test(director): the boot stretch hands the keyboard over

[wine-fall]

Roadmap line 0 — "the listener's first typed line is silently dropped" (issue
#145) — no longer reproduces under its own recorded conditions. This retires it
as cannot-reproduce, not fixed, and spends most of its effort making sure
that distinction survives the deletion.

The evidence

Three runs on 2026-09-04 reproducing the August 25 setup exactly: the real
Claude brain, a seeded persona.md (so first-run onboarding is skipped),
--plain, piped stdin, one line typed after the second beat aired plus a 3s
settle. Run B's shape — a single line, so there is no second line to confuse
the result.

run	first line echoed	settings.json
1	yes	not read (the probe quit too early)
2	yes	musicEnabled: false after 35.2s
3	yes	musicEnabled: false after 5.0s

musicEnabled has exactly one writer in the tree, src/steer-tools.ts — the
Director's steer tool. So runs 2 and 3 are read at a deterministic seam: the
line did not merely echo, it crossed the steer path.

Why this is not "fixed"

No cause was ever identified. lineReader's settled guard predates the
August 25 loss by a week, and LineQueue's take/peek semantics have not
changed since — so the guard is pinned by test here, never credited with the
cure. Either an unrelated change closed the window, or the timing drifted, or
the original conditions held a variable nobody wrote down.

That makes the pin more worthwhile, not less: a race that closed without a
known cause returns silently — no error, no log line, just a listener's first
sentence going missing.

What landed

• test/director-steer.test.ts — the hand-over the issue asked for: once a
  pre-broadcast reader (the crash-report offer, the setup conversation) has
  settled, the next line typed reaches the Director's steer path, on a real
  LineQueue. test/guide.test.ts already pins the reader's own guard on the
  quit and eof arms; the seam between them was not covered.
  Mutation-verified — delete settled from src/guide.ts and this test goes
  red with the steer brain never seeing the line, which is the reported
  symptom exactly.
• specs/spec01/01-core-loop.md §3.3 — the violated contract now carries
  its own history, so a first line going missing again is recognisable as this
  defect returning rather than a fresh one.
• ROADMAP.md — line 0 and its §0 are deleted; the disposition is recorded
  under "Not on this roadmap". Line 4's eval half (Steer tool-choice eval (Ollama): switch/end/reply routing has no repeatable LLM-in-the-loop test [spec 11 §5] #98) is now read first among
  the P2s, being the only line overdue rather than upcoming.

Peer review

codex exec review (gpt-5.6-sol, xhigh): 1 finding, applied. It observed
that the new test is green on the base revision — it exercises a guard that was
already there, not the unexplained failure — and that deleting line 0 risked
dropping the record from planning.

Both halves are correct as facts. The remedy it proposed (keep line 0 until a
base-revision-red test drives the recorded startup path) is not reachable: you
cannot write a red test for a defect that will not reproduce, so the item would
sit at the top of the roadmap permanently undischargeable. The response instead
states the limitation in the places a future reader will actually look — the
test's own comment, the commit message, and the spec — and keeps the record in
two places rather than keeping a phantom P0.

Verification

• vitest: 1413 passed, 1 skipped, 69 files
• tsc --noEmit: clean
• oxlint src test scripts .github/scripts: clean
• The repro harness lives in the gitignored scratch/ and is not committed.

AI coding brief

Original request — the user asked what was left on the roadmap, then, on
seeing a P0 they had no memory of, asked what it actually was so they could try
to reproduce it and "otherwise just mark it done". The why: an unremembered P0
blocking every by-ear judgement above it is worse than no P0.

Manual interventions — the user asked for concrete commands plus an
investigation the agent could run itself, rather than a description handed back
to them. Mid-run they asked directly whether the bug was fixed, which is what
forced the cannot-reproduce / fixed distinction to be made explicit instead of
being quietly rounded to "done".

Retro — two things cost a round each and were self-inflicted. The first
probe quit the moment it saw the ⌨ echo, before the brain could act, which
proved nothing: ⌨ has two printers (director.ts on the steer path,
guide.ts when a reader consumes), so only the settings.json seam decides.
Asking "which seam proves this, and who else can write it" before running would
have skipped that. The first test draft also fired the esc pulse synchronously,
before the read registers on a microtask — test/guide.test.ts already had the
setImmediate line, so reading the neighbouring tests first would have avoided
it. Both are the same lesson the repo already states: read the seam, not the
narration.

🤖 Generated with Claude Code

https://claude.ai/code/session_018nkjmN4XQdDbffSRnnAaEi