty. Port strict review-verdict parsing from shell to Go #589 and Port exact-head merge approval selection from shell to Go #590 concern distinct predicates.
Approach: The existing evaluator and command are suitable starting points, but their output contract does not directly replace the shell caller contract.
Dependencies: Native blocked-by list is empty, matching “None blocking”; native parent is Replace complex shell with tested Go, starting with gates and merging #588. No implementation dependency on Port the gate runner and pass cache from Python to Go #713–Replace the commit and pull request workflow guards with Go hooks #719 was identified. Rollout prerequisites from the epic remain omitted.
Size: S and one PR are plausible for this leaf migration, including generated output and focused integration coverage.
Verify plan: Every AC ID appears beside a proposed check, but several checks do not establish the full stated outcome.
Risks: Unsafe acceptance is identified; reverting the PR supplies an undo.
Operator authority: None satisfies the required form. Clarify separately how the epic’s owner-approved protection prerequisite will be satisfied.

Gaps

Acceptance criteria contain implementation prescriptions and multiple outcomes. AC4 mandates a Go command and deletion of a named function; AC5 mandates a specific command and generated implementation absence. AC1 combines contract publication with exhaustive difference accounting. Move implementation choices into Approach and separate each observable acceptance outcome.
The compatibility evidence is narrower than the migration risk. AC2 checks only formerly green rollups, and its Verify entry replays existing Bats fixtures. Require contract cases for all identified differences, with expected old and chosen outcomes, including formerly failing and pending inputs. Otherwise, a failure becoming success can escape that comparison. Existing Go tests already provide readable evidence for ties, repeated contexts, unknown states, and required-check behavior (internal/checks/checks_test.go:63, internal/checks/checks_test.go:78, internal/checks/checks_test.go:94).
Approach omits the caller output and exit-status transition. Shell emits per-check TSV with states and workflow-qualified names; Go emits summary JSON, with unreadable input returning 5 instead of 3 (packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:49, internal/agentrepo/command.go:45). Preflight formats those TSV rows and reports superseded runs; merge execution uses them for failure and pending diagnostics (packages/agents/source/skills/merge-pr/shared/scripts/preflight.sh:101, packages/agents/source/skills/merge-pr/shared/scripts/squash-merge.sh:203). State which outward diagnostics remain and which deliberately change. Verify both callers’ output and refusal behavior against the branch binary.
AC3 lacks merge-path failure verification. Go table cases alone do not prove callers refuse unsafe inputs or handle an unavailable command. Add integration checks for empty, unreadable, truncated, and unknown-state rollups plus command failure, asserting no merge write occurs. Define whether “unknown-state rollup” includes an unknown state in a superseded entry: Go validates verdicts after selecting latest entries (internal/checks/checks.go:148, internal/checks/checks.go:257).
AC6’s benchmark does not establish its threshold by itself. Specify the benchmark invocation, measurement boundary, and explicit comparison proving a 500-check evaluation finishes under one second.
Dependencies omit rollout protections required by Replace complex shell with tested Go, starting with gates and merging #588. Identify how consumers receive the required binary before updated callers, and how owner-approved protection mapping covers changed Go enforcement paths before rollout. Current configured golden and owner patterns omit those Go paths (.agents/agent-skills.conf:4, .agents/agent-skills.conf:9). If another issue supplies either prerequisite, name it and add the appropriate native blocker; otherwise distinguish work included here from a rollout hold.
Generation and consumer verification are incomplete. AC5 covers generated-skill consistency, but the plan omits the full generation → token measurement → deployment sequence and consumer migration documentation. Add those checks and document adoption steps or explain “No consumer impact.” Replacing this judge also does not remove jq from the surrounding merge path (packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:19).

Split

No split required. Keep the contract, evaluator adjustments, caller transition, focused regression coverage, generated output, and adoption guidance together. Reference independently delivered rollout prerequisites explicitly.

Definition of done

No item is fully covered by the current AC and Verify plan; implementation and merge evidence remain pending.

Round: 1 of 3

Refinement: NOT READY — native prerequisite and caller compatibility gaps remain

Checked

Problem: Supported. Shell treats unknown states as green and evaluates every status context; Go rejects unknown states and selects the latest context (packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:69, internal/checks/checks.go:180, internal/checks/checks.go:257).
Acceptance criteria: Eight correctly numbered AC entries, including performance and consumer criteria. AC structure and verification gaps remain below.
Scope and non-goals: Explicit. Fits Feature 3 of Replace complex shell with tested Go, starting with gates and merging #588 without duplicating Port the gate runner and pass cache from Python to Go #713–Replace the commit and pull request workflow guards with Go hooks #719 or the predicates in Port strict review-verdict parsing from shell to Go #589 and Port exact-head merge approval selection from shell to Go #590.
Approach: Now addresses diagnostic adaptation, exit-status mapping, deletion, ordered generation, and adoption documentation.
Dependencies: Native parent is Replace complex shell with tested Go, starting with gates and merging #588; native blocked-by list is empty. This conflicts with the stated merge prerequisite.
Size: S, one PR, remains plausible for this leaf migration.
Verify plan: Every AC ID has a check. The benchmark now names an invocation and threshold comparison. Caller coverage remains incomplete.
Risks: Unsafe acceptance, diagnostic changes, and premature rollout are identified; reverting the PR supplies an undo.
Operator authority: None satisfies the required form. The separate owner-approved protection hold is now explicit; current patterns still omit replacement Go paths (.agents/agent-skills.conf:4, .agents/agent-skills.conf:9).

Round 1’s benchmark, generation sequence, adoption guidance, and identification of rollout protections are addressed. The remaining gaps follow.

Gaps

Record the stated merge prerequisite natively. Dependencies says this change “must not merge ahead of” Ship the Go binaries as release assets installed through mise's github backend #333, which remains open, but declares “Native blocked-by: none.” Add Ship the Go binaries as release assets installed through mise's github backend #333 as a native blocker. Keeping the protection update as an explicitly separate rollout hold is consistent with Replace complex shell with tested Go, starting with gates and merging #588.
Resolve superseded-run reporting before deleting its implementation. Approach deletes mp_superseded, but AC5 covers only failing or running checks and their state/name output. Preflight also prints the number of omitted earlier runs, with an existing assertion requiring that message (packages/agents/source/skills/merge-pr/shared/scripts/preflight.sh:110, tests/skills/merge-pr/scripts/merge-workflow.bats:75). Require preservation through the replacement, or an explicit contract decision to change that diagnostic, with integration coverage.
Verify the complete compatibility table and existing Go consumer. Approach makes the contract table a fixture, but AC2’s replay is limited to existing merge Bats rollups; AC3 adds selected cases. Require execution of every contract row against its recorded old and chosen outcomes, including malformed inputs, superseded unknown states, and head movement. Repeated status contexts already demonstrate a concrete unsafe migration: Go accepts an older ERROR followed by newer SUCCESS, while shell retains the ERROR (internal/checks/checks_test.go:78, packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:89). Also retain regression coverage for unattended evaluation: ma_gates already invokes the same Go command with required names and expected head (packages/agents/source/skills/merge-pr/shared/scripts/merge-agent.sh:169). Shared evaluator changes must preserve that consumer’s required-check and head-binding behavior.
Cover command failure beyond absence, and describe preflight accurately. AC4 covers a judge that cannot run, but its Verify entry tests only a missing command. Add a failing command stub, including partial success-looking output followed by nonzero exit, and assert no merge write. AC5 also speaks of each caller’s “refusal” status, although preflight currently reports failed checks and exits successfully (tests/skills/merge-pr/scripts/merge-workflow.bats:86); squash execution refuses with status 1 (packages/agents/source/skills/merge-pr/shared/scripts/squash-merge.sh:56). State and verify those distinct outward contracts.
Finish separating observable AC from implementation. AC1 still combines contract publication with exhaustive difference accounting; AC6 combines implementation deletion with reproducible generation and names internal functions indirectly through its Verify entry. Move deletion prescriptions into Approach and give independently verifiable outcomes separate AC entries. There is room within the twelve-criterion limit.

Split

No split required. Keep the compatibility contract, evaluator changes, caller adapters, integration coverage, generated output, and adoption documentation together. Track #333 natively and retain the separate protection rollout hold.

Definition of done

Ticks indicate coverage in the proposed AC and Verify plan, not completed implementation evidence.

Round: 2 of 3

Refinement: NOT READY — compatibility and failure contracts still leave unsafe or contradictory outcomes

Checked

Reviewed #720, its two previous verdicts, #588, native dependencies, #333, #544, and sibling scopes #713–#722. Code reads used only origin/main at 56543d2. No files, checks, or external state were changed.

Problem: The two evaluators differ as described. The statement that preflight “never refuses” needs qualification.
Acceptance criteria: Eleven correctly numbered AC<n>: entries, within the limit. AC7 supplies a non-functional criterion. AC1 and AC6 still mix acceptance outcomes with implementation and verification prescriptions.
Scope and non-goals: Explicit; fits Feature 3 of Replace complex shell with tested Go, starting with gates and merging #588. No substantive overlap with Port the gate runner and pass cache from Python to Go #713–Replace the commit and pull request workflow guards with Go hooks #719, Replace the merge workflow guard with a Go hook #721, or Replace the rebase and conflict-resolution guards with Go hooks #722.
Approach: Addresses diagnostic listings, superseded counts, caller adapters, ordered generation, and documentation. Remaining semantic conflicts are below.
Dependencies: Native parent is Replace complex shell with tested Go, starting with gates and merging #588; Ship the Go binaries as release assets installed through mise's github backend #333 is now a native blocker; Port the gate runner and pass cache from Python to Go #713 is natively blocked by Settle the check judge contract and divergence corpus before porting it #720. The separate owner-protection rollout hold is explicit and consistent with Replace complex shell with tested Go, starting with gates and merging #588. Consumer installation and version sequencing remain incomplete.
Size: M and one PR are plausible for this leaf migration, subject to the eventual size gate.
Verify plan: Every AC ID appears beside a check. Some checks do not establish the complete promised behavior.
Risks: Unsafe acceptance, diagnostic drift, unattended tightening, and premature rollout are identified. Reverting the PR supplies an undo.
Operator authority: None satisfies the required form.

Round 2’s native #333 blocker and superseded-run diagnostic gaps are closed. Full-table execution and unattended regression coverage are now required. Failure-command coverage and separation of the two caller contracts improved, but remain incomplete.

Gaps

AC1 and AC6 still fail the requested acceptance-criterion structure. AC1 combines publication of a contract, its contents, and how tests execute it. AC6 combines deletion of an implementation with reproducible generation. Move test machinery and deletion prescriptions into Approach/Verify; express compatibility against the contract and reproducible generated output as separate observable outcomes. This was an explicit Round 2 finding and remains unresolved.
AC4’s unknown-state contract conflicts with the proposed exit mapping, and superseded unknown states remain unspecified. Go currently returns Failed/1 for an unknown status, conclusion, or context state, rather than Unreadable/5 (internal/checks/checks.go:180, internal/checks/checks_test.go:82, internal/checks/checks_test.go:120). Approach maps 1 to the ordinary failing-check path, but AC4 requires “could not be read.”

Go also selects latest runs before judging states (internal/checks/checks.go:148, internal/checks/checks.go:155). An older unknown-state CheckRun followed by a newer successful run therefore becomes green. AC4’s unqualified “rollup with an unknown state” requires refusal, but neither the required corpus classes nor Verify explicitly cover this superseded case. State whether validation covers every entry, define the attended diagnostic mapping, and test unknown status, conclusion, and context state—including superseded entries. Preserve or explicitly account for the existing unattended interface.
The compatibility corpus still omits a concrete failure-to-green difference: timestamp ordering. Shell compares startedAt strings; Go parses timestamps and compares instants (packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:71, packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:88, internal/checks/checks.go:110, internal/checks/checks.go:269).

For two completed runs of the same check, a failure at 2026-10-01T02:00:00+02:00 and success at 2026-10-01T01:00:00Z make shell choose failure and Go choose success. That violates AC3, yet none of AC1’s required classes forces this case into the corpus. Invalid timestamps are another existing difference already demonstrated by internal/checks/checks_test.go:158. Add timestamp-format and ordering cases, and specify a rule that cannot weaken the shell outcome. A corpus-only difference detector cannot discover classes absent from its corpus.
“Preflight never refuses” is factually wrong and could erase existing safety behavior if implemented literally. Preflight calls mp_require_reviews before reporting checks and runs unattended gates before reaching the check section (packages/agents/source/skills/merge-pr/shared/scripts/preflight.sh:94). Required reviews cause exit 75; an existing regression explicitly requires this (tests/skills/merge-pr/scripts/merge-workflow.bats:385).

Limit AC5 and Approach to the attended check-reporting section: failing or pending checks do not themselves make that section refuse. Preserve existing review, environment, and unattended-gate refusals. Verify both the reporting behavior and these existing refusal paths.
AC4’s failure messages cannot be established solely by inserting the proposed judge adapter. Both callers parse the saved response before invoking the judge. Preflight formats state and reads review status first; squash execution reads head and base fields first (packages/agents/source/skills/merge-pr/shared/scripts/preflight.sh:78, packages/agents/source/skills/merge-pr/shared/scripts/squash-merge.sh:116, packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:39). A truncated or malformed live response can exit there with a parser error before the promised “could not be read” message.

Missing or unusable commands can also fail earlier in ma_context, with an existing diagnostic that AC9 promises to preserve (packages/agents/source/skills/merge-pr/shared/scripts/merge-agent.sh:28, tests/skills/agent-merge-pr/workflow.bats:12). Clarify which entry points and input boundaries AC4 covers, reconcile its generic diagnostic with AC9, and ensure integration tests reach and assert the intended boundary rather than merely observing an earlier refusal.
The subprocess verification still omits Round 2’s success-looking partial-output case. Verify AC4 now names an unmapped exit status and garbage output, but does not explicitly exercise a command that emits a valid-looking success result and then exits with a mapped nonzero status, such as 1 or 2. Garbage output does not establish that the adapter checks process status when parsing succeeds.

Add this disagreement case, require nonzero status to prevent success, and assert the diagnostic and absence of a merge call. Also reject a zero-status response lacking the required result fields. The command’s result and process status are distinct parts of its interface (internal/agentrepo/command.go:68, internal/agentrepo/command.go:74, internal/agentrepo/command.go:96).
Binary publication is now blocked correctly, but installation of a compatible binary is still not secured. Ship the Go binaries as release assets installed through mise's github backend #333 explicitly excludes consumer installation; feat(distribution): migrate consumer tool pins to release assets #544 owns that work. Existing adoption guidance permits no agentrepo, and the shared loader accepts releases from v0.16.0 upward (docs/agent-skills-migration.md:9, packages/agents/source/lib/lib.sh:140). Those releases do not provide the proposed diagnostic listing.

Add feat(distribution): migrate consumer tool pins to release assets #544 as a native prerequisite, or specify and verify an alternative installation sequence. In either case, identify the release/capability required by the new callers and ensure it is installed before deployment. Publishing the first archive alone does not establish that consumers have the replacement interface. AC8’s documentation inspection does not prove installation or compatibility.

