[MAOLVLHL]

What this changes

• Claude Code's tool_result line now becomes a tool_end for the call it answers, carrying status: "error" when the CLI sets is_error.
• Both adapters now emit the CLI's own id as tool_call_id — tool_use.id / tool_use_id on Claude, item.id on Codex.
• Every record carries an agent_run_id, so two runs streaming into one log stay distinguishable even when a CLI hands out the same tool-call id twice.
• close() closes any call that is still open as status: "abandoned", so a killed run no longer leaves a call that looks like it is still running.
• The workflow manual documents the lifecycle.

Nothing downstream changed. The cockpit view and the --progress formatter render records generically, so the new fields are additive for both.

Why

Fixes #182. A Claude Code run produced tool_start records with no matching tool_end: the adapter treated the user line as output echo, which is where stream-json delivers tool_result blocks. The result itself carries only tool_use_id, so recording the CLI's id was a prerequisite for pairing anything at all — and it also lets a log separate two concurrent runs instead of collapsing them when the CLI reuses an id.

I did not infer the tool name inside the mapper, because a tool_result never names the tool; the sink remembers the tool_use it saw, and an unmatched result still lands as a readable tool_end with tool: null rather than being dropped. I also did not add a read-side join pass: the log has to carry ids for any consumer to pair calls, and a post-processing pass would still have no way to tell a killed call from a slow one. The abandoned status comes from the sink because it is the only place that knows a run ended.

The 300s default agent timeout is the case that motivated the last part: the CLI never emits its final result, so without closing open calls on shutdown the log claims a tool ran forever.

Tests

• npm test run locally — everything passes except swebench harness isolation (evidence diff filter + per-run uv cache + predictions), which fails identically on unmodified main (405ffa5) on this machine for want of uv/docker. That failure is pre-existing and environmental, not related to this change.
• New or updated tests cover this change — five cases added to test/agent-stream.test.js, all driven by the existing claude.ndjson fixture: result pairing, is_error mapping to error, an orphan result staying readable, shutdown closing an open call as abandoned, and two concurrent sinks not colliding on the same CLI id. node --test test/agent-stream.test.js → 17 pass, 0 fail.
• npm run drift:scan — not applicable: no workflow or atom changed.

AI assistance

Written with an AI coding agent (DeepSeek Harness) working from my own read of the streaming path. I reviewed the full diff and ran the suite locally before opening; the one failing test was confirmed pre-existing by re-running it against unmodified main.

[surlymochan]

Merged — thank you! Squash commit 37a1b1a, CI green (verify + ledger). Verified independently before merging: checked out the PR head and ran test/agent-stream.test.js 17/17 locally, and reviewed the diff (additive record fields only, 5 new tests, docs updated). The pairing fix is exactly what the issue described. 🙏