Voice roster reads: sweep, digest, last reply, screen (WI-1268)

[thedancingdeveloper]

What

The voice assistant can read the session roster. session.sweep is the roster question ("what's going on", "what needs me"); a named session goes to session.digest, session.last_reply, or session.screen. All four are curated reads with no approval. list_sessions stays a bounded name/role lookup (≤32) and spells a non-worker as name (role), so "oversight" resolves. The sweep the assistant actually runs is forced to digest mode with screen_lines clamped to 5.

Why

A spoken "what's going on" was answered from an unbounded session list or a raw tail, and never saw the blocked report, the last reply, or the screen. WI-1268.

Does not reimplement WI-1266 (session.list pagination), which is PR #968. This branch does not depend on it: the curated tools pass ids through to core, and the engine's own lookup still wants a UUID.

Closes nothing on GitHub; the item is WI-1268 in Vogt.

Verification

• Python core: uv run pytest, uv run mypy, uv run ruff check ., uv run ruff format --check .
• Docs: uv run python scripts/check_docs.py
• PWA (web/): pnpm typecheck, pnpm test, pnpm test:browser (Playwright)
• Engine (engine/): cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test -p vogt-engine-server --lib assistant::tests
• Not applicable — this change doesn't touch code these gates cover

Also uv run pytest tests/test_voice_capabilities.py --no-cov (4 passed). Full cargo test --workspace was not run; the assistant lib tests (78) were.

Docs

• docs/ENGINE.md §6.5.1: the four reads are Voice-readable, and list_sessions is the bounded name/role lookup.
• Not applicable

[thedancingdeveloper]

Review — WI-1268 voice roster reads (head 93d73f7)

Reviewed the diff in a detached checkout of that SHA; did not touch the author's worktree or the dirty checkout on wi-1044. uv run pytest tests/test_voice_capabilities.py --no-cov: 4 passed.

What checks out:

• The four reads are in CURATED_READS, the comment sits above the array so _const no longer invents a quoted operation, and the new regression asserts each is Voice-readable in ENGINE.md §6.5.1, present in CURATED_READS, non-mutating, and on the MCP transport. The §6.5.1 rows and the list_sessions tools row match the code.
• sweep_args rewrites only session.sweep: it forces mode=digest, leaves every other argument (attention, limit, fields, states) alone, and the existing curated-read loop asserts the rewritten arguments are exactly what the core is called with. Other operations pass through unchanged.
• list_sessions filters by case-insensitive name and role, caps at MAX_VOCABULARY_SESSIONS (32), and now returns role. vocabulary_session spells a worker as the bare name and anything else as name (role), with the unit test covering worker, oversight and a blank name. Both role variants exist on SessionRole.
• WI-1266 (session.list pagination) is correctly left out.

Blocker: the screen-lines clamp drops the screen whenever the number is not an integer

let lines = object
    .get("screen_lines")
    .and_then(Value::as_u64)
    .unwrap_or(0)
    .min(5);

Value::as_u64 returns None for any JSON number with a fractional part, and the unwrap_or(0) then stores 0. sweep_sessions treats screen_lines == 0 as "ask for no screen at all" (screen_tail=[]), so the roster the assistant reads back has no screen. Tool-call arguments are model-produced JSON, not the Rust struct the unit test builds: {"screen_lines": 5.0} is the shape that arrives, and it clamps to 0. The prompt makes this worse rather than better — it says "keep screen_lines at 5 or fewer" and never tells the model to pass a positive value, so a compliant model that omits it also gets 0 via the same default. The test only exercises screen_lines: 40 (an integer) and a missing key, so both paths are green.

Use as_u64().or_else(|| as_f64().map(|n| n as u64)), and default the missing key to 5 rather than 0, since this rewrite exists to make the screen speakable, not to suppress it.

blockers (head 93d73f7)

[thedancingdeveloper]

Re-review — head 199ec23

The blocker from the previous review is fixed. sweep_args now reads screen_lines through as_u64 and falls back to as_f64 for a finite, non-negative whole number, so a model-sent 5.0 clamps to 5 instead of being dropped. Anything else — missing, fractional, negative — defaults to 5 rather than 0, and an explicit integer 0 is still honoured as "no screen". The prompt now says omitting the field still reads five lines. The rewrite still touches only session.sweep and still forces digest mode.

Verified: cargo test -p vogt-engine-server --lib sweep_the_assistant_runs passes, including the new cases for 3.0, 5.0, 2.5, -1, 0 and a missing key. Nothing else in the diff changed since 93d73f7.

no blockers (head 199ec23)