Split

No mandatory split identified. Keep the contract, corpus, evaluator changes, caller adapters, generated output, and adoption documentation together. Keep binary distribution, consumer installation, and the owner-protection rollout update explicitly coordinated with their separate delivery work.

Definition of done

The definition of done for #720 (Feature). Tick each item and give its evidence.

Round: 3 of 3

Status: redesigned after round 3 of 3 (#720 (comment))

Done:

Split the issue. This issue is now a Spike that settles the check judge contract (exit mapping including unknown and superseded states, start-time ordering by instant, which refusals precede the check section, and the divergence corpus) and records the findings in a comment, with each AC answered.
The caller migration moved to the new Feature Replace the shell check judge in the merge scripts with the Go evaluator #723, blocked by this Spike, Ship the Go binaries as release assets installed through mise's github backend #333 and feat(distribution): migrate consumer tool pins to release assets #544. Both are sub-issues of Replace complex shell with tested Go, starting with gates and merging #588.
Port the gate runner and pass cache from Python to Go #713 is now blocked by Replace the shell check judge in the merge scripts with the Go evaluator #723 instead of this issue.
Both bodies were rewritten in the Definition of Ready format with at most 12 ACs, each one observable outcome, and every AC cited in the Verify plan.

Why: round 3 found that one issue mixed a contract that was not yet settled with the migration that depends on it, and that the failure and compatibility rules were unsafe or contradictory in places. The Spike settles the contract first.

Next: refinement round 1 of 3 for this Spike and for #723.

Blocked on: nothing for this Spike. #723 is blocked on this Spike, #333 and #544.

Refinement: NOT READY — bound the sweep and make release and refusal evidence achievable

Checked

Read #720, all four comments, #588, native links, #723, #333, #544, #713 and the additional native dependent #719. Code references are to main at d41369e7. No files, checks, repository tooling, posts or operator commands were run.

Problem: The stated differences are supported: shell normalization, grouping and string ordering appear at packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:69; Go validation, instant ordering and verdict selection appear at internal/checks/checks.go:75, internal/checks/checks.go:110, internal/checks/checks.go:180 and internal/checks/checks.go:257. Caller reporting, early parsing and review refusal also match the description.
Acceptance criteria: Eleven correctly numbered AC<n>: entries, within the limit. Each describes a findings artifact or recorded experiment; AC10 covers performance. The implementation-deletion prescriptions from the former Feature are gone.
Scope and non-goals: Explicitly limited to research, findings and follow-ups.
Approach: Suitable comparative research, subject to the gaps below.
Dependencies: No native blockers for the Spike. Replace the shell check judge in the merge scripts with the Go evaluator #723 is natively blocked by Settle the check judge contract and divergence corpus before porting it #720, Ship the Go binaries as release assets installed through mise's github backend #333 and feat(distribution): migrate consumer tool pins to release assets #544.
Size: S is plausible once the sweep has a defined bound.
Verify plan: Every AC ID appears; several checks need strengthening.
Risks: Unsafe acceptance and incomplete corpus coverage are identified; no product change ships.
Operator authority: None is appropriate.

Gaps

AC3 leaves the sweep’s domain undefined, and Verify AC4 does not prove its universal claim. “Every combination” does not specify which known states, which state fields, which entry types, or which timestamp values constitute the finite sweep. Status contexts use createdAt, while CheckRuns use startedAt (internal/checks/checks.go:110); a generic “start-time spelling” axis leaves that distinction to the researcher.

Require the findings to enumerate the domains, generation rule and expected total before reporting coverage. For AC4, two example rows cannot establish the outcome for any pair whose string and instant orders disagree. Require an explicit general decision rule, supported by the defined sweep and code reasoning. Also reconcile AC3’s accepted-exception option with Approach’s unconditional “never greener” rule.
AC7’s Verify entry does not cover all the refusals its table must inventory. It names the two entry scripts, pull-request.sh and the review regression, but early stops also come from the shared loader and unattended helpers. Examples include the loader’s version refusal at packages/agents/source/lib/lib.sh:151, command usability at packages/agents/source/skills/merge-pr/shared/scripts/merge-agent.sh:28, and unattended gates at packages/agents/source/skills/merge-pr/shared/scripts/merge-agent.sh:118.

Define the boundary as the shell check-reporting/judging section, distinguish the relevant invocation modes, and require tracing the helpers invoked before it. Otherwise “each refusal” can omit existing stops while passing the listed verification. The exit-75 review regression itself is correctly identified at tests/skills/merge-pr/scripts/merge-workflow.bats:385.
AC9 requires release evidence that may depend on the implementation this Spike explicitly excludes. The current evaluator can choose success for the offset example that AC4 requires to remain non-green (internal/checks/checks.go:269). Naming and demonstrating a release that implements the chosen future contract therefore cannot be assumed achievable before Replace the shell check judge in the merge scripts with the Go evaluator #723.

Allow an explicit “no compatible release exists yet” finding. In that case, require the capability contract, the issue responsible for implementing and releasing it, and the installation sequencing through feat(distribution): migrate consumer tool pins to release assets #544. Verify that recorded decision rather than requiring a published compatible release to close an otherwise unblocked Spike.
D4 has no verification step. The plan does not require gitleaks, check-leaks.sh with the repository’s private-strings list, or leak review of the findings and follow-up posts. Add those checks and their recorded results to the completion plan. The recorded corpus and command output are part of the material being published.

Split

Keep the Spike/Feature split. #720 should settle and record the contract and corpus; #723 should implement and verify the migration. Release production and installation remain coordinated through #333 and #544. No further split or pull request is required for this Spike.

Definition of done

Ticks indicate coverage by the AC and Verify plan, not completed execution.

D4 Unmet in plan: the two leak scans and leak review of each post are absent.
spike-1 Approach requires a findings comment; AC11 and Verify AC11 check its contents.
spike-2 AC11 requires an answer to every AC; the corresponding Verify entries require the supporting tables, recordings, references and measurements.
spike-3 Scope and Approach require filing follow-up issues; AC11 and Verify AC11 require links from the findings comment, including Replace the shell check judge in the merge scripts with the Go evaluator #723.

PR items D1–D3 and D5–D10 do not apply.

Round: 1 of 3

Refinement: NOT READY — clarify the sweep’s state fields, complete the caller contract, and finish D4 coverage

Checked

Read the current #720 body, all comments including the Spike’s round 1 verdict, #588, native links, #723, #333, #544 and #719. Code was read from main at d41369e7.

Problem: Supported by code. Shell normalization and string ordering appear at packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:69; Go validation, state handling and instant ordering appear at internal/checks/checks.go:75, internal/checks/checks.go:180 and internal/checks/checks.go:269. Early parsing, caller reporting and exit-75 review refusal also match.
Acceptance criteria: Twelve correctly numbered AC<n>: entries. They describe findings artifacts or recorded experiments, with non-functional coverage in AC10 and AC12. The remaining contract and coverage gaps are below.
Scope and non-goals: Research, findings and follow-ups are clearly separated from implementation.
Approach: Comparative research is appropriate. AC4 now requires a general ordering rule, and Approach’s “never greener” rule has no exceptions.
Dependencies: No native blockers for the Spike. Replace the shell check judge in the merge scripts with the Go evaluator #723 is natively blocked by Settle the check judge contract and divergence corpus before porting it #720, Ship the Go binaries as release assets installed through mise's github backend #333 and feat(distribution): migrate consumer tool pins to release assets #544. AC9’s “no compatible release exists yet” option closes round 1’s release-evidence blocker.
Size: S is plausible for the bounded research and findings.
Verify plan: Every AC ID appears. AC7 now traces the loader and unattended helpers, identifies invocation modes and fixes the boundary at the shell check section. Its exit-75 regression is present at tests/skills/merge-pr/scripts/merge-workflow.bats:385.
Risks: Unsafe acceptance and incomplete corpus coverage are identified; no product change ships.
Operator authority: None is appropriate.

Gaps

AC3 still leaves the CheckRun state domain ambiguous. The two entry types and seven timestamp relations now bound much of the sweep, but a CheckRun has two independently relevant fields: status and conclusion. “Every state spelling for its type” does not specify whether their domains are crossed, or whether one field varies while the other stays fixed. The judges branch on status before reading conclusion (packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:75; internal/checks/checks.go:184).

Specify the per-field domains and their combination rule, including which lowercase and missing-field variants participate. Otherwise, different generators can report different expected totals while claiming full coverage. Verify AC3 should check that field-level product, rather than only agreement between the researcher’s chosen total and recorded total.
The tables do not require the complete chosen caller contract that Replace the shell check judge in the merge scripts with the Go evaluator #723 consumes. AC1 requires the shell’s exit status, but only the chosen merge-path outcome. AC6 similarly maps command responses to outcome categories without requiring each caller’s process status, report or diagnostic. AC5 supplies messages and statuses for unknown states; AC7 supplies them for earlier stops. That leaves cases such as empty or missing rollups and malformed evaluator output without a required chosen status and diagnostic for each caller.

These distinctions matter: preflight reports unreadable or empty checks at packages/agents/source/skills/merge-pr/shared/scripts/preflight.sh:104, while squash execution refuses at packages/agents/source/skills/merge-pr/shared/scripts/squash-merge.sh:215. Replace the shell check judge in the merge scripts with the Go evaluator #723 AC3 and AC4 explicitly depend on the contract naming caller behavior.

Require the relevant table rows to record preflight and squash outcomes separately, including process status and output/message, and make Verify inspect those fields. This completes the findings artifact without prescribing implementation.
Round 1’s D4 gap is only partly closed. AC12 and Verify AC12 now require gitleaks and the repository leak check with the private-strings list. Neither requires the separate leak review of each post named in D4.

Add that review for the complete findings comment and every follow-up body, including embedded corpus and command output, and require its clean result to be recorded.

Split

Keep the existing split: #720 settles the contract and corpus; #723 implements the migration. Distribution and installation remain with #333 and #544. No further split or pull request is required for this Spike.

Definition of done

Ticks indicate coverage by the AC and Verify plan, not completed execution.

D4 Unmet in plan: both scans are covered by AC12 and Verify AC12, but leak review of each post is still absent.
spike-1 Approach requires one findings comment; AC11 and Verify AC11 inspect it.
spike-2 AC11 requires every AC to be answered with the corresponding tables, recordings and evidence; the gaps above must be resolved for that evidence to be sufficient.
spike-3 Scope and Approach require filing follow-up issues; AC11 and Verify AC11 require links from the findings comment, including Replace the shell check judge in the merge scripts with the Go evaluator #723.

A Spike needs no pull request. PR items D1–D3 and D5–D10 do not apply.

Round: 2 of 3

Refinement: READY — all three round 2 gaps are closed

Checked

Read the current #720 body, its round 2 verdict, #723, #333, #544 and #719. Code references below are from origin/main at bfd9f84.

Problem: Supported by the shell’s normalization and selection logic at packages/agents/source/skills/merge-pr/shared/scripts/pull-request.sh:69 and Go’s parsing, verdict and timestamp logic at internal/checks/checks.go:75, internal/checks/checks.go:180 and internal/checks/checks.go:269.
Acceptance criteria: AC1–AC12 remain stable, correctly numbered and within the twelve-AC limit. Each specifies an observable research artifact or measurement; AC10 and AC12 cover non-functional requirements. AC3 now defines independent field domains and crosses them on both entries. Recounting gives status 6 + 6 + 1 + 1 = 14, conclusion 9 + 9 + 1 + 1 = 20, and context state 5 + 5 + 1 + 1 = 12. Thus 7 × ((14 × 20)² + 12²) = 548800 + 1008 = 549808, exactly the stated total. Omissions and excluded alternatives are explicit.
Scope and non-goals: The deliverables remain findings, corpus, measurements and follow-up issues. Product implementation belongs to Replace the shell check judge in the merge scripts with the Go evaluator #723.
Approach: The fixed fields, ordered pairs and seven timestamp pairs make the sweep reproducible. The ordering rule, superseded-state decision and “never greener” requirement are compatible.
Dependencies: The Spike has no native blocker. Replace the shell check judge in the merge scripts with the Go evaluator #723 is natively blocked by this Spike, distribution in Ship the Go binaries as release assets installed through mise's github backend #333 and installation in feat(distribution): migrate consumer tool pins to release assets #544. The downstream sequence in Replace the commit and pull request workflow guards with Go hooks #719 agrees. AC9 permits documenting that no compatible release exists yet.
Size: S remains plausible for generated research and a findings comment; implementation is separately scoped.
Verify plan: Every AC ID has a named proving check. AC3 verifies the field products and complete case keys, including duplicates and omissions. AC1, AC5, AC6 and AC7 require separate preflight and squash results with numeric caller process statuses, exact stdout and stderr, messages, invocation modes and earlier-stop boundaries; their Verify entries inspect those fields. This matches the reporting/refusal distinction at packages/agents/source/skills/merge-pr/shared/scripts/preflight.sh:101 and packages/agents/source/skills/merge-pr/shared/scripts/squash-merge.sh:203. The exit-75 regression exists at tests/skills/merge-pr/scripts/merge-workflow.bats:385.
Risks: Unsafe acceptance, incomplete coverage and deliberate behavior changes have corresponding checks. No product change ships.
Operator authority: None is consistent with the research scope.

Gaps

None. The field-domain ambiguity, incomplete caller contracts and missing per-post leak review are closed. No new readiness gap or contradiction found.

Split

Keep the existing split: #720 settles the contract and corpus; #723 implements migration. Distribution and installation remain with #333 and #544.

Definition of done

Ticks indicate AC and Verify-plan coverage; execution remains pending.

D4 AC12 and Verify AC12 require gitleaks, check-leaks.sh with the private-strings list, and a separate leak review of every complete final post, with clean results recorded and changed text rechecked.
spike-1 Approach requires one findings comment; AC11 and Verify AC11 inspect it.
spike-2 AC11 requires every AC’s answer; AC1–AC10 and their Verify checks supply the evidence.
spike-3 Scope and Approach require filing follow-up issues; AC11 and Verify AC11 require their links in the findings comment.

A Spike needs no pull request. PR items D1–D3 and D5–D10 do not apply.

Round: 3 of 3

Findings for #720: conservative check evaluation contract

Research head: f130498bd7383fe02f8b9d21965f555866abd98c.
I fetched origin and used an isolated worktree. No production files
changed. No commit, PR, merge or release forms part of this Spike.

The recommended contract validates every entry, keeps every status
context, and takes the worse check-run verdict selected by either
text order or instant order. A caller accepts success only when both
the process status and a complete, consistent result say success.

Evidence and notation

All source locations below refer to the research head. Let M mean
packages/agents/source/skills/merge-pr/shared/scripts/ and L mean
packages/agents/source/lib/. Paths written after those prefixes
concatenate with them. This avoids ambiguous generated-copy
references.

M/pull-request.sh:69-110: shell judge and superseded count.
internal/checks/checks.go:75-177: parse and evaluation precedence.
internal/checks/checks.go:180-206,257-286: verdict and selection.
internal/agentrepo/command.go:55-101,107-119: command protocol.
M/preflight.sh:40-120: preparation, review and check reporting.
M/squash-merge.sh:63-218: preparation and check refusal.
M/merge-agent.sh:19-30,118-180: unattended context and gates.
L/lib.sh:140-183,679-714,830-885: loader, stamps and API errors.
L/agent.sh:56-65: interrupted operations and detached HEAD.

