crewmate-tactics

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

Crewmate tactics — recipes (crewrift tier)

On-demand recipes (14). Trigger→action heuristics; pull the relevant one when its situation arises.

1. Frame a CvC policy as four throughput loops; spend hearts as frontier ammo

crewrift

Build behavior so every action increases connected friendly junction count, prevents loss of connected junctions, or keeps the heart/gear/resource pipeline alive -- think discovery/economy/expansion/defense, not individual resource gain or wandering. Treat hearts as scarce frontier ammunition: spend them only on junctions one or two hops from the current net:cogs network and on hub-near junctions early (align radius 25), never on isolated far targets, deep Clips territory before a scrambler clears it, or stale junction memories. Exploit role HP modifiers beyond their obvious use: scouts get +400 HP (limit 500) and +100 energy (limit 120) so they survive ~500 ticks deep in enemy territory -- use that durability for recon, chokepoint-blocking, and target intel, not pure exploration; scramblers get +200 HP (limit 300); aligners and miners get none (limit 100). _{sources: archive/cogames_playground/alpha_cog/docs/cogs-v-clips-strategy.md, opencode:ses_219ac1095fferkumOQK8Ph6LTk}

2. Deficit-aware mining: target the scarcest element, tune K, commit per trip

crewrift

Because a heart needs 7 of EVERY element, a surplus in one cannot offset a deficit in another -- read team:* hub inventory and route miners to the most-depleted element (relative to the 7-each recipe plus planned gear costs), and don't camp gear stations the hub can't afford. Deficit-driven mining has a tuned local-optimum K (number of targeted elements), not 'more is better': on the 4-element economy the sweep was K=1:+0.34, K=2:+0.37, K=3:+0.03 vs substrate (K=3 barely filters, diluting the signal). Implement with a per-trip stable target re-picked only after a deposit, chosen via softmax over hub deficits (RESOURCE_TEMP=0.1 makes a 1-heart deficit ~2x weight, a 4-heart deficit ~16x) to steer toward scarcity without mid-trip thrash. Note: 'extractor loyalty' locking and forced post-deposit re-evaluation were byte-identical NO-OPs because the deficit ranking barely changes between bursty deposits -- before building any stale-signal mitigation (caching/locking/forced re-eval), verify the signal actually changes during the window you mean to protect. _{sources: alpha_cog/AGENTS.md, archive/cogames_playground/alpha_cog/docs/cogs-v-clips-strategy.md, claude-code:425dc11e-1c8d-4d12-bf44-adf01b6fa796, codex:019e194d-3a17-7eb2-9df9-7bae24112862 (+2)}

3. Heart economy: preload gated on hub surplus; hub camping is queueing, not waste

crewrift

The quantified cost of dying is heart waste, not death itself: 33-39% of all hearts produced are lost when carriers die, each lost heart costing ~5 miner trips to replace -- so spend/drop hearts before risky pushes rather than avoiding deaths (bounded upside: halving waste yields only ~7 extra alignments/episode). Aligners preloading 2 hearts per hub trip (chaining two alignments) shipped +1.90 mean on 50 seeds at 10k, but ONLY when gated on hub element surplus of at least 2x the heart recipe; without the gate they camp the hub when hearts are scarce and starve teammates (and a fixed preload of 2 becomes a liability in high-death environments, wasting 2 hearts of upfront mining per death). Conversely, hub camping during a true drought is OPTIMAL, not idle waste: with ~6 aligners competing for bursts of only 4-5 hearts, leaving forfeits priority on the next burst -- every mechanism that pulled aligners away during a drought LOST score. So: when hearts are scarce, top off and camp; when NOT scarce, top off and leave immediately rather than waiting for a full restock. Verify apparent idleness is not optimal queueing before giving idle agents busywork. _{sources: opencode:ses_1fb299595ffeQdQJTUCJtJD6Yu, opencode:ses_1ffba12c8ffegXTM7TLJ7oigMW, opencode:ses_1fbb083dfffeFNp623MAOdar2j, codex:019e1357-835f-7442-bd46-7e911617621f}

