human-rule delegates in one call, and human:* protects authorship only — rulings are inputs the machine executes - #166
Conversation
…rder, and human:* protects authorship only Mechanism (#111): --rework/--rework-file/--park on human-rule and human-rule-issue (CLI + MCP). A delegation posts the trusted 'Rework note @<anchor>: ...' work order beside the ruling in ONE call, in the exact prefix form the producer's trusted-comments marker read accepts - round-tripped through that same function in a unit test. Ruling and order stay two records: provenance vs spendable instruction, ruling always first. A bare reject/design refuses; park is the explicit minority spelling, never the default. Model (the 2026-07-31 rulings): the human namespace's protection is authorship only - no AI actor writes a human:* label, none removes one as an override - and clearing-by-execution is modeled. pr_human_sacred narrows: human:design with its work order executed (head pushed past the pin) is consumable; worklist routes an ordered ruling as rework-ruling; the re-vet's record-verdict clears the spent label as the ruling's completion. human:close-candidate and retired human:reject stay absolute parks (their consumers are human-close and migrate-reject). VET_PROTOCOL stays 4: the vet function is unchanged, only the record gate and the label sweep moved. Closes #111 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Warning Review limit reached
Next review available in: 2 minutes You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: ASSERTIVE Plan: Pro Plus Run ID: 📒 Files selected for processing (8)
WalkthroughChangesHuman ruling delegation
Estimated code review effort: 4 (Complex) | ~60 minutes Sequence Diagram(s)sequenceDiagram
participant Human
participant HumanRule
participant GitHub
participant Worklist
participant Producer
participant Vetting
Human->>HumanRule: submit ruling with --rework or --park
HumanRule->>GitHub: record ruling and optional pinned Rework note
GitHub->>Worklist: expose ruling state
Worklist->>Producer: route delegated rework
Producer->>GitHub: push requested changes
GitHub->>Vetting: provide executed design delegation
Vetting->>GitHub: record verdict and clear spent label
Possibly related PRs
Suggested reviewers: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
…rts do Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…tier-format the prose, and the vetter prompt stops claiming every human:* row is excluded - the_marker_is_the_prompts_exact_spelling used .expect on campaign-prompt.txt, which the flake package's fileset excludes - the same graceful let-else bail every other prompt-conformance test uses - pinned pre-commit (prettier-rainix) over the changed markdown - review-prompt.txt line 6 said human:*-labelled PRs are ALREADY EXCLUDED from unvetted; after #111 an EXECUTED human:design delegation is listed, and the prompt now says what that row means Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
QA evidence — test + adversarial-mutation pass for the #111 delegation change. Suite: full Round-trip (the issue's own check): Mutation pass (from the committed baseline, one mutant at a time, whole suite per mutant, source restored between — all 9 KILLED):
Authorship protection re-verified after the narrowing: |
153 commits of `main` merged in (merge base e9029ed), 88 of them touching pr-review-report-rs/src/main.rs. `main` won everywhere it moved alone; every region both sides changed is resolved as the UNION of the two intents, never a winner. Nothing about what #111 does was changed. pr-review-report-rs/src/main.rs (2 hunks) - worklist_row's JSON row. #111 replaces the `has_human_override` local with the delegation-aware pair, and main added `reviewDecision` to the same object. Union: `humanOverride` keeps its LABEL-PRESENCE meaning (computed straight from `has_human_override(detail)`), `humanWorkOrder` rides beside it, and main's `reviewDecision` field is kept verbatim with its comment — 2z reads the approved set off the row and re-deriving it with `gh search prs` is the cost that field exists to remove. - End of file: append/append. #111 adds `mod delegation_111_tests`, main adds `mod work_tokens_tests` (the `work-tokens` subcommand). Both modules kept. pr-review-report-rs/src/main.rs — minimal_args, a real semantic interaction `minimal_args` is NEW on main; its `human_rule` / `human_rule_issue` arms passed `ruling: "reject"` with no work order. Under #111 that is no longer a minimal call, it is an INVALID one: `ruling_work` refuses a bare reject because a reject IS a send-back. `each_refusal_names_an_argument_that_actually_narrows_it` then failed on