The evidence block at the end is a lossless recording, not a link to
local files. Its metadata holds every hand-written input, both judges'
exact stdout, stderr and numeric status, full caller recordings,
unattended recordings, benchmark samples and the sweep manifest.
The binary matrix records both exact outputs for every sweep key.
The decoder reconstructs all 549808 input/output records. No fixture
requires access to the research machine.

P means attended preflight, S means squash with an explicit
operator confirmation. U means unattended preflight. G means
squash without explicit confirmation, including unattended squash.
These are fixture invocation modes, not confirmation grants.

For AC1 and AC6, earlier prerequisites and later gates pass, with all
GitHub writes stubbed. P reaches its check-reporting section and
continues to process exit 0. S either refuses there or reaches a
successful stubbed merge and process exit 0. The check section alone
doesn't establish that a real merge is authorized. AC7 overrides this
boundary whenever a real earlier stop applies.

The following executable definitions specify complete caller streams.
m is the decoded evidence metadata. Concatenation adds no implicit
whitespace. Every newline is explicit. Empty stderr means "".
pfx includes the checks heading. sfx ends just before judging.
ptail and stail include all otherwise successful later output.
Those tails come from the recorded isolated fixture, not a proposed
change to production diagnostics.

r = {(x['id'], x['caller']): x for x in m['callers']}
p = r['success', 'preflight']['stdout']
s = r['success', 'squash-merge']['stdout']
pfx, rest = p.split('SUCCESS        ci / test\n', 1)
ptail = rest.split('\nsummary: all green.\n', 1)[1]
sfx, stail = s.split('open, ready, mergeable, and green\n', 1)
summary = {
    0: 'all green.\n',
    1: ('FAILING. Do not merge. Hand this to the '
        'fix-pr skill instead.\n'),
    2: ('still running. Wait through the watch-pr '
        'skill before merging.\n'),
}
err = {
    1: ('squash-merge: #7 has a failing check. '
        'Hand it to the fix-pr skill.\n'),
    2: ('squash-merge: #7 has a check still running. '
        'Wait through the watch-pr skill.\n'),
    4: ('squash-merge: #7 has no checks reported; '
        'the result is inconclusive.\n'),
    5: 'squash-merge: the checks of #7 could not be read.\n',
}
def line(state, name):
    return f'{state:<14} {name}\n'
def P(code, rows=(), superseded=0):
    if code in (4, 5):
        body = ('no checks reported\n' if code == 4 else
                'the check rollup could not be read\n')
    else:
        body = ''.join(line(state, name) for state, name
                       in sorted(rows, key=lambda x: x[1]))
        if superseded:
            body += (f'({superseded} earlier run(s) left out: '
                     'a later run of the same check replaced '
                     'each)\n')
        body += '\nsummary: ' + summary[code]
    return (0, pfx + body + ptail, '')
def S(code, rows=()):
    if code == 0:
        return (0, sfx + 'open, ready, mergeable, and green\n'
                + stail, '')
    body = ''.join(line(state, name) for state, name in rows)
    return (1, sfx, body + err[code])
F = [('FAILURE', 'ci / test')]
Q = [('QUEUED', 'ci / test')]
C = [('ERROR', 'test'), ('SUCCESS', 'test')]

Each tuple is (caller process status, stdout, stderr).
P(5) is an unreadable report and continuation, not green.
S(5) refuses with no merge call. P(4) reports empty and continues.
S(4) refuses. P(1,...) and P(2,...) report and continue.
S(1,...) and S(2,...) print the relevant rows and refuse.
The failure diagnostic deliberately drops “judged by its latest run”:
that phrase would misdescribe the conservative two-order decision.

AC1, AC2 and AC5: input contract and recorded comparison

Table keys are exact id values in m['hand']. That record holds the
saved input bytes, command arguments, shell stdout/stderr/status and
current Go stdout/stderr/status. Both commands ran on each input.
All shell stderr streams were empty. Go stderr was empty except for
the required-check hints explicitly recorded under AC8.

The last column marks deliberate changes to the attended shell path.
yes includes stricter interpretation or changed diagnostics.
parser identifies an earlier stop, not an adapter result.

Key	Shell	Go	P result	S result	Change
h00	3	5	parser/5	parser/5	no
h01	4	5	P(5)	S(5)	yes
h02	4	5	P(5)	S(5)	yes
h03	4	4	P(4)	S(4)	no
h04	0	0	P(0,OK)	S(0)	no
h05	1	1	P(1,F)	S(1,F)	message
h06	2	2	P(2,Q)	S(2,Q)	no
h07	0	1	P(5)	S(5)	yes
h08	0	1	P(5)	S(5)	yes
h09	0	1	P(5)	S(5)	yes
h10	0	1	P(5)	S(5)	yes
h11	0	1	P(5)	S(5)	yes
h12	0	1	P(5)	S(5)	yes
h13	0	1	P(5)	S(5)	yes
h14	0	5	P(5)	S(5)	yes
h15	0	5	P(5)	S(5)	yes
h16	0	1	P(1,F,1)	S(1,F)	yes
h17	1	1	P(1,F,1)	S(1,F)	yes
h18	2	2	P(2,Q,1)	S(2,Q)	no
h19	1	0	P(1,C)	S(1,C[:1])	message
h20	1	0	P(1,F,1)	S(1,F)	yes
h21	0	5	P(5)	S(5)	yes
h22	0	0	P(5)	S(5)	yes
h23	0	0	P(5)	S(5)	yes
h24	0	0	P(5)	S(5)	yes

OK = [("SUCCESS", "ci / test")]. Input keys:

h00: malformed.
h01: missing.
h02: null.
h03: empty.
h04: success.
h05: failure.
h06: pending.
h07: unknown_status.
h08: unknown_conclusion.
h09: unknown_context.
h10: lower_status.
h11: lower_conclusion.
h12: lower_context.
h13: missing_status.
h14: empty_name.
h15: unknown_type.
h16: tie.
h17: tie_reverse.
h18: unstarted.
h19: repeated_context.
h20: offset.
h21: invalid_time.
h22: superseded_status.
h23: superseded_conclusion.
h24: superseded_context.

For h00, both callers exit 5 before evaluating checks. The exact
stdout and stderr are in m['callers'] for malformed, with raw {
as the response. Stderr is the parser diagnostic below. The preflight
first stops at its state formatter. Squash first stops in mp_field.
This result stays unchanged in the chosen contract.

jq: parse error: Unfinished JSON term at EOF at line 1, column 1

Missing and null rollups deliberately change from empty to unreadable.
Their check-section reports are P(5) and S(5), not the old empty
reports. An actual empty array remains P(4) and S(4).

Unknown status, conclusion and context state all mean unreadable,
even on superseded entries. Validate every raw entry before selection,
including a conclusion that a non-completed status would otherwise
hide. Lowercase spellings and a missing CheckRun status are also
unreadable. This preserves Go's refusal and tightens shell acceptance.
Don't silently uppercase them and loosen the unattended path.

Accept the six uppercase CheckRun statuses and nine uppercase
conclusions in AC3. An absent, null or empty conclusion is allowed and
means pending when status is COMPLETED. An absent, null or empty
context state means PENDING. Recognized pending statuses still ignore
a recognized conclusion for verdict purposes. Null/empty CheckRun
status, other field types, empty names and unknown entry types are
unreadable. A context accepts only the five uppercase context states,
plus the missing-value convention above. These validation rules also
cover entries excluded by time ordering.

AC3: finite sweep, coverage and results

The domain manifest was written before execution. Generation order is
CheckRun then StatusContext, first-entry index, second-entry index,
then timestamp-pair index. CheckRun indices enumerate status as the
outer axis and conclusion as the inner axis. Omission is absence of
the JSON property. It isn't null or an empty string.

Status: COMPLETED, IN_PROGRESS, PENDING, QUEUED, REQUESTED, WAITING,
then those six lowercase spellings, UNKNOWN_STATE, then omission.
Conclusion: ACTION_REQUIRED, CANCELLED, FAILURE, NEUTRAL, SKIPPED,
STALE, STARTUP_FAILURE, SUCCESS, TIMED_OUT, then those nine
lowercase
spellings, UNKNOWN_STATE, then omission.
Context state: ERROR, EXPECTED, FAILURE, PENDING, SUCCESS, then
those five lowercase spellings, UNKNOWN_STATE, then omission.
Seven timestamp pairs, in the issue's order: equal, earlier, later,
offset-reversed, invalid first, invalid second, both invalid.
m['manifest']['times'] records every literal timestamp.
Fixed head head, CheckRun workflow ci, name test, or context
test. No required names or expected-head argument. No mixed types
or extra fields. JSON array order is first then second.

Expected total, fixed before running:
7 * ((14 * 20)**2 + 12**2) = 549808.
That is 548800 CheckRun pairs and 1008 context pairs.

The shell batch ran the exact jq program extracted from
mp_judge_checks. Its TSV rows went through the same failure-first,
pending-second status reduction. The Go batch called the actual
agentrepo.Run command handler for each input, with fresh buffers and
no constraints. Batching avoids over a million process launches.
The unmodified shell function and built CLI separately matched both
streams and statuses on 103 fresh-process samples, plus every hand
fixture. This distinguishes command recordings from an independent
reimplementation of either judge.

The coverage check independently reconstructed all mandated field
products and all expected keys, compared each entry's complete fields
with its index and checked timestamp values, order and fixed fields.
It found exactly 549808 unique keys, zero duplicates, zero missing
keys and zero unexpected keys. All 280 crossed CheckRun entries and
all 12 context entries participate in both positions.

Chosen results: 382 success, 2168 failing, 11994 pending and 535264
unreadable. Zero shell failures or pending results become success.
Current Go yields success for 120 shell-failing and 624 shell-pending
cases. The chosen rule prevents all 744. The full three-way counts
are in m['summary']['cross'], whose rows are
[shell status, current Go status, chosen status, count].

AC4: time ordering rule

First reject every invalid timestamp, even on a superseded entry.
For each CheckRun group (workflowName, name), compute both the
shell text-order maximum and the parsed-instant maximum. Keep every
entry tied at either maximum. Evaluate those candidates and choose
failure over pending over success, with deterministic state/name
ordering for equal verdicts. The selected display row represents
that worst candidate. Count one displayed row per CheckRun group.

Use the shell's exact text key: missing/empty start or a start that
begins 0001 maps to 9999-12-31T23:59:59Z. For instant order,
missing/empty start or the exact zero timestamp outranks started
runs. Other timestamps must parse as RFC3339Nano. This preserves the
shell's unusual zero-prefix case without letting it choose green
over a conflicting instant-order failure.

For any ordering disagreement, the candidate union includes both
maxima. Its worst verdict cannot be green if either maximum fails or
is unfinished. Ties retain all candidates, so array order cannot
weaken the result. The argument applies to groups larger than pairs.
For h20's offset example, the failure wins and gives P(1,F,1) and
S(1,F). The sweep found zero unsafe offset-reversed cases.

Keep every StatusContext entry, in its existing array order for
otherwise equal display names. Don't discard an older failing or
pending context. Validate createdAt, but don't use it to discard
contexts. This preserves the shell contract for repeated contexts.

Retain the old superseded count: number of CheckRuns minus the number
of distinct (workflowName, name) groups. Its legacy wording remains
exactly as in P above, including for ties. It counts collapsed display
rows, not proof that every omitted timestamp is strictly earlier.

AC6: evaluator process and result contract

Keep numeric outcomes 0 through 6 and their precedence: unreadable,
head moved, empty, missing required, failed, pending, success.
Require an additive schema: 1, checks array and nonnegative
integer
superseded on every result. Each check has string state and name
and a verdict of ok, pending or fail. Preserve existing
outcome, sorted unique names, skipped and error semantics.
For unreadable input require a nonempty error and empty checks.
For missing required checks, keep the missing names in names.

Validate the whole stdout as exactly one JSON object with no trailing
non-whitespace bytes. Verify types, outcome/status agreement, listing
verdicts, outcome precedence and consistency of summary names and
counts. For success require a nonempty all-ok listing, no error or
non-success names, and process exit 0. This rejects empty success and
partial success-looking output. Don't infer success from an exit
status alone or from a JSON fragment.

Let R(n) mean the complete, internally consistent schema-1 result
for outcome n, ending with one newline. Specific result examples are
specified below in executable form. Unless stated otherwise evaluator
stderr is empty. Caller stdout/stderr/status use the full P/S tuples,
not the evaluator status. All these rows occur at the check section
unless the early-boundary column says otherwise.

Eval exit	Eval stdout	Category	P	S	Boundary
0	R(0)	green	P(0,OK)	S(0)	checks
1	R(1)	failing	P(1,F)	S(1,F)	checks
2	R(2)	running	P(2,Q)	S(2,Q)	checks
3	R(3)	failing	P(1,MR+OR)	S(1,MR)	checks
4	R(4)	empty	P(4)	S(4)	checks
5	R(5)	unreadable	P(5)	S(5)	checks
6	R(6)	unreadable	P(5)	S(5)	checks
7, 127, 137	any	unreadable	P(5)	S(5)	checks
other nonzero	any	unreadable	P(5)	S(5)	checks
absent	absent	unreadable	P(5)	S(5)	see AC7
0	`garbage\n`	unreadable	P(5)	S(5)	checks
0	`{}\n`	unreadable	P(5)	S(5)	checks
0	`[]\n`	unreadable	P(5)	S(5)	checks
0	help text	unreadable	P(5)	S(5)	checks
0	`{"outcome":0}\n`	unreadable	P(5)	S(5)	checks
0	R(0) + garbage	unreadable	P(5)	S(5)	checks
1	R(0)	unreadable	P(5)	S(5)	checks
2	R(0)	unreadable	P(5)	S(5)	checks
other nonzero	R(0)	unreadable	P(5)	S(5)	checks
0	missing result field	unreadable	P(5)	S(5)	checks

The absent-command row describes P/S with the loader allowing absence.
U/G instead stop at ma_context, as AC7 specifies. Loader version
refusals precede both. Mapped 1 or 2 with R(0) remains unreadable,
not a valid failed/pending result. Evaluator stderr is captured and
must never substitute for a result. Keep recognized skipped-check
hints as described in AC8. For protocol/command failures, report the
stable unreadable message rather than copying arbitrary subprocess
stdout into the caller report.

MR = [('MISSING_REQUIRED', 'test')]
OR = [('SUCCESS', 'ci / other')]
def R(n):
    rows = {0: [('ok', 'SUCCESS', 'ci / test')],
            1: [('fail', 'FAILURE', 'ci / test')],
            2: [('pending', 'QUEUED', 'ci / test')],
            3: [('ok', 'SUCCESS', 'ci / other')]}.get(n, [])
    value = dict(schema=1, outcome=n,
                 checks=[dict(verdict=v, state=s, name=k)
                         for v, s, k in rows], superseded=0)
    if n in (1, 2, 3):
        value['names'] = ['test']
    if n == 5:
        value['error'] = 'invalid rollup'
    return json.dumps(value, separators=(',', ':')) + '\n'

For outcome 3, append MISSING_REQUIRED diagnostic rows for missing
names to the display listing. The P/S row above includes both the
reported successful check and
the missing-check diagnostic in preflight. Squash prints only the
missing-check diagnostic before refusing. Head-moved status 6 remains
a
separate evaluator outcome and refuses as unreadable in the attended
adapter. U preserves the earlier gate's status 6.

