Radio: three calls the person can make, and four teammates who actually go #102

Pushed to vesper/63-radio-three-calls-the-person-can-make-an; bun test is green (394 pass, 0 fail).

<kbd>Z</kbd> opens a radio panel — the same DOM overlay the buy menu uses, three big buttons, digits <kbd>1</kbd>–<kbd>3</kbd>, and <kbd>Z</kbd>/<kbd>Esc</kbd>/Close to leave — with go A, go B and hold, plus a radio button on the touch pad. A call is one line of plain data on the world (world.radio: the side, the word, the time) and the override lives in src/client/plan.ts: ordersFor hands a bot the called site while the call stands and the rolled site when it expires, so bots.ts changed only where it reads the plan — it asks for its orders and never learns the word radio. Only the caller's own side hears it, one call is live at a time, a call stands for twenty seconds and startRound clears it, and the call lands on the HUD message line as You: go A.

Worth a look: chooseGoal in bots.ts, where hold leaves the site alone and points the goal at a walkable spot a stride from where the bot stands — that is what stops a push without leaving anyone frozen in the open — and the fact that a planted or dropped bomb still outranks anything called. A carrier obeys "go A" exactly as it obeys the plan; a bot already walking re-paths on the next tick because the site is part of the objective string repath already compares.

The arena's numbers are unchanged, which was the risk in the issue: with world.radio null the per-tick read is one null check, bun run scripts/train.ts --bench reports 12k steps/s both before and after (11–12k on the first, colder run either way), and bun run scripts/measure.ts prints the identical table on both sides of the change — 89 rounds, 60% plant rate, detonated 38 · time 24 · elimination 15 · defused 12, median 123s. A test also pins ordersFor with no call to siteFor across 60 seeds × 6 rounds × both sides × five slots.

Not in here, as the issue asked: no audio for the calls, no bots calling anything themselves, no chat box, and the roll itself is untouched.

To try it: start a match, take the bomb or stand at a site, press <kbd>Z</kbd> and call the other site — the four teammates turn round on the next tick and the message line says what you said; wait twenty seconds and they drift back to the round's own plan.

🤖 Generated with Claude Code

Closes #63

Review submitted on pull #102 as changes requested, anchored to the new hold branch in src/client/bots.ts.

What I did: read the full diff at 334f17a, ran bun test in a scratch clone (394 pass, 0 fail) and mutation-checked each new behaviour (per-tick readOrders, hold branch, isCall guard, startRound clear, KeyZ) — every one turns the suite red when removed. Booted the app from that head locally (/ in 17 ms, security headers intact, radio overlay and pad button in the first paint) because the preview host returned 404 on every path; the screenshots confirm it deployed and the menu copy is right.

The blocking finding: chooseGoal's CT branch still returns world.map.sites[contact ?? brain.site], so rotateTo outranks a live call — I reproduced it (CT side, seed 31, call "A", one CT death near B: all four teammates keep brain.site === "A" and walk to B), and CT is the default side, so the README's "whatever the round rolled" is false there. Secondary, low: openRadio can stack over the pause overlay and re-grabs pointer lock on close while paused.

My clone is untouched (git status clean, nothing pushed).

Both findings fixed and pushed as 76d5d69.

The blocking one. chooseGoal's CT branch returned sites[contact ?? brain.site], so rotateTo outranked a live call and one teammate dying near a site cancelled the radio for the rest of the round — on the default side. ordersFor now reports whether the site came from a person (Orders.called), readOrders copies it to brain.ordered, and the contact rotation is only consulted while no call stands. The flag is also part of repath's objective string, so a call that happens to agree with the rolled site still turns round a bot that had already rotated off it — without that, the repro's "call A when A was rolled" case waits for the 3–6 s repath timer. The new test in bots.test.ts is sentinel's repro (CT side, seed 31, a CT body and feed entry at B, then radioCall(world, "A")); it fails with either half of the fix removed, and it also pins the fall-back: once the call expires the side rotates to B again, which is the rotation still being there rather than being deleted.

The low one. openRadio now refuses to open when !running, when the pause overlay is up, or when the match-over card is up, so <kbd>Z</kbd> on a paused match cannot stack the radio over it and Escape cannot hand the pointer lock back into a stopped world. Closing is unguarded, which is what keeps Escape working for a panel that was opened legitimately.

Numbers are unchanged, as before: scripts/measure.ts prints the same 89 rounds, 60% plant rate and detonated 38 · time 24 · elimination 15 · defused 12, and bun run scripts/train.ts --bench reports 12k steps/s — the new called flag is a constant false on the silent path. bun test is 395 pass, 0 fail. The README's Playing and The bots sections now say the call outranks the contact rotation, which is what the code does.

I did not touch the preview 404 — nothing in this change serves routes, and a local boot of this head answers / in 15 ms with the three security headers and the #radio overlay in the first paint.

🤖 Generated with Claude Code

Approved pull #102 at 76d5d69, anchored to the brain.ordered ? null : rotateTo(...) line in src/client/bots.ts.

Both round-one findings are genuinely fixed, and I verified them rather than taking the commit message's word: my CT repro (seed 31, teammate dead at B, then radioCall "A") now sends all standing teammates to A, and openRadio refuses to open over the pause overlay or the summary card. bun test is 395 pass / 0 fail in a scratch checkout, and reverting the guard, the :called objective segment, or the called flag each turns the suite red. The app boots from this head (/ in 17 ms, security headers intact, radio overlay and pad button in the first paint); the preview host itself still 404s, which I noted as infrastructure rather than a defect. My clone is clean and nothing was pushed.

<kbd>Z</kbd> opens a radio panel — the same DOM overlay the buy menu uses, three big buttons, digits <kbd>1</kbd>–<kbd>3</kbd>, <kbd>Z</kbd>/<kbd>Esc</kbd>/Close to leave, and a radio button on the touch pad — with go A, go B and hold. A call is one line of plain data on the world (world.radio: the side, the word, the time) and the override lives in src/client/plan.ts: ordersFor hands a bot the called site while the call stands, says whether a person named it, and falls back to the rolled siteFor when it expires. bots.ts changed only where it reads the plan. Only the caller's own side hears it, one call is live at a time, it stands twenty seconds, startRound clears it, and it lands on the HUD message line as You: go A.

Since the review at 334f17a, two rounds of fixes. The blocking one: chooseGoal's CT branch let rotateTo's contact outrank a live call, so one teammate dying near a site cancelled your radio for the rest of the round on the default side — the contact is now only consulted while no call stands, and :called is part of repath's objective so a call that agrees with the rolled site still turns round a bot that had rotated off it. The low one: openRadio refuses to open over the pause overlay or the match-over card. This push adds nothing to the code — it settles sentinel's last non-blocking note by deciding that the dead keep the radio, and writing that into the README with a test rather than leaving it an accident of where a check is not.

Worth a look: the hold branch at the top of chooseGoal, which points the goal at a walkable spot a stride away so a held side repositions rather than freezing in the open, and the fact that a planted or dropped bomb still outranks anything called.

The arena's numbers are unchanged, which was the risk named in the issue: with world.radio null the per-tick read is one null check, scripts/measure.ts prints the identical table on both sides of the change (89 rounds, 60% plant rate, detonated 38 · time 24 · elimination 15 · defused 12, median 123s), and bun run scripts/train.ts --bench reports 12k steps/s either way. bun test is 396 pass, 0 fail.

Not in here, as the issue asked: no audio for the calls, no bots calling anything themselves, no chat box, and the roll itself is untouched.

To try it: start a match as CT, press <kbd>Z</kbd> and call a site — the four teammates commit to it and stay committed even after one of them dies somewhere else; wait twenty seconds and they go back to reading the round for themselves.

🤖 Generated with Claude Code

Merged main and pushed; bun test is 447 pass, 0 fail on the merged tree.

The radio itself is unchanged: <kbd>Z</kbd> opens a three-call panel — go A, go B, hold — a call is one line of data on the world (world.radio), plan.ts's ordersFor reads it in front of the rolled siteFor and reports that a person said so, which is what puts it ahead of the defence's own contact rotation; bots.ts only asks for its orders.

The conflicts and how I read them:

• input.ts was the real one. #101 replaced the fixed KEY_ACTIONS map with a bindable ACTIONS list, so rather than keeping radio as a hardcoded key I added it to that list with KeyZ as its default and a Radio label in bind.ts. It now appears in the Controls disclosure, the settings round-trip and the "already bound to…" collision rule like every other action — 17 actions, and the README's "the same 16 actions" became 17.
• Main's binding tests used KeyZ as their example of "a fresh code nothing else holds", which the radio now holds. Their meaning is the freshness, not the letter, so they use KeyG; that is five tests in input.test.ts, settings.test.ts and bind.test.ts. A new test in bind.test.ts pins the radio's place in the bindable set.
• bots.ts was two sets of brain fields landing in the same line — heardAt from the bots-hearing change, slot/holding/ordered from the radio — and both are kept.
• *game.ts and README's Playing*** both kept: #101's "every key above is a default, rebind under Controls" alongside the radio paragraph, which now says the key is a binding like any other. The radio panel's own hint no longer prints "Z" — it names Escape and Close, which no rebind can move.

With the radio silent this branch is main exactly: scripts/measure.ts prints main's own table (88 rounds, 61% plant rate, detonated 40 · time 23 · elimination 14 · defused 11) — my earlier numbers moved because main's hearing change moved them, not the merge — and bun run scripts/train.ts --bench reports 12k steps/s. The app boots from the merged head: / answers in 8 ms with the #radio overlay and the radio pad button in the first paint.

🤖 Generated with Claude Code

Approved pull #102 at 14d40ab, anchored to the new radio: "KeyZ" default in src/client/input.ts.

The merge with #101 turned the radio into a proper bindable action rather than a hardcoded key, and I verified that end to end: bun test is 447 pass / 0 fail in a scratch checkout; changing the default key, dropping radio from ACTIONS, removing its label, deleting the onRadio dispatch branch, or reverting the brain.ordered guard each turns the suite red. I also probed the stored-settings migration — an old save that already bound Z keeps it and the radio falls to the first unclaimed default, 17 actions and 17 distinct codes — and booted the app from this commit (15 ms on /, security headers intact, overlay and pad button in the first paint, bundle carries the wiring).

Noted without blocking: the preview host has 404'd at all three heads, so I used a local boot each time, and the pull description is stale (394 pass, Z as a fixed key, no mention of the rotation precedence, the dead-player radio, or the #101 merge). My clone is clean and nothing was pushed.