crewmate-tactics

name	crewmate-tactics
description	Use for crewmate tactics recipes in scripted Coworld policy optimization.

Crewmate tactics — recipes (loop tier)

On-demand recipes (6). Trigger→action heuristics; pull the relevant one when its situation arises.

1. When a sort-key penalty is inert, expand the candidate set, not the key

loop

When sort-key penalties over a candidate set prove inert, the problem is usually that the candidate SET is too small (0-1 items), not the key -- the next value is expanding/changing the set (add threatened held targets, high-connectivity targets, preemptive frontier captures, blacklisted trap targets) rather than adding another sort key over the same tiny set. Relatedly, a binary danger/safety penalty that fires uniformly on every candidate has zero effect on relative sort order: when the gating condition is ubiquitous (e.g. every candidate is near some enemy cell), use a distance-PROPORTIONAL penalty that discriminates between candidates instead of a binary in/out threshold. And always make the choice an explicit scored, priority-tiered ranking over threat/value (with hostile-presence detection and an abort guard for risky actions like a key exchange) rather than grabbing the first element (occupants[0]). _{sources: opencode:ses_1fb2002bcffeK2Ldmjf34wdpM3}

2. Architecture for a multi-role rule-based player

loop

(1) Separate strategy-selection from per-role priority logic via a ROLE_EVALUATORS table mapping role -> pure evaluator evaluate(strategic_state, belief_state, action_memory) -> ModeDirective(mode_name, params), so each role is its own testable function and the dispatcher stays trivial. (2) Encode each role as an explicit ordered priority chain (P1, P2, P3 ... Default) of guard-condition -> mode, written as numbered priorities not nested if/else, so coverage gaps are visible. (3) Drive strategy off a small named set of high-level modes (idle, scout, probe_target, probe_systematic, hold_position, coordinate_cross_room, seek_leadership), with a lower layer turning modes into concrete actions, keeping the strategy layer free of pathing/input detail. (4) Build incrementally behind a shared stub evaluator that always returns a safe default mode (e.g. 'scout') so the architecture compiles and tests pass before any real per-role logic exists. Make gear/role switching deliberate: switching clears the current gear and charges hub element costs, so switch only on explicit triggers (e.g. miner->aligner when hub element surplus is high and active aligners are few), and gate retooling on live territory safety -- if the desired role differs from current gear and the current cell is not friendly, enter a retreating state before routing to a gear station, or the unsafe agent may no-op or die en route. _{sources: codex:019e081d-ff5d-7111-90ab-9b441207a6ee, archive/cogames_playground/alpha_cog/docs/cogs-v-clips-strategy.md, archive/cogames_playground/bulbacog/designs/OUTER_LOOP_AND_MODES.md}

3. Roles are leased jobs; role-gate stuck-recovery (re-pick helps aligners, hurts miners)

loop · ⚠ session-derived, unverified

Treat roles as temporary jobs, not identities: give each selected role a short lease that a new role must beat by a switch margin to take over, preventing oscillation; allow stalled-mode and death/gear-loss as hard overrides, but block gear changes in live-unsafe territory (retreat first) and let cargo-carrying or heart-holding specialists finish their current task before retooling. Role-gate stuck-recovery: target re-picking on a stuck agent helps aligners/scramblers (any junction alignment is productive) but HURTS miners, because for a miner the deficit signal that drove the original target is load-bearing -- blacklisting the #1-deficit extractor concentrates miners on #2 and devastates spread-wins seeds, so let stuck miners keep a conservative bounded explore instead of re-picking. _{sources: archive/cogames_playground/bulbacog/designs/STRATEGY.md, opencode:ses_219f1f7a8ffeQVQYqola5M5a9n}

4. cogs_vs_clips role economy: per-agent baseline, shared phase machine, miner-heavy open

loop · negative result

A workable scripted cogs_vs_clips baseline gives each agent an independent per-agent state machine with no shared cross-agent state: pick a role (miner/scout/aligner/scrambler), pathfind to the depot to acquire matching gear if missing, execute the role's primary action, fall back to noop or local exploration when blocked; per-agent intent persists via a stateful wrapper and thresholds/movement biases live in overridable hyperparameters. For coordinated play, drive role distribution from one phase machine with a single shared state per policy and hysteresis on stock thresholds (e.g. hub-rich entry 210 / exit 180): phases progress GEAR_UP_AND_SCATTER -> EXPAND -> HOLD -> ENDGAME, reassigning roles deterministically by sorted agent id on boundaries, with a dual threshold for emergency pivots (moderate while the specialist is alive, higher when dead, since a single threshold either misses the pre-collapse window or panics). Open miner-heavy (prioritize gatherers until ~4 trusted miner claims) to bring the heart economy online -- the failure mode otherwise is support/aligner agents idling while resources sit uncollected -- and keep a miner floor through mid-game (retire miners only with a large resource buffer, sufficient hearts, and real defensive demand). Keep wandering scouts at zero standing demand and reassign equipped scouts (scout map memory is private and a scout doesn't directly score, so it's a rare fallback only when the team has no captures, no hearts, and poor map info). For a 2-of-6 underdog rescue, add a 'contested' lane mode that collapses all owned agents onto the single hottest contested lane when num_owned is below a small threshold. _{sources: players_checkouts/players/players/cogsguard/baseline/README.md, optimizers/.cursor/skills/scripted-navigation/SKILL.md, optimizers/.cursor/skills/scripted-navigation/reference.md, archive/cogames_playground/bulbacog/designs/STRATEGY.md (+3)}