4. Put capability/state gates ahead of tactical branches: give Crewrift ghosts a terminal mode

crewrift · negative result · ⚠ session-derived, unverified · see related: S001 (other tier)

A dead/eliminated Crewrift agent (ghost) cannot report bodies and needs no threat-avoidance, yet ghosts keep doing their own tasks. Without an explicit terminal mode, a ghost fixates on a body it can no longer report and cycles hunting->fleeing on its own corpse (~every 240 ticks), wasting compute and self-defeating. Fix in RuleBasedStrategy: select a crewmate_ghost mode when self_role=="dead" BEFORE the report/flee branches, gate Report/Flee on 'can actually report' (live-crewmate / unknown state only), and default ghosts to task-doing (Normal/task mode). General principle: put capability/state gates (dead -> terminal mode) AHEAD of situational tactical branches in the mode selector, so an agent never enters logic it lacks the capability to use. _{sources: claude-code:4430a077-5411-426b-a87a-75512ec4300f, claude-code:91f456a7-41e8-4a89-98e9-44e379160b15, codex:019e7174-3079-7351-b8ea-b1bfcb917ff4, codex:019ea961-a6a9-7f42-ba2e-c8b65df11c22}

5. Crewborg flee thrash was a threshold defect: add sticky hysteresis, suspect the consumer

crewrift · negative result · ⚠ session-derived, unverified

A crewborg oscillation bug (flip-flopping between normal navigation and flee_from with no imposter visible) was a downstream THRESHOLD defect, not a Bayesian-suspicion error: crewborg's flee_from acts on the RETAINED last-known roster position of a target (not current visibility), so an off-screen threat still drives movement, and combined with a hard 60px flee threshold and no hysteresis the player flip-flopped per tick. Fix in strategy/rule_based.py with sticky Flee hysteresis: enter Flee at 60px, STAY in Flee until the active threat is beyond 100px, and stop fleeing a stale last-known position after 48 unseen ticks; clear sticky-flee state on meeting, report-body override, death, imposter role, or non-playing phases. Preserve the hysteresis when touching flee/approach behavior -- per-tick flip-flopping is the exact failure it exists to prevent. General lesson: suspect the threshold that CONSUMES a posterior before suspecting the posterior itself. _{sources: codex:019e8fc7-ebad-7b00-bc95-df1727641ce0, codex:019e9354-431e-7ae1-95a8-358bb4cf53de}

6. Guard evidence-free fallback tiers and never-satisfiable target loops

crewrift · negative result · ⚠ session-derived, unverified · see related: S002 (other tier)

Two Crewrift crewmate target-selection failures: (1) a 'geometry fallback' tier picking the nearest eligible station by pure distance can fire on the very first localized frame before any belief evidence (icons, radar pings) accumulates, committing with zero evidence -- gate evidence-free fallback tiers behind a warm-up or require at least weak evidence first. (2) No retry limit plus a completion signal that can never fire = infinite re-lock loop: an agent selected a station, held the action button, waited for the target icon to vanish, timed out because the icon was not its own, re-selected the same nearest station, and looped forever (losing ~65% of the match). Any action mode whose success depends on an external state change needs a per-target attempt counter / backoff / blacklist so a never-satisfiable target is abandoned. Relatedly, companion-follow behavior must degrade gracefully: pick a valid target (reject self/known-hostile/dead), perform the cover activity only while the target stays visible and in range, and interrupt it to reacquire the last-seen position when sight is lost -- 'never lose sight of the companion' is the core invariant. _{sources: opencode:ses_1ff325156ffei2UiDygaVOZo5O, opencode:ses_21a4f6bb3ffepgWI170Xv63XFW, codex:019e196e-93c2-7523-8da7-96f782ea578c}

