fix(qwen): give sidecar lines the shape Entire's compact reader needs

[suhaanthayyil]

The defect

Qwen's sidecar is already JSONL and get-transcript-position already reports the record count, so Entire's line-offset scoping (transcript.SliceFromLine, cmd/entire/cli/transcript/parse.go:130) already lands on record boundaries. The line count was never the problem.

The line shape was. Measured against cli@main, qwen's transcript.jsonl came out at 0 bytes.

Entire compacts an external agent's transcript with the generic JSONL reader in cmd/entire/cli/transcript/compact/compact.go. Shipping a compact-transcript method does not avoid it — checkpoint/persistent.go:1136 always runs the generic compactor over the raw transcript.

That reader keeps a line only when both hold, and a native sidecar record failed both:

half	what cli reads	native record
kind	normalizeKind (compact.go:197) reads a top-level type, falling back to role	has neither — the discriminator is event (UserPromptSubmit, PostToolUse, …), so every line was dropped
content	parseMessage (compact.go:617) reads content from a top-level message object	message is a string (Qwen's notification text), so even a line that passed the first check would carry nothing

qwen is not listed in #67 — its table implies qwen is fine. It is not; I have commented there.

The change

Each record keeps its native fields and gains a projection built from them: a top-level type, a timestamp under the key parseLine reads (compact.go:391), and the message wrapper.

event	projected as	content
UserPromptSubmit	user	text block with the prompt
PreToolUse	assistant	tool_use with type/id/name/input — the only fields stripAssistantContent (compact.go:704) preserves
PostToolUse / PostToolUseFailure	user	tool_result with snake_case tool_use_id and a string content (compact.go:665)
Stop / StopFailure	assistant	text block with the final message or error details
everything else	not projected	dropped, as before

Qwen fires PreToolUse before a call and PostToolUse after it, so that pairing is Claude Code's own tool_use/tool_result shape: inlineToolResults (compact.go:454) folds the output into the preceding tool_use. Each hook still costs exactly one line, so position == line count is preserved. An unpaired tool_result is dropped by compact.go:288, which is cli's own behaviour for Claude Code; install-hooks registers both hooks together, so the pair is present in any session Entire set up.

message is now raw JSON, and that is load-bearing

sidecarRecord.Message was string. With an object under that key, json.Unmarshal in readSidecarRecords fails — and on main that error aborts the whole sidecar read, so prompts, modified files and summaries all disappear. Typing it as json.RawMessage is what makes a projected record round-trip.

This also interacts with #65, which changes readSidecarRecords to skip unparseable lines instead of erroring: without this type change, a projected line would be silently skipped there. The two changes touch different hunks of transcript.go and merge cleanly in either order.

Qwen's notification text is the only thing that used message, and it is only sent on events that are never projected, so it is written exactly as before — there is a test pinning that. A projected record that also carried one would keep it as the first text block of its content.

Back-compat

The projection is derived data and is ignored on read, so a sidecar written by an older build still parses and later hooks append projected records alongside it. Test: a legacy line plus a freshly projected one, read together.

Measured

A full hook sequence (session-start, notification, prompt, pre-tool-use, post-tool-use, stop, session-end) driven through ParseHook, then run through cli@main's transcript.SliceFromLine + compact.FullWithBoundary:

before  start=0  raw 1032B / 7 lines   transcript.jsonl=0B
after   start=0  raw 1674B / 7 lines   transcript.jsonl=574B, 3 lines all carrying content
after   start=3  boundary=1, delta=439B

The compacted lines now read:

{"type":"user","ts":"2026-05-20T12:00:02Z","content":[{"text":"Create hello.txt"}]}
{"type":"assistant","ts":"2026-05-20T12:00:03Z","id":"tool-1","content":[{"id":"tool-1","input":{"content":"hi","file_path":"hello.txt"},"name":"write_file","result":{"output":"{\"ok\":true}","status":"success"},"type":"tool_use"}]}
{"type":"assistant","ts":"2026-05-20T12:00:05Z","content":[{"text":"Created hello.txt","type":"text"}]}

Lifecycle and notification records are correctly dropped rather than emitted empty.

Why the tests are shaped this way

Reverting only the message wrapper — keeping the top-level type — reproduces the other failure mode exactly: 468 bytes across 4 lines, every content array empty. A non-zero byte count would call that fixed. The tests therefore assert the message text, and they drive real hook payloads through ParseHook and read what the agent actually wrote, rather than comparing against a checked-in expected constant.

Mutations tried, all caught, all compiling:

mutation	caught by
drop the message wrapper (keep type)	3 contract tests
drop the top-level type (keep the wrapper)	3 tests
tool_use_id -> toolUseId	tool-shape test
tool_use input -> arguments	tool-shape test
message typed as a string again	3 tests (the read-back guard)
PostToolUse carries no tool output	3 tests
notification message dropped	notification test

go test ./... in agents/entire-agent-qwen: 33 pass (was 23).

Overlap with my other open qwen PRs

• fix(qwen): read sidecar records larger than 64 KB and skip torn lines #65 (sidecar scanner buffer) touches readSidecarRecords in transcript.go and adds transcript_test.go. This PR touches appendSidecar in the same file and adds sidecar_jsonl.go + jsonl_test.go. Different hunks, different new files — no conflict. The semantic interaction is described above.
• fix(qwen): resolve entire through PATH in installed hook commands #66 (hook command PATH) touches hooks.go / hooks_test.go only — no overlap.

[ashtom]

Reviewed for functional and security issues. This projection doubles large hook payloads, making the old 64 KiB reader limit even easier to hit. I integrated the reviewed #65 branch, including the larger-record support and physical-line offset fix, so this PR works independently. The original projection remains intact. Module race tests and lint pass, including the >64 KiB and malformed-line regressions. Current main was merged for the shared CI toolchain fix. No unresolved finding in this PR; fresh CI has 36/36 successful checks and GitHub reports MERGEABLE/CLEAN. Prefer merging #65 before this PR to keep the review history clear.