The failure

Both spec-entry cases failed on a claim about a handoff block their walks emitted correctly. The block was missing from the record, so the asserter had nothing to judge and rightly refused an unevidenced step. Four walks, two cases, one cause.

It isn't the parser. Running the hook's own extraction over the settled transcript returns every turn including the missing one — so the file didn't hold that turn yet when the hook read it.

The hook runs inside the stop sequence, not after it. The agent has stopped thinking; the runtime is still writing down the last thing it said.

The fix

The turn isn't lost, just not yet on disk — the stop payload carries it as last_assistant_message, the runtime handing it over at exactly the moment it's unreadable anywhere else. It's appended as the closing turn when not already present.

Reading the transcript later from assert would also have worked, but it splits the record across two components and leaves the world holding a pointer rather than the walk. This keeps the world self-contained.

Appending-when-absent is not deduplication. Two identical turns genuinely in the transcript both survive — a step that ran twice is something the asserter must see.

Truncation, reconsidered alongside it

The caps were never about disk (the logs are throwaway) but about the asserter's prompt, which every recorded action is read into. That justifies trimming a file body no claim rests on; it doesn't justify trimming evidence.

• walk log — caps nothing at all
• command output — 2000 → 10000, a ceiling that bounds a runaway without ever reaching real output
• file-read bodies — 400 unchanged, the one place content is genuinely incidental

The one known live difference, written down

An inline !`command` directive is substituted when a skill loads live. A walk reads prose as a file, so the walker takes the fallback the prose supplies for exactly that case — correct, expected, and no longer reported as a marker.

The asserter is also told what it costs: the primary arm is the one real sessions almost always take and it is never exercised here, so a claim resting on the substituted value is unprovable rather than passed. Named as the only such exemption — anything else resembling an environment quirk is a finding, not a licence.

Test plan

• 81/81 across the prose suites, including four new cases: the closing turn arrives and in order, a turn the transcript already caught isn't double-recorded, a genuine repeat survives, and the walk log truncates nothing
• Both spec-entry cases want a re-run to confirm they now pass on their own evidence

🤖 Generated with Claude Code

Stack

1. docs(design): prose-tests programme design log #544
2. feat(prose-tests): the framework — cases, worlds, runner, skill #545
3. test(prose): feature happy-path corpus — five worlds, seven cases #546
4. test(prose): bugfix corpus — the investigation-centric surfaces #548
5. test: retry recursive teardown removals — kill a class of phantom failures #549
6. fix(entry-skills): close the handoff fences — six files render their arms wrong #550
7. docs: a contributing page for working on the system #551
8. fix(entry-skills): every handoff arm says to invoke the skill #552
9. fix(implementation): environment setup belongs to the setup reference alone #553
10. fix(prose-tests): the asserter is told which substitutions were armed #554
11. feat(prose-tests): the mid-flow substitution, and a world only prose can describe #555
12. test(prose): claims assert consequences, not what was displayed #556
13. feat(prose-tests): record everything the agents do, results included #557
14. fix(discussion-entry): the handoff reports the source it actually had #558
15. fix(prose-tests): the stop hook records, and names the model that walked #559
16. fix(prose-tests): command output was never actually recorded #560
17. feat(prose-tests): judge the walk as told, not the summary returned #561
18. feat(prose-tests): decide in code what an agent should not be deciding #562
19. test(prose): a case starts where a session starts #563
20. feat(prose-tests): walk on Sonnet, judge on Opus, escalate a failure #564
21. test(prose): give the eight read-only cases something that can fail #565
22. test(prose): only walks that can be observed, and checks that survive the trip #566
23. fix(prose-tests): the verdict names only the model the record names #567
24. test(prose): discovery, walked to the point where work first exists #568
25. fix(prose-tests): the asserter judges which of prose or walker was at fault #569
26. docs(conventions): a step whose reference routes every exit still signposts #570
27. test(prose): discovery's epic arm, to the same durability boundary #571
28. fix(prose-tests): keep the walk's closing turn, and name the one live difference #572 👈 current
29. fix(prose-tests): prescribe the cross-check, drop a check that cannot fire, lint the scope #573
30. test(prose): declare the prose each case actually walks #574
31. fix(investigation): the symptom interview never ran for work shaped in discovery #575
32. feat(prose-tests): conduct — playing a person where a script cannot reach #576
33. fix(prose-tests): what a write put in a file is evidence, not incident #577
34. fix(investigation): the interview starts from the carrier, not over it #578
35. test(prose): quick-fix scoping, the whole definition stage in one walk #579
36. fix(prose-tests): the world log is the walker's record #580
37. fix(prose-tests): a confirmed failure reports both runs' checks #581
38. fix(prose-tests): each prescribed command runs as written #582
39. fix(scoping): context gathering reads the whole carrier, asks only the gaps #583