fix(explore): exclude unready todos from dispatch suggestions

[Jim-jimu]

Goal And Delivered Outcome

• Reproduced defect: Explore can select B while its completion dependency A is unfinished, suggest acquiring B's execution lease, and omit B's waiting condition from the Agent's turn context.
• Before → after: both planners now exclude unready Todos from current dispatch, worker bundles and resource allocation. Waiting conditions remain visible in bounded diagnostics. Completing A makes B selectable on the next plan; handing A to another Agent does not.
• Intended base: main. Standalone reproduced defect; no issue is closed. Complements the execution-admission checks in fix(todos): enforce canonical completion dependencies during execution #5884.

Implemented Against

• Specification: docs/reference/canonical-lease-renew.md (handover and completion-dependency admission) and docs/reference/protocols/task-graph-projection-v0.md (typed Todo topology), revision f03a9acf28b726949a855ab54e2f4f1d06799a3f.

Criterion	Disposition	Implementation / validation
Unfinished completion dependencies prevent execution	Implemented in planning suggestions	Both planners reuse todo_item_is_actionable_open; readiness regressions and public CLI journeys
Handover preserves dependencies	Preserved	Transfer A, observe B still waiting, complete A, then acquire and complete B
Successor lineage is not a completion dependency	Preserved	Lineage compatibility case; no new graph interpretation

• Self-check: reviewed planner, context and execution owners; reproduced the defect on the base; verified isolated File/SQLite CLI behavior. Shared readiness projection and exclusions between the existing planners without adding a dependency evaluator.

Scope And Continuation

• Complete within this scope: Todo/worker branch selection, waiting diagnostics, bounded Agent context and empty-selection guidance. No successor required.
• Additive packet fields preserve actionability, status and resume conditions. The Todo plan records the total rejection count; compact context includes up to three rejected candidates and an omission count.
• Existing Explore opt-in, analysis-only mode, quota, claims and execution leases retain their authority. General graph scheduling and priority inheritance are outside scope.

Validation

• Tested revision: 48b1898f3f30e80ea8c5d376e591e56ed14bc10e, including upstream f03a9acf. Older broad runs are identified below.
• Run state: finished
• Input classes: synthetic, public_fixture

Check kind	Result	Public-safe evidence / limitation
regression_parity	passed	Same 25 new regression cases: 23 target-defect failures and 2 compatibility passes on unmodified production at f03a9acf; all 25 pass with the fix, included in the 304 below
unit / integration	passed	Final revision: 304 Python tests across Explore, canonical dependency execution, cold-source imports and backup retention; command below
real_entrypoint / real_backend	passed	tests/control_plane/test_canonical_dependency_execution.py: public CLI with isolated canonical File and SQLite 3.51.3 stores; rejected B handover, successful A handover, unchanged wait, A completion, replanning and B completion
unit / integration	passed	Final revision: 36 TypeScript tests in cold_source_import.test.ts, todo_execution_dependency.test.ts and content_digest_single_owner.test.ts; Node 22.22.3, concurrency 1
static	passed	Final revision: Ruff, mypy, TypeScript typecheck, diff check, semantic advisory and CLI output-budget regression
integration	passed	Final revision: risk-selected premerge with --from-git-diff --git-diff-base origin/main: 5 direct and 19 selected checks passed, no manual holds; includes public-boundary checks
unit / integration	passed	Historical full TypeScript run on the feature contents committed as 81d0382e, before upstream integration: 4,252 passed, 32 PostgreSQL-dependent skips, 0 failed
unit / integration	failed	Historical full Python attempt before integration stopped at approximately 52%: 10,406 passed, 263 skipped, 7 failed. All seven failures reproduce on unmodified f03a9acf and the final revision in the two files listed below

Final focused Python command:

LOOPX_USAGE_PING=0 .venv/bin/python -m pytest -q \
  tests/test_explore*.py tests/capabilities/test_explore*.py \
  tests/control_plane/test_canonical_dependency_execution.py \
  tests/control_plane/test_cold_source*.py \
  tests/control_plane/test_chat_cold_source_import.py \
  tests/control_plane/test_backup_registered_source_retention.py

• Existing failures: tests/cli_commands/test_periodic_report_intent_capability_chain.py and tests/cli_commands/test_todo_complete_settlement_capabilities.py assert missing report intents / settlement completion evidence. They are outside this diff and remain unresolved.
• Coverage includes false/missing readiness, completion and Monitor-change conditions, resource capacity, shared-scope bundles, lineage, foreign ownership, analysis-only/off behavior and bounded diagnostic omissions.
• Gaps: the full Python attempt was interrupted for upstream integration and was not completed; the full TypeScript suite was not rerun after integration. No live-model, packaged App/Lark or PostgreSQL journey was measured. No provider storage or execution-admission rule changes.

Frontend / Visual Evidence

• UI impact: none. Agent-facing CLI/context projections change; no visual surface is modified.
• Before / after, states / viewports and attention review: N/A.
• Source data: none.

Type of Change

• Bug fix
• Documentation update
• Test update

LoopX Area

• Capability or extension (Explore)
• Control plane (Agent context projection)

Technical Direction

• Core control-plane hardening: align Explore dispatch suggestions with existing Todo readiness.

Shared-authority RFC fixture impact

N/A: no provider migration, runtime-routing or shared-authority RFC completion claim. Existing File/SQLite fixtures exercise the affected public readers and unchanged execution lifecycle.

Boundary Checklist

• The diff and this description contain no private state, credentials, raw traces, internal links or local machine paths.
• No maintainer-owned benchmark work was duplicated.
• The change stays within the reproduced planning defect.
• UI impact is marked none.
• Every contribution commit includes a DCO Signed-off-by trailer.

[loopx-agent]

动机

使用 Explore 安排下一步工作的 Agent 与操作者。

前置 Todo A 尚未完成时，旧版可能建议领取等待 A 的 B，造成无效调度；本版把 B 留在诊断中，A 转交仍须等待，A 真正完成后 B 才重新可选。

已验证两个 planner 和 turn-context 的等待诊断、File/SQLite 真实 claim/lease/handoff/completion，以及关闭状态的完整输出对照。

不授予 spawn、claim、lease、quota 或外部发布权限，不声称模型实际采用、全套 CI 或长期运行已验收。

改动思路

共用既有 readiness 和 typed projection，可在一个规则位置维护两 planner，避免新的并行调度状态。

当前边界是正确的计划与等待诊断、现有测试维护和 CI 日志可观察性，不启动执行或扩张权限。

沿用统一的 Todo readiness，先排除不能行动的候选，再评分、分配资源和组 worker bundle。等待状态进入有界诊断；实际执行继续由原 quota/claim/lease 入口裁决。

规范依据：docs/reference/canonical-lease-renew.md，不可变版本 ec21f7da8b9a861fa73477f0ad69e8a4f30e5389。已在阅读实现之前读此版本；对应当前有界改动的接受项如下：

• todo_done: Actual prerequisite done releases the canonical completion wait; transfer and successor lineage do not. implemented.
• todo_dependency_pending: Unsatisfied wait fences execution, and release remains available. implemented.
• resume_ready=true: Read the same current task after actual prerequisite completion before normal acquisition. implemented.

并对照同版本 docs/reference/protocols/task-graph-projection-v0.md：successor lineage 不是完成依赖。

具体改动

审查了全部 13 文件（+336/-80），包括最终加入的 workflow 和 periodic-report 测试维护。两 planner 共享 readiness/exclusion helper；TS turn-context 保留等待条件和总省略数。原 periodic 规则已明确普通完成不产生 milestone；基线七个旧期望失败在相同本地命令复现，本版对齐既有规则，实际 native successor 正例仍通过。CI job ceiling 从 45 改为 60 分钟且输出改为 verbose，未删 collection、test deadline、coverage 或汇总失败门槛；真实两 worker 的分片命令在其余测试仍等待时已输出失败名称。

独立验证：627 focused +71 adjacent Python tests; 158 semantic checks; TS typecheck; premerge five direct checks plus catalog/risk/boundary checks. Baseline native readiness oracle fails both File/SQLite; H passes. Base original periodic tests: seven stale expectation failures; current corresponding tests pass.

额外的完整 public CLI packet 对照覆盖 off、analysis-only、等待、完成和大量无关 Todo；off 仅归一化临时路径及独立验证的路径 hash 后完全一致。File/SQLite 原生 A 转交仍等待，A 完成后 B 领取/完成。另一个拥有更多 fallback 的基线 fixture 本来就没选 B，记录为合法分支，未伪称其失败。

未来改动可维护性检查：本次已把 readiness/exclusion 收敛为两 planner 共同使用的既有规则适配，不需要再建一套调度状态。

对主干的风险

No current material blocker found. 不授予 spawn、claim、lease、quota 或外部发布权限，不声称模型实际采用、全套 CI 或长期运行已验收。 Hosted four-shard full timing not qualified; job ceiling increase preserves tests/coverage. Author description still cites an older tested head, so this review judges the complete current head independently.

未查询、等待或继承 GitHub CI。全仓套件、真实模型对 guidance 的采纳和长期运行未做；现有 UI/消息展示入口没有新增操作，不声称 packaged frontend/Lark 验收。

我的整体评价

该精确 head 的当前有界结果成立，未发现阻塞缺陷，批准代码评审。合并仍是独立权限与仓库策略动作；本次不合并、不升级运行中的安装。

Reviewer: model_agent; runtime_reported; model=gpt-6.1-sol; provider=OpenAI; reasoning_effort=xhigh; execution_observation_id=de1d923ded17a66b1ad1ad14fa37aee1f64208c77ede42beac3b565c69abe4a5

English verdict: APPROVE - exact head 22d5827. Independently validated the whole current diff, real owning entrypoints/backends, recovery and baseline defect sensitivity; no CI was consulted. Parent profile/product qualification and merge authority remain separate.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=xhigh

APPROVE：未发现当前变更造成的阻塞问题；现有 CLI 输出预算仍为单独的 premerge hold。精确 head d8f2fd85b412b6d81200a2feb2179fbb427e6bfa，全差异基线 5e5f0576b15587a9233ce7a2d9571940bf22ad94。本次重新判断完整 16 文件，没有继承旧 head 22d5827553568202f4b502512005f6985ce1e8ad 的批准。

动机

使用 Explore 安排下一步工作的 Agent 与操作者。前置 Todo A 尚未完成时，旧版可能建议领取等待 A 的 B，造成无效调度；本版把 B 留在诊断中，A 转交仍须等待，A 真正完成后 B 才重新可选。已验证两个 planner 和 turn-context 的等待诊断、File/SQLite 真实 claim/lease/handoff/completion，以及关闭状态的完整输出对照。 这是完成现有计划入口的可用修复，不授予 spawn、claim、lease、quota 或外部发布权限，不声称模型实际采用、全套 CI 或长期运行已验收。

改动思路

从真实 Todo 读模型取得当前等待状态，在排序、worker bundling 和资源分配之前应用既有 todo_item_is_actionable_open；等待者只进入拒绝诊断。共用既有 readiness 和 typed projection，可在一个规则位置维护两 planner，避免新的并行调度状态。 TS 仍负责有界 turn-context 投影，Python 仍是既有 planner provider，没有第二套依赖解析或手工同步状态。当前边界是正确的计划与等待诊断、现有测试维护和 CI 日志可观察性，不启动执行或扩张权限。 转交 A 只是换负责人，只有完成 A 才能解除 B 的等待；新的计划也不能替代随后的 quota、claim 和 lease 检查。

具体改动

关键代码讲解

_candidate_readiness 直接复用当前规范谓词，保留 status、resume_when 和 resume_ready。_execution_exclusion 共用于两个 planner；Monitor 继续使用自己的非探索 lane，其他不可执行 Todo 在容量计算之前排除。_build_worker_branch_candidates 不把等待任务并入共同 scope 的工作包。projectExploreTurnContext 只取前三个拒绝项，并从完整 rejection count 派生遗漏数；拒绝项没有领取/租约建议。

基于 docs/reference/canonical-lease-renew.md 在 5e5f0576b15587a9233ce7a2d9571940bf22ad94 的既有规格，todo_done、todo_dependency_pending、resume_ready=true 三项均 implemented：等待条件限制执行，转交保留依赖，前置任务真正完成后重新读回再领取。该文档及 task-graph-projection 契约与上轮基线字节相同。本次覆盖 README、两 planner、TS 投影和相关回归；还检查完整差异中的 CI shard 45→60 分钟作业上限及 verbose 失败名输出、双语验证文档、过期 periodic milestone 测试对齐、真实 mutation locator 修复和归因测试纯格式清理。测试级超时、coverage、worker 数和执行权限均未放宽。三条 Explore 生产模块与旧 head 字节相同，新增验证维护也重新运行。

对主干的风险

独立验证 872 项当前 focused/改动测试和 667 项语义/census 测试通过，TS typecheck、Ruff、diff hygiene 通过。使用隔离的真实 File/SQLite 原生 CLI 复核依赖等待→A 转交后仍等待→A 完成→B 可执行，未改活动 Goal。相同三任务 oracle 在精确基线两个 provider 都失败；带额外 fallback 的合法场景两版都可正确选择其他工作，不能拿它伪造缺陷复现。默认关闭时三个实际入口完整输出在仅核对并归一化 fixture 路径/路径摘要后相同；analysis-only 仍不给命令。缺 readiness、拒绝 acceptance、共同 scope 和诊断遗漏均已覆盖。

语义与 CI 对齐

预检查 advisory 未检测到受支持的新词汇 carrier，独立完整语义检查通过；该观察不能替代契约解释。这里复用既有状态词汇与 typed projection，没有新 authority 或机器义务。原生 premerge 五条直接检查通过，19 项选中检查中 18 项通过，CLI 输出预算失败。相同命令在 5e5f0576b15587a9233ce7a2d9571940bf22ad94 和 d8f2fd85b412b6d81200a2feb2179fbb427e6bfa 的四个 Todo-list JSON 行以相同增长数值失败，故归为 pre_existing_unrelated；这些是未改的 Todo-list reader 路径，Explore 的实际变更不变量有独立通过证据。预算未被放宽，premerge 仍保留红状态；没有查询或等待 CI，也不宣称全套/hosted shard 时长已验收。

我的整体评价

long_horizon 与 user_experience 均 improved：等待任务不再占执行容量，前置任务完成后正常恢复，独立任务仍能继续。完整设计保持 proportionate，未来维护已通过共用 readiness/exclusion helper 改善；没有理由再增加新调度器。额外 mutation locator 测试保护当前真实 mutation runner，格式整理不改变归因规则，整体没有新增状态机制。APPROVE 与单独的 premerge/合并资格 hold 同时成立；全套测试、实际模型采纳、packaged frontend/Lark 和长期运行仍未由本次验证，不能把计划正确当成这些目标完成。

English verdict: APPROVE - d8f2fd8; full current diff reviewed, canonical readiness reused and real File/SQLite lifecycle plus feature-off parity validated. 872 focused and 667 semantic tests pass. Existing CLI budget failures reproduce identically at base/head and remain a separate premerge hold; CI not consulted.