7. Loop tactics: focus on imposter/voting, scale role mix proportionally, avoid no-op waits

crewrift · ⚠ session-derived, unverified

crewborg crew play is already near-optimal in self-play (perfect 48/48 task completion, no losses), so concentrate crewborg optimization on imposter and voting behaviors where almost all available fitness gain sits. When down-scaling a game from its default player count, scale the role mix and other count-coupled config proportionally, not just the player count: an 8-player 2-imposter/6-crew (1:3) default must become 1 imposter / 3 crew at 4 players (per-player tasks, cooldowns, vote timer can stay default, but forgetting the role ratio silently breaks balance). For exploratory perturbation, prefer deterministic-but-pseudo-random firing (a 1-in-N pseudo-random check) over fixed-cadence (every 8th tick) to avoid predictable/exploitable periodicity while staying reproducible. And delete 'hope the other agent does X' branches: a who_should_move=="hope_partner_moves" value was a no-op that burned hysteresis time before falling back -- when you cannot influence another agent, do something useful locally and trust shared-strategy convergence. _{sources: claude-code:8ec27261-9517-4365-944b-2880873bda47, claude-code:dfa4502e-c404-40e8-bf86-8d15c72ba5af, codex:019dfe75-7dd6-7dd3-90af-84bb28fda415, opencode:ses_1fb48dd11fferLGXS1vWdHa0wS}

8. Crewrift crew defense: report bodies, hoard the button, stay grouped

crewrift

Any Crewrift meeting (body report OR emergency-button call) resets EVERY imposter's kill cooldown, so a crewmate's solo trip to the once-per-game button is the field's strongest crew defense. Corpus of 1,875 league games: ~92.5% have a button call (mean 2.15/game), crew button-calls outnumber imposter ones ~3740 to 298. Rank reporting a visible body ABOVE accusing a suspect: a report opens a meeting on the spot without spending the button; reserve the one-shot button (Accuse mode) for a sustained active tailer, committing through the walk even if the tail briefly lapses (caveat: imposters can press it to blend in, and once button calls are exhausted imposters kill more freely if crew chased the button instead of tasks). Stay with groups of 2-3 and avoid being alone in low-traffic rooms (Electrical, Reactor); if a player follows you into an isolated room, leave -- isolation is when imposters kill. Group-and-watch (move as one group, watch everyone task section by section) makes any kill immediately reveal the killer and lets survivors vouch; the crewmate's core tradeoff is time on tasks vs time near others, so track what players do for later voting reference. _{sources: coworlds/coworld-crewrift/players/notsus/README.md, personal_labs/crewrift_lab/crewrift/crewborg/docs/designs/button-runne, personal_labs/crewrift_lab/crewrift/crewborg/design.md, bitworld/among_them/bot-policies/sidecar/prompts/crewmate.md (+1)}

9. Filter task targets by assignment evidence, latch resolved tasks, retract speculation same-frame

crewrift · negative result

When a game assigns each player a private subset of objectives (e.g. 8 of 40 stations), a proximity-only selector is fundamentally broken: it chases the nearest objectives regardless of ownership, revisiting points and completing things that aren't its job. Filter candidates by ASSIGNMENT evidence with an ordered fallback that prefers high confidence over speculation: visible task icon (definitely assigned and here) > confirmed-mandatory > closest mandatory > checkout task > radar-projection candidate (speculative, only when no icon evidence) > home rally point. Add a per-task 'resolved' latch so the list shrinks monotonically, treating both terminal outcomes identically: a task confirmed someone else's and one you completed both latch resolved ('never visit the same spot twice'); skip resolved stations before radar matching, clear a latch only if an icon genuinely renders again, reset all latches at round start. A speculative flag from ambiguous radar must be dropped on the SAME frame whenever the task's inspection rect is fully visible and neither a strict nor fuzzy icon match fires -- do NOT wait out a slow N-frame miss counter to retract a belief current vision already contradicts, or the bot loiters revisiting others' tasks. _{sources: opencode:ses_21e25877effeMecaawB2dUys7v, bitworld/among_them/players/SMART_BOT_GUIDE.md, bitworld/among_them/players/evidencebot_strategy.md, players_checkouts/players/users/james/personal_cogs/among_them/guided_ (+1)}

10. Task completion: external confirmation, Navigate->Hold->Confirm machine, target hysteresis

crewrift

Never mark a game objective complete on a local timer alone; confirm by direct evidence specific to YOUR action (the task icon disappearing while you still stand on the rect), not a team-wide aggregate -- a shared progress bar looks like completion but any other player bumping it during your hold produces false positives (~22% per hold observed); fall back to a team aggregate only when no specific signal exists, and gate an icon-vanished conclusion on having previously seen ~90% per-task progress so an occlusion/edge flicker doesn't false-complete. Failure mode: a crewmate that latched on its hold countdown hitting zero falsely believed it succeeded on UNASSIGNED tasks (server silently ignored its presses) and ground through the whole map. Structure completion as an explicit Navigate -> Hold -> Confirm machine: navigate to the station's passable centre; enter Hold only inside the exact bounding rect (match the server's no-margin containment); hold slightly above the server's acceptance window (e.g. 74 ticks vs ~72) -- padding just above avoids premature release without lengthening every task; Confirm by watching the icon disappear, with a deadband ('task ready' state) tolerating the icon briefly vanishing mid-hold. Add target hysteresis to avoid thrashing: lock a target for a commit window (~48 ticks/2s), re-evaluate ~every 24 ticks, switch only to a stronger evidence tier or the same tier at under half the current distance, and break the lock immediately if the target becomes completed or resolved-not-mine. _{sources: opencode:ses_21e25877effeMecaawB2dUys7v, bitworld/among_them/players/evidencebot_strategy.md, bitworld/among_them/players/italkalot.md, personal_labs/crewrift_lab/crewrift/crewborg/design.md (+2)}

11. Mode/reflex architecture: role-gated full mode switches, externally-detected reporting, ghost handling

crewrift

In a mode/reflex state machine a reflex should be a FULL mode switch (run the old mode's onExit, install the new directive, run the new mode's onEnter with fresh scratch, then run its decide() that same tick), not a one-tick override. Role-gate reflexes with mutually exclusive role conditions so the same world event routes by role -- seeing a body routes a crewmate to reporting but an imposter to fleeing -- and neither fires for the wrong role. Model body-reporting as an interrupt-driven mode entered only via a reflex (or LLM directive) when an unknown body appears, fired only from task-completing mode so meeting and higher-priority modes aren't interrupted; do not detect report success inside reporting mode (let the server starting a meeting trigger the voting-screen reflex into meeting mode). Gate body-reporting with a legality check (alive, non-ghost crewmate): imposters can't self-report on the default path, ghosts can't report, and even an LLM-issued directive must be rejected by that check. For ghost players, force task-completing mode regardless of any LLM directive (ghosts keep completing their own tasks and that capacity counts), disable body-abandonment, and use straight-line steering with no walk-mask pathfinding since ghosts pass through walls. On agent death (gear and carried items destroyed, no respawn/teleport) reset role to unknown and clear the objective but PRESERVE spatial localization and position so the agent navigates from its death location without re-localizing. _{sources: opencode:ses_21b475c59ffeOFBSt0jbwsGras, players_checkouts/players/users/james/personal_cogs/among_them/guided_, players_checkouts/players/users/james/personal_cogs/among_them/guided_, archive/cogames_playground/bulbacog/designs/MEMORY.md}

12. Derive task catalogs by role-play, then audit every mode after finding one half-finished one

crewrift · ⚠ session-derived, unverified · see related: S003 (other tier)

Derive a policy's task catalog by role-playing a full game per role and asking 'what must I do as this role at each game phase?', not by brainstorming a flat list of action verbs (which misses phase-specific tasks). Once you find one half-finished behavior mode, audit every other mode for the same defect -- finding one is strong signal. After a hold-forever task bug, a full audit found: reporting mode steering at stale body coordinates with no success check; fleeing/reporting modes standing still until a TTL expired; and several modes (fear, investigating, alibi_building, sabotage_watching) that were pure no-op stubs reachable only via explicit LLM directive. Take inspiration from an existing janky policy but rewrite the implementation rather than copying it verbatim, since a policy already flagged janky is unfit to drop into a new mode. _{sources: opencode:ses_209613a85ffeqO1Gq9NEDobBYi, opencode:ses_21a4f6bb3ffepgWI170Xv63XFW, opencode:ses_21e58aef7ffeoJ7uZeDgGBCfIh}

13. Persephone offense and coordination: cheap-reveal-first probing, mobility asymmetry, clock-aware risk, identical-code convergence

crewrift

In Persephone's Escape probe team affiliation with the cheap reveal first: a color/team exchange (~3-4s) triages allies before the more expensive role exchange; escalate to a mutual role exchange only with confirmed allies and keep a confirmed-non-targets list so you never re-probe a ruled-out player. Budget interactions against the round clock with config-adaptive thresholds read from the parsed schedule (intro Panel 3 yields (duration_secs, hostage_count) rows -- do NOT hardcode a 15s round): a full probe cycle costs ~9-13s (~6-9s skipping color exchange), so short rounds yield only 1-2 interactions and target selection is the single most impactful decision (probing the wrong player is unrecoverable) -- prefer nearby, cooperative-seeming, unknown-team targets, and express urgency (CALM/PRESSING/PANIC) as fractions of round duration AND rounds-remaining so a 5-round game's Round 3 reads mid-game. Exploit role-mobility asymmetry: the partner with no positional requirement (Cerberus, Demeter) is the mobile half and should volunteer as hostage to cross rooms while the co-located partner (Hades, Persephone) avoids moving; when all confirmed same-team players in your room are grunts, the partner is in the other room by elimination, so cross (preferring a room Hades is NOT in). Escalate risk as the clock runs out: with the exchange incomplete in the final round accept any risk including global-chat identity reveal, since the game is lost without it; a grunt's only final-round job, if the exchange is undone, is helping the two key roles meet. Cross-room comms are impossible, so teams rely on implicit coordination that converges ONLY when all agents run identical code -- every role's logic must still be correct when allies don't make the expected complementary move. Demeter must search louder than Cerberus (she seeks a hiding Persephone, not an actively-searching Hades) and should R.OFFER every confirmed Nymph immediately as a signal only Demeter would send. A grunt with no mechanical objective (the Shade) is the most flexible operative: relay intel, run interference, volunteer as hostage at no cost, decoy by falsely claiming the key role, and seek leadership; divide defensive grunts across leadership/probing/decoy/coverage roles. _{sources: players_checkouts/players/users/james/personal_cogs/persephone/agents/, opencode:ses_1fbfc5c8fffesCp5Ilc2Hs3Q6X, personal_cogs/persephone/agents/eurydice/DESIGN.md, personal_cogs/persephone/agents/eurydice/DEMETER_STRATEGY.md (+3)}

14. Throttle and gate entry-request presses so a retry doesn't cancel a pending request

crewrift · negative result

A repeated entry-request button press can CANCEL a still-pending request, so bound and throttle entry-request retries (e.g. a 72-tick interval with a global anti-cancel guard across task signatures) instead of pressing every tick while navigating. Also gate the entry-request task on belief preconditions (target recently seen in a chatroom and within interaction proximity) before pressing, to avoid accidentally creating a private chat in open space. _{sources: orpheus/DESIGN.md (Orpheus agent framework design)}