docs: record Coding session stall diagnosis

.project-docs/30-worklog/tasks/20260901-diagnose-session-stall-c398db33.md

@@ -0,0 +1,69 @@
+# Task: Diagnose in-app Coding session stall
+
+## Identity
+
+- Task ID: 20260901-diagnose-session-stall-c398db33
+- Mode: Feature
+- Branch: main
+- Worktree: /Users/inmanx/Documents/makelore
+- Base commit: 8062b45103be45ae036167d3e9890c3ca9ba3e80
+- Owner: codex
+- Status: Completed
+
+## Scope
+
+- Diagnose whether the visible Makelore Code run in project `肉鸽小游戏`, Agent
+  `小小明`, Conversation `新对话` is still doing useful work or is stuck.
+- Correlate the supplied screenshot with the live Electron/Pi process, Main log,
+  project metadata, and persisted Pi session timeline.
+- Make no product-code change and do not abort, recover, or otherwise mutate the
+  user's live Conversation.
+
+## Intent And Constraints
+
+- Treat the screenshot's earlier Token Point failures separately from the latest
+  `继续` turn; do not infer that a historical error is the current terminal cause.
+- Preserve Main ownership of runtime state and the rule that accepted work is not
+  automatically replayed.
+- Use only read-only runtime diagnostics. Do not expose credentials, raw upstream
+  request details, or unrelated Conversation content.
+
+## Outcome
+
+- Confirmed that the visible run is stuck in its terminal-settlement path rather
+  than continuing model or tool work.
+- The latest turn began at 11:50:10 local time. All observed upstream requests for
+  that turn returned HTTP 200, its tools completed, and Pi persisted a final
+  assistant message with `stopReason: stop` at 11:55:38.
+- The screenshot was captured at 12:02:14 while the UI still displayed
+  `处理中 12 分 3 秒` and the stop control. At 12:10:08 the Pi session and Main log
+  were still unchanged, the shared Agent Server remained alive but idle at 0% CPU,
+  and it held no TCP connection.
+- The earlier `词元点数余额不足` rows are historical failed branches and did not
+  cause the latest run to wait. The effective failure boundary is after final
+  `message_end`: the run did not become idle, so `agent_settled` was either not
+  emitted/delivered or did not take effect in Main projection.
+- No live run, process, project file, or product source was changed.
+
+## Verification
+
+- Correlated project/Agent/Conversation IDs from the local schema-v2 metadata.
+- Compared screenshot mtime, Pi JSONL mtime, Main log mtime, and wall-clock time.
+- Inspected the final persisted Pi message and the preceding completed tool results.
+- Checked the shared Agent Server process state, CPU time, and open TCP sockets.
+- Checked ADR-006 and the current event projector: `message_end` completes the
+  assistant node, while authoritative `agent_settled` is required to set the run
+  idle and release run ownership.
+
+## Follow-ups
+
+- If a fix is requested, reproduce a long real tool-use turn whose final assistant
+  message is persisted but settlement is missing, then harden the Agent Server/Main
+  terminal handshake without replaying the accepted prompt.
+- Add regression coverage for final `message_end` followed by missing/delayed
+  `agent_settled`, including safe recovery and release of run/background ownership.
+
+## Promotion Candidates
+
+- None. This task records a live diagnosis only; any product behavior change needs a
+  separate implementation task and verification.