bug: Claude Code resume starts a new raw.jsonl instead of continuing the active recording

[galexy]

Symptom

When a Claude Code conversation is resumed (via `/resume` or via reopening the same session), `ox agent prime` fires again and creates a new recording session directory for the same agent_id, orphaning the prior in-flight recording in a "local only" state. The two halves of the conversation end up in two separate `raw.jsonl` files that are never stitched together, and the first half never gets finalized or uploaded.

Live repro (happened to this user today)

Agent_id `OxCh8K` was active from 2026-04-21T15:25:31Z (original Claude session start) through 2026-04-21T17:10Z. Conversation was paused and later resumed at 2026-04-22T00:00:49Z.

Two recording directories now exist under `repo_019c5812-...//.sageox/cache/sessions/`:

Dir	raw.jsonl lines	size	First/last entry	Status
`2026-04-21T15-25-galexy-OxCh8K/`	383 (header + 382 entries)	420 KB	15:25 → 17:10 (whole #525 bug-fix session including PR merge)	`.recording.json` missing, marked "local only" by `ox session list`, no meta.json, no summary.*, never uploaded
`2026-04-22T00-00-galexy-OxCh8K/`	20 (header + 19 entries)	16 KB	00:00 → now (just the continuation turns)	active recording

Both headers carry `agent_id: OxCh8K`. The agent_instances registry shows exactly one OxCh8K entry created 2026-04-21T08:30:34-07:00 — so the agent identity is stable across the resume. Only the recording file is not.

Expected

On resume, `ox agent prime` should either:

1. Resume the existing recording — reopen the same `raw.jsonl` and `source_offset`, keep appending. Preserves one contiguous session on disk and on the ledger.
2. Finalize the prior recording and start a new one — run the full stop → summarize → upload pipeline for the first half, then open a fresh recording for the continuation. Both halves reach the ledger, linked by parent/continuation metadata.

Either is fine. What must not happen is the current behavior: orphan the first half in "local only" with no meta.json and no summary, silently lose it.

Actual

• `.recording.json` for the first session is gone (cleared by some hook between 17:10Z and 00:00Z — candidate culprits: `ClearRecordingStateForAgent` on SessionEnd / Clear hook, or daemon orphan/liveness reap).
• No `meta.json`, `summary.json`, or `session.md` was written — the daemon's async finalize was never triggered.
• `ox session list` shows the session as `✗ local only` — the content is on disk but the ledger never learned about it.
• When the Claude session resumed, `ox agent prime` created `2026-04-22T00-00-galexy-OxCh8K/` from scratch instead of continuing the existing one.

What this user had to do to recover

Manual steps: `mkdir ledger/sessions/`, copy raw.jsonl + context-trace.jsonl + raw-session.md → session.md, `ox session regenerate --summary`, `ox session upload`. Works but is not something end users should need to discover.

Related

• bug: resolveAgentIDFromStates reuses agent_id from sole active recording regardless of workspace or PID #528 (recently merged) hardened agent_id collision handling — this is a different failure mode: same agent_id across a resume, but the recording file lifecycle doesn't follow.
• Possibly related to Claude Code hook ordering: which hook fires on Claude `/resume` vs. on initial SessionStart, and does the SessionEnd hook run when the user closes vs. when they /resume elsewhere?

Acceptance criteria

• After a Claude Code `/resume`, `ox session list` shows one session for the Claude conversation, not two — either resumed in place or finalized-then-continued.
• If the prior half is finalized-and-continued, it MUST upload to the ledger (with its own `meta.json` + summary) rather than being left in "local only" state.
• Integration test that simulates a resume (start a recording, stop Claude Code, re-prime with the same Claude session UUID) and asserts exactly one of the two expected outcomes.