5. cogs_vs_clips tactics: bottleneck mining, junction capture priority and legality, success detection

loop

Route miners to the heart-pipeline bottleneck element -- the element with the lowest pooled team stock, since buildable hearts = min(stocks)//heart_cost_per_element. Spread mild scarcity across nearby miners (avoid duplicating a visible miner's announced resource when levels are close; pick the next-lowest), but under severe deficits restrict all affected miners to that pool until projected claims close the gap. Prioritize capture/expansion targets in order: local capturable neutral, directional neutral expansion, neutral frontier repairs between friendly and enemy territory, then clean-enemy only when neutral work is exhausted or valuable enough to justify creating a contested junction; take cheap local legal captures opportunistically over distant stale ones, and treat equidistant targets as unequal via a territorial-coverage bonus (kept non-negative and capped, returning 0 when no anchors exist so it never overrides base distance before territory is established). Enforce action legality from exact junction ownership: align accepts neutral and clean-enemy (connected or disconnected) but must REJECT contested junctions (the align handler cannot clear enemy tags there), while scramble/cleanup accepts clean-enemy and contested; include contested junctions in BOTH friendly-network and enemy-threat queries since they carry both net tags. Detect action success by state change, not assumption: scramble/align succeeded when heart count decreases (consumes 1 heart), mine succeeded when cargo increases, move succeeded when last_action_executed matches intent. Do not add 'go recharge' logic for failed energy-limited moves -- energy auto-regenerates every step (fully near aligned buildings like the nexus AOE) -- just retry the failed action up to a small max (default 3). _{sources: archive/cogames_playground/bulbacog/designs/STRATEGY.md, optimizers/.cursor/skills/scripted-navigation/SKILL.md, optimizers/.cursor/skills/scripted-navigation/reference.md, archive/cogames_playground/bulbacog/designs/OUTER_LOOP_AND_MODES.md (+1)}

6. CoGsGuard vibe indirection, role-mix control, and gear/vibe debugging

loop

The CoGsGuard policy uses a 'vibe' indirection layer: each agent's behavior (miner/scout/aligner/scrambler/gear/default/heart) is selected by a settable vibe field, so an external coordinator steers by writing the vibe instead of editing the policy. Default and heart vibes noop; 'gear' picks a smart role then switches its vibe to it; miner/scout/aligner/scrambler each acquire needed gear then perform. Agents change vibe dynamically, so debug the current vibe per agent, not a static role. Control the initial mix via policy-URI query params (e.g. role?miner=4&scrambler=2&gear=1&aligner=2&scout=1); vibes assign by agent id in fixed order scrambler -> aligner -> miner -> scout -> gear; default is scrambler=1, miner=4 with the remainder defaulting to 'gear'; agents beyond the count get none; use role_cycle (comma-separated) for a repeating pattern by index or role_order for exact per-agent ordering. Agents must buy role-specific gear from gear stations using cogs-commons resources before executing (miner = 3 carbon/1 oxygen/1 germanium/1 silicon, scrambler = 1/3/1/1; bonuses like +40 cargo or +200 HP), so commons depletion gates role activation. Debug: a missing initial vibe is usually URI params not parsed (often shell quoting of policy_uri) or agent id exceeding the count -- confirm via the startup 'Initial vibe assignment' log; an agent that bumps a gear/depot station but stays NO_GEAR usually means the commons lacks the required resources or the agent isn't aligned with the station's commons (check commons inventory). Give a safe fallback rather than crashing on a missing precondition: skip _calculate_deficits and idle when heart_recipe is still unknown after N steps instead of raising and killing the episode. _{sources: players/cogsguard/role/AGENTS.md, players_checkouts/players/players/cogsguard/role/README.md, players_checkouts/players/players/cogsguard/role/CLAUDE.md, players_checkouts/players/docs/findings/baseline-policy-recipe-discove}