[FEATURE] Persist recap / post_turn_summary to session JSONL for downstream parsing

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code's session recap (a.k.a. post_turn_summary / "away summary",
introduced in the /recap family of changes — disable via /config or
CLAUDE_CODE_ENABLE_AWAY_SUMMARY=0) generates a short, useful summary
of what's happened in the session so far and renders it as a grey
overlay in the CLI when you return after being away.

The summary is good — concise, factual, exactly the kind of thing a
downstream tool would want for cross-session analytics, dashboards, or
chat-history search. But it's currently render-only: it doesn't land in
the per-session JSONL under ~/.claude/projects/<encoded-cwd>/<uuid>.jsonl,
so anything outside the live CLI has to re-derive the same summary
itself (we're piping transcripts through a local LLM merger for exactly
this purpose; Claude Code is already doing the work, we'd just like to
read the result).

Adjacent issues:

• #34610 (closed) requested the /recap command itself — feature

shipped despite the not-planned close.

• #48084 asks for docs of /recap + away-summary controls (docs gap).
• #53516 asks for a stable, documented schema of JSONL line types

generally — this would be one new entry in that schema.

• #55023 / #55863 are platform-specific render bugs for the overlay.
• #57743 asks for UI/visual separation of summaries — orthogonal.

None of these request persistence. This one does.

Proposed Solution

When the recap fires (auto on return after the away threshold, or
manually via /recap), append a line to the active session's JSONL with
a stable, parseable shape — mirroring the existing synthetic-entry
pattern used by ai-title, last-prompt, and file-history-snapshot:

{
  "type": "recap",
  "recap": "<the same text rendered in the overlay>",
  "trigger": "away" | "manual" | "post_turn",
  "leafUuid": "<uuid of the assistant turn the recap was computed from>",
  "sessionId": "<sessionId>",
  "timestamp": "<RFC3339>"
}

Notes:

• trigger lets consumers tell auto-fires apart from operator-invoked

ones (the operator-invoked ones are usually the more curated ones).

• leafUuid ties the recap to a specific point in the transcript so

downstream tools can scope "recap of work up to X" cleanly. Matches
the existing last-prompt shape.

• No change to the rendered UX. Persisting in addition to rendering is

additive.

• Opting out: respect the existing CLAUDE_CODE_ENABLE_AWAY_SUMMARY=0

/ /config switch — if recap is disabled, no JSONL entry either.

Why it's worth doing

• Downstream summarizers can skip their own work. Fleet tooling

that ingests JSONL (chat search, session dashboards, multi-host
rollups) gets a first-class human-readable summary per session
without running a second LLM pass over the transcript. In our setup
this would let us drop a planned Gemma-based ASR-merge/summary step
entirely.

• No new surface to maintain. The recap is already produced and

rendered; this is one additional writeLine next to the existing
rendering step.

• Schema-friendly. Falls inside the lane #53516 already opened — if

Anthropic ever stabilises the JSONL line-type schema, recap is a
natural member.

Alternatives considered

• Read the rendered overlay from the TUI — fragile, terminal-only.
• Re-summarize the transcript downstream — duplicates work the CLI

already does, with a different model and different quality.

• Use ai-title as a proxy — too short (single line, no progress

detail).

Environment

• Claude Code 2.1.150, Linux.
• Recap confirmed working as a CLI overlay in this version.