@tea-agent/loop-agent 0.39.0-beta.8 → 0.39.0-beta.9

package/docs/templates/spec-registry.schema.json

@@ -0,0 +1,45 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://loop-agent.local/schemas/spec-registry.schema.json",
+  "title": "Frontend specification registry",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["version", "roots"],
+  "properties": {
+    "version": { "const": 1 },
+    "roots": {
+      "type": "array",
+      "maxItems": 32,
+      "items": {
+        "type": "object",
+        "additionalProperties": false,
+        "required": ["path"],
+        "properties": {
+          "path": {
+            "type": "string",
+            "minLength": 1,
+            "pattern": "^(?!/)(?!.*//)(?!.*\\\\)(?!.*(?:^|/)\\.{1,2}(?:/|$)).+$"
+          },
+          "scope": {
+            "type": "array",
+            "items": { "type": "string", "minLength": 1 },
+            "uniqueItems": true
+          },
+          "mode": { "enum": ["normative", "advisory"] },
+          "description": { "type": "string" }
+        }
+      }
+    },
+    "discovery": {
+      "type": "object",
+      "additionalProperties": false,
+      "properties": {
+        "enabled": { "type": "boolean", "default": true },
+        "maxDepth": { "type": "integer", "minimum": 0, "maximum": 8, "default": 3 },
+        "maxFiles": { "type": "integer", "minimum": 1, "maximum": 200, "default": 100 },
+        "maxBytes": { "type": "integer", "minimum": 1, "maximum": 10485760, "default": 1048576 },
+        "timeoutMs": { "type": "integer", "minimum": 1, "maximum": 10000, "default": 2000 }
+      }
+    }
+  }
+}

package/harness.json

@@ -82,9 +82,9 @@
   "executors": {
     "pi": {
       "description": "Pi 负责规划、评审、诊断；当 DAG toolProfile=write 时也可做有界写入。模型按复杂度三档配置，支持 provider/model 字符串或带 thinking 的对象；斜杠前为 Pi provider，后为 modelId，勿只写裸 modelId。",
-      "LOW": "deepseek/deepseek-v4-flash",
-      "MED": "deepseek/deepseek-v4-flash",
-      "HIGH": "deepseek/deepseek-v4-flash"
+      "LOW": "wizard-local/minimax-m3",
+      "MED": "wizard-local/gpt-5.6-sol",
+      "HIGH": "wizard-local/gpt-5.6-sol"
     }
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tea-agent/loop-agent",
-  "version": "0.39.0-beta.8",
+  "version": "0.39.0-beta.9",
   "type": "module",
   "bin": {
     "loop-agent": "bin/loop-agent.js",

package/skills/frontend-bounded-implement/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: frontend-bounded-implement
 description: >-
-  Use only for frontend implementation and repair writers after the canonical
+  Use only for frontend implementation writers after the canonical
   contract gate has authorized a bounded write set.
 ---
 
@@ -11,7 +11,6 @@ description: >-
 
 - For implementation, use the run-owned
   `contracts/frontend-implementation-contract.json`.
-- For repair, also use `contracts/frontend-repair-assessment.json`.
 - Treat requirement `expectedOutcome`, interaction `trigger` /
   `expectedBehavior`, target files, UI states, Mock/API strategy, and verification
   targets as closed decisions. Do not rediscover or reinterpret them.
@@ -25,7 +24,7 @@ description: >-
 - Start the bounded edit after confirming the contract and targets. Optional
   context is not a reason to postpone implementation.
 - If behavior is still ambiguous or a required target is outside `writeSet`, stop
-  with `IMPLEMENTATION_OUTCOME: blocked`; do not compensate with broader scouting.
+  without writing and explain the blocker; do not compensate with broader scouting.
 
 ## Implementation Discipline
 
@@ -39,15 +38,10 @@ description: >-
 - Run only the focused checks already frozen by the task. Leave full verification
   to downstream deterministic nodes.
 
-## Required Outcome
+## Completion Summary
 
-The first non-empty response line must be exactly one of:
-
-- `IMPLEMENTATION_OUTCOME: changed` — a non-empty bounded diff was produced.
-- `IMPLEMENTATION_OUTCOME: already-satisfied` — the contract was already met and
-  no file changed.
-- `IMPLEMENTATION_OUTCOME: blocked` — implementation cannot proceed without
-  guessing or exceeding the write boundary.
-
-Then report changed files, requirements/interactions implemented, tests or focused
-checks attempted, and residual risks. Never report `changed` for an empty diff.
+Briefly report changed files, implemented requirements/interactions, focused checks
+attempted, blockers, and residual risks. This summary is human-readable only: the
+runtime derives implementation state from persisted write events, run delta, write
+guard, requirement coverage, and focused checks. Do not use a mandatory
+`IMPLEMENTATION_OUTCOME:` response-text protocol.

package/skills/frontend-design-review/SKILL.md

@@ -9,31 +9,24 @@ references:
 
 # Frontend Design Review
 
-For frontend design review nodes. Read the checklist; audit contract, scout, Mock
-strategy, effective plan, task bounds, and traceable design evidence.
-Knowledge base, OpenSpec, and `ai_workspace/` are parallel sources. Query the
-knowledge-base connector when available; always search/read `openspec/schemas/**`,
-`openspec/project-specs/**`, and `ai_workspace/**` before accepting conventions.
-Connector format is TODO: never invent results.
+For frontend design review nodes. Read the checklist; audit the canonical contract,
+bound Mock strategy, task bounds, and traceable design evidence supplied by the
+upstream context. Do not rediscover task sources, search specification roots, or
+replace frozen evidence with nearby-code conventions. Use only exact, bound evidence
+paths when additional reading is necessary; never invent connector results.
 
-## Verdict Contract
+## Authoritative Terminal
 
-First non-empty line: exactly `VERDICT: pass` or `VERDICT: request-revision`.
-Any blocker requires request-revision. An initial pass is the effective verdict;
-there is no intermediate first-design shell gate.
+Complete the review with exactly one committed typed tool call:
 
-In standard/full topology, revision and final review run only after initial
-`VERDICT: request-revision`. Verify every Required Plan Correction against the
-complete revision; reject remaining, incomplete, or new gaps. An initial pass uses
-the original plan and skips both conditional nodes.
+- `approve_design` when no blocking design defect remains;
+- `request_design_changes` when a blocker exists. It must include a typed
+  `issueCategory`, at least one `evidenceRef`, and non-empty findings.
 
-Small-risk topology instead removes initial review and revision, then runs
-`frontend-final-design-review-pi` as the sole review of the original plan. Do not
-expect initial findings or a revision there.
-
-The deterministic gate prefers final-review output and falls back to initial review
-only when no current-run final output exists. Existing malformed/non-pass primary
-output fails closed.
+The typed terminal is the only authoritative outcome. Do not require or emit a
+`VERDICT:` line, JSON verdict, revision patch, or a second final-design-review
+verdict. A requested design change ends this run; the recovery controller creates
+any bounded follow-up run from the appropriate phase.
 
 ## Blocking Conditions
 
@@ -51,29 +44,15 @@ Knowledge-base absence is advisory if relevant rules from either openspec
 specification directory were read and applied. Block skipped fallback, unresolved
 conflict, or unresolved UI decisions.
 
-## Method And Output
+## Method And Human-readable Output
 
 Map criteria to steps/files/states/checks, audit paths/evidence, classify Blocking
 or Advisory, and never edit files.
 
 ### Spec Evidence Rules
 
-Run `grep`/`find`, then explicit `read` calls for applicable specs/checklist. Only
-successful paired reads count as “已读取规范文件”; summaries do not. List each path/
-section in `Checked Items`; search/read both openspec specification directories
-before accepting conventions.
-
-```markdown
-VERDICT: pass
-
-## Findings
-- None blocking.
-
-## Required Plan Corrections
-- None.
-
-## Checked Items
-- ...
-```
-
-Each correction names its criterion/section, gap, and exact plan change.
+Read only the canonical contract and any exact evidence paths bound in its upstream
+context. Successful paired reads are the only proof that a specification was read;
+summaries do not substitute for them. A concise human-readable summary may list
+checked items and findings, but it is never a control protocol. Each requested
+change names its criterion/section, gap, and required plan change.

package/skills/frontend-implementation/SKILL.md

@@ -5,9 +5,8 @@ description: >-
 references:
   - path: references/node-contracts.md
     required: true
-    # Current file is ~6.3KB; cap aligned with actual size so the required
-    # reference injects untruncated.
-    maxChars: 7000
+    # Keep the required node contract intact under the bounded-skill test budget.
+    maxChars: 9000
   - path: references/design-spec.md
     required: true
     maxChars: 3500

package/skills/frontend-implementation/references/design-spec.md

@@ -55,6 +55,7 @@ and conflicts. `openspec fallback` includes terms, paths/headings/lines, rules,
 - Cite knowledge-base or openspec spec evidence for component/token choices;
   label weaker repo fallback.
 - Make deviations and unresolved gaps explicit.
-- Contract `uiComponentChoices[]` per UI purpose: `specified` needs
-  `specReference {path,section,line}`; `new` needs null ref + rationale; never
-  silently replace a spec-defined component.
+- Contract `uiComponentChoices[]` per UI purpose: `specified` needs an OpenSpec
+  `specReference {path,section,line}`; a task-source-mandated `new` component
+  needs a PRD `specReference {path,section,line}` plus rationale; never silently
+  replace a spec-defined component.

package/skills/frontend-implementation/references/node-contracts.md

@@ -8,7 +8,7 @@ Pre-write nodes are read-only. Preserve IDs, labels, commands, language, require
 - **`frontend-contract-pi`**（仅 cited 且候选非空）：在既有 Markdown contract 的末尾输出一个 `frontend-openspec-selection-v1` JSON block，对 generation-frozen specRoots 候选逐一分类为 `required` / `relevant` / `irrelevant`；显式 `requiredReadPaths` 和任务源文件引用必须为 `required`。只读，不改 task.json，不新增节点。
 - **`frontend-scout-pi`**: routes, components, tokens, data/API/Mock, scripts, tests, assets; fact vs inference vs gap. Start from the generation-frozen OpenSpec capability context and contract selection block; then perform bounded task-term searches only for unresolved relevant rules. Do not recursively read every candidate directory. Output stack, styling, conventions, state/data, test entry points, reuse, risks.
 - **`frontend-plan-pi`**: AC → observable `expectedOutcome`; interactions → `trigger`+`expectedBehavior`; include steps, in-bound files, UI states, reuse, deps, Mock/API, frozen verify entrypoints and integration gaps. DAG freezes a protected skeleton (`schemaVersion`, `sourceBinding`, `riskLevel`, `targets.files`, `mockApi.productionDefaultOff`). The model commits the plan decision ledger incrementally through typed `record_*` tools (target-surface / component-choice / state-flow / data-flow / mock-api / design-deviation / dependency) and closes with exactly one `finalize_plan`; narrative prose stays free-form context and is never the compile authority. `frontend-design-policy-shell` assembles the canonical contract from committed facts (validated, fail-closed — no fenced-JSON text parsing) and materializes `frontend-implementation-contract-v1`; unresolved optional symbols are dropped and audited. `frontend-design-policy-shell` deterministically renders plan.md. IDs+paths ≠ behavior semantics. Applicable UI states need behavior/implementation/verification; non-applicable states need reasons. Active Mock needs frozen commands; `auto` without confirmed capability/command → `not-needed`; `required` without command → generation-time blocked.
-  - `uiComponentChoices[]` (optional): `{purpose, component, decision, specReference, rationale}`; `specified` needs `specReference{path,section,line}`; `new` needs null ref+rationale.
+  - `uiComponentChoices[]` (optional): `{purpose, component, decision, specReference, rationale}`; `specified` needs an OpenSpec `specReference{path,section,line}`; a task-source-mandated `new` component needs a PRD `specReference{path,section,line}` plus rationale; `reuse-existing` may omit the ref.
 - **`frontend-design-policy-shell`**: read-only deterministic design check (UI coverage, evidence resolvable/fresh, Mock strategy bound to frozen commands, component choice `specReference`, targetPaths boundary, dependency authorization, UI state ownership, requirement coverage); passes by materializing the single authoritative `frontend-implementation-contract-v1`. Blocked ends the run without a model reformat step.
   - `verificationTarget.commandLabel` must be a **frozen command label** from `verifyEvidence.commandLabels`; others fail closed (`invalid-output`). Empty frozen set (no `run.json`) skips check.
   - `mockApi.strategy !== "not-needed"` fails closed when `mockCommandLabels` empty (`no authorized Mock verification commands`).
@@ -16,14 +16,15 @@ Pre-write nodes are read-only. Preserve IDs, labels, commands, language, require
   - Component conformance: missing `uiComponentChoices` (frozen bucket non-empty + UI-visible work) → `component-choices-missing`; bad/uncited `specReference.path` → `component-spec-reference-invalid` / `component-spec-not-cited`; at least one `decision=specified` choice must bind to the frozen component/theme bucket → `component-spec-choice-unbound`.
   - Canonical serialization (UTF-8) >64KB → `retryable-invalid` + `contract-too-large`.
 - **`frontend-writer-admission-shell`**: sole write authorization for `frontend-implement-pi`. Derives and freezes the concrete writeSet (implementation targets ∪ non-static verification target files), pre-writer baseline, and admission digest (Mock strategy, allowed strategies digest, frozen command digest, `productionDefaultOff`, real-integration expectation); writes `contracts/frontend-writer-admission-result.json` (schemaId `frontend-writer-admission-shell-v1`). The scheduler checks before queueing and before executing the writer: only `admitted` releases it, otherwise SKIPPED + `frontend-prewrite-not-authorized`. Admission blocked (`candidate-contract-invalid` / `prewrite-blocked`) is forced `failed` at aggregation.
-- **`frontend-implement-pi`**: sole regular exclusive writer; uses `frontend-bounded-implement` (not this skill). Stay in `writeSet`; real requests default-on. Atomic handler/intercept/adapter with tests. Stop on forbidden paths/guesses. First line `IMPLEMENTATION_OUTCOME: changed|already-satisfied|blocked`; runtime checks against attributed diff. Optional mock-verify when frozen; static+behavior always; behavior must prove page consumption. `not-needed` keeps real integration pending unless real backend has fresh evidence.
+- **`frontend-implement-pi`**: sole regular exclusive writer; uses `frontend-bounded-implement` (not this skill). Stay in `writeSet`; real requests default-on. Atomic handler/intercept/adapter with tests. Stop without writing on forbidden paths or unresolved ambiguity. Runtime derives implementation state from persisted write events, run delta, write guard, requirement coverage, and focused checks; any response summary is non-authoritative. Optional mock-verify when frozen; static+behavior always; behavior must prove page consumption. `not-needed` keeps real integration pending unless real backend has fresh evidence.
 
 ## Typed terminal protocol
 
 - `frontend-plan-pi` commits incremental `record_*` facts plus one `finalize_plan` (plan).
 - `frontend-design-review-pi` commits `approve_design` | `request_design_changes` (request carries typed issueCategory + evidenceRefs + findings).
 - `frontend-review-pi` commits `approve_review` | `request_review_changes` (request carries typed issueCategory + evidenceRefs + findings).
-- These committed typed tools replace `VERDICT:` first-line parsing; the verdict is consumed only as deterministic data input by downstream admission/closeout.
+- These committed typed tools are the only design/review terminals. They replace
+  `VERDICT:` first-line parsing; response text is human-readable context only.
 
 ## Contract / trace / stages (M1–M2)
 

package/skills/frontend-review/SKILL.md

@@ -9,17 +9,27 @@ references:
 
 # Frontend Review
 
-Use for `frontend-review-pi`; read the findings guide first. Required inputs are
-original task/reference material, effective plan/design branch, implementation summary,
-and `contracts/frontend-review-context.json`. That canonical context binds the validated
-implementation contract, effective verification trace, repair assessment, optional
-`frontend-lint-assessment-v1`, and run-owned diff. The diff is authoritative.
+Use for `frontend-review-pi`; read the findings guide first. The required input is
+`contracts/frontend-review-context.json`, its bound diff summary/fragments, and only
+the current source files named by that diff. The canonical context binds the validated
+implementation contract, effective verification trace, optional
+`frontend-lint-assessment-v1`, and run-owned diff. Do not reopen original task sources,
+plans, design-review prose, OpenSpec, or `ai_workspace/`; the bound diff is
+authoritative.
 
-## Verdict Contract
+## Authoritative Terminal
 
-The first non-empty line must be exactly `VERDICT: pass` or
-`VERDICT: request-revision`. Any Critical/Important finding, failed or missing
-required check, forbidden write, or unmet acceptance criterion forces revision.
+Complete the review with exactly one committed typed tool call:
+
+- `approve_review` when no Critical/Important finding remains;
+- `request_review_changes` for a defect, missing required evidence, forbidden
+  write, or unmet acceptance criterion. It must include a typed `issueCategory`,
+  at least one `evidenceRef`, and non-empty findings.
+
+The typed terminal is the only authoritative outcome. Do not require or emit a
+`VERDICT:` line or response-text JSON verdict. A request for review changes ends
+this run and is handled by the recovery controller rather than a same-run repair
+branch.
 
 ## Review Scope
 
@@ -38,11 +48,9 @@ required check, forbidden write, or unmet acceptance criterion forces revision.
   baseline-matched diagnostics on unchanged files, and none on writer-changed files.
   Report debt, never `lint passed`; `failed`/`unavailable` blocks. Typecheck, build,
   and test still require successful final exits.
-- Component/design claims need both knowledge-base and local openspec evidence. Query
-  the connector when available and always read task-relevant files under
-  `<repoRoot>/openspec/schemas/`, `<repoRoot>/openspec/project-specs/`, and `<repoRoot>/ai_workspace/`.
-  Connector format is TODO; never invent evidence. Only successful `read` calls are
-  observable as "已读取规范文件".
+- Evaluate component/design claims against the evidence already bound in the canonical
+  contract and review context. Do not use a repository-wide search or connector query
+  to replace that frozen evidence.
 - Treat shell exit status as authoritative. Do not edit files.
 
 ## Evidence And Output

package/skills/frontend-review/references/review-findings.md

@@ -14,24 +14,23 @@
   `baseline-debt` only with a valid `frontend-lint-assessment-v1`; this is not a
   passed lint result. A changed-file lint diagnostic, new unmatched diagnostic,
   command drift, missing baseline, timeout, or unparseable output is Important.
-- Treat knowledge base, `openspec/schemas/`, `openspec/project-specs/`, and `ai_workspace/` as
-  parallel sources. Record connector/query/source/time plus openspec search terms and
-  matched paths/headings; label fallback or unavailable accurately.
+- Treat the design evidence already bound in the canonical review context as the
+  source of truth. Do not launch a connector query or repository-wide OpenSpec/
+  `ai_workspace` search to replace frozen evidence; label a missing bound evidence
+  reference accurately instead.
 
 ## Review Sequence
 
 1. Establish changed-file inventory and write boundaries.
-2. Compare original requirement with derived contract/constraints.
+2. Compare the bound contract with the approved constraints and cumulative diff.
 3. Map each criterion to code, states, tests, and evidence.
 4. Inspect interactions, state/data/API behavior, failure paths, and regressions.
 5. Check mock selection, contract-to-fixture mapping, activation/default path,
    handler/fixture/adapter and consumer diff, verification, production imports, and
    the real-integration gap.
-6. Check design evidence, responsive/a11y behavior, dependencies, and maintenance.
+6. Check bound design evidence, responsive/a11y behavior, dependencies, and maintenance.
 7. Classify findings and derive the verdict mechanically.
 
-Skipping either directory is Important when design compliance affects acceptance.
-
 ```text
 - [Critical|Important|Minor] path:line — Problem; impact; required correction; evidence.
 ```

package/skills/loop-agent/references/hybrid-dag.md

@@ -52,6 +52,13 @@ loop-agent dag execute --dag <temp-dir>/hybrid-dag.json --init-only --canvas-pat
 - 顶层：`objective`、`successCriteria`、`globalConstraints`、`defaults`、`skillsByRole`、`executorModels`
 - 每 node：`role`、`skills`、`writePolicy`、`writeSet`、`piStep`、`shell`、`outputContract`、`executor`（`pi` | `shell` | `static`）；Pi 写入节点额外声明 `toolProfile: "write"`；安全只读 Pi 节点可选声明 `retryPolicy`（生成器自动注入默认值）
 
+**Source binding v2（frontend source fidelity ledger）**：
+
+- `dagSourceBindingSchema` 是 `schemaVersion` discriminated union：v1（`schemaVersion: 1`，strict 拒绝 ledger 字段）与 v2（`schemaVersion: 2`，`ledgerPath`/`ledgerSha256`/`inputDigest`/`requirementToFragments` 全部必填）。
+- 仅 `frontend-implementation` 且存在可解析 requirement/acceptance 引用时产出 v2；绑定完全来自 `task advance` 已持久化的 `source/requirement-ledger.json`（intake/apply/prewrite 同一份受管账本），缺失/损坏/漂移 fail closed（`SOURCE_FIDELITY_LEDGER_MISSING`/`SOURCE_FIDELITY_LEDGER_STALE`），不静默降级为 v1。
+- v1 不得携带 optional ledger 字段冒充 v2；prewrite 门禁对 v2 校验 ledger 字节 sha256、inputDigest 新鲜度与 requirement→fragment 覆盖，任一失败以稳定 `source-fidelity-*` 码阻断 writer。
+- ledger `sourceDocuments[].path` 与 DAG 侧 `serializeDagTaskSourcePath`/`resolveDagTaskSourcePath` 单一边界一致（repo-root relative），路径转换不得二次进入 digest 计算。
+
 **Model 生成 DAG 的 template 卫生**：
 
 - 执行前将所有 `REPLACE/WITH/...` placeholder 换为具体 repo path。

package/dist/worker/console/static/assets/channel-AbFtVnFm.js

@@ -1 +0,0 @@
-import{ag as o,ah as n}from"./mermaid.core-BlvpGVG8.js";const t=(a,r)=>o.lang.round(n.parse(a)[r]);export{t as c};

package/dist/worker/console/static/assets/classDiagram-JCYQIIEL-n1yf2Jcv.js

@@ -1 +0,0 @@
-import{s as a,c as s,a as e,C as t}from"./chunk-GF5L2VYU-B4n-03Qm.js";import{_ as i}from"./mermaid.core-BlvpGVG8.js";import"./chunk-5VM5RSS4-BzRFrN8j.js";import"./chunk-XXDRQBXY-9boOMji4.js";import"./chunk-KBJHAD2P-B_afMhHB.js";import"./chunk-2GRJ4B5K-Dd4g2nti.js";import"./index-DBh_x_Ek.js";var n={parser:e,get db(){return new t},renderer:s,styles:a,init:i(r=>{r.class||(r.class={}),r.class.arrowMarkerAbsolute=r.arrowMarkerAbsolute},"init")};export{n as diagram};

package/dist/worker/console/static/assets/classDiagram-v2-OCEON4UE-n1yf2Jcv.js

@@ -1 +0,0 @@
-import{s as a,c as s,a as e,C as t}from"./chunk-GF5L2VYU-B4n-03Qm.js";import{_ as i}from"./mermaid.core-BlvpGVG8.js";import"./chunk-5VM5RSS4-BzRFrN8j.js";import"./chunk-XXDRQBXY-9boOMji4.js";import"./chunk-KBJHAD2P-B_afMhHB.js";import"./chunk-2GRJ4B5K-Dd4g2nti.js";import"./index-DBh_x_Ek.js";var n={parser:e,get db(){return new t},renderer:s,styles:a,init:i(r=>{r.class||(r.class={}),r.class.arrowMarkerAbsolute=r.arrowMarkerAbsolute},"init")};export{n as diagram};