AC7: earlier stops, modes and exact stream rules

No earlier gate is removed or relaxed. Full process recordings for
malformed JSON and review/PR-state cases are in m['callers'] and
m['early']. They include both streams, numeric status and
reachability.
The remaining rows below are source-traced contracts. They aren't
claims that every external failure was injected into a live workflow.

For the source-traced rows, use the exact source prefixes below to
complete stdout. Messages in the diagnostic catalogue end in one
newline. An ordinary dt_die adds no prefix. Squash's die adds
squash-merge: . Both exit 1. Empty stdout/stderr is explicitly "".
The table's tuple notation is (status, stdout, stderr, reached).
pass means the other caller proceeds to P(0,OK) or S(0), with its
full recorded stdout, empty stderr, status 0 and reached=yes.

Pgh = '\n== gh ==\nok\n'
Ptarget = (Pgh + '\n== target ==\nlocal branch: topic\n'
           'pull request: #7\n')
Pstate = Ptarget + '\n== state ==\n'
Pbefore = pfx.removesuffix('\n== checks ==\n')
Sd = '\n--- checking SQUASH_AGENTMSG\n'
Ss = (Sd + 'drafted for #7\n'
      '\n--- checking the review stamp\n')
Sr = Ss + 'reviewed at head head, onto main at base\n'
Sa = Sr + '\n--- checking the gh account\n'
# Concrete failing subprocess fixture: exit 2, empty stdout.
E = 'fixture failure\n'
# U's EXIT trap, with no mail transport configured:
def report(n, status):
    return (f'REPORT (blocker): Merge workflow for #{n} held '
            f'with status {status}. Inspect the preserved '
            'draft, review, and merge result before retrying.\n')

With U enabled, append report(number,status) to stdout on every
nonzero exit after pull-request.sh installed its trap. Before
preflight assigns its number, number is unknown. During loader
failure, the trap isn't installed, so no report appears. G appends it
only when DEVTOOLS_UNATTENDED=1. A configured reporting transport
keeps its existing output and failure behavior instead. This is
orthogonal to the judge result and doesn't change the exit status.

The table uses ordinary P/S tuples and explicitly states mode changes.
All rows preserve their current stop and message. Source suffixes
refer to the catalogue immediately following the table.

First stop	P tuple	S tuple	Modes
loader version read	(1,"",D1,no)	same	all
loader version floor	(1,"",D2,no)	same	all
git root read	(128,"",E,no)	(128,"",E,no)	all
worktree readiness	(1,"",D3,no)	pass	P/U
argument parser	pass	(1,"",D4,no)	S/G
missing commit-lint	pass	(1,"",D5,no)	S/G
missing gh	(0,PG,"",no)	(1,"",D6,no)	P/S
missing jq	(0,PJ,"",no)	(1,"",D7,no)	P/S
viewer auth error	(2,Pgh,E,no)	(2,"",E,no)	all
viewer empty	(1,Pgh,D8,no)	(1,"",D8,no)	all
rate limit	(76,Pgh,RL,no)	(76,"",RL,no)	all
PR lookup error	(2,Pgh,E,no)	(2,"",E,no)	all
no target	(0,PN,"",no)	(1,"",D9,no)	all
missing draft	pass	(1,Sd,D10,no)	S/G
wrong draft number	pass	(1,Sd,D11,no)	S/G
snapshot read error	(2,Ptarget,E,no)	(2,Ss,E,no)	all
malformed snapshot	saved h00	saved h00	P/S
malformed fields	(5,Pstate,E,no)	(5,Ss,E,no)	P/S
base read error	pass	(2,Ss,E,no)	S/G
missing stamp	pass	(1,Ss,D12,no)	S/G
stale stamp	pass	(1,Ss,D13,no)	S/G
expected head mismatch	pass	(1,Ss,D14,no)	S/G
account mismatch	pass	(1,Sa,D15,no)	S/G
unknown mergeability	saved early	saved early	all
closed PR	saved early	saved early	all
draft PR	saved early	saved early	all
conflicting PR	saved early	saved early	all
required review	saved early/75	saved early/75	all
changes requested	saved early/75	saved early/75	all
unknown review enum	saved early/1	saved early/1	all
not CLEAN	saved early	saved early	all
unknown fork flag	saved early	saved early	all

PG is "\n== gh ==\ngh is not installed\n" followed by
"Stop here and tell the operator.\n". PJ replaces gh with jq.
PN is Pgh followed by the no-target block in preflight lines 66-74,
with branch topic and an empty open-PR list. Its exact bytes are:

PN = (Pgh + '\n== target ==\nlocal branch: topic\n'
      'pull request: NONE for this branch, and none given.\n'
      'Pass the number as the skill argument, or open one '
      'with the pr skill.\n\nopen pull requests here:\n')

Diagnostic catalogue, using quoted Python strings for exact bytes.
binary is the selected executable path, here is harness in the
recorded fixture, and the example old version is v0.15.0. These are
parameters of the existing diagnostic, not text to hardcode in #723.

D1 = E + f'Could not read agentrepo version: {binary}\n'
D2 = 'agentrepo found v0.15.0; required v0.16.0 or newer\n'
D3 = ('worktree setup: mise config is untrusted. Run:\n\n'
      '  mise trust\n')
D4 = 'squash-merge: usage: squash-merge.sh [--confirmed] [number]\n'
D5 = ('squash-merge: commit-lint is not installed; add '
      '"go:github.com/tbhb-dev/devtools/cmd/commit-lint" = '
      '"<devtools-version>" to mise.toml [tools]\n')
D6 = 'gh is not installed\n'
D7 = 'jq is not installed\n'
D8 = 'GitHub returned no viewer login\n'
D9 = ('squash-merge: no open pull request for topic. '
      'Pass a number explicitly.\n')
D10 = ('squash-merge: no squash message drafted. '
       'Run squash-message.sh first.\n')
D11 = 'squash-merge: SQUASH_AGENTMSG was drafted for #8, not #7.\n'
D12 = ('squash-merge: no independent review has cleared this '
       'message. Run it, resolve what\nit returns, then retry:'
       '\n\n  bash harness/review.sh 7\n')
D13tail = ('squash-merge: the reviewed message and pull request '
           'are no longer the ones about to\nmerge. Review '
           'again against the current draft and branch:'
           '\n\n  bash harness/review.sh 7\n')
D14 = 'squash-merge: Expected head differs from the reviewed head\n'
D15 = ('squash-merge: gh is signed in as agent, not the '
       'configured other. Nothing was merged.\n')

D1 includes the failed version command's exact stderr, empty when
that command printed none. D3 also has source-defined alternatives,
both P/U process exit 1
before the check section and S/G pass: worktree setup: Vale styles are missing. Run:\n\n mise run devtools:worktree-setup\n, or
worktree setup: could not check mise trust:\n plus captured output
and newline. Other Vale configuration failures use
worktree setup: Vale could not load the repository configuration:\n
plus captured output and newline.

D4 also covers malformed expected-head syntax with
squash-merge: Expected head must be a full commit SHA\n.
D13 is the exact dt_require_stamp diagnostic followed by D13tail:
changed draft, changed bound field, changed repository configuration,
or changed trusted policy authority. Each ends in newline. The
field-change template is The {field} changed after the review (reviewed {old}, now {new}). Run the review script again. The other
literal messages are in L/lib.sh:687-709. Failures reading the stamp
or its digest keep the underlying stderr before D13tail and exit 1.

For external reader/parser errors, E means the exact captured stderr,
not a replacement message. Exit 2 above is the concrete external
failure fixture, and 128 is the Git failure fixture. Other subprocess
statuses propagate numerically unchanged. Parser status is jq's 5,
with its actual diagnostic and any partial stdout retained. An
arbitrary external program can emit arbitrary bytes, so this family
is parameterized by its captured bytes and numeric status. The h00
record gives a concrete complete parser example without that freedom.

RL is exactly the classifier message at L/lib.sh:860-864, plus
No automatic retry was made.\n, followed by the original error.
Secondary, primary GraphQL and primary REST have separate messages.
All exit 76 and retain original stderr. The first failing read decides
which stdout prefix applies. Later API reads use their prefix at that
point, not necessarily Pgh/S's empty prefix. No retry or budget query
belongs to this Spike.

U/G-specific earlier helpers follow. These modes execute ma_context.
P and explicitly confirmed S skip it and instead take their normal
report/merge paths, unless a common earlier gate above already stops.
For U, ma_gates runs before the check section. For G, ma_gates first
runs AFTER the check section and lint, at squash line 244. Don't move
its refusal earlier while replacing only the judge.

Context stop	U status/stdout	G status/stdout	stderr
operation marker	1/report(unknown,1)	1/report(7,1)	C1
detached HEAD	1/report(unknown,1)	1/report(7,1)	C2
unborn HEAD	1/report(unknown,1)	1/report(7,1)	C3
dirty worktree	1/report(unknown,1)	1/report(7,1)	C4
no configured login	1/report(unknown,1)	1/report(7,1)	C5
missing uv	1/report(unknown,1)	1/report(7,1)	C6
absent agentrepo	1/report(unknown,1)	1/report(7,1)	C7
unusable agentrepo	1/report(unknown,1)	1/report(7,1)	C8
unattended --confirmed	U normal	1/report(7,1)	C9

The --confirmed flag belongs to squash. The paired U invocation is
ordinary preflight, which reports and exits 0 with its normal stdout
and empty stderr when its gates pass. All other context stops have
reached=no for U/G. Plain G without the
unattended flag has empty stdout instead of report(...). Underlying
Git/config read failures propagate their exit status and stderr,
with the same empty prefix and trap rule. C6 includes Bash's
command -v uv output only if that builtin emits one in the selected
shell. Bash in this environment emits none for the lookup.

C1 = ('HELD: MERGE_HEAD is in progress; finish or abort it '
      'first.\n')
# Substitute CHERRY_PICK_HEAD, REVERT_HEAD, rebase-merge,
# or rebase-apply for the other marker cases.
C2 = 'HELD: HEAD is detached.\n'
C3 = ('Create the initial commit through the interactive '
      'commit skill first.\n')
C4 = ('Commit the work first; unattended merging requires '
      'a clean tree.\n')
C5 = 'gh_login must be configured on the machine\n'
C6 = 'uv is required for the YAML policy reader\n'
C7 = 'agentrepo is required for check judgement\n'
C8 = 'a working agentrepo is required for check judgement\n'
C9 = ('squash-merge: Unattended workflows refuse --confirmed; '
      'use pre-approval or delegation.\n')

agent-merge-pr/shared/scripts/merge.sh sets unattended mode and
forwards to squash. With two numeric arguments it converts the second
into --expected-head. Its --confirmed refusal therefore comes from
the shared squash caller after ma_context, not from the wrapper.

Before U/G reach their initial response parsing, mp_view also runs
ma_reviews. A malformed raw response stops at its mp_field with
status 5, stdout Ptarget for U or Ss for G, and the exact jq stderr.
A failed review-pages read propagates its numeric status and stderr.
A mismatched review-page head stops with status 5 and the jq error
PR head changed while reading review authors; the check section
isn't reached. P/S don't take this extra reader and continue normally
on a valid snapshot with the same paginated-review fault.

U's pre-check gates have the following first-stop catalogue. For
each row, U has reached=no. Its stdout is Pbefore plus any earlier
gate output plus the trap report. The stderr is the listed diagnostic
with newline. P and S pass on the same isolated gate-only fault.
G reaches the checks first, then stops at the same gate after lint:
its stdout is sfx plus the green line and the running-linters heading,
plus any prior gate output and its optional trap report. The process
status equals U's. This records the other caller even though its stop
lies beyond the check-section boundary.

First U gate stop	Status	Diagnostic/source
PR state/draft/readiness	1	A1
actual login mismatch	1	A2
reviewer login [bot] suffix	1	A3
no required reviewer	1	A4
invalid reviewer type	1	A5
exact-head approval absent	1	A6
empty owner-review label	1	A7
owner-review label present	1	A8
pinned diff acquisition	1	A9
policy inspection/evaluation	1	A10
owner approval absent	1	A11
requirements unavailable	1	A12
evaluator non-success	1..6	AC6/AC8 JSON and hint
readiness not executable	1	A13
readiness exits nonzero	2	exact E fixture

A ma_gates review refusal is status 1. It isn't the status-75
mp_require_reviews refusal. That distinction stays unchanged.
All ma_context checks run again inside ma_gates before this list.

A1 = 'PR is not open, ready, MERGEABLE, and CLEAN\n'
A2 = 'gh is signed in as other, not agent\n'
A3 = ('required_reviewer must use the GraphQL login reviewer, '
      'without [bot]\n')
A4 = 'required_reviewer must be configured on the machine\n'
A5 = 'required_reviewer_type must be Bot or User when set\n'
A6 = 'WAITING_REVIEW: reviewer must approve the current head\n'
A7 = 'owner_review_label must not be empty\n'
A8 = 'needs-owner: the PR has the owner-review label\n'
A9 = 'Could not fetch the pinned base and head.\n'
A11 = ('needs-owner: golden files or loosened gates require '
       'merge_owner approval on the current head\n')
A12 = 'No required checks are known for the base\n'
A13 = 'merge_ready_cmd must name an executable\n'

A9 variants: unavailable repository URL, incomplete pinned histories,
failed complete diff, or failed name-status list. Preserve the exact
messages in M/pull-request.sh:140-151 and any preceding Git stderr.
A10 is merge policy unreadable: {reason}\n, from policy.py's main
exception handler, or a propagated uv/Python process failure. Reasons
cover invalid policy patterns, repository paths, APM lock fields,
review/findings shape, ruleset fields and missing required checks.
No raw policy value gets added to the replacement judge diagnostic.
Missing requirements from the policy reader use
merge policy unreadable: No required checks are known for the base
plus newline, rather than the later empty-output check's A12.
For owner-required policy, the existing JSON findings precede A11 on
stderr. A successful owner override instead prints its existing
owner override: {login} approved {head}\n on stdout.
Configuration parsing and Git/filesystem reads at any of these points
retain their numeric status and exact captured streams as above.

The malformed/required-review recordings and source trace establish
that preflight's CHECK SECTION reports failing and running checks
without refusing. They do not assert that preflight never refuses.
The existing review regression is at
tests/skills/merge-pr/scripts/merge-workflow.bats:385.

AC8: unattended fixtures and hints

None of the existing check-bearing fixtures under
tests/skills/agent-merge-pr/ changes its semantic outcome, numeric
status or stderr hint under this contract. The schema/listing fields
are additive stdout changes. The old summary fields keep their exact
values. The evidence includes direct evaluator replays of every
check-shape variant with required names and expected head supplied.
These are evaluator replays, not a claim that the entire Bats suite
ran through a future implementation.

