fix(resume): a replayed transcript had lost every tool result

[ericleepi314]

What

Resuming a session redrew its tool calls with nothing under them. Captured by running a session, killing it, and replaying it with /sessions <id>:

before                                  after
⏺ Read(src/domain/usage.ts)             ⏺ Read(src/domain/usage.ts)
                                          ⎿  Read 3 lines
⏺ Bash(ls -1 src | head -12)            ⏺ Bash(ls -1 src | head -12)
                                          ⎿  __tests__
                                             app
                                             App.tsx
                                             …

Every result the session had ever produced was gone.

Why

rehydrate folded tool/call into a row and never looked at tool/result at all, and the transcript builder discarded the row's text on purpose — reasonably, back when that text was the model-facing payload nobody wants pasted into a transcript.

It no longer is. The card projections (#15, #17) now produce a real ⎿ summary, and the harness persists each card's structured shape in the session log precisely so presentResult reproduces the identical card from a stored event:

the read tool projects them through its output.presentationMeta (persisted with the session log) and presentResult narrows that metadata back into this view on live and replay paths alike

So replay presents its results the same way the live path does, and the row's text is that summary.

Details

• presentResult takes the call's name and arguments as parameters instead of reading them off the live maps — that is what lets replay share it.
• Results are matched to their call by id, not by arrival order, because parallel calls interleave their results in the log.
• A write or edit cannot draw its diff card on the flat replay trail, so it gets the card's own header as text — Wrote 3 lines to notes.txt, Added 1 line, removed 1 line — rather than the "file has been updated successfully" boilerplate addressed to the model.

Test

• npm run typecheck, npm test — 150 files / 1953 passing
• node scripts/verify-boundary.mjs
• harnessClient.test.ts: the stored log now carries a tool/result, and the rehydrated row asserts its presented text
• messages.test.ts: the tool row's text reaches the trail line (the fixture's 'ignored raw result' was the old contract, now 'Found 3 files')
• live: a two-pass probe — run tools, kill, /sessions <id> — showing the replayed trail matching the live one

🤖 Generated with Claude Code