Intent単位の自律実行と品質修復ループを導入する - #2170
Closed
j5ik2o wants to merge 28 commits into
Closed
Conversation
|
Warning Review limit reached
Next review available in: 22 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. 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: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Plus Run ID: ⛔ Files ignored due to path filters (218)
📒 Files selected for processing (142)
📝 WalkthroughWalkthroughChangesIntent自律性と品質修復
Estimated code review effort: 5 (Critical) | ~120 minutes Possibly related PRs
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 84
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (5)
docs/reference/06-hooks-and-tools.ja.md (1)
256-273: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick winStop フックの説明を Intent autonomy に同期してください。
docs/reference/06-hooks-and-tools.mdとdocs/reference/06-hooks-and-tools.ja.mdのautonomous Construction／自律的ConstructionをIntent autonomyに更新してください。semiとfullは block cap 8、human-wait carve-out の抑制はfullのみです。フック概要表と no-progress カウンターの説明も更新対象です。🤖 Prompt for 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. In `@docs/reference/06-hooks-and-tools.ja.md` around lines 256 - 273, Synchronize the Stop hook documentation with Intent autonomy in docs/reference/06-hooks-and-tools.ja.md lines 256-273 and docs/reference/06-hooks-and-tools.md lines 258-275, updating the hook overview and no-progress counter descriptions from autonomous Construction to Intent autonomy. Document that both semi and full use a block cap of 8, while only full suppresses the human-wait carve-outs; apply the equivalent changes in each language.packages/framework/core/tools/amadeus-state.ts (1)
2921-2922: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick winエラーメッセージ末尾の案内が退役した概念を指しています。
本文は「(autonomous Construction is exempt)」で終わります。本 PR は Construction 自律モードを Intent 自律モードへ置き換えました。利用者はこの案内から次の操作を導けません。正しい操作は Intent 自律モードの選択です。
amadeus-lib.tsのAUTONOMY_LADDER_EXITと同じ表現に揃えてください。🔧 修正案
error( - `Refusing to ${verb} "${slug}": a real human has not acted at this gate since it opened. The approval gate requires a typed human turn before it can commit. Acknowledge the gate as a human, then ${verb}. (autonomous Construction is exempt)` + `Refusing to ${verb} "${slug}": a real human has not acted at this gate since it opened. The approval gate requires a typed human turn before it can commit. Acknowledge the gate as a human, then ${verb}. (Intent autonomy selected with \`amadeus-bolt set-autonomy --mode semi|full\` is exempt)` );🤖 Prompt for 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. In `@packages/framework/core/tools/amadeus-state.ts` around lines 2921 - 2922, Update the approval-gate error message in error() to replace the retired “autonomous Construction is exempt” guidance with the Intent autonomy-mode selection wording used by AUTONOMY_LADDER_EXIT in amadeus-lib.ts.packages/framework/core/tools/amadeus-advisory-choice.ts (1)
890-897: 🩺 Stability & Availability | 🟠 Major | 🏗️ Heavy lift
openを提示時のcheckpointとintentRunに限定してください。異なる
checkpointの pending が混在すると、Rationaleが提示内容と一致せず、正しい回答でもrecordProtectedAdvisoryChoiceがfalseになります。hold が解除されないケースを回帰テストに追加してください。🤖 Prompt for 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. In `@packages/framework/core/tools/amadeus-advisory-choice.ts` around lines 890 - 897, 提示時のcheckpointとintentRunに一致するpendingだけを、recordProtectedAdvisoryChoice内のopenフィルタで選択してください。既存のclosedAt、timestamp、acceptsFreshChoice条件は維持し、異なるcheckpointのpendingを除外してRationaleの照合対象を限定してください。異なるcheckpointが混在する場合でも正しい回答でholdが解除される回帰テストを追加してください。packages/framework/core/tools/amadeus-orchestrate.ts (1)
3593-3612: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win未完了の per-unit directive にも production autonomy を接続してください。
Line 3612 は directive を直接 emit します。したがって、
semiまたはfullでもintent_autonomy_mode、autonomy_auto_approve、quality_repair、intent_grant_idが設定されません。Line 3589 の最終 unit 経路はrouteMainWorkflowDirectiveを呼びます。未完了 unit の emit も同じルーティングを通してください。
修正案
- emit(directive); + emit(routeMainWorkflowDirective(directive, stateContent, codekbCtx));🤖 Prompt for 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. In `@packages/framework/core/tools/amadeus-orchestrate.ts` around lines 3593 - 3612, Route the unfinished per-unit directive through routeMainWorkflowDirective instead of emitting it directly at the directive.unit path. Match the final-unit path’s routing so semi and full modes populate intent_autonomy_mode, autonomy_auto_approve, quality_repair, and intent_grant_id, while preserving the existing gate, next_stage removal, and unit assignment behavior.tests/unit/t211-swarm-batch-progress.test.ts (1)
277-305: 📐 Maintainability & Code Quality | 🟠 Major | ⚡ Quick win引数の語彙を Intent モードへ揃えてください。
codegenStateWithAutonomyの引数は旧 Construction 語彙(""、"gated"、"autonomous")のままです。一方、テスト説明は新しい Intent 語彙(none、semi、full)を使います。328行以降ではseedAutonomyProject(..., "gated")が「semi」を意味します。この二重語彙は誤読を招きます。将来の修正で誤ったモードを設定する危険があります。引数を
"none" | "semi" | "full"の union 型にしてください。型を絞れば、未知の値の分岐も不要になります。♻️ 提案する修正
-function codegenStateWithAutonomy(autonomy: string): string { +type IntentAutonomy = "none" | "semi" | "full"; + +function codegenStateWithAutonomy(autonomy: IntentAutonomy): string { const base = autonomousCodegenState("code-generation").replace( "- **Status**: Running", [ "- **Status**: Running", "- **Last Updated**: 2026-07-28T00:00:00Z", "- **Completed**: 4", "- **Last Completed Stage**: infrastructure-design", ].join("\n"), ); - if (autonomy === "") { + if (autonomy === "none") { return base .replace("- **Intent Autonomy Mode**: full", "- **Intent Autonomy Mode**: none") .replace( "- **Construction Autonomy Mode**: autonomous", "- **Construction Autonomy Mode**: unset", ); } - if (autonomy === "gated") { + if (autonomy === "semi") { return base.replace("- **Intent Autonomy Mode**: full", "- **Intent Autonomy Mode**: semi"); } - if (autonomy === "autonomous") return base; - return base - .replace("- **Intent Autonomy Mode**: full\n", "") - .replace( - "- **Construction Autonomy Mode**: autonomous", - `- **Construction Autonomy Mode**: ${autonomy}`, - ); + return base; }
seedAutonomyProjectの引数型も同じ union へ変更し、呼び出し側を"none"/"semi"/"full"に置き換えてください。🤖 Prompt for 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. In `@tests/unit/t211-swarm-batch-progress.test.ts` around lines 277 - 305, Update codegenStateWithAutonomy and seedAutonomyProject to accept the Intent-mode union "none" | "semi" | "full"; replace callers using the legacy values with the corresponding new values, remove the unknown-value fallback branch, and preserve the generated state mappings for each supported mode.
🤖 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 `@amadeus/spaces/default/memory/project.md`:
- Line 309: project.md の重複した訂正規則を統合し、Line 309 と Line 315
のうち1件だけを正本として残してください。本文と cid
が同一であることを確認し、もう一方の重複エントリを削除して規則が一度だけ読み込まれる状態にしてください。
In `@docs/guide/02-your-first-workflow.ja.md`:
- Around line 166-171: 新しい対話例のコードフェンスを、docs/guide/02-your-first-workflow.ja.md
の166-171行と docs/guide/02-your-first-workflow.md の164-169行で、それぞれ未指定の開始フェンスから text
指定の開始フェンスに変更してください。
In `@docs/guide/04-phases-and-stages.md`:
- Around line 293-298: Separate and unify the Quality Repair and Code Generation
failure transition contracts. In docs/guide/04-phases-and-stages.md:293-298 and
packages/framework/core/amadeus-common/protocols/stage-protocol.md:127-133,
define quality findings as observe-quality → repair/replanned → READY or
REPAIR_STALLED, and execution failures as BOLT_FAILED → halt-and-ask
retry/skip/abort. In docs/guide/16-worked-examples.ja.md:357-374 and
docs/guide/16-worked-examples.md:301-312, classify the broken SES mock
explicitly and revise its transition order accordingly; describe sibling
preservation and failed Bolt reruns only within the appropriate Quality Repair
behavior.
In `@docs/guide/10-state-and-audit.ja.md`:
- Around line 95-100: 監査イベント分類とlegacy
grant移行の記載が全言語の正典間で不一致です。docs/guide/10-state-and-audit.ja.md:95-100 と
docs/guide/10-state-and-audit.md:95-100 にcanonicalなtransaction
event（LOOP_MONITOR_EVENT_SET_COMMITTEDなど）を追加し、GATE_AUTHORIZATION_SELECTEDをreplay/doctor専用の履歴観測として記載してください。docs/reference/12-state-machine.ja.md:185-196
と docs/reference/12-state-machine.md:181-196 も同じ分類へ同期し、standing grant選択をlive
emitterとして扱わず、GRANT_REVOKEDはread-only observationではなく明示的な履歴取消しとして区別してください。
In `@docs/guide/16-worked-examples.ja.md`:
- Around line 310-314: docs/guide/16-worked-examples.ja.md の日本語本文に追加された “Intent”
および “Intent-scoped” の表記を、ガイドの規則に従って小文字の “intent” 表記へ統一してください。該当する本文箇所（Bolt 1
周辺と指摘された追加箇所）だけを変更し、コード、コマンド、識別子、その他の用語は変更しないでください。
- Around line 331-339: すべての指定された fenced code block
に適切な言語識別子を追加してください。docs/guide/16-worked-examples.ja.md の autonomy
menu（331-339）と実行結果（365-370）、docs/guide/16-worked-examples.md の autonomy
menu（279-287）と実行結果（305-310）は
text、docs/reference/12-state-machine.ja.md（189-192）および
docs/reference/12-state-machine.md（189-192）の command block は shell を指定してください。
In `@docs/guide/glossary.ja.md`:
- Line 39: 日本語本文の正規表記を小文字のintentに統一する。docs/guide/glossary.ja.mdの39-39行ではIntent
auditとIntent-scopedをintent表記に、132-132行ではIntent自律レベルをintent表記に変更する。docs/guide/workshop-mode.ja.mdの188-189行、254-255行、435-435行では、本文中のIntent
autonomyとIntent modeを小文字のintent表記へ変更し、コード識別子・見出し・画面表示は変更しない。
In `@docs/guide/glossary.md`:
- Line 17: Update the protocol and reference term lists in
docs/guide/glossary.md to include autonomy-mode, then regenerate
docs/reference/04-stage-protocol.ja.md from the updated Projection Manifest
rather than editing the generated file directly.
- Line 60:
Markdown表内のリテラル区切り文字をエスケープし、生成元を修正して再生成してください。docs/guide/glossary.md:60 と
docs/guide/glossary.ja.md:82 では
autonomous\|gated、docs/harness-engineering/08-construction-and-swarm.md:37 と
docs/harness-engineering/08-construction-and-swarm.ja.md:20 では none\|semi\|full
になるよう生成元を更新してください。docs/reference/04-stage-protocol.ja.md:681 は glossary
の生成元修正による再生成で反映し、直接編集しないでください。
In `@docs/guide/workshop-mode.ja.md`:
- Around line 254-258: Update docs/guide/workshop-mode.ja.md lines 254-258 to
instruct users to update the working tree before resuming, using git pull
--ff-only or an explicit fast-forward instead of only git fetch --all. Apply the
same replacement in docs/guide/workshop-mode.md line 190; both references should
ensure the current branch and amadeus-state.md receive the latest projection.
In `@docs/harness-engineering/08-construction-and-swarm.ja.md`:
- Around line 32-39: docs内の旧自律ラダーからamadeus-bolt set-autonomyへ進む説明を更新し、none /
semi / fullとIntent-scoped
grantを正規のIntent契約として扱う。旧コマンド経路を削除するか、replay・診断専用で実行経路ではないことを明記し、現在の説明と矛盾しないようにする。
In `@docs/reference/04-stage-protocol.md`:
- Line 794: Update the “Ladder prompt (legacy)” table cell to prevent the
autonomous|gated text from being parsed as a column separator, either by
escaping the pipe within that cell or replacing it with equivalent non-pipe
wording while preserving the meaning.
- Line 783: docs/reference/04-stage-protocol.md の Bolt 定義を直接編集せず、対応する用語定義を
docs/guide/glossary.md で更新し、glossary:projection の生成処理を実行して投影内容を反映してください。
- Line 809: Update the English and Japanese mode tables to state that semi
automatically approves only ordinary in-phase stage-gates, while the
walking-skeleton gate always waits for human approval. Apply the corresponding
contract clarification at docs/reference/04-stage-protocol.md:809,
docs/guide/02-your-first-workflow.ja.md:162-171,
docs/guide/02-your-first-workflow.md:162-169,
docs/guide/04-phases-and-stages.ja.md:215 and 227-228, and
docs/reference/04-stages/construction.ja.md:44-60 and
docs/reference/04-stages/construction.md:44-60; update
docs/reference/03-orchestrator.md#L437 as well. Do not modify the
already-correct walking-skeleton description or regression tests.
In `@packages/framework/core/amadeus-common/protocols/stage-protocol.md`:
- Around line 97-133: Clarify the general human-approval rules to state that the
active Intent autonomy mode takes precedence for Construction stage and phase
gates and question decisions: none remains human-controlled, semi auto-approves
only within phases, and full uses the Intent grant and decide-question flow
without prompting the user. Reconcile the question interaction requirements so
semi/full do not both require and suppress the same prompt, while preserving
human interaction at phase boundaries and for any parked, conflict,
human-required, or aborted result.
- Line 822: Remove the deprecated Ladder prompt flow and standardize the current
Intent mode and grant-check procedure. In
packages/framework/core/amadeus-common/protocols/stage-protocol.md lines 801 and
810, replace the old references and definition with the Intent mode and grant
confirmation; in docs/reference/12-state-machine.ja.md lines 181-184 and
docs/reference/12-state-machine.md lines 181-184, remove or mark historical the
walking-skeleton ladder and autonomy-ladder exit paths and align both language
versions with the current mode-selection procedure.
In `@packages/framework/core/hooks/amadeus-stop.ts`:
- Around line 147-148: `intentAutonomyMode(stateContent)` だけで行っている `full`
判定を更新し、現在の Intent UUID に紐づく canonical projection の grant を参照してください。grant
が存在し、`active` で、revoke または complete 済みでない場合にのみ `full` として停止抑制を有効にし、それ以外は `full`
として扱わないでください。
In `@packages/framework/core/knowledge/amadeus-shared/state-template.md`:
- Around line 102-104: Update the Construction Autonomy Mode value in the state
template to one of the supported values: unset, autonomous, or gated, replacing
the invalid “[internal scheduling projection]” placeholder. Leave Intent
Autonomy Mode, Intent Grant, and State Version: 7 unchanged.
In `@packages/framework/core/otel/event-registry.ts`:
- Around line 196-207: Update the AUTO_DECISION_REVIEWED optional-attribute
definitions in packages/framework/core/otel/event-registry.ts:196-207 and
packages/framework/core/knowledge/amadeus-shared/audit-format.md:250-250 in the
same change, adding Event Identity and Projection Revision to both locations’
Optional lists.
In `@packages/framework/core/tools/amadeus-advisory-choice.ts`:
- Around line 921-932: Update the pending and receipt searches in the advisory
cancellation flow around storeResult to compare the complete identity, including
advisoryInstance, intentRun, and checkpoint, using the existing
identityKey(identity) matching convention. Do not rely on advisoryInstance
uniqueness; ensure both the pending validation and latest receipt selection only
use records belonging to the requested identity.
- Around line 417-420: Extract the repeated revoked-status and identity-matching
predicate used at the filtering sites into a single shared helper, then replace
the duplicated conditions at the five referenced locations with calls to that
helper. Ensure the helper accepts the candidate and comparison item and
preserves the current behavior.
In `@packages/framework/core/tools/amadeus-audit.ts`:
- Around line 89-93: Update the PRESENCE_PROTECTED_EVENTS denylist to include
LOOP_MONITOR_EVENT_SET_COMMITTED, QUALITY_REPAIR_TRANSACTION_COMMITTED,
INTENT_AUTONOMY_TRANSACTION_COMMITTED, AUTO_DECISION_REVIEWED, and
INTENT_COMPLETION_TRANSACTION_COMMITTED, ensuring append and append-raw reject
forged review and completion-seal events.
In `@packages/framework/core/tools/amadeus-autonomy-review-production.ts`:
- Around line 100-117: Update parseReviewPayload to validate all
principal-related payload fields, including reviewPrincipalRef and any other
fields consumed by parsedReviewEvent, with typeof value === "string" checks
before returning the payload. Keep digest validation and existing field checks,
and do not rely on String() coercion to handle missing values.
- Around line 86-88: Update the adapter’s stableId and reviewExtension digest
logic to reuse exported domain-module helpers from amadeus-autonomy-review.ts
instead of duplicating their implementations. Export the existing stable ID
helper and extension-head calculation, then make stableId and reviewExtension
delegate to them so both paths always produce identical digests for
validatePersistedReviewState.
- Around line 289-310: Update the human-turn selection in authorizeHumanReview
so it only accepts a HUMAN_TURN occurring after the latest
AUTO_DECISION_REVIEWED event for the same review flow. Do not reuse a turn
already consumed by a prior decision; return PROVENANCE_REQUIRED when no newer
HUMAN_TURN exists, while preserving the existing binding and sourceAuditRevision
construction for the valid turn.
In `@packages/framework/core/tools/amadeus-bolt.ts`:
- Around line 1027-1035: Update the autonomy-mode projection around
schedulingMode and the related readAutonomyMode/owedBatchGate consumers so semi
preserves the batch-end approval gate, while only full maps to autonomous.
Ensure the persisted representation is recognized as semi rather than falling
back to the non-swarm path, and preserve existing none behavior. Add a
regression test covering set-autonomy --mode semi through the next batch and
verifying approval is required.
In `@packages/framework/core/tools/amadeus-directive.ts`:
- Around line 617-620: Move validation for the four autonomy
fields—intent_autonomy_mode, autonomy_auto_approve, intent_grant_id, and
quality_repair—out of the shared checkRunStageShared flow and into the
run-stage-only validation path. Match the existing standing_grant_id handling so
dispatch-subagent treats these fields as unknown keys while run-stage continues
validating them.
- Around line 847-859: Update checkOptionalEnum to match the existing checkEnum
pattern by casting the allowed array to readonly string[] before calling
includes, rather than casting value to T; preserve the current validation and
error behavior.
In `@packages/framework/core/tools/amadeus-harness-registry.ts`:
- Around line 38-43: 簡略化のため、PackageHarnessId と SelfInstallHarnessId の
RegistryRow ベースの条件型を Extract に置き換えてください。各型で対応する face プロパティが true の行を Extract
し、その結果から id の文字列型を取得する構成にし、現在の型の結果を維持してください。
In `@packages/framework/core/tools/amadeus-intent-autonomy-production.ts`:
- Around line 96-120: Update coordinatorFor and
createAuditIntentAutonomyRepository so the audit shards read by coordinatorFor
are passed into the repository and reused, instead of the repository calling
readIntentAutonomyTransactionsFromAudit to read them again. Preserve the
existing legacyModeHistory and legacyStandingGrants projection behavior while
eliminating the duplicate read and associated repeated event scanning.
- Around line 413-416: grant 不在時の fingerprint 導出が grantScope、production の
scopeFingerprint/normFingerprint、commitProductionQuestionDecision
で分散しているため、共通のフォールバック導出関数を追加し、各箇所から再利用してください。scope の入力キーと norm の導出規則を統一し、既存の
grant 由来 fingerprint はそのまま優先してください。
- Around line 785-795: Remove the unnecessary non-null assertion from the
monitorLatchIdentity access in the alreadyResumed qualityResumeAlreadyCommitted
call, using the guaranteed string type on
ProductionQualityResumeContext.envelope while leaving the surrounding resume
validation unchanged.
- Around line 171-178: qualityState など該当する3箇所で、固定の
emptyQualityPluginProjection(...) を使わず、永続化された実際の QualityPluginProjection を復元して
resolveIntentQualityActivation に渡してください。none モードで人間がオプトインした状態を保持し、quality repair
が誤って disabled にならないようにします。
- Around line 219-226: Update latestHumanTurnId and the related
validHumanContext/issueGrant flow to use a block-level HUMAN_TURN provenance
identity instead of only its timestamp. Include the audit block digest together
with the intent, space, and shard identifiers, and validate that complete
identity before issuing the full grant.
In `@packages/framework/core/tools/amadeus-intent-autonomy-replay.ts`:
- Around line 34-46: Update transactionShape to validate that beforeProjection
exists and satisfies the legal AutonomyProjection shape, and that every element
in events has the expected transaction-event structure before asserting
IntentAutonomyTransaction. Reuse the existing projection/event validation
helpers visible in the module where available, and retain the digest and
non-empty-events checks only after all fields pass validation.
- Around line 117-127: Update replayIntentAutonomyAudit and the
transaction-loading path so duplicate intentUuid/expectedRevision branches do
not propagate intent-autonomy-revision-conflict and halt replay. Handle the
conflict as a diagnostic event or convert it to the repository’s fail-closed
default projection, preserving replay for unaffected transactions. Add coverage
for an audit containing such a branch and verify replay remains non-throwing.
- Around line 151-169: Simplify the transactionLock definition passed to
createMemoryIntentAutonomyRepository by removing the double cast around
withAuditLock. Adjust the callback typing or withAuditLock signature as needed
so transactionLock remains a generic synchronous callback without casts, while
preserving the existing lock-depth behavior and onCommit audit emission.
In `@packages/framework/core/tools/amadeus-intent-autonomy-runtime.ts`:
- Around line 405-413: Normalize IntentAutonomyCommitReceipt.transactionDigest
across both idempotent paths by using the digest of the corresponding
transaction read from repository.readTransactions(intentUuid), not
autonomyDigest(existing) or autonomyDigest(before). Update the receipt
construction at
packages/framework/core/tools/amadeus-intent-autonomy-runtime.ts:405-413 and
:702-712, and centralize receipt generation through the existing receipt() flow
so all paths return the transaction-level digest.
- Around line 655-665: Update the coordinator’s resume method to catch
exceptions from validateResumeCondition, clearRepairStalledLatch, and commit,
matching the existing park error-handling pattern. Convert any thrown failure
into the method’s { error } result contract so resumeProductionQuality and other
callers do not receive uncaught exceptions.
In `@packages/framework/core/tools/amadeus-intent-autonomy.ts`:
- Around line 726-728: Update the condition in the effect-classification guard
to remove the unreachable prohibitedEffects.includes check and its unnecessary
ProhibitedEffectClassification cast; retain the existing rejection when
effect.classification is not "workflow-reversible".
- Around line 663-674: In the decision flow around the policy filtering and
decisionRecord creation, bind the grant already validated near the current-grant
check to a local variable and use it instead of projection.currentGrant!. Handle
policyOptions.size > 1 explicitly as a policy conflict by parking or returning
invalid, preventing fallback to norm, history, or solo-election; retain
deterministic policy decisions only when exactly one option remains.
- Around line 502-506: In the semi-mode gating logic, update the internalGate
condition to rely only on occurrence.kind being "stage-gate"; remove the
occurrence.phase !== "phase-boundary" check while preserving the existing
human-command provenance and human-required return behavior.
In `@packages/framework/core/tools/amadeus-intent-completion.ts`:
- Around line 939-943: Update the missing calculation in the receipt aggregation
flow to include only harness IDs with zero receipts, not IDs with duplicate
receipts. Keep duplicate receipt IDs handled exclusively by the rejected set in
the surrounding byHarness logic, so awaitingLiveCapability does not treat
received harnesses as missing.
In `@packages/framework/core/tools/amadeus-loop-monitor-replay.ts`:
- Around line 132-143: Update validEvent in the replay module to validate each
event type’s required payload fields, especially requiring a valid delivery
object with a valid partition for LOOP_DELIVERY_OBSERVED. Reject malformed event
payloads during decode so foldLoopMonitorEventSets and applyLoopDelivery only
receive structurally valid events and preserve LoopMonitorReplayIncompleteError
handling.
- Around line 96-100: Update writeJsonAtomic to clean up the generated temporary
file when writeFileSync or renameSync fails. Preserve the existing atomic write
behavior and ensure cleanup is attempted before propagating the original error.
- Around line 203-211: Unify receiptId derivation in receiptFor with the shared
logic used by createMemoryLoopMonitorRepository and loopReceipt, ensuring
SHA-256 results use the first 32 characters so permits remain interoperable
across repositories. Replace the hard-coded 71 slice endpoint with the shared
function or named length constant, preserving the existing receipt fields.
In `@packages/framework/core/tools/amadeus-loop-monitor-runtime.ts`:
- Around line 22-46: 正規化 JSON とダイジェスト関連の実装を共有 module（例:
amadeus-canonical.ts）へ集約し、正本を 1
つにしてください。packages/framework/core/tools/amadeus-loop-monitor-runtime.ts の 22-46
行ではローカルの isRecord、bytewise、canonicalJson、sha256、stableId を削除し、共有 module の
canonicalJson、sha256、stableId を import
してください。packages/framework/core/tools/amadeus-quality-repair.ts の 12-36
行でもローカル実装を削除して共有 module を import し、同ファイル 814-820 行の qualityDigest /
qualityStableId 再輸出ラッパーを廃止して利用側を共有 symbol に向けてください。
In `@packages/framework/core/tools/amadeus-loop-monitor.ts`:
- Around line 624-626: In the historyLimit function and its related history
handling near the referenced logic, add a concise comment documenting that
history is intentionally a pattern-detection window rather than the complete
duplicate-detection set, and that older upstreamEventIdentity values may be
evicted. Do not change the existing limit or duplicate-detection behavior.
- Around line 617-622: Update edgeConflict and the corresponding
deliveryReplayDecision logic to use the same causal handling as
amadeus-loop-monitor-replay.ts: treat predecessorDeliveryId === null as one
partition-level root predecessor rather than comparing null values as ordinary
delivery IDs. Ensure root conflicts are evaluated from the complete causal
state, not only the retained history, so the result does not change after
historyLimit truncation.
In `@packages/framework/core/tools/amadeus-quality-repair-replay.ts`:
- Around line 32-49: Update validQualityEvent to validate the required
projection structure before accepting a decoded transaction: require
qualityScopeId, epoch, observationSequence, and workflowExecutionState with
their expected object and primitive types, including the nested epoch. Keep
decodeQualityRepairTransaction and the existing event-type validation unchanged,
while rejecting projections that are missing required fields or contain invalid
stalledLatch/workflowExecutionState values.
- Around line 77-110: Update createAuditQualityRepairRepository and its callers
so repositories are cached by projectDir, intent, and space, reusing the loaded
repository across observations instead of scanning audit shards on every
creation. Preserve existing transaction locking and commit-event behavior, and
ensure distinct scopes do not share cached repositories.
- Around line 65-74: Update the identity comparison in the transaction loop of
the quality-repair replay flow to use the existing qualityDigest
canonicalization instead of JSON.stringify. Preserve the transactionId lookup,
conflict error, and deduplicated return behavior while aligning comparison
semantics with amadeus-quality-repair-runtime.ts.
In `@packages/framework/core/tools/amadeus-quality-repair-runtime.ts`:
- Around line 599-609: In the stalled-latch check around
repository.readProjection, evaluate prior !== null before accessing
prior.stalledLatch. Reorder the existing conditions so the null guard is
explicit and the evidenceFingerprint comparison remains unchanged.
- Around line 490-498: Update readByPartition to call
repository.readTransactions() only once, then build a
qualityScopeId-to-partition-key index in a single pass while preserving
foldQualityProjection’s projection semantics. Use that index to resolve the
requested partition and return the matching QualityRuntimeProjection without
repeatedly copying or rescanning transactions.
- Around line 614-673: Move the QUALITY_SNAPSHOT_OBSERVED append in the
transaction callback to occur only after loop.observeDelivery and any follow-up
action have succeeded. Ensure every CONFLICT, INCOMPLETE, and threshold
reservation mismatch returns before committing the staged quality event, or
explicitly discards that staged event, so failed observations do not advance
observationSequence, epoch.window, or consecutiveNonProgress.
In `@packages/framework/core/tools/amadeus-quality-repair.ts`:
- Around line 125-130: 未使用の QualityRequiredOutputDescriptor 抽象と requiredOutputs
フィールドを削除してください。qualityContributionContent から requiredOutputs
の生成を除き、compileQualityContribution の requiredOutputs
に関する検証・参照も削除して、現在の空配列のみを許容する経路を不要な型なしで維持してください。
In `@packages/framework/core/tools/amadeus-state.ts`:
- Around line 2878-2901: productionStageAutonomy と
commitProductionStageGateDecision
に渡す共通入力を1つのオブジェクトへ抽出してください。walkingSkeleton、phaseBoundary、autonomyDigest(graph)
をそのオブジェクト内で一度だけ計算し、両方の呼び出しへ同じ入力を渡して occurrence の一致と不要な再計算の回避を実現してください。
- Around line 3838-3841: Remove the unused grantId declaration and the
unreachable conditional that adds fields["Grant Id"] in the surrounding
delegation-state construction; also remove the associated shape-compatibility
comment as requested, while leaving the humanPresenceGuardDisabled and
humanActedSinceGate logic unchanged.
- Around line 2862-2865: handleApprove の実装では有効な standing grant が grant-backed
経路に入り、resolveStandingGrantRouteReceipt、validateStandingGrantWithinLedger、approveUnderLock(...,
{ grantId }) が実行されるため、対象コメントの「到達しない」という記述を削除し、実際の grant-backed
承認動作を説明する内容に更新してください。
- Around line 2570-2577: Update the completion flow around
emitWorkflowCompletionAuditRows and commitProductionIntentCompletion to perform
a read-only completable check within the same lock before emitting completion
audit rows. When the check rejects with workflow-not-completable, exit without
writing the completion audit, then preserve the existing commit/state-update
flow only after audit succeeds. Add a test covering this rejection path and
verifying that no completion audit is emitted.
In `@packages/framework/core/tools/amadeus-utility.ts`:
- Around line 317-320: Update renderAutonomyStatus so autonomy === null is
labeled as unknown or unavailable rather than “none”; preserve the existing mode
rendering for non-null autonomy and keep the text output semantically consistent
with the JSON autonomy: null representation.
- Around line 362-367: Wrap the entire autonomy status projection flow around
readProductionAutonomyProjection and projectIntentAutonomyStatus in exception
handling. When either call throws, fall back to the same output as
renderAutonomyStatus(null); preserve the existing null-projection behavior when
no exception occurs.
- Around line 670-673: Update the Bun.spawnSync call used to obtain
piVersionOutput for the pi executable to include a 5000-millisecond timeout,
preventing an unresponsive pi --version invocation from blocking the --doctor
flow indefinitely.
In `@packages/framework/harness/claude/skills/amadeus/SKILL.md`:
- Line 63: Escape the Markdown table-cell ambiguity in the finalize argument
descriptions by replacing the inline reason alternatives with a single <reason>
placeholder, then list the allowed values outside the table. Apply this
consistently at packages/framework/harness/claude/skills/amadeus/SKILL.md:63-63,
packages/framework/harness/codex/skills/amadeus/SKILL.md:61-61, and
packages/framework/harness/kimi/skills/amadeus/SKILL.md:63-63, preserving the
existing reason semantics.
In `@packages/framework/harness/cursor/commands/amadeus.md`:
- Line 45: Update the shared directive dispatcher described by the run-stage
entries in packages/framework/harness/cursor/commands/amadeus.md:45-45 and
packages/framework/harness/opencode/commands/amadeus.md:45-45 to handle
await-advisory-choice with formal-check and advisory-decision, stop when
required, and invoke next again when continuing. In both files, distinguish
Approve from Request Changes: report only an Approve decision, while returning
Request Changes to the conductor’s Keep/Modify/Redo loop without reporting.
In `@packages/framework/harness/kiro-ide/skills/amadeus/SKILL.md`:
- Line 92: In both
packages/framework/harness/kiro-ide/skills/amadeus/SKILL.md:92-92 and
packages/framework/harness/kiro/skills/amadeus/SKILL.md:92-92, update the
quality-repair flow around the directive.quality_repair branch to create a
carrier, run observe-quality, and branch on repair, replanned, and parked
outcomes; require resume-quality before any retry or resumption, while recording
stalled repairs as REPAIR_STALLED in Core state.
In `@tests/.coverage-patch-allowlist.json`:
- Around line 3836-3856: 同一ファイル内の異なる関数で coverage allowlist の fingerprint
が重複しています。isPendingQuestionStop と isConversationalStop などの fingerprint
生成処理を更新し、関数名または周辺コンテキストをハッシュ入力に含めて selector ごとに一意になるようにしてください。既存エントリの
fingerprint も新方式で再生成し、重複とドリフト検知の弱さを解消してください。
- Around line 5683-5693: amadeus-quality-repair-runtime.ts の
recordEvidence、resume、normalizeObservation、handleStatus
に対する広範なカバレッジ免除を削除し、実際に未到達な防御的分岐・失敗分岐だけを個別の targetLines として残してください。正常系は既存の
t428・t429 などの決定的テストでカバーし、必要な残余分岐も対応するテストで到達させてください。
In `@tests/integration/t-advisory-human-choice-boundaries.test.ts`:
- Line 453: Update the advisoryReportHoldReason assertion to use the in-scope
pending fixture, passing pending.identity.checkpoint instead of the outer
identity.checkpoint; leave the expected unresolved advisory choice text
unchanged.
- Around line 467-472: Update the assertion around
revokeMisattributedAdvisoryChoice to verify the specific rejection reason for
the existing evidence-trail branch, not only ok: false. Match the returned
reason against the established symbol or value used for “inspection evidence
exists,” while preserving the current projectDir, advisoryInstance, and
eventIdentity inputs.
In
`@tests/integration/t427-loop-monitor-five-harness-projection.integration.test.ts`:
- Around line 14-40: Centralize the harness cohort, ROOT resolution, and
dist/core byte-comparison helper in tests/harness/. In
tests/integration/t427-loop-monitor-five-harness-projection.integration.test.ts:14-40,
pass TOOL_FILES to the shared helper; in
tests/integration/t430-quality-repair-five-harness-projection.integration.test.ts:14-40
and
tests/integration/t433-intent-autonomy-five-harness-projection.integration.test.ts:14-42,
remove local currentHarnessCohort definitions and use the shared cohort/helper;
in
tests/integration/t434-intent-completion-five-harness-projection.integration.test.ts:8-29,
replace its tuple array and loop with the helper and add manifest.coreDirs
validation.
In `@tests/integration/t429-quality-repair-runtime.integration.test.ts`:
- Line 177: snapshots に明示的な要素型を付け、kind と permit の参照が型チェックされるようにしてください。Line 240 の
thresholdResult と同じ方針で、実際に格納するスナップショットのフィールド型を定義または既存型を再利用し、暗黙の any[] をなくしてください。
In `@tests/integration/t432-intent-autonomy-runtime.integration.test.ts`:
- Around line 256-264: The test named “walking skeleton is human in semi and
grant-backed in full” only exercises fullRuntime and asserts decided. Add a
semiRuntime scenario that verifies walking-skeleton requires human judgment,
while preserving the existing fullRuntime grant-backed assertion; alternatively,
rename the test to describe only the full-mode behavior.
- Line 240: Replace the literal expectedProjectionRevision value in the
createAutonomyProjection assertion with initial.projectionRevision, matching the
existing commandFull expectation and keeping the test aligned with the generated
initial revision.
In `@tests/integration/t434-intent-completion-live-seam.integration.test.ts`:
- Around line 30-46: Update the Bun.spawn invocation in the harness execution
flow to set a child-process timeout shorter than the test’s normal 30-second
limit, using the supported timeout option rather than a 60_000-millisecond
value. Before JSON.parse in the receipt handling, validate that trimmed stdout
is non-empty and throw an error containing descriptor.id when no receipt output
is produced.
In `@tests/integration/t435-intent-autonomy-production.integration.test.ts`:
- Around line 29-44: Update the run function’s environment setup to explicitly
set AMADEUS_SKIP_HUMAN_PRESENCE_GUARD for both guard branches: use the disabled
value when guard is true and the enabled value otherwise. Keep the existing
artifact-guard setting and tool invocation unchanged so default and guarded
paths no longer depend on inherited environment state.
- Around line 452-489: Set an explicit 120000 ms timeout on the heavy test named
“full repairs repeated quality failures, replans once, and durably parks a
stalled loop” by passing it as the test callback’s third argument. Keep the
existing test logic and assertions unchanged.
In `@tests/unit/t211-swarm-batch-progress.test.ts`:
- Around line 295-297: Update the gated branch in the test fixture helper around
the autonomy-mode replacement to explicitly document that Construction Autonomy
Mode remains autonomous, and clarify in the relevant test name or annotation
that the semi Intent Autonomy Mode case depends on that value for fan-out.
In `@tests/unit/t33.test.ts`:
- Around line 370-387: Strengthen the rejected legacy-mode tests in
tests/unit/t33.test.ts lines 370-387 by snapshotting state and audit before and
after each autonomous and gated rejection, asserting that Intent Autonomy Mode
is unchanged and no INTENT_AUTONOMY_TRANSACTION_COMMITTED event is added. Apply
the same post-failure non-mutation checks to
tests/integration/t393-birth-autonomy-field.integration.test.ts lines 85-90
after the gated invocation.
In `@tests/unit/t426-loop-monitor.test.ts`:
- Around line 124-142: Update the cases in the test named “fails the whole
compile on unknown fields and invalid schema invariants” to include a
descriptive label with each invalid manifest, then iterate over the labeled
cases and include that label in the assertion context so failures identify which
input unexpectedly returned ok: true.
In `@tests/unit/t427-loop-monitor-runtime.test.ts`:
- Around line 101-109: Update resultFor to accept the graph instance as an
argument and derive constraintFingerprint from that instance instead of
recompiling via graph(). Update each resultFor call site to pass the same graph
used by reserveJudge, preserving matching fingerprints for custom thresholds.
In `@tests/unit/t428-quality-repair.test.ts`:
- Line 220: Update the test setup around recordQualityReplan to advance the
epoch state through that helper instead of directly assigning
replanSinceLastProgress. Preserve the intended state for the subsequent
assertions and use the same recordQualityReplan path already established earlier
in the test.
In `@tests/unit/t434-intent-completion.test.ts`:
- Around line 297-331: テストの前提条件失敗時に無言で PASS
しないよう、tests/unit/t434-intent-completion.test.ts の 297-331 行および同ファイルの
230、232、460、462 行にある早期 return
を、エラー詳細を含む例外送出へ置き換えてください。tests/unit/t433-autonomy-review-observability.test.ts の
92-104 行では、first.value.nextCursor が null でないことを先に検証し、first.ok が偽の場合は早期 return
せず例外を送出してください。
- Around line 424-428: テスト内の skipped 作成で fixture.validations[0]
をスプレッドせず、buildFiveHarnessEvidence が返す rawReceipt を基に RawIntentLiveReceipt
のフィールドを明示的に構築してください。outcome は skipped、observation は null
とし、CommittedValidatedIntentLiveReceipt 固有の検証フィールドを含めないようにします。
---
Outside diff comments:
In `@docs/reference/06-hooks-and-tools.ja.md`:
- Around line 256-273: Synchronize the Stop hook documentation with Intent
autonomy in docs/reference/06-hooks-and-tools.ja.md lines 256-273 and
docs/reference/06-hooks-and-tools.md lines 258-275, updating the hook overview
and no-progress counter descriptions from autonomous Construction to Intent
autonomy. Document that both semi and full use a block cap of 8, while only full
suppresses the human-wait carve-outs; apply the equivalent changes in each
language.
In `@packages/framework/core/tools/amadeus-advisory-choice.ts`:
- Around line 890-897:
提示時のcheckpointとintentRunに一致するpendingだけを、recordProtectedAdvisoryChoice内のopenフィルタで選択してください。既存のclosedAt、timestamp、acceptsFreshChoice条件は維持し、異なるcheckpointのpendingを除外してRationaleの照合対象を限定してください。異なるcheckpointが混在する場合でも正しい回答でholdが解除される回帰テストを追加してください。
In `@packages/framework/core/tools/amadeus-orchestrate.ts`:
- Around line 3593-3612: Route the unfinished per-unit directive through
routeMainWorkflowDirective instead of emitting it directly at the directive.unit
path. Match the final-unit path’s routing so semi and full modes populate
intent_autonomy_mode, autonomy_auto_approve, quality_repair, and
intent_grant_id, while preserving the existing gate, next_stage removal, and
unit assignment behavior.
In `@packages/framework/core/tools/amadeus-state.ts`:
- Around line 2921-2922: Update the approval-gate error message in error() to
replace the retired “autonomous Construction is exempt” guidance with the Intent
autonomy-mode selection wording used by AUTONOMY_LADDER_EXIT in amadeus-lib.ts.
In `@tests/unit/t211-swarm-batch-progress.test.ts`:
- Around line 277-305: Update codegenStateWithAutonomy and seedAutonomyProject
to accept the Intent-mode union "none" | "semi" | "full"; replace callers using
the legacy values with the corresponding new values, remove the unknown-value
fallback branch, and preserve the generated state mappings for each supported
mode.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
j5ik2o
force-pushed
the
codex/intent-autonomy
branch
from
August 4, 2026 10:05
aed5414 to
0b477e9
Compare
Compile declarative loop manifests into a harness-neutral state machine, persist canonical event sets with a repairable replay index, and gate Judge dispatch and latch recovery on durable receipts. Project the shared implementation to the five supported harnesses.
j5ik2o
force-pushed
the
codex/intent-autonomy
branch
from
August 4, 2026 11:03
0b477e9 to
079a4ce
Compare
Member
Author
This was referenced Aug 4, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
概要
Intent単位で
none/semi/fullの自律レベルを選択できるようにし、完全自律ではIntent終端まで進行できるCore契約を導入します。PRの作成・レビュー・マージはAI-DLC Coreから疎結合のまま維持します。主な変更
検証
bun tests/run-tests.ts --ci --verbose: 808 files、10,806 assertions、失敗0bun run build: rebase後の最新mainで成功Closes #2067
Closes #2095
Closes #2096
Note
Low Risk
Changes are markdown/JSON workflow records under
amadeus/spaces/default/with no application runtime or auth logic touched in the shown diff.Overview
This diff adds the completed AI-DLC paper trail for initiative
260803-intent-autonomy(#2095 → #2096 → #2067), rather than changing Core runtime code in the files shown.It introduces
amadeus/spaces/default/intents/260803-intent-autonomy/amadeus-state.md, marking the workflow completed atbuild-and-testwith Construction Autonomy Mode: autonomous, mirror receipts through workflow completion, and project sync to Done.It registers nine new elections in
elections.jsonand full ballot/ledger/tally packages underamadeus/spaces/default/elections/: early-stage learning picks (umbrella Intent + dependency order, Issue acceptance as Must-have, static RE evidence vs unverified live harness behavior) and §13 stage elections—mostly zero-confirm “0 learnings to persist”, plus NFR Design choice 6 (design from complexity/partition/atomicity without inventing SLOs) after ballot amendments and discussion holds.codekb/amadeus/re-scans/260803-intent-autonomy.mdis a new differential refresh that documents observed baseline gaps (GAP-01–21) and explicitly states target Issue vocabulary was not present at the scanned commit—useful context if reviewers compare this artifact set to the product PR that claims Loop Monitor / Quality Repair implementation elsewhere.Reviewed by Cursor Bugbot for commit b1033df. Configure here.
Summary by CodeRabbit
none・semi・full）を選択可能にしました。