Relevant fixtures: workflow.bats:76,223,233,242,253,267,
loosens.bats:28, confirmation.bats:33-34,161, and the base success
fixture in test_helper.bash. The missing-requirements policy case
stops before evaluation with A12, unchanged. Missing/unusable/pinned
runner cases in workflow.bats:12-41 keep the AC7 boundaries.

Variant	Outcome/status	stdout summary	stderr hint
empty	empty/4	{"outcome":4}	""
skipped/current	failed/1	test/test	H(test,no)
skipped/lagging	failed/1	test/test	H(test,yes)
skipped/unknown	failed/1	test/test	H(test,unknown)
skipped/missing	missing/3	test/validate	H(validate,yes)
required NEUTRAL	failed/1	names=test	""
updated SUCCESS	green/0	{"outcome":0}	""

def H(name, differs):
    return (f'skipped required check: {name}\n'
            f'merge base differs from base commit: {differs}\n'
            'run: gh pr update-branch 7\n')

In the compact rows, test/test and test/validate mean
names/skipped, each a single-name array.

The full byte-exact old JSON and stderr are in m['unattended'].
Add the new fields without removing those summaries. Keep exact
required-name matching across every workflow. Required NEUTRAL and
SKIPPED fail, optional NEUTRAL and SKIPPED pass. Keep missing-required
precedence and the skipped hint. Head mismatch remains 6 and missing
expected-head data remains unreadable/5, as recorded in m['hand'].

Go unit fixtures reordered statuses and status rerun in
internal/checks/checks_test.go deliberately change success/0 to
failed/1. Unknown status/conclusion/context fixtures change failed/1
to unreadable/5. New coverage must add superseded invalid states,
lowercase fields, offset disagreements and command protocol faults.
These tightenings apply to both callers of the shared evaluator.

AC9: capability and release sequence

No compatible release exists yet. The current main implementation,
and the latest local tag v0.22.1, don't provide the proposed schema,
complete listing or superseded count. Main also still implements the
unsafe context/offset selection demonstrated above. Don't invent a
minimum released version for capabilities that don't exist.

Missing promised outputs: all-entry validation failures for unknown
and lowercase fields, conservative dual-order selection, all status
contexts, full workflow-qualified listing, superseded count, schema
marker and consistent protocol-failure adaptation. #723 implements
these evaluator and caller changes and documents the first release
that contains them. #333 supplies binary distribution. #544 installs
the compatible release before its dependent callers are deployed.

With no agentrepo, P reports unreadable/0 and S refuses unreadable/1
at the new adapter, while U/G retain their earlier C7 refusal/1.
A release below v0.16.0 still fails the shared loader with D2/1.
An older accepted release lacking schema 1 or its required fields
cannot be green: P reports unreadable/0, S refuses unreadable/1.
U/G must validate the same result protocol at their gate and refuse
with unreadable/5. An unusable executable can stop earlier at D1/C8.
Don't add a second shell judge as a fallback.

AC10: measured time and proposed limit

Generate 500 distinct successful CheckRuns, fixed head head,
workflow
ci, names test-000 through test-499, all started at
2026-10-01T01:00:00Z. Save once. Build outside the timed region.
Time each subprocess invocation through its completion, capturing
both streams, for 20 runs. All runs exited zero. The evidence stores
all 40 samples, in milliseconds.

Shell function via Bash: median 28.179033 ms.
Built Go command: median 1.954330 ms.

Commands: bash judge.sh 500.json (the unchanged sourced shell
function) and agentrepo checks evaluate --file 500.json.
The timer used Python time.perf_counter_ns(). These current-judge
measurements exclude the not-yet-implemented schema validation and
caller adapter. #723 must measure that complete boundary separately.
Keep its proposed limit: median strictly below 1000 ms for 20 runs,
including invocation, whole-response validation and caller verdict,
with compilation excluded. Both existing medians are below the limit.

AC11: ownership, open questions and follow-up work

Every AC has a section in this comment. The migration belongs to
#723. It should
commit the contract, decoder and corpus before deleting the shell
judge, then implement the additive protocol and caller adapters.
Its tests should compare both caller streams and statuses, including
earlier-stop reachability. Update docs/checks-evaluator.md and the
consumer adoption documentation in that same feature.

Further work already has owners: #333 distribution, #544 installation,
#588 replacement-path protection, then #713 gate migration. #715,
#719, #721 and #722 retain their existing ordering and rollout holds.
No additional implementation issue is needed for findings already
inside #723. No duplicate issue was filed. No feature PR is
appropriate
for this Spike.

Open implementation choices that don't alter the contract: which
release number first carries it, how #723 stores the compact sweep
fixture, and whether its eventual diff requires the split already
specified in #723. The evaluator precedence, validation scope,
ordering rule, caller messages and failure mapping above are
decisions,
not questions deferred to the builder.

AC12: publication checks

PUBLICATION HOLD: the required private-strings list isn't available
in this session. No findings comment or follow-up body has been
posted. Gitleaks found no leaks in this complete draft and its decoded
metadata and 549808-record corpus. The repository check-leaks.sh also
returned zero, but warned that no private values were loaded. That
patterns-only result does not satisfy the required private-list scan.

A separate manual leak-review pass found no credentials, private
identifiers, local home paths or prohibited coordination references
in the complete draft or its decoded evidence. All fixture identifiers
are synthetic. No follow-up post exists. Recheck the final text after
adding the missing private-list scan result and before publication.

Recorded evidence and decoder

Save the following base64 block as evidence.b64. Decode it with the
code below. The XZ payload starts with an unsigned little-endian
8-byte metadata length, then UTF-8 JSON metadata, then four bytes per
case: unsigned little-endian shell-dictionary index, one-byte
Go-dictionary index, one-byte proposed evaluator outcome.
Dictionary entries are [status, stdout, stderr].

Uncompressed SHA-256:
ee8e74ba494a5979c91eab8a563edf0c2ef758c2e1e0638fccd4891f70e5aa72.
The matrix SHA-256 is in m['summary']['matrix_sha256'].

import base64, hashlib, itertools, json, lzma, struct
from pathlib import Path
raw = lzma.decompress(base64.b64decode(
    Path('evidence.b64').read_text()))
size, = struct.unpack_from('<Q', raw)
m = json.loads(raw[8:8 + size])
matrix = raw[8 + size:]
assert len(matrix) == 549808 * 4
assert hashlib.sha256(matrix).hexdigest() == (
    m['summary']['matrix_sha256'])
Path('metadata.json').write_text(json.dumps(m, indent=2))
d = m['manifest']
index = 0
seen = set()
with Path('records.ndjson').open('w') as output:
    for kind in ('CheckRun', 'StatusContext'):
        entries = []
        if kind == 'CheckRun':
            for s, c in itertools.product(d['status'],
                                          d['conclusion']):
                e = dict(__typename=kind,
                         workflowName='ci', name='test')
                for k, v in [('status', s), ('conclusion', c)]:
                    if v is not None:
                        e[k] = v
                entries.append(e)
            field = 'startedAt'
        else:
            for s in d['state']:
                e = dict(__typename=kind, context='test')
                if s is not None:
                    e['state'] = s
                entries.append(e)
            field = 'createdAt'
        axes = (enumerate(entries), enumerate(entries),
                enumerate(d['times']))
        for (i, a), (j, b), (k, (x, y)) in (
                itertools.product(*axes)):
            key = (kind, i, j, k)
            assert key not in seen
            seen.add(key)
            rollup = dict(headRefOid='head', statusCheckRollup=[
                dict(a, **{field: x}), dict(b, **{field: y})])
            si, gi, chosen = struct.unpack_from(
                '<HBB', matrix, index * 4)
            record = dict(key=key, input=rollup,
                          shell=m['dictionary'][0][si],
                          go=m['dictionary'][1][gi],
                          chosen=chosen)
            output.write(json.dumps(record) + '\n')
            index += 1
assert index == len(seen) == 549808

The chosen outcomes are a research model. They aren't represented as
implemented behavior. Replaying the recorded current outcomes uses
the cited head's shell function and built command. The Go package and
command tests passed with go test ./internal/checks ./internal/agentrepo.
The future implementation must pass the model and caller contract.
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Findings for #720: conservative check evaluation contract

Original recording head: f130498bd7383fe02f8b9d21965f555866abd98c.
Fresh replay head: 44be762a34eec4a832e1ed608c77ebf188357719.

The corpus and original caller recordings extend the existing
research comment.
The judge corpus, hand fixtures and unattended evaluator fixtures were
independently replayed on the fresh head.

The recommended contract validates every entry, keeps every status
context, and takes the worse check-run verdict selected by either
text order or instant order. A caller accepts success only when both
the process status and a complete, consistent result say success.

Evidence and notation

Source locations in the inherited findings refer to the original recording head.
Caller supplements below use the fresh replay head. Let M mean
packages/agents/source/skills/merge-pr/shared/scripts/ and L mean
packages/agents/source/lib/. Paths written after those prefixes
concatenate with them. This avoids ambiguous generated-copy
references.

M/pull-request.sh:69-110: shell judge and superseded count.
internal/checks/checks.go:75-177: parse and evaluation precedence.
internal/checks/checks.go:180-206,257-286: verdict and selection.
internal/agentrepo/command.go:55-101,107-119: command protocol.
M/preflight.sh:40-120: preparation, review and check reporting.
M/squash-merge.sh:63-218: preparation and check refusal.
M/merge-agent.sh:19-30,118-180: unattended context and gates.
L/lib.sh:140-183,679-714,830-885: loader, stamps and API errors.
L/agent.sh:56-65: interrupted operations and detached HEAD.

The evidence block at the end is a lossless recording, not a link to
local files. Its metadata holds every hand-written input, both judges'
exact stdout, stderr and numeric status, full caller recordings,
unattended recordings, benchmark samples and the sweep manifest.
The binary matrix records both exact outputs for every sweep key.
The decoder reconstructs all 549808 input/output records. No fixture
requires access to the research machine.

P means attended preflight, S means squash with an explicit
operator confirmation. U means unattended preflight. G means
squash without explicit confirmation, including unattended squash.
These are fixture invocation modes, not confirmation grants.

For AC1 and AC6, earlier prerequisites and later gates pass, with all
GitHub writes stubbed. P reaches its check-reporting section and
continues to process exit 0. S either refuses there or reaches a
successful stubbed merge and process exit 0. The check section alone
doesn't establish that a real merge is authorized. AC7 overrides this
boundary whenever a real earlier stop applies.

The following executable definitions specify complete caller streams.
m is the decoded evidence metadata. Concatenation adds no implicit
whitespace. Every newline is explicit. Empty stderr means "".
pfx includes the checks heading. sfx ends just before judging.
ptail and stail include all otherwise successful later output.
Those tails come from the recorded isolated fixture, not a proposed
change to production diagnostics.

r = {(x['id'], x['caller']): x for x in m['callers']}
p = r['success', 'preflight']['stdout']
s = r['success', 'squash-merge']['stdout']
pfx, rest = p.split('SUCCESS        ci / test\n', 1)
ptail = rest.split('\nsummary: all green.\n', 1)[1]
sfx, stail = s.split('open, ready, mergeable, and green\n', 1)
summary = {
    0: 'all green.\n',
    1: ('FAILING. Do not merge. Hand this to the '
        'fix-pr skill instead.\n'),
    2: ('still running. Wait through the watch-pr '
        'skill before merging.\n'),
}
err = {
    1: ('squash-merge: #7 has a failing check. '
        'Hand it to the fix-pr skill.\n'),
    2: ('squash-merge: #7 has a check still running. '
        'Wait through the watch-pr skill.\n'),
    4: ('squash-merge: #7 has no checks reported; '
        'the result is inconclusive.\n'),
    5: 'squash-merge: the checks of #7 could not be read.\n',
}
def line(state, name):
    return f'{state:<14} {name}\n'
def P(code, rows=(), superseded=0):
    if code in (4, 5):
        body = ('no checks reported\n' if code == 4 else
                'the check rollup could not be read\n')
    else:
        body = ''.join(line(state, name) for state, name
                       in sorted(rows, key=lambda x: x[1]))
        if superseded:
            body += (f'({superseded} earlier run(s) left out: '
                     'a later run of the same check replaced '
                     'each)\n')
        body += '\nsummary: ' + summary[code]
    return (0, pfx + body + ptail, '')
def S(code, rows=()):
    if code == 0:
        return (0, sfx + 'open, ready, mergeable, and green\n'
                + stail, '')
    body = ''.join(line(state, name) for state, name in rows)
    return (1, sfx, body + err[code])
F = [('FAILURE', 'ci / test')]
Q = [('QUEUED', 'ci / test')]
C = [('ERROR', 'test'), ('SUCCESS', 'test')]

Each tuple is (caller process status, stdout, stderr).
P(5) is an unreadable report and continuation, not green.
S(5) refuses with no merge call. P(4) reports empty and continues.
S(4) refuses. P(1,...) and P(2,...) report and continue.
S(1,...) and S(2,...) print the relevant rows and refuse.
The failure diagnostic deliberately drops “judged by its latest run”:
that phrase would misdescribe the conservative two-order decision.

AC1, AC2 and AC5: input contract and recorded comparison

Table keys are exact id values in m['hand']. That record holds the
saved input bytes, command arguments, shell stdout/stderr/status and
current Go stdout/stderr/status. Both commands ran on each input.
All shell stderr streams were empty. Go stderr was empty except for
the required-check hints explicitly recorded under AC8.

The last column marks deliberate changes to the attended shell path.
yes includes stricter interpretation or changed diagnostics.
parser identifies an earlier stop, not an adapter result.

Key	Shell	Go	P result	S result	Change
h00	3	5	parser/5	parser/5	no
h01	4	5	P(5)	S(5)	yes
h02	4	5	P(5)	S(5)	yes
h03	4	4	P(4)	S(4)	no
h04	0	0	P(0,OK)	S(0)	no
h05	1	1	P(1,F)	S(1,F)	message
h06	2	2	P(2,Q)	S(2,Q)	no
h07	0	1	P(5)	S(5)	yes
h08	0	1	P(5)	S(5)	yes
h09	0	1	P(5)	S(5)	yes
h10	0	1	P(5)	S(5)	yes
h11	0	1	P(5)	S(5)	yes
h12	0	1	P(5)	S(5)	yes
h13	0	1	P(5)	S(5)	yes
h14	0	5	P(5)	S(5)	yes
h15	0	5	P(5)	S(5)	yes
h16	0	1	P(1,F,1)	S(1,F)	yes
h17	1	1	P(1,F,1)	S(1,F)	yes
h18	2	2	P(2,Q,1)	S(2,Q)	no
h19	1	0	P(1,C)	S(1,C[:1])	message
h20	1	0	P(1,F,1)	S(1,F)	yes
h21	0	5	P(5)	S(5)	yes
h22	0	0	P(5)	S(5)	yes
h23	0	0	P(5)	S(5)	yes
h24	0	0	P(5)	S(5)	yes

OK = [("SUCCESS", "ci / test")]. Input keys:

h00: malformed.
h01: missing.
h02: null.
h03: empty.
h04: success.
h05: failure.
h06: pending.
h07: unknown_status.
h08: unknown_conclusion.
h09: unknown_context.
h10: lower_status.
h11: lower_conclusion.
h12: lower_context.
h13: missing_status.
h14: empty_name.
h15: unknown_type.
h16: tie.
h17: tie_reverse.
h18: unstarted.
h19: repeated_context.
h20: offset.
h21: invalid_time.
h22: superseded_status.
h23: superseded_conclusion.
h24: superseded_context.

For h00, both callers exit 5 before evaluating checks. The exact
stdout and stderr are in m['callers'] for malformed, with raw {
as the response. Stderr is the parser diagnostic below. The preflight
first stops at its state formatter. Squash first stops in mp_field.
This result stays unchanged in the chosen contract.

jq: parse error: Unfinished JSON term at EOF at line 1, column 1

Missing and null rollups deliberately change from empty to unreadable.
Their check-section reports are P(5) and S(5), not the old empty
reports. An actual empty array remains P(4) and S(4).

Unknown status, conclusion and context state all mean unreadable,
even on superseded entries. Validate every raw entry before selection,
including a conclusion that a non-completed status would otherwise
hide. Lowercase spellings and a missing CheckRun status are also
unreadable. This preserves Go's refusal and tightens shell acceptance.
Don't silently uppercase them and loosen the unattended path.

Accept the six uppercase CheckRun statuses and nine uppercase
conclusions in AC3. An absent, null or empty conclusion is allowed and
means pending when status is COMPLETED. An absent, null or empty
context state means PENDING. Recognized pending statuses still ignore
a recognized conclusion for verdict purposes. Null/empty CheckRun
status, other field types, empty names and unknown entry types are
unreadable. A context accepts only the five uppercase context states,
plus the missing-value convention above. These validation rules also
cover entries excluded by time ordering.

AC3: finite sweep, coverage and results

The domain manifest was written before execution. Generation order is
CheckRun then StatusContext, first-entry index, second-entry index,
then timestamp-pair index. CheckRun indices enumerate status as the
outer axis and conclusion as the inner axis. Omission is absence of
the JSON property. It isn't null or an empty string.

Status: COMPLETED, IN_PROGRESS, PENDING, QUEUED, REQUESTED, WAITING,
then those six lowercase spellings, UNKNOWN_STATE, then omission.
Conclusion: ACTION_REQUIRED, CANCELLED, FAILURE, NEUTRAL, SKIPPED,
STALE, STARTUP_FAILURE, SUCCESS, TIMED_OUT, then those nine
lowercase
spellings, UNKNOWN_STATE, then omission.
Context state: ERROR, EXPECTED, FAILURE, PENDING, SUCCESS, then
those five lowercase spellings, UNKNOWN_STATE, then omission.
Seven timestamp pairs, in the issue's order: equal, earlier, later,
offset-reversed, invalid first, invalid second, both invalid.
m['manifest']['times'] records every literal timestamp.
Fixed head head, CheckRun workflow ci, name test, or context
test. No required names or expected-head argument. No mixed types
or extra fields. JSON array order is first then second.

Expected total, fixed before running:
7 * ((14 * 20)**2 + 12**2) = 549808.
That is 548800 CheckRun pairs and 1008 context pairs.

The shell batch ran the exact jq program extracted from
mp_judge_checks. Its TSV rows went through the same failure-first,
pending-second status reduction. The Go batch called the actual
agentrepo.Run command handler for each input, with fresh buffers and
no constraints. Batching avoids over a million process launches.
The unmodified shell function and built CLI separately matched both
streams and statuses on 103 fresh-process samples, plus every hand
fixture. This distinguishes command recordings from an independent
reimplementation of either judge.

The coverage check independently reconstructed all mandated field
products and all expected keys, compared each entry's complete fields
with its index and checked timestamp values, order and fixed fields.
It found exactly 549808 unique keys, zero duplicates, zero missing
keys and zero unexpected keys. All 280 crossed CheckRun entries and
all 12 context entries participate in both positions.

Chosen results: 382 success, 2168 failing, 11994 pending and 535264
unreadable. Zero shell failures or pending results become success.
Current Go yields success for 120 shell-failing and 624 shell-pending
cases. The chosen rule prevents all 744. The full three-way counts
are in m['summary']['cross'], whose rows are
[shell status, current Go status, chosen status, count].

Fresh-head verification reconstructed all mandated keys, rejected duplicates,
and matched all 549808 shell triples, Go triples and chosen outcomes.
All 30 hand fixtures and 101 fresh-process sweep samples also matched.
Zero shell-failing or pending cases became green.

AC4: time ordering rule

First reject every invalid timestamp, even on a superseded entry.
For each CheckRun group (workflowName, name), compute both the
shell text-order maximum and the parsed-instant maximum. Keep every
entry tied at either maximum. Evaluate those candidates and choose
failure over pending over success, with deterministic state/name
ordering for equal verdicts. The selected display row represents
that worst candidate. Count one displayed row per CheckRun group.

Use the shell's exact text key: missing/empty start or a start that
begins 0001 maps to 9999-12-31T23:59:59Z. For instant order,
missing/empty start or the exact zero timestamp outranks started
runs. Other timestamps must parse as RFC3339Nano. This preserves the
shell's unusual zero-prefix case without letting it choose green
over a conflicting instant-order failure.

For any ordering disagreement, the candidate union includes both
maxima. Its worst verdict cannot be green if either maximum fails or
is unfinished. Ties retain all candidates, so array order cannot
weaken the result. The argument applies to groups larger than pairs.
For h20's offset example, the failure wins and gives P(1,F,1) and
S(1,F). The sweep found zero unsafe offset-reversed cases.

Keep every StatusContext entry, in its existing array order for
otherwise equal display names. Don't discard an older failing or
pending context. Validate createdAt, but don't use it to discard
contexts. This preserves the shell contract for repeated contexts.

Retain the old superseded count: number of CheckRuns minus the number
of distinct (workflowName, name) groups. Retaining the "earlier/later"
diagnostic for ties is intentional: its bytes stay exactly as in P.
The number counts collapsed display rows; it does not assert that
every omitted timestamp is strictly earlier.

AC6: evaluator process and result contract

Keep numeric outcomes 0 through 6 and their precedence: unreadable,
head moved, empty, missing required, failed, pending, success.
Require an additive schema: 1, checks array and nonnegative
integer
superseded on every result. Each check has string state and name
and a verdict of ok, pending or fail. Preserve existing
outcome, sorted unique names, skipped and error semantics.
For unreadable input require a nonempty error and empty checks.
For missing required checks, keep the missing names in names.

Validate the whole stdout as exactly one JSON object with no trailing
non-whitespace bytes. Verify types, outcome/status agreement, listing
verdicts, outcome precedence and consistency of summary names and
counts. For success require a nonempty all-ok listing, no error or
non-success names, and process exit 0. This rejects empty success and
partial success-looking output. Don't infer success from an exit
status alone or from a JSON fragment.

Let R(n) mean the complete, internally consistent schema-1 result
for outcome n, ending with one newline. Specific result examples are
specified below in executable form. Unless stated otherwise evaluator
stderr is empty. Caller stdout/stderr/status use the full P/S tuples,
not the evaluator status. All these rows occur at the check section
unless the early-boundary column says otherwise.

Eval exit	Eval stdout	Category	P	S	Boundary
0	R(0)	green	P(0,OK)	S(0)	checks
1	R(1)	failing	P(1,F)	S(1,F)	checks
2	R(2)	running	P(2,Q)	S(2,Q)	checks
3	R(3)	failing	P(1,MR+OR)	S(1,MR)	checks
4	R(4)	empty	P(4)	S(4)	checks
5	R(5)	unreadable	P(5)	S(5)	checks
6	R(6)	unreadable	P(5)	S(5)	checks
7, 127, 137	any	unreadable	P(5)	S(5)	checks
other nonzero	any	unreadable	P(5)	S(5)	checks
absent	absent	unreadable	P(5)	S(5)	see AC7
0	`garbage\n`	unreadable	P(5)	S(5)	checks
0	`{}\n`	unreadable	P(5)	S(5)	checks
0	`[]\n`	unreadable	P(5)	S(5)	checks
0	help text	unreadable	P(5)	S(5)	checks
0	`{"outcome":0}\n`	unreadable	P(5)	S(5)	checks
0	R(0) + garbage	unreadable	P(5)	S(5)	checks
1	R(0)	unreadable	P(5)	S(5)	checks
2	R(0)	unreadable	P(5)	S(5)	checks
other nonzero	R(0)	unreadable	P(5)	S(5)	checks
0	missing result field	unreadable	P(5)	S(5)	checks

The absent-command row describes P/S with the loader allowing absence.
U/G instead stop at ma_context, as AC7 specifies. Loader version
refusals precede both. Mapped 1 or 2 with R(0) remains unreadable,
not a valid failed/pending result. Evaluator stderr is captured and
must never substitute for a result. Keep recognized skipped-check
hints as described in AC8. For protocol/command failures, report the
stable unreadable message rather than copying arbitrary subprocess
stdout into the caller report.

MR = [('MISSING_REQUIRED', 'test')]
OR = [('SUCCESS', 'ci / other')]
def R(n):
    rows = {0: [('ok', 'SUCCESS', 'ci / test')],
            1: [('fail', 'FAILURE', 'ci / test')],
            2: [('pending', 'QUEUED', 'ci / test')],
            3: [('ok', 'SUCCESS', 'ci / other')]}.get(n, [])
    value = dict(schema=1, outcome=n,
                 checks=[dict(verdict=v, state=s, name=k)
                         for v, s, k in rows], superseded=0)
    if n in (1, 2, 3):
        value['names'] = ['test']
    if n == 5:
        value['error'] = 'invalid rollup'
    return json.dumps(value, separators=(',', ':')) + '\n'

For outcome 3, append MISSING_REQUIRED diagnostic rows for missing
names to the display listing. The P/S row above includes both the
reported successful check and
the missing-check diagnostic in preflight. Squash prints only the
missing-check diagnostic before refusing. Head-moved status 6 remains
a
separate evaluator outcome and refuses as unreadable in the attended
adapter. U preserves the earlier gate's status 6.

AC7: earlier stops, modes and exact stream rules

No earlier gate is removed or relaxed. Full process recordings for
malformed JSON and review/PR-state cases are in m['callers'] and
m['early']. They include both streams, numeric status and
reachability.
The catalogue is supplemented by fresh-head caller executions for stamp,
rate-limit, diff, policy and gate streams. Exact bytes are embedded below;
external commands use concrete failure fixtures.

For the source-traced rows, use the exact source prefixes below to
complete stdout. Messages in the diagnostic catalogue end in one
newline. An ordinary dt_die adds no prefix. Squash's die adds
squash-merge: . Both exit 1. Empty stdout/stderr is explicitly "".
The table's tuple notation is (status, stdout, stderr, reached).
pass means the other caller proceeds to P(0,OK) or S(0), with its
full recorded stdout, empty stderr, status 0 and reached=yes.

Pgh = '\n== gh ==\nok\n'
Ptarget = (Pgh + '\n== target ==\nlocal branch: topic\n'
           'pull request: #7\n')
Pstate = Ptarget + '\n== state ==\n'
Pbefore = pfx.removesuffix('\n== checks ==\n')
Sd = '\n--- checking SQUASH_AGENTMSG\n'
Ss = (Sd + 'drafted for #7\n'
      '\n--- checking the review stamp\n')
Sr = Ss + 'reviewed at head head, onto main at base\n'
Sa = Sr + '\n--- checking the gh account\n'
# Concrete failing subprocess fixture: exit 2, empty stdout.
E = 'fixture failure\n'
# U's EXIT trap, with no mail transport configured:
def report(n, status):
    return (f'REPORT (blocker): Merge workflow for #{n} held '
            f'with status {status}. Inspect the preserved '
            'draft, review, and merge result before retrying.\n')

With U enabled, append report(number,status) to stdout on every
nonzero exit after pull-request.sh installed its trap. Before
preflight assigns its number, number is unknown. During loader
failure, the trap isn't installed, so no report appears. G appends it
only when DEVTOOLS_UNATTENDED=1. A configured reporting transport
keeps its existing output and failure behavior instead. This is
orthogonal to the judge result and doesn't change the exit status.

The table uses ordinary P/S tuples and explicitly states mode changes.
All rows preserve their current stop and message. Source suffixes
refer to the catalogue immediately following the table.

First stop	P tuple	S tuple	Modes
loader version read	(1,"",D1,no)	same	all
loader version floor	(1,"",D2,no)	same	all
git root read	(128,"",E,no)	(128,"",E,no)	all
worktree readiness	(1,"",D3,no)	pass	P/U
argument parser	pass	(1,"",D4,no)	S/G
missing commit-lint	pass	(1,"",D5,no)	S/G
missing gh	(0,PG,"",no)	(1,"",D6,no)	P/S
missing jq	(0,PJ,"",no)	(1,"",D7,no)	P/S
viewer auth error	(2,Pgh,E,no)	(2,"",E,no)	all
viewer empty	(1,Pgh,D8,no)	(1,"",D8,no)	all
rate limit	(76,Pgh,RL,no)	(76,"",RL,no)	all
PR lookup error	(2,Pgh,E,no)	(2,"",E,no)	all
no target	(0,PN,"",no)	(1,"",D9,no)	all
missing draft	pass	(1,Sd,D10,no)	S/G
wrong draft number	pass	(1,Sd,D11,no)	S/G
snapshot read error	(2,Ptarget,E,no)	(2,Ss,E,no)	all
malformed snapshot	saved h00	saved h00	P/S
malformed fields	(5,Pstate,E,no)	(5,Ss,E,no)	P/S
base read error	pass	(2,Ss,E,no)	S/G
missing stamp	pass	(1,Ss,D12,no)	S/G
stale stamp	pass	(1,Ss,D13,no)	S/G
expected head mismatch	pass	(1,Ss,D14,no)	S/G
account mismatch	pass	(1,Sa,D15,no)	S/G
unknown mergeability	saved early	saved early	all
closed PR	saved early	saved early	all
draft PR	saved early	saved early	all
conflicting PR	saved early	saved early	all
required review	saved early/75	saved early/75	all
changes requested	saved early/75	saved early/75	all
unknown review enum	saved early/1	saved early/1	all
not CLEAN	saved early	saved early	all
unknown fork flag	saved early	saved early	all

PG is "\n== gh ==\ngh is not installed\n" followed by
"Stop here and tell the operator.\n". PJ replaces gh with jq.
PN is Pgh followed by the no-target block in preflight lines 66-74,
with branch topic and an empty open-PR list. Its exact bytes are:

PN = (Pgh + '\n== target ==\nlocal branch: topic\n'
      'pull request: NONE for this branch, and none given.\n'
      'Pass the number as the skill argument, or open one '
      'with the pr skill.\n\nopen pull requests here:\n')

Diagnostic catalogue, using quoted Python strings for exact bytes.
The fixture executable is agentrepo, here is harness, and the
example old version is v0.15.0. These fixture values are not text to
hardcode in #723.

D1 = E + 'Could not read agentrepo version: agentrepo\n'
D2 = 'agentrepo found v0.15.0; required v0.16.0 or newer\n'
D3 = ('worktree setup: mise config is untrusted. Run:\n\n'
      '  mise trust\n')
D4 = 'squash-merge: usage: squash-merge.sh [--confirmed] [number]\n'
D5 = ('squash-merge: commit-lint is not installed; add '
      '"go:github.com/tbhb-dev/devtools/cmd/commit-lint" = '
      '"<devtools-version>" to mise.toml [tools]\n')
D6 = 'gh is not installed\n'
D7 = 'jq is not installed\n'
D8 = 'GitHub returned no viewer login\n'
D9 = ('squash-merge: no open pull request for topic. '
      'Pass a number explicitly.\n')
D10 = ('squash-merge: no squash message drafted. '
       'Run squash-message.sh first.\n')
D11 = 'squash-merge: SQUASH_AGENTMSG was drafted for #8, not #7.\n'
D12 = ('squash-merge: no independent review has cleared this '
       'message. Run it, resolve what\nit returns, then retry:'
       '\n\n  bash harness/review.sh 7\n')
D13tail = ('squash-merge: the reviewed message and pull request '
           'are no longer the ones about to\nmerge. Review '
           'again against the current draft and branch:'
           '\n\n  bash harness/review.sh 7\n')
D14 = 'squash-merge: Expected head differs from the reviewed head\n'
D15 = ('squash-merge: gh is signed in as agent, not the '
       'configured other. Nothing was merged.\n')

D1 includes the failed version command's exact stderr, empty when
that command printed none. D3 also has source-defined alternatives,
both P/U process exit 1
before the check section and S/G pass: worktree setup: Vale styles are missing. Run:\n\n mise run devtools:worktree-setup\n, or
worktree setup: could not check mise trust:\n plus captured output
and newline. Other Vale configuration failures use
worktree setup: Vale could not load the repository configuration:\n
plus captured output and newline.

D4 also covers malformed expected-head syntax with
squash-merge: Expected head must be a full commit SHA\n.
D13 is the exact dt_require_stamp diagnostic followed by D13tail:
changed draft, changed bound field, changed repository configuration,
or changed trusted policy authority. Each ends in newline. The
field-change template is The {field} changed after the review (reviewed {old}, now {new}). Run the review script again. The literal prefixes and full
caller streams are captured as stamp-draft, stamp-field,
stamp-config and stamp-policy below. Unrelated trusted-ref movement
with unchanged policy content/status/authority no longer refuses.
Repository-config comparison applies only in unattended mode. Failures reading the stamp
or its digest keep the underlying stderr before D13tail and exit 1.

For external reader/parser errors, E means the exact captured stderr,
not a replacement message. Exit 2 above is the concrete external
failure fixture, and 128 is the Git failure fixture. Other subprocess
statuses propagate numerically unchanged. Parser status is jq's 5,
with its actual diagnostic and any partial stdout retained. An
arbitrary external program can emit arbitrary bytes, so this family
is parameterized by its captured bytes and numeric status. The h00
record gives a concrete complete parser example without that freedom.

RL is the complete stderr in rate-secondary, rate-graphql or
rate-rest below, including the original fixture error. All statuses
are 76; P stdout is Pgh and S stdout is explicitly empty. For U/G,
append the trap report to stdout. Later failing reads use their prefix
at that read.

U/G-specific earlier helpers follow. These modes execute ma_context.
P and explicitly confirmed S skip it and instead take their normal
report/merge paths, unless a common earlier gate above already stops.
For U, ma_gates runs before the check section. For G, ma_gates first
runs AFTER the check section and lint, at squash line 244. Don't move
its refusal earlier while replacing only the judge.

Context stop	U status/stdout	G status/stdout	stderr
operation marker	1/report(unknown,1)	1/report(7,1)	C1
detached HEAD	1/report(unknown,1)	1/report(7,1)	C2
unborn HEAD	1/report(unknown,1)	1/report(7,1)	C3
dirty worktree	1/report(unknown,1)	1/report(7,1)	C4
no configured login	1/report(unknown,1)	1/report(7,1)	C5
missing uv	1/report(unknown,1)	1/report(7,1)	C6
absent agentrepo	1/report(unknown,1)	1/report(7,1)	C7
unusable agentrepo	1/report(unknown,1)	1/report(7,1)	C8
unattended --confirmed	U normal	1/report(7,1)	C9

The --confirmed flag belongs to squash. The paired U invocation is
ordinary preflight, which reports and exits 0 with its normal stdout
and empty stderr when its gates pass. All other context stops have
reached=no for U/G. Plain G without the
unattended flag has empty stdout instead of report(...). Underlying
Git/config read failures propagate their exit status and stderr,
with the same empty prefix and trap rule. C6 includes Bash's
command -v uv output only if that builtin emits one in the selected
shell. Bash in this environment emits none for the lookup.

C1 = ('HELD: MERGE_HEAD is in progress; finish or abort it '
      'first.\n')
# Substitute CHERRY_PICK_HEAD, REVERT_HEAD, rebase-merge,
# or rebase-apply for the other marker cases.
C2 = 'HELD: HEAD is detached.\n'
C3 = ('Create the initial commit through the interactive '
      'commit skill first.\n')
C4 = ('Commit the work first; unattended merging requires '
      'a clean tree.\n')
C5 = 'gh_login must be configured on the machine\n'
C6 = 'uv is required for the YAML policy reader\n'
C7 = 'agentrepo is required for check judgement\n'
C8 = 'a working agentrepo is required for check judgement\n'
C9 = ('squash-merge: Unattended workflows refuse --confirmed; '
      'use pre-approval or delegation.\n')

agent-merge-pr/shared/scripts/merge.sh sets unattended mode and
forwards to squash. With two numeric arguments it converts the second
into --expected-head. Its --confirmed refusal therefore comes from
the shared squash caller after ma_context, not from the wrapper.

Before U/G reach their initial response parsing, mp_view also runs
ma_reviews. A malformed raw response stops at its mp_field with
status 5, stdout Ptarget for U or Ss for G, and the exact jq stderr.
A failed review-pages read propagates its numeric status and stderr.
A mismatched review-page head stops with status 5 and the jq error
PR head changed while reading review authors; the check section
isn't reached. P/S don't take this extra reader and continue normally
on a valid snapshot with the same paginated-review fault.

U's pre-check gates have the following first-stop catalogue.
For these isolated faults, P and explicitly confirmed S continue with
P(0,OK) and S(0): status 0, their full defined stdout and explicitly
empty stderr. U stops before checks; G stops after checks and lint.
The supplement's gate function gives exact full U/G tuples.
Ordinary failures have no earlier gate stdout. HISTORICAL readiness
failures add exactly {"outcome":0}\n; owner overrides add their
recorded line before that JSON. CHOSEN FUTURE readiness failures add
R(0) from AC6 instead. Captured rows below specify historical streams.

First U gate stop	Status	Exact diagnostic/recording
PR state/draft/readiness	1	A1
actual login mismatch	1	A2
reviewer login [bot] suffix	1	A3
no required reviewer	1	A4
invalid reviewer type	1	A5
exact-head approval absent	1	A6
empty owner-review label	1	A7
owner-review label present	1	A8
pinned diff acquisition	1	A9
policy inspection/evaluation	1	A10
owner approval absent	1	A11
requirements unavailable	1	A12
evaluator non-success	1..6	AC6/AC8 JSON and hint
readiness not executable	1	A13
readiness exits nonzero	2	exact E fixture

A ma_gates review refusal is status 1. It isn't the status-75
mp_require_reviews refusal. That distinction stays unchanged.
All ma_context checks run again inside ma_gates before this list.

A1 = 'PR is not open, ready, MERGEABLE, and CLEAN\n'
A2 = 'gh is signed in as other, not agent\n'
A3 = ('required_reviewer must use the GraphQL login reviewer, '
      'without [bot]\n')
A4 = 'required_reviewer must be configured on the machine\n'
A5 = 'required_reviewer_type must be Bot or User when set\n'
A6 = 'WAITING_REVIEW: reviewer must approve the current head\n'
A7 = 'owner_review_label must not be empty\n'
A8 = 'needs-owner: the PR has the owner-review label\n'
A9 = 'Could not fetch the pinned base and head.\n'
A11 = ('needs-owner: golden files or loosened gates require '
       'merge_owner approval on the current head\n')
A12 = 'No required checks are known for the base\n'
A13 = 'merge_ready_cmd must name an executable\n'

A9's five variants are diff-url, diff-history, diff-fetch,
diff-diff and diff-names below, including the fixture Git stderr.
A10's concrete validation cases cover empty/invalid patterns,
repository paths, APM lock scalars, reviews, findings and applied rules.
policy-required distinguishes the reader's missing-requirements
message from A12's later empty-output refusal. policy-process
records process exit 2 and E. owner-deny includes the JSON finding
before A11; owner-override includes the finding on stderr and
override on stdout. No raw policy value enters the judge diagnostic.

On the fresh head ma_gates uses agentrepo merge approval.
No qualifying approval prints nothing before A6/A11.
approval-invalid records the missing-reviews command diagnostic
before A6; the same command error precedes A11 for the owner check.
Caller status remains 1, U before checks and G after checks/lint.

The malformed/required-review recordings and source trace establish
that preflight's CHECK SECTION reports failing and running checks
without refusing. They do not assert that preflight never refuses.
The existing review regression is at
tests/skills/merge-pr/scripts/merge-workflow.bats:385.

Exact caller supplements

These recordings execute fresh-head preflight and squash caller bodies with
their real stamp, rate classifier, diff and gate helpers. Prerequisite
readers pass; external failures and merge writes are stubbed, and filesystem
inputs use memory descriptors. Passing squash stops at the write boundary.
The real Python policy validator and exception handler supply validation
diagnostics; the lock-scalar case isolates that reader.
fixture failure\n is the injected process stderr, with explicitly empty
process stdout and exit 2. The snapshot has title test, author author,
URL https://example.invalid/pr/7, empty commits/files/body,
one successful ci / test CheckRun, and the values shown in stdout.

Decode this zlib/base64 block as ac7_b64 with the Python below. Rows map each mode to
[process_status, stdout_index, stderr_index], indexing streams.
These are HISTORICAL current-shell recordings, with the descriptor
label normalized to fd 3; "" is explicitly empty.
Every nonzero row refuses. U rows have reached=no; gate G rows have
reached=yes. Stamp/rate/review-page rows have reached=no.
For stamp faults P/U pass; S uses G's tuple without its trap report,
except stamp-config, which passes in S. For gate/policy/diff faults
P/S pass as defined above. Plain G removes the trap report only.
CHOSEN FUTURE stops and mode differences stay the same; evaluator
stdout gains the AC6 schema-1 fields. The reconstruction below defines
the future streams separately from the historical recordings.

eJztWVtv2zYU/iuE97AOkHPxJU489MHbjCxAL1mSrg9VYdASbXGRSJWk4hhF/vvOISlbdJ02zdIh
3fKQALS+c3jOd24U9bGljWK00K3hu1Ys2u02STKWXHIxJ+d/vBmd/z4ZHY9fXbw8P45FqujMsJTM
pCI/DGKxgTcZI4pdcbYg2tCijMXZ+PT12QV5Ns1lcsnUT0Pykqk5IwupLme5XHhNJGN5ShbcZCho
Kk32d8iJ0CVLjNVaKqaZuoKtrQmR3yYiVKSksCoBUeWGTBmoxJVRS7BpJxatqLXhCFhMxRyUoTeK
cKO9vh1yVonAjUTx0hA6p1yAKv2hojpr2w2HDRxDI7SmYAYaVFZ5Dk8+VEyDLFgjJMkl7KisjBRM
EzqVFfgmY2G1wc5uR7uV+6+d70mlFBPGeW71TxUVSTZE/gmZgkUkowqU6l3vBvwysI5fgHzGaLrh
ccPDZysPZJ5GYOnCCvz0n6PijJVScyPVkiRSzPi8UtRwKTaouSURvmO3rdJS5jyxnhvQFNVlBtVH
K5NJxc3ybkR872kQi+fPyTwjz5/HQl7anyAPGMl5wc2QHHPzezUlmgFTKYVcWT/82bunGVjBC0a4
Fj8acgklI8hMycI+n1GeVwqsf0s52GZIziiYDR6SgosKlPkOhWDBrk1NC/Qy7bf/UZNUJsBIesU1
+xnJmlbpHPYFJNgE3bLUO+SVxODJAtI4cQ2PLEBHQVMGkTlWtMz+eDHc6ov1e7vrpeIFgr18Qygi
XMwYUJ+u3U1kUWAA8A/X8BhyKueCfSVdr0XiOLFbQfg0KanWLMVeT51yT4KdEzQw8IUV+gjQAtIE
Z5HdeGTIjeMMOhu7gmzbRfTE7nFnAkenJw0WCLtOGEv9FJzxawP235ZHNZln4/OLf4vJizVTCUXg
lDkKFcs5neZLJ5MxBRW7wCxFR1BFJisFjxdcpDAGrCEgaJqkYZLatMQ+KkW+/BKL8+zuBG7Wpl0b
qtAV/A0OETSvax06RcmTWDRby9CfSkAMGxyzUoabHDwlBgDQsPB3WL0+Hb+KhWt+Q98EAYs6IThG
EkykWNjhKXk6tFPRtyYgEVS8HJ8dj0e/vBiTXfLri/EI1PkembKEaxgt6Pnp2es/x7/FApgFHcaU
eri7y67hdJSzHS6uaM7T3VLtDh7NYelXWYF+IaF1Ad/GVWXQwNVqlO74uD3wsXE1QyD5bATwXwT5
5uOCv0OXZ1tVQQLRJJGVwGjzuQA1KAF5O2f424YIGiRLJlyfgUaxCrEjbK4YE7WUqoSo98GC5aZd
6DmZA0360QTQl1TdEWKxjuiMmSSrrYcMhOiWXCBFSKfV7/jmGsLLmf4KjdsUfVF8WvE83WpQymez
L4rn3B8JGkLQUvLa8P95R/kYt+BMBcyyuDXcu3k0GWoRE1tvk6QAiQrCKGiBiQPjgSWVQUaemkuj
uTzWWD6V2AOFpfNth8BTKX2DUnrYmL2L67DtmGsTt6K45d7JgRMyl3nKhB1ucet9DOOOpboNrwAM
SqHx0L7U51Jqhoxaj22lcdjO9V0rQ2hZKgmZDNEKXqpdyTwVtmNJwkuj4ils0GSNpd6A/0QmPbWK
21vFd50FDuGv/iqBZLhigVd23xJSx5q2l3L+RsHfBrhoPDWCRxLMp4r8hy/ZDx2OBymKe3vTf1Bv
/vow9BeOzyDqs5R0h3tgy+lZ+PVokcHMsCF0V6zuft4Wqf42SfpI+Lm1k566n0pqDFPCj9YJLDPt
32qlvYBlRWmWn1cFtcRUwQU1titTRRP8DoO3vhATe+OGn6v2Pq9FVTleFE9SrpwBsDmHEk2ZZ7e+
uvu8mhPXAj1N+m7gGReYGHdEwxDNObhqLb6byOj0JV5KXuLRhWgoMKqsIOQp15bqoMmsRoELs58G
gIe6BXzQtySwoyKrw7cwvNH3Q3Lie6WntNKOzPoLRC7noKTGRDYP8VvWu6k07z+naMpW3yLxA6w7
iBc0ybhg2+UmZlmylfAv6LEibyC1oTbhTAeRt3JvRycXJ6+OJ2fjP0/Gb4ck3NYfX7Yf++0Jx+83
yemU5bckcvACgqqwWwCT9oMe/tr2pWyVWJGvO3cwCHqFxwWXF/ULy5BoQUudgT3WMPymacdVnayg
U1Gw8R+Q8OW6cGX2VWC0NLVFfMd0r+UUm9nbsbuKBRRbLfcSLKEFFQx60L32vZfQ3dtAyspcLkEO
e+09ugfeHIPYe6gxudCt4ceWnThtOx5wedwavtuP9qL99zeRfzbjMGWazzrrZ66Omw+764fOqObD
Hj7Eb2Lt1edZfHwKjwcHUT86AMPO3WIAixo8x47zIQ+gh03o4QoK48wEuKMm7ghxeFvehhMwwt5Y
w/bB333AOTP3O7iqge6bwDIEdwNwdwW2nwVCaC+A9lZQ/Bci+wGyv0LaGgihBwHUEYXNv72+RW4I
DKL9w7XAEa5WAu7cUIM7UQfCW1MBqw6ubnx7bMOQD4noBER0umto/e7Y0NyLOv215gNcAdwlSbsu
iFD9IFA/aOChiyVMNw0/DAw/8ob75tjGHleD+1EX8rQGw6qDq4Zud6wJTOkGnna7n8LBhTm7DoWC
4Hd7oVCYKN0g/N2QG+tDCA9SoHvQgNfnkRAfUNltUmn7bQg+DMCHDTB2khB7FGBthY32A0hvrwmB
FUI6ISQov56NxqgbQjoBxDahUS+EBEHq2SCN+iEkCEnPhmR0EEKCSPRsJEaDEBKw37Psjw5DSEB4
b+B42fA64Llnea7nfdu/jocCAdm9o80oMhPA+wHx/b1P4e31cA4lg3j097dIwnjekAkC1O9sqW4r
HAoFIet3twmtJ3IoGoSy39sm6mZyKBaEt7+tCTmxUCqIeP9goyQ+red+EP/+YFMAJ3EoECQDrG5u
bv4GBqmEJw==

import base64, hashlib, json, zlib
ac7_raw = zlib.decompress(base64.b64decode(ac7_b64))
assert hashlib.sha256(ac7_raw).hexdigest() == (
    'cdd12dd359577feda3c694c6009995034f668968d9e3508eacd6d35de2550d26')
ac7 = json.loads(ac7_raw)

def captured(case, mode):
    status, out, err = ac7['rows'][case][mode]
    return status, ac7['streams'][out], ac7['streams'][err]

u_before = captured('diff-url', 'U')[1].removesuffix(report(7, 1))
g_before = captured('diff-url', 'G')[1].removesuffix(report(7, 1))
historical_eval_ok = '{"outcome":0}\n'
eval_ok = R(0)  # CHOSEN FUTURE success bytes, including checks/superseded.
finding = '["fixture.txt","changes a golden file"]\n'
override = 'owner override: owner approved head\n'
def gate(mode, status, stderr, evaluator='', owner=False):
    out = u_before if mode == 'U' else g_before
    if owner:
        out += override
    return status, out + evaluator + report(7, status), stderr

A1–A8 and A12 use their identically named captured rows.
A9 uses the five diff-* rows; A10 uses policy-*;
A11 uses owner-deny; A13 uses ready-executable.
Readiness exit 2 uses ready-status. owner-override follows
an approved owner finding with readiness exit 2 to record both streams.
CHOSEN FUTURE evaluator non-success uses
gate(mode,n,hint,R(n)) for n=1..6, with AC8's preserved stderr hint;
every byte of the complete AC6 result precedes the trap report.
Success evaluation before readiness prints exactly eval_ok = R(0).
Historical evaluator summaries, including AC8's, are not future results.
For readiness recordings, reconstruct the chosen future tuples as:

def future_captured(case, mode):
    status, out, err = captured(case, mode)
    if case in ('ready-executable', 'ready-status', 'owner-override'):
        assert out.count(historical_eval_ok) == 1
        out = out.replace(historical_eval_ok, eval_ok)
    return status, out, err

All other captured gate/policy/diff streams, statuses and stderr stay
unchanged. gate also reconstructs historical evaluator streams when
passed their historical JSON explicitly.
For reviews-head, jq's descriptor label is normalized to fd 3
in both prose and decoded stderr; status is 5 for both callers and neither reaches
checks. P/S continue on the same isolated paginated-review fault.
Other external errors retain their captured numeric status and bytes.

AC8: unattended fixtures and hints

None of the existing check-bearing fixtures under
tests/skills/agent-merge-pr/ changes its semantic outcome, numeric
status or stderr hint under this contract. The schema/listing fields
are additive stdout changes. The old summary fields keep their exact
values. The evidence includes direct evaluator replays of every
check-shape variant with required names and expected head supplied.
These are evaluator replays, not a claim that the entire Bats suite
ran through a future implementation.

Relevant fixtures: workflow.bats:76,223,233,242,253,267,
loosens.bats:28, confirmation.bats:33-34,161, and the base success
fixture in test_helper.bash. The missing-requirements policy case
stops before evaluation with A12, unchanged. Missing/unusable/pinned
runner cases in workflow.bats:12-41 keep the AC7 boundaries.

Variant	Outcome/status	stdout summary	stderr hint
empty	empty/4	{"outcome":4}	""
skipped/current	failed/1	test/test	H(test,no)
skipped/lagging	failed/1	test/test	H(test,yes)
skipped/unknown	failed/1	test/test	H(test,unknown)
skipped/missing	missing/3	test/validate	H(validate,yes)
required NEUTRAL	failed/1	names=test	""
updated SUCCESS	green/0	{"outcome":0}	""

def H(name, differs):
    return (f'skipped required check: {name}\n'
            f'merge base differs from base commit: {differs}\n'
            'run: gh pr update-branch 7\n')

In the compact rows, test/test and test/validate mean
names/skipped, each a single-name array.

The full byte-exact old JSON and stderr are in m['unattended'].
Add the new fields without removing those summaries. Keep exact
required-name matching across every workflow. Required NEUTRAL and
SKIPPED fail, optional NEUTRAL and SKIPPED pass. Keep missing-required
precedence and the skipped hint. Head mismatch remains 6 and missing
expected-head data remains unreadable/5, as recorded in m['hand'].

Go unit fixtures reordered statuses and status rerun in
internal/checks/checks_test.go deliberately change success/0 to
failed/1. Unknown status/conclusion/context fixtures change failed/1
to unreadable/5. New coverage must add superseded invalid states,
lowercase fields, offset disagreements and command protocol faults.
These tightenings apply to both callers of the shared evaluator.

AC9: capability and release sequence

No compatible release exists yet. The fresh main implementation,
and the latest local tag v0.23.0, don't provide the proposed schema,
complete listing or superseded count. Main also still implements the
unsafe context/offset selection demonstrated above. Don't invent a
minimum released version for capabilities that don't exist.

Missing promised outputs: all-entry validation failures for unknown
and lowercase fields, conservative dual-order selection, all status
contexts, full workflow-qualified listing, superseded count, schema
marker and consistent protocol-failure adaptation. #723 implements
these evaluator and caller changes and documents the first release
that contains them. #333 supplies binary distribution. #544 installs
the compatible release before its dependent callers are deployed.

With no agentrepo, P reports unreadable/0 and S refuses unreadable/1
at the new adapter, while U/G retain their earlier C7 refusal/1.
A release below v0.16.0 still fails the shared loader with D2/1.
An older accepted release lacking schema 1 or its required fields
cannot be green: P reports unreadable/0, S refuses unreadable/1.
U/G must validate the same result protocol at their gate and refuse
with unreadable/5. An unusable executable can stop earlier at D1/C8.
Don't add a second shell judge as a fallback.

AC10: measured time and proposed limit

Generate 500 distinct successful CheckRuns, fixed head head,
workflow
ci, names test-000 through test-499, all started at
2026-10-01T01:00:00Z. Save once. Build outside the timed region.
Time each subprocess invocation through its completion, capturing
both streams, for 20 runs. All runs exited zero. The evidence stores
all 40 samples, in milliseconds.

Original shell median: 28.179033 ms; original Go median: 1.954330 ms.
Fresh-head shell median: 13.267504 ms; fresh-head Go median: 1.880917 ms.

Commands: bash judge.sh 500.json (the unchanged sourced shell
function) and agentrepo checks evaluate --file 500.json.
The timer used Python time.perf_counter_ns(). These current-judge
measurements exclude the not-yet-implemented schema validation and
caller adapter. #723 must measure that complete boundary separately.
Keep its proposed limit: median strictly below 1000 ms for 20 runs,
including invocation, whole-response validation and caller verdict,
with compilation excluded. Both existing medians are below the limit.

AC11: acceptance answers and follow-up

AC	Answer
AC1	Contract rows distinguish each input class and both callers.
AC2	Saved rollups include both judges' exact outputs and statuses.
AC3	All 549808 required field-product cases replayed; zero unsafe green results.
AC4	Union of text and instant maxima preserves the worse verdict.
AC5	Every entry is validated, including superseded unknown states.
AC6	Process/result table defines both callers for all required command cases.
AC7	Earlier-stop table and historical recordings give caller statuses and streams (descriptor label normalized); separate reconstruction defines future AC6 streams. Required review still stops with 75.
AC8	Existing unattended fixtures and their outcomes and hints are listed.
AC9	No compatible release yet; #723 implements and releases it, #333 distributes it, #544 installs it.
AC10	Twenty-run measurements and the proposed limit are recorded.
AC11	This table answers each criterion and identifies the implementation dependency.
AC12	See the final AC12 section.

The caller migration and protocol implementation belong to
#723,
distribution to #333,
installation to #544,
and replacement-path protection to
#588.
No additional follow-up issue is needed.

Recorded evidence and decoder

Save the following base64 block as evidence.b64. Decode it with the
code below. The XZ payload starts with an unsigned little-endian
8-byte metadata length, then UTF-8 JSON metadata, then four bytes per
case: unsigned little-endian shell-dictionary index, one-byte
Go-dictionary index, one-byte proposed evaluator outcome.
Dictionary entries are [status, stdout, stderr].

Uncompressed SHA-256:
ee8e74ba494a5979c91eab8a563edf0c2ef758c2e1e0638fccd4891f70e5aa72.
The matrix SHA-256 is in m['summary']['matrix_sha256'].

import base64, hashlib, itertools, json, lzma, struct
from pathlib import Path
raw = lzma.decompress(base64.b64decode(
    Path('evidence.b64').read_text()))
size, = struct.unpack_from('<Q', raw)
m = json.loads(raw[8:8 + size])
matrix = raw[8 + size:]
assert len(matrix) == 549808 * 4
assert hashlib.sha256(matrix).hexdigest() == (
    m['summary']['matrix_sha256'])
Path('metadata.json').write_text(json.dumps(m, indent=2))
d = m['manifest']
index = 0
seen = set()
with Path('records.ndjson').open('w') as output:
    for kind in ('CheckRun', 'StatusContext'):
        entries = []
        if kind == 'CheckRun':
            for s, c in itertools.product(d['status'],
                                          d['conclusion']):
                e = dict(__typename=kind,
                         workflowName='ci', name='test')
                for k, v in [('status', s), ('conclusion', c)]:
                    if v is not None:
                        e[k] = v
                entries.append(e)
            field = 'startedAt'
        else:
            for s in d['state']:
                e = dict(__typename=kind, context='test')
                if s is not None:
                    e['state'] = s
                entries.append(e)
            field = 'createdAt'
        axes = (enumerate(entries), enumerate(entries),
                enumerate(d['times']))
        for (i, a), (j, b), (k, (x, y)) in (
                itertools.product(*axes)):
            key = (kind, i, j, k)
            assert key not in seen
            seen.add(key)
            rollup = dict(headRefOid='head', statusCheckRollup=[
                dict(a, **{field: x}), dict(b, **{field: y})])
            si, gi, chosen = struct.unpack_from(
                '<HBB', matrix, index * 4)
            record = dict(key=key, input=rollup,
                          shell=m['dictionary'][0][si],
                          go=m['dictionary'][1][gi],
                          chosen=chosen)
            output.write(json.dumps(record) + '\n')
            index += 1
assert index == len(seen) == 549808

The chosen outcomes are a research model. They aren't represented as
implemented behavior. Replaying the recorded current outcomes uses
the cited head's shell function and built command. The Go package and
command tests passed with go test ./internal/checks ./internal/agentrepo.
The future implementation must pass the model and caller contract.
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AC12

Scans and leak review on the text above, before this section was added: gitleaks 8.30.1 (detect --no-git), exit 0, 0 hits; the devtools leak checker with the private-strings lists loaded, exit 0, 0 hits. Independent reviews: three rounds; the third found no blockers and a clean full-text leak review. The text above sha256 4875c68c0af60640b8f0fd23498acfebaa4462ec2ef8643af0f26d16c52a144e.

Settle the check judge contract and divergence corpus before porting it #720

Description

Problem

Acceptance criteria

Scope and non-goals

Approach

Dependencies

Size

Verify plan

Risks

Operator authority

Activity

tbhb-agent commented on Oct 25, 2026

Checked

Gaps

Split

Definition of done

tbhb-agent commented on Oct 25, 2026

Checked

Gaps

Split

Definition of done

tbhb-agent commented on Oct 25, 2026

Checked

Gaps

Split

Definition of done

tbhb-agent commented on Oct 25, 2026

tbhb-agent commented on Oct 25, 2026

Checked

Gaps

Split

Definition of done

tbhb-agent commented on Oct 25, 2026

Checked

Gaps

Split

Definition of done

tbhb-agent commented on Oct 25, 2026

Checked

Gaps

Split

Definition of done

tbhb-agent commented on Oct 26, 2026

Evidence and notation

AC1, AC2 and AC5: input contract and recorded comparison

AC3: finite sweep, coverage and results

AC4: time ordering rule

AC6: evaluator process and result contract

AC7: earlier stops, modes and exact stream rules

AC8: unattended fixtures and hints

AC9: capability and release sequence

AC10: measured time and proposed limit

AC11: ownership, open questions and follow-up work

AC12: publication checks

Recorded evidence and decoder

tbhb-agent commented on Oct 31, 2026

Findings for #720: conservative check evaluation contract

Evidence and notation

AC1, AC2 and AC5: input contract and recorded comparison

AC3: finite sweep, coverage and results

AC4: time ordering rule

AC6: evaluator process and result contract

AC7: earlier stops, modes and exact stream rules

Exact caller supplements

AC8: unattended fixtures and hints

AC9: capability and release sequence

AC10: measured time and proposed limit

AC11: acceptance answers and follow-up

Recorded evidence and decoder

AC12

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships