imposter-tactics

name	imposter-tactics
description	Use for Crewrift-specific imposter tactics recipes when optimizing a player.

Imposter tactics — recipes (crewrift tier)

On-demand recipes (10). Trigger→action heuristics; pull the relevant one when its situation arises.

1. Relax exact-equality kill gates to thresholds, or the imposter never fires

crewrift · negative result · ⚠ session-derived, unverified

Baseline imposters that kill only when (kill icon lit AND EXACTLY one non-teammate visible AND within range) silently fall through to wander/follow whenever 2+ crewmates are in view, and opponents exploit it by clustering; an exact-equality witness count (max_witnesses==0 / 'exactly 1 visible') almost never holds in a 4-player game. Relax single-condition equality gates to thresholds or direction+distance cutoffs (kill an isolated target even if a second crewmate is in peripheral range). When a code path never executes in replays, suspect an equality guard that should be a range before adding new logic. _{sources: claude-code:0919285d-1374-4827-b596-6e7107d95526, opencode:ses_1f7b6dc30ffeDmQtlUZI25HAjh, opencode:ses_21ffeb5ffffeA1Pk6sdyHDzwQW, opencode:ses_224eeb40bffeZqmBRcrjFgPFul}

2. Filter known fellow imposters out before target selection AND witness counting

crewrift

In the default 2-imposter game, imposters otherwise repeatedly target and chase each other (0px separation) and get stuck, and a partner standing near a victim blocks a kill, because the witness/target filter checks a knownImposterColors set that is never populated at runtime (so max_witnesses==0 is unsatisfiable and the hunt set includes an un-killable teammate). Capture teammate colors during Crewrift RoleReveal (an imposter viewer sees only its teammates' role icons, emitted at id base 9500 -- color sprites OUTSIDE the normal 1000+ player id range) and use them to exclude teammates and their corpses. Treat known-imposter evidence as additive and also infer it behaviorally: repeated failed close-range strikes where the kill button stays ready and no body appears mean the target is un-killable (the server rejecting an illegal kill is a reliable signal), so record it as a teammate. Also keep a killed-colors set and reject already-dead colors (observed: same color struck twice), and verify the exclusion in episode traces rather than assuming it works. _{sources: players_checkouts/players/users/james/personal_cogs/among_them/guided_, claude-code:4430a077-5411-426b-a87a-75512ec4300f, claude-code:91f456a7-41e8-4a89-98e9-44e379160b15, codex:019e08db-c4f1-7ec1-8f76-108fbbe21008 (+5)}

3. Gate the Hunt strike on reachability and a real sighting, and tune range to the engine

crewrift

Filter Hunt targets by reachability BEFORE applying the isolation preference: a fallback that picks an unreachable isolated target over a reachable non-isolated one deadlocks Hunt and wastes the turn even when a reachable crewmate exists. Drive the actual strike on a pixel-level in-view sighting, not an occupancy belief (room belief drives the pre-kill search, but you cannot strike what you cannot see); when the target leaves view, keep a short ~48-tick (~2s) memory window and navigate toward its last-seen world position with normal movement (not a blind kill-strike), then clear it. Match action thresholds to the server's real mechanics: the strike range was bumped 16px->20px to match the actual kill range so behavior stops silently failing at the boundary -- tune attack/interact range, pickup radius, and similar gates against the engine's true value, not a guess. _{sources: claude-code:25488cd9-61d5-4231-bd87-e2744b6db043, claude-code:4430a077-5411-426b-a87a-75512ec4300f, codex:019e7174-3079-7351-b8ea-b1bfcb917ff4, opencode:ses_206fd87a6ffel8Y8ooSpgw4Aeo (+1)}

4. Confirm a kill with two signals, time confirmation from the strike, and never fall through to re-acquire

crewrift · negative result

Confirm a kill landed by requiring BOTH a new body within ~30 world px of the strike point AND the kill-ready cooldown flipping true->false; do NOT infer a kill from a drop in visible-crewmate count (a crewmate walking off-screen is a false positive). Start the post-action confirmation window only AFTER the action is emitted (distance<=range AND button pressed), never when the target is first sighted: a 12-tick KillConfirm started at a ~68px sighting always expires before the bot closes the ~16 ticks to 20px range, abandoning every kill. After any strike attempt, unconditionally transition to a cover/flee state for a fixed period; a failed confirmation that falls through into ordinary target search causes runaway pursuit (the bot resets its tracker and chases a fresh victim while the kill is actually on server cooldown). Also require the kill-edge detector's previous phase to have been Playing, since a meeting resets killCooldown and the reset otherwise looks like a fresh kill edge that spuriously fires post-kill Evade. _{sources: claude-code:067e7439-cbd1-4b26-9b3f-d054bde3aa40, claude-code:4430a077-5411-426b-a87a-75512ec4300f, codex:019e08bb-a4ea-7c83-a8d6-2652eaa1a3b8, codex:019e7174-3079-7351-b8ea-b1bfcb917ff4 (+3)}

5. Make post-kill Evade local, edge-triggered, self-terminating, and corpse-aware

crewrift

Run a dedicated post-kill Evade window BEFORE any body-reporting logic (vent if possible else move off the body), to avoid instant self-reports and clear the corpse before Search/Hunt/Pretend resume; keep Evade LOCAL -- sending the imposter to the furthest globally-reachable point breaks it (it never leaves Flee), so just leave the immediate vicinity (prefer-evade ~72 ticks / 3s) and drop back to cover, choosing the nearest station >=24px from the body in the away hemisphere (dot((station-self),(self-body))>=0). Commit the recovery plan at strike time rather than re-deriving it next tick. Enter flee only on the hunting->fleeing reflex when a body NEWLY appears (edge-triggered on body-count increase), restricted to alive non-ghost imposters (guard isGhost as well as alive), with a cooldown (~96 ticks) and exit on a timer (default ~240 ticks / 10s) OR once >=48px from the body. Critically, remember bodies the bot created itself and suppress fleeing from any body whose color matches the bot's own, or it will repeatedly flee its own victim's (or its own ejected) corpse and waste time. _{sources: personal_labs/crewrift_lab/crewrift/crewborg/design.md, players_checkouts/players/users/james/personal_cogs/among_them/guided_, players_checkouts/players/users/james/personal_cogs/among_them/guided_, claude-code:73fa0a3d-37fb-4408-bae8-245db6e62c56 (+3)}

6. Make fake-task cover audience-conditional and mimic the real task lifecycle

crewrift

An imposter that only acts when a crewmate is visible will freeze once the crew disperses, and standing idle at a station is itself a tell. Process arrival BEFORE re-evaluating occupancy (a recurring bug retargeted the next tick without ever idling), and effectively disable the re-evaluation TTL (~10000 ticks) so the imposter pretends until it is time to Search rather than thrashing every N ticks. Walk station-to-station and structurally mimic the crewmate task lifecycle: a fake-hold (press A, no movement, ~60 ticks / 2.5s) then a linger (still, no buttons, 36 ticks) -- instantly walking off after 'finishing' looks unnatural, and a fake A-press completes nothing since the server only credits a player at its OWN assigned stations. Roll the probabilistic fake task ONLY on the tick of entering a task's radius (debounced by previous-nearby-task index, p1/12, lasting 90-180 ticks, then 240-tick cooldown) -- never per-frame, which thrashes. Never anchor in the spawn room (everyone co-located) or rooms with no station, never start an unwatched hold (abandon it the instant the last crewmate leaves and re-dispatch toward crew-dense rooms), and end the hold early on a witness swap (fire at most once per loiter, avoid re-picking within ~30px). After a kill, start a fake-task cooldown so the bot does not work next to a fresh body (an incriminating adjacency). Expose a compact mode summary to the LLM (locked station, remaining loiter/fake-hold ticks, whether witness-swap fired) so it can tune cover risk via parameters, not raw scratch state; and distinguish cover modes by their ANCHOR (station vs companion vs room) before merging them. _{sources: bitworld/among_them/players/evidencebot_strategy.md, personal_labs/crewrift_lab/crewrift/crewborg/design.md, claude-code:940b2eb9-3677-4c05-8754-c9cfa163f531, claude-code:f70c9801-7ee7-499b-97f9-4fd0848a6b8e (+5)}

7. Imposter core loop and strict behavior priority

crewrift

Blend in by walking station-to-station faking tasks (stand where a task would be done, press action); kill before crewmates finish all tasks because cooldowns are long; immediately after a kill flee far (vent if available, else corridors away); never target or accuse a known teammate. Strict-priority policy: (1) handle a body in view -- self-report a fresh own-kill, else flee others' bodies; (2) kill a lone visible non-teammate (no other non-teammates in view) when the button is lit and in range; (3) hunt a lone crewmate; (4) continue/start a fake task; (5) follow a crewmate; (6) wander. Kill only when alone with the target and no witnesses, prefer low-traffic rooms (Electrical, Reactor, Lower Engine, Nav, Shields), avoid high-traffic multi-entrance rooms (Cafeteria, MedBay, Admin), move away fast, and record the kill tick+position for self-report logic. _{sources: bitworld/among_them/bot-policies/sidecar/prompts/imposter.md, bitworld/among_them/players/evidencebot_strategy.md, bitworld/among_them/players/how_to_make_a_bot.md, bitworld/among_them/players/italkalot.md (+4)}

8. Self-report your own kill to seize the meeting, and bank alibis during cooldown

crewrift

Self-reporting a body you killed can be the strongest imposter play: it opens the meeting immediately, lets you flush your pre-queued accusation first, and frames you as the helpful discoverer. Gate it tightly -- match the body to YOUR recent kill via stored kill tick+position within a tight ~20-frame window and slightly loose radius (sprite offset), clear the kill marker after reporting to avoid double-reporting, pre-load chat that frames you as the finder while accusing a random innocent (deliberately NOT the most-recently-seen color, which is likely the victim), do it only with 5+ players remaining, and never if another player might have seen you near the body. During the long cooldown, proactively build alibis: be seen by others before sneaking off, fake tasks by standing at task areas briefly, and remember which players saw you where to cite as alibi evidence in meetings. Note: an abandoned body may take ~200 ticks to be found, and the resulting meeting resets your cooldown -- self-reporting controls meeting timing and maximizes your lead. _{sources: bitworld/among_them/bot-policies/sidecar/prompts/imposter.md, bitworld/among_them/players/evidencebot_strategy.md, claude-code:25488cd9-61d5-4231-bd87-e2744b6db043, opencode:ses_224eeb40bffeZqmBRcrjFgPFul}

9. Imposter meeting play: deflect with real cues first, route fabrications through identical templates, enforce role secrecy

crewrift

Compute suspicion for the imposter too but score ONLY non-teammates (never a known teammate; a ghost holds no suspicion), reading it as 'how sus this crewmate looks to the table,' and keep the legacy clear-leader vote rule for deflection (mis-ejecting crew is the imposter's GOAL, not a risk). Meeting priority, chat and vote coupled: (1) proactively accuse+vote a non-teammate who genuinely looks sus using REAL cues (strongest because it is not a lie); else (2) bandwagon onto whoever others suss, citing only SAFE hard-to-disprove FABRICATED cues (lurking on a vent, near a body, tailing me) and never a bold falsifiable witnessed-kill claim; else (3) skip at the deadline. Self-report or report-and-blame only SOMETIMES (always doing so is a detectable pattern -- inject randomness). Route an imposter's fabricated accusations through the IDENTICAL chat template/format as a crewmate's real ones (e.g. ' sus: <reason 1>, <reason 2>'), because any formatting difference is itself a tell; the imposter can safely weaponize chat precisely because its own crewmate logic ignores chat and is immune to the same manipulation. Enforce role secrecy at the prompt/schema level: never emit the role-revealing word, pass known allies as 'safe colors' that must never be accused, require every statement consistent with the FULL chat log and claiming only locations in the agent's own history, and reject LLM-proposed false self-role claims unless the objective explicitly permits deception. _{sources: bitworld/among_them/players/evidencebot_strategy.md, bitworld/among_them/players/mod_talks/DESIGN.md, bitworld/among_them/players/mod_talks/LLM_VOTING.md, coworlds/coworld-crewrift/players/notsus/README.md (+4)}

10. Specialize policy by role at the decision boundary, not in a shared payload

crewrift

Asymmetric roles (crewmate vs imposter, dead/ghost vs alive) need different objectives and tactics: a single role-agnostic system prompt with the role only in the data payload, or one shared 'pick a target' function, loses the strategic asymmetry the game rewards (crewmate votes on evidence, imposter votes to deflect/survive, deception chat needs its own module). Encode role/state capabilities into per-mode default directives -- a dead/ghost crewmate can still complete its own tasks so its default should beeline to tasks, while imposters default to hunting. Concretely in guided_bot the intended defaults are GUIDED_BOT_LLM_DISABLE=0 and GUIDED_BOT_LLM_GAMEPLAY_DIRECTIVES=0 (LLM controls meeting chat/voting, NOT non-meeting movement); if the LLM appears to affect movement outside meetings, confirm GUIDED_BOT_LLM_GAMEPLAY_DIRECTIVES=0. _{sources: claude-code:7b965041-3e1e-4d77-bc3a-a48662331630, claude-code:d4cde1e2-f4b5-4bd6-bf05-dd48002edf58, opencode:ses_21e58aef7ffeoJ7uZeDgGBCfIh, players/users/james/personal_cogs/among_them/guided_bot/coworld/DC_DEB}