its own guard ("refused for some OTHER reason, so this test proved nothing") — the guard working exactly as designed. Fixed at the root: both arms now carry `rework`, which is what the minimum for those two tools now is. The test is unweakened and the doc comment's claim that the arms state what a minimal call looks like stays true. campaign-prompt.txt - step 2(a): main added `reviewDecision` to the row's field list; #111 added `rework-ruling` to the `nextAction` list. Both, with `rework-ruling` ahead of `parked-skip`, matching `next_action`'s arm order. - step 3: main rewrote 3b for the run budget, #111 added the DESIGN RULINGS DELEGATE THE SAME WAY (#111) paragraph to step 3. Both. review-prompt.txt Step 1 was rewritten by both. Kept main's RUN BUDGET paragraph whole and swapped in #111's replacement for the exclusion sentence — "PRs a human still HOLDS (an absolutely-parking `human:*` label, …, or a `human:design` whose work order is not yet executed)" plus the EXECUTED-delegation sentence — because that sentence is the one #111 changes and the budget is the one main changes. Main's rewritten `counts` bullet is kept as-is; #111 never touched it. CLAUDE.md Main's transitions table is a superset (state-load, await, preflight, sol-toolchain, open_pr, push, work-tokens, already-fixed …). Kept it whole and replaced exactly one row — `human-rule`, the only row #111 rewrites. plugins/human-fsm/README.md Kept main's table including its new `/ncc` row; took #111's `/reject` and `/design` rows (the `--rework`/`--park` spellings). .claude-plugin/marketplace.json + plugins/human-fsm/.claude-plugin/plugin.json Version bumps on both sides: 0.6.0 → 0.7.0 (this branch) vs 0.6.0 → 0.9.2 (main). Resolved to 0.10.0 in BOTH files — this branch's bump was a minor one for a new capability, so the same intent applied over main's 0.9.2 is 0.10.0. Main's descriptions (which name /ncc) are kept. `plugin-version-lockstep` passes: `ok human-fsm 0.10.0`. Verified after the merge - pr_human_sacred stays FIRST in verdict_plan, ahead of the no-sha refusal, with the spent-`human:design` clearing after it. - next_action's order is human_parked → ReworkRuling → modeled-ai:* park → deploy → CI classifier. Main inserted no arm above the two human ones. - unvetted_fetch's cheap pre-filter reads `pr_absolutely_parked`; the `--queue` presentability filter still reads `has_human_override` (label presence), which is what that view wants and what the branch left it as. - Every PrSignals field main carries is populated in worklist_row. - VET_PROTOCOL stays 4 — main did not bump it and neither does this. - labels_to_remove still strips only `ai:*`; no vetter-reachable vocabulary names a `human:*` label. - #209's `every_org_wide_enumeration_withholds_archived_repos` passes unchanged: #111 adds no org-wide enumeration, so the filtered/exempt sets need no edit. - #207's `cc_gate` keeps `RepoArchived` as its first arm ahead of the sacred arm; #111 touches PR sacredness only, not the issue-side classifier. cargo build clean; cargo test 1053 passed / 0 failed (975 bin + 78 across the six integration targets); rainix-rs-static (cargo fmt --check + cargo clippy --all-targets --all-features -D warnings -D clippy::all, after touching main.rs) exit 0 with no warnings; pre-commit run --all-files all hooks Passed (denofmt re-aligned the two markdown tables). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Base merge:
|
| Gate | Result |
|---|---|
cargo build |
clean |
cargo test |
1053 passed, 0 failed, 0 ignored — 975 (bin) + 3 + 9 + 47 + 8 + 7 + 4 |
rainix-rs-static (cargo fmt --all --check + cargo clippy --all-targets --all-features -- -D warnings -D clippy::all) |
exit 0, 0 warnings — run after touch pr-review-report-rs/src/main.rs so a warm target could not mask a lint |
pre-commit run --all-files |
exit 0, all 11 hooks Passed (denofmt re-aligned the two markdown tables on its first pass) |
plugin-version-lockstep |
ok human-fsm 0.10.0 |
Nothing was left unresolvable as a union.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 7
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@campaign-prompt.txt`:
- Line 45: Update the producer instructions in campaign-prompt.txt to remove
executable raw gh operations, including gh search prs, gh pr view, gh pr checks,
gh run view, and gh pr checkout. Replace each with the corresponding sanctioned
pr-review-report workflow or subcommand for discovery, inspection, checks, run
status, and checkout/state transitions. Preserve raw gh command names only where
the text explicitly forbids their execution, and ensure all GitHub input/output
uses typed pr-review-report operations.
