clowder

Work ONE 3-5 ticket related cluster to a verified colