[linkerlin]

Summary

未完成 run 的 tick 保险丝与连上时的常驻说明对齐到 #20 的信封：码在 loop_fuse=pending_redirect_loop，保险丝的 redirect_reason 用同一个码。裸 pending_redirect_loop 不再被写成 next_action 的并列值，error 字段仍然不出现。

The unfinished-run tick fuse and the connect-time instructions now match the #20 envelope. The code is loop_fuse=pending_redirect_loop, and the fuse redirect_reason uses that same code. A bare pending_redirect_loop is no longer written as a peer of next_action, and error stays absent.

What changed

• 常驻 standing_instructions() 把 or pending_redirect_loop 改成 or loop_fuse=pending_redirect_loop。不写 error=pending_redirect_loop，也不把裸码当成 next_action。全文 1195 字，仍小于 1200。
• 保险丝（closed）在无软缺口的早返回和有软缺口的返回上都写 redirect_reason=pending_redirect_loop。dispatch_reason 仍是 pending_solidify。loop_fuse、ok=false、error 缺席、session_not_failed=true、preserve 三元组、tick_again=false 不动。有软缺口时仍不写 gate_soft_continue。
• 第一次重定向仍是 redirect_reason=pending_solidify，没有 loop_fuse，ok=true。next_action 两次都是 swarm_solidify。
• 仪器、swarm_tick 参数说明、README（英/中）、quickstart、AGENTS、DEBUG #68 写同一口径。DEBUG #61 的 tools/list 缓存没有做假修复。
• 不写 Parent baseline，不放宽固化门，不打开 val 密封，不调用 skip_validation。

How to test

1. uv run pytest tests/test_mcp_host_usability_pending_smoke.py tests/test_mcp_host_usability_p0.py tests/test_mcp_host_usability_p1.py tests/test_mcp_host_usability_p2.py tests/test_mcp_host_usability_polish.py tests/test_swarm.py tests/test_mcp_server.py tests/test_mcp_protocol.py tests/test_open_box_setup_doctor.py tests/e2e/test_swarm_full_e2e.py tests/test_usability_external_state_e2e.py tests/test_swarm_instrument_soft_copy.py tests/test_swarm_beginner_guidance.py -m "not slow and not llm"
2. 期望：190 passed，3 deselected（llm/slow）。本地已得到这个结果。
3. uv run ruff check src/evolver/swarm.py src/evolver/mcp_server.py tests/test_mcp_host_usability_pending_smoke.py tests/test_mcp_server.py
4. uv run ruff format --check 同上四个文件
5. uv run mypy src/evolver/swarm.py src/evolver/mcp_server.py

常驻说明含 loop_fuse=pending_redirect_loop，不含裸 or pending_redirect_loop、error=pending_redirect_loop、next_action=pending_redirect_loop。保险丝：loop_fuse=pending_redirect_loop、redirect_reason=pending_redirect_loop、无 error、ok=false、session_not_failed=true、preserve 三元组仍在；软缺口省略 gate_soft_continue。第一次重定向：redirect_reason=pending_solidify，没有 loop_fuse。boot / status 仍是 gate_soft_continue=true，没有 session_not_failed。只查 error == pending_redirect_loop 的宿主模型会错过保险丝；查 loop_fuse 或保险丝 redirect_reason 的模型会去固化。

Risk

Low — 公开 MCP 返回体只在保险丝上把 redirect_reason 从 pending_solidify 改成 pending_redirect_loop。dispatch_reason、ok、error 缺席、next_action、session_not_failed 与 preserve 三元组没变。常驻说明换了一句，长度从 1196 降到 1195。固化门、Parent baseline、val 密封都没有动。

仍在的宿主摩擦

• 把缺席的 gate_soft_continue 当成「继续 tick」（只在严格 false 时才停），并且不看 tick_again / next_action / ok 的宿主，仍可能再 tick。保险丝上 ok=false 仍能停住只看 ok 的宿主。
• 把 ok=false 本身当成整段会话失败、且不读 session_not_failed 与 loop_fuse 的宿主，仍会在第二次 tick 上停掉会话。ok=false 留着，是因为只看 ok 的宿主靠它停止空转。
• 宿主 tools/list 缓存滞后（DEBUG #61）没有做假修复。安全信号在返回体上。

Harness/evaluator governance

Upstream governance surface: MCP swarm host contract (standing_instructions, unfinished-run tick fuse redirect_reason / loop_fuse, instrument text, and tool schema). Solidify gates, Parent baseline, and val seals are unchanged.
Downstream EvoX impact: N/A -- not a harness/evaluator governance change
Rollout-local scope: N/A -- not a harness/evaluator governance change
Promotion boundary: N/A -- not a harness/evaluator governance change
Evaluator mismatch sets: N/A -- not a harness/evaluator governance change
Non-regression evidence: pending-smoke host models plus the 190-test host-usability / swarm / MCP / e2e suite (not slow and not llm)
Fix-severity review: low
Owner approval: N/A -- not a harness/evaluator governance change
Security boundary: host-facing MCP envelope and connect-time instructions only; no new data, tool, or secret path
Rollback: revert this commit
Live promotion: no
Autonomous evaluator self-editing: no

Self-check

• Tests added or updated to cover the new behavior; gates pass locally on the host-usability suite named above, plus ruff and mypy on the touched modules.
• No new runtime dependencies.

  

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.