In `@plugins/human-fsm/commands/design.md`:
- Around line 24-34: Update the changed fenced command blocks in the design
documentation, including the blocks around the rework, park, and line 52
commands, by adding an appropriate language tag such as text to each opening
fence so markdownlint MD040 passes.
- Around line 49-54: Preserve the selected issue-side disposition across the
human-fsm documentation. In plugins/human-fsm/commands/design.md lines 49-54,
update the human-rule-issue fallback to pass the same --park or --rework <WORK
ORDER> choice instead of always using --rework. In plugins/human-fsm/README.md
lines 17-18, document the required disposition arguments for both reject and
design flows.
In `@pr-review-report-rs/src/main.rs`:
- Around line 15401-15432: Extract the duplicated work-order disposition logic
into a shared ruling_work_parts helper taking the subject &Value, anchor, and
&RulingWork, returning (Option<String>, bool, &'static str), then use it at both
parallel call sites with prj and j respectively. Extract the dry-run formatting
match into dry_run_work_line(body, deduped), and replace both inline blocks with
these helpers while preserving the existing wording and behavior.
- Around line 53999-54035: Add a malformed pinned-note case to
design_delegation_reads_the_pin_against_the_head using an empty anchor, such as
“Rework note @: do X” or whitespace-only anchor text. Assert it returns
Delegation::Parked for both the current and changed heads, preserving the
fail-safe behavior of rework_note_anchor when it returns None.
In `@README.md`:
- Around line 177-181: Update the README passage describing sanctioned human:*
label removals to clarify that the two listed paths are normal runtime paths,
while explicitly acknowledging the one-time migrate-reject transition that
removes retired human:reject. Preserve the existing explanations for author
removal and clearing-by-execution.
- Around line 113-116: Update the public command syntax documentation so
disposition flags are advertised only for the reject and design rulings: in
README.md lines 113-116, split or annotate both human-rule tables accordingly,
and in CLAUDE.md line 119 narrow human-rule syntax to reject/design. Keep
close-candidate and keep-open documented without --rework or --park.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro Plus
Run ID: ba7fa610-91d1-4b8c-b9dc-66e149fc737a
📒 Files selected for processing (10)
.claude-plugin/marketplace.jsonCLAUDE.mdREADME.mdcampaign-prompt.txtplugins/human-fsm/.claude-plugin/plugin.jsonplugins/human-fsm/README.mdplugins/human-fsm/commands/design.mdplugins/human-fsm/commands/reject.mdpr-review-report-rs/src/main.rsreview-prompt.txt
…as given /design offers --rework and --park, and the fallback for the `is an ISSUE, not a pull request` refusal named only --rework — so a reader who chose --park, hit the refusal and copied what was in front of them delegated what they meant to park. The fallback now shows both spellings and says to carry the one already chosen, with the mechanical difference under it: --rework posts the trusted work order beside the ruling, --park posts no order at all. The plugin README's /reject and /design rows spell the issue-side call with its flags for the same reason, each with the rule that binds it — --rework required on either subject for reject, exactly one of the two for design. reject.md is deliberately unchanged: `ruling_work`'s "reject" arm errors on (None, _) and has no (None, true) arm, so --park is unspellable there and its fallback has exactly one disposition to carry. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both public tables advertised `[--rework …|--park]` as part of every ruling's syntax. `ruling_work`'s `_` arm refuses both on close-candidate and keep-open — the verb is its own disposition — so the tables named flags the code rejects, and the two disagreed with each other besides (README's human-rule-issue row carried them, CLAUDE.md's did not). Both syntax columns now stop at "<note>", and the rule is stated where it holds: reject REQUIRES --rework (or --rework-file), design takes exactly one of --rework / --park, and close-candidate / keep-open take neither. The prose under the README table gained the same missing half — it stated reject's requirement and design's exclusive choice and said nothing about the two verbs that refuse both. The same section claimed exactly TWO sanctioned removals of a `human:*` label while `migrate-reject` is a third. They are not the same kind of move, so the count is split rather than bumped: two RUNTIME removals (the author superseding their own ruling, and clearing-by-execution), plus a MIGRATION over the fixed population still carrying the `human:reject` #133 retired. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`command_contract` reads every line inside a command's fenced blocks as something the caller runs, and the opening fence's info string was one of them — so ```text refused as "not a transition of this binary", which is a markdown lint answered with a contract violation. The info string is dropped with the fence it belongs to. Nothing else about the rule moves: every remaining line must still be a `pr-review-report` transition or a slash command, and a raw gh/git state change is still refused wherever it sits. `a_fence_language_tag_is_markdown_and_buys_a_command_nothing` pins both halves, because "skip the info string" is one careless edit from "skip the first command": four fence spellings of the same block agree, a raw `gh pr edit` smuggled in below a legal transition is still refused, and a block whose only line is the tag still runs nothing. With that, design.md's and reject.md's command blocks carry `text` tags (MD040) — this branch rewrites the command inside every one of them. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`rework_note_anchor` returns None when the parsed anchor is empty, and
no case drove that branch: `design_delegation` was tested with a
well-formed pin and with a legacy unpinned note, so deleting the guard
left the whole suite green.
The guard is load-bearing. An empty pin names a head no push can ever
match, so a note reading as pinned-but-empty would classify as
Delegation::Executed on the FIRST push past it — `pr_human_sacred` stops
treating the PR as parked and `verdict_plan` clears a `human:design` no
producer was ever ordered to work.
`design_delegation_reads_the_pin_against_the_head` now covers both the
parser and the classifier, at a matching AND a moved head. The
whitespace spelling is there on purpose: "Rework note @ : do X" parses
to " ", non-empty until trim() runs, which is a different intermediate
from the bare "@:".
Mutation: with the `if anchor.is_empty()` guard deleted, the parser
assertion fails on Some("") and the classifier assertion fails with
Executed where Parked is required.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
human_rule_pr_apply and human_rule_issue_apply built rework_body, computed note_deduped, took the report tail and formatted the dry-run line in byte-for-byte parallel, differing only in which subject had been fetched (prj vs j). Two sites means a change to the disposition wording or the dedup rule has to land twice, and one missed site is two different human-facing reports for the same disposition. `ruling_work_parts(subject, anchor, work)` returns the three values, and `dry_run_work_line(body, note_deduped)` formats the dry-run tail; both sit beside `ruling_work_report`, where the disposition's other pure piece already lives. Each call site collapses to one line, and the fetch difference — the one thing that actually differs — is what is left visible. rework_body stays owned, since the write path consumes it with .filter(|_| !note_deduped); the dry-run path borrows it with as_deref(). No clone added, no behaviour changed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ork-ruling kept in the enumeration, design-delegation paragraph reinserted, work-order priority above the parked classifier, 0.10.0
`denofmt` recomputes a markdown table's column widths from its widest cell, so shortening the `human-rule` / `human-rule-issue` syntax columns (9d4c1ce) and spelling the issue-side flags into the plugin README's `/reject` and `/design` rows (7b85ad6) moved every pipe in three tables. The prose paragraph under the README table rewraps for the same reason — its first line absorbed the sentence 9d4c1ce ended. `bdc9d1af` was `deno fmt --check` clean, so this is entirely churn from this branch's own edits, not a pre-existing red. Content unchanged: the diff is padding, pipe alignment and wrap points only. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CodeRabbit review addressed — 6 fixed, 1 rejectedHead is now
The one rejection: the edit to Two of the six needed more than the reported symptom:
QA on the composed treeEverything below was run against
Mutation, re-run on the composed treeDeleting One commit beyond the merge
|
|
Reviewed 54809bd: ready Verified before merge: 20 SUCCESS / 1 skipped, MERGEABLE/CLEAN, 0 unresolved threads, 1062 tests, rainix's PINNED clippy (0.1.94 via 53e96a7d) and 11/11 pre-commit hooks clean. Re-verified mergeable AFTER #204 landed on main. I located each fix in the merged tree rather than relaying the report. Both halves of #111 land here. The MECHANISM: The MODEL, from the two rulings on the issue: All five review fixes confirmed present in the tree: The empty-anchor test is the one worth singling out, because it records the CONSEQUENCE rather than the mechanics: a note that read as pinned-but-empty would name a head no push can ever match, so the FIRST push past it would classify the delegation as executed and let the re-vet clear a Two caveats recorded rather than hidden, neither blocking. One fix beyond the merge, worth knowing it was not cosmetic: a |
…ject Review of the live population found the note-as-order fallback cannot fire on it. All six open human:design PRs (rain.erc4626.words#233 #166, raindex#2778 #2721 #2720, cyclo.site#412) predate the marker-emitting ruling transition: each carries the label alone, with no trusted 👤 human comment and no Rework note, so no order is recoverable for any of them. Mapping those to ai:reject would assert an answer nobody gave AND land them in the producer's queue as orderless rejects — a state campaign-prompt.txt defines as inert ('a reject-labelled PR WITHOUT any trusted note is PARKED for a human'), so six unanswered design questions would go neither answered nor asked. So migrate_design_pr_plan maps by whether the ANSWER is recoverable: - Answered (a trusted order live at head, or a trusted ruling whose note can be posted as one) -> ai:reject + that order, as before. - Unanswered (neither) -> ai:design, the machine's own state for 'a human must answer this', which is the /nd queue. MigrateOrder loses its Unrecoverable arm: it now describes only where a recoverable order comes from, and recoverability decides the TARGET. Also: - strands_flag_error states the reason that always holds. A comment-only design ruling writes no label and strands nothing, so the message leads with the anchor rationale (it would pin to close-candidate @<at>, the record of judging the flag, while answering something else) and keeps the stranding note for rulings that do write a label. - last_human_ruling_note only carries a SEND-BACK verb's note forward. A torn human-close leaves a trusted close ruling on an OPEN subject, and its note is a close reason, not an order to hand a producer. - CLAUDE.md states that issue-side design writes no label. Refs #219 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2 participants
Closes #111
Both halves of the issue land here: the delegation mechanism (the original issue text) and the model it plugs into (the 2026-07-31 update comment carrying the two human rulings). The rulings, verbatim, are the spec for the second half:
Mechanism — the ruling transition carries the work order
human-rule(CLI and MCP, PR and issue) now takes--rework "<order>"/--rework-file <path>/--park:👤 human/Ruled <anchor>: …) is provenance — pinned, historical, true forever. The work order (Rework note @<anchor>: <order>) is an instruction, spent once executed. A delegation posts both, ruling FIRST (a PR may carry a ruling with no order — the explicit park — but never an order with no ruling; the step order is a tested property, and a failure between the two is reported loudly as "currently reads as PARKED", never silently).REWORK_MARKERis the constant"Rework note", andrework_note_commentis the only writer. A unit test round-trips the emitted comment through the producer's own verification function —trusted_comments(_, Some(REWORK_MARKER)), the exactstarts_withreadtrusted-comments --marker 'Rework note'performs — and asserts a spoofed author is filtered. The hand-typedRework Note/rework notesilent-park failure is unconstructible.@<head sha>on a PR; the flag/issue anchor on an issue), so ruling and order go stale together and a note written for a superseded tree is distinguishable from one for the current tree.rejectordesignREFUSES, naming the legal spellings.rejectrequires--reworkand offers no--parkat all — a reject IS a send-back; one not worth reworking is a different ruling, and the refusal names them (design --park,close-candidate,human-close).designtakes exactly one of--rework/--park.close-candidate/keep-openrefuse both flags — the verb is its own disposition, and theclose-candidaterefusal nameshuman-closeas its consumer.Model — the human namespace protects AUTHORSHIP ONLY
Per the rulings: no AI actor writes a
human:*label; none removes one as an override. Absolute parking was the ruled-out overreaction — a ruling is an INPUT the machine executes, and clearing-by-execution is now modeled:pr_human_sacrednarrows: sacred = an absolutely-parking label (human:close-candidate, retiredhuman:reject), a native APPROVED/CHANGES_REQUESTED review, a👤 humanruling pinned to the CURRENT head, or an un-executedhuman:design. An EXECUTED delegation (trustedRework notepinned only to superseded heads — the producer pushed past it) is consumable.worklistrows classify bydesign_delegation(Parked/Ordered/Executed) and anOrderedrow routes as the newnextAction: rework-ruling, above every CI-derived action.Executedfalls through to ordinary fleet duty;Parkedstaysparked-skip.human:designfrom labels alone (only the absolute set is decidable without comments); an executed delegation is un-vetted by the ordinary anchor-staleness path, andrecord-verdict's plan clears the spenthuman:designalongside the staleai:*— the completion of the ruling, through the normal rework → un-vetted → re-vet flow. Authorship protection is intact and tested: the vetter's vocabulary reaches nohuman:*label, the sharedlabels_to_removestill never stripshuman:*, and the un-executed/parked/absolute shapes all still refuse (exit 3).main.rs~:285 comment enshrining the old model ("its meaning is ABSOLUTE rather than carved-out … nothing removes one but the human superseding their own ruling") is rewritten to describe the current behaviour, as are the CLAUDE.md invariants, the README state diagram + human-transitions section, and thevet_actionordering rationale.campaign-prompt.txtlearns the shape: ahuman:designPR whose trustedRework noteis pinned to the CURRENT head is a work order (mirroring the reject-with-rework-note language, dispatched viarework-ruling); parked-without-note remains parked-for-human; the producer never touches ahuman:*label — execute-and-push is its entire part.Consuming-transition table (every human:* state → its consumer)
human:design+ order @Head (delegated)worklist→rework-ruling→ execute per note → push; the push un-vets, and the vetter'srecord-verdictclears the spent label (clearing-by-execution)human:design--park(explicit)human-rulere-rule /human-close) — the surviving pure park, reachable only by its explicit spellinghuman:close-candidate(PR or issue)human-close— a DECIDED close awaiting the human's own terminal edge. This is the smallest honest model, stated rather than left dead silently: closing is the human's act by standing ruling (the cron never closes), so there is no machine execution to model; what changed is that the tool's refusals and docs now NAMEhuman-closeas this state's consumerhuman:reject(retired, #133)migrate-reject(existing) — sacred and bucketed until migratedhuman:keep-open(issue)flag-close-candidaterefuses over it and the issue stays in the producer's ordinary queue. Nothing waits on anyone, so there is no dead state to exitIssue-side
design/rejectrulings take the same--rework/--parkspellings (the order posts on the issue, pinned to its anchor). The producer'suncovered-issuesstate-load already includes human-ruled issues, so an issue-side executable ruling is reachable work; the human-queue backlog VIEW's exclusion ofhuman:*issues is a view, not a gate, and is unchanged.Deliberately unchanged
VET_PROTOCOLstays 4.vetted_at_headis untouched and the vet function (PR-at-head → verdict) is unchanged; what moved is WHEN a verdict may be recorded (the sacred gate) and which spent labels the write sweeps. A protocol-4 verdict is still a value of the current function, so no recompute of the fleet is warranted.classify_lane/HUMAN_DECISION_LABELSstill bucket by label (the label is the state); a delegated design PR shows in the human-decisions lane until the re-vet clears it. The dash re-homing ofhuman:design(the HUMAN ACTION → producer action owner flip on rain-org-health) is a separate One reader for the changed-file list, and it says whether the list is whole #148-shaped follow-up, gated on this landing — deliberately not touched here.has_human_overridekeeps its label-presence meaning for the queue's presentability filter and lane domination (a mid-flow PR is not presentable either way).QA
the_emitted_work_order_passes_the_producers_own_verification,the_marker_is_the_prompts_exact_spelling,the_ruling_and_the_work_order_stay_two_records_and_the_ruling_leads,an_identical_work_order_dedups_and_a_changed_one_does_not,park_or_delegate_is_an_explicit_choice_never_a_default,the_cli_carries_the_work_order_spellings,design_delegation_reads_the_pin_against_the_head,the_human_namespace_partitions_into_absolute_and_consumable,sacredness_narrows_to_authorship_and_currency,the_re_vet_clears_an_executed_design_as_completion,worklist_routes_an_ordered_ruling_to_the_producer_and_parks_the_rest,no_ai_actor_writes_a_human_label_and_only_execution_clears_one— each fails on base (the delegation symbols do not exist on base, and base'sverdict_plan/pr_human_sacred/next_actionreturnRefuseHuman/sacred/parked-skipwhere these assert the consuming behaviour); MCP surface cases extended ina_human_ruling_validates_or_reaches_no_effectfail on base (bare reject validated there).trusted_comments+ the--marker 'Rework note'prefix read the campaign prompt mandates), the issue text's three requirements (fail-closed marker, same-anchor pin, explicit park) and the update comment's two verbatim rulings; the label-set partition is asserted againstPR_SACRED_LABELSitself, not a re-typed list.human-rule, one call, two distinct records) AND — per the 2026-07-31 update comment — the model (authorship-only protection, live consuming transitions for everyhuman:*state, explicit-only parking, the ~:285 comment rewrite, the campaign-prompt shape); covered all of both.Closes #111.Co-Authored-By: Claude noreply@anthropic.com
Summary by CodeRabbit
New Features
Documentation
human-fsmplugin to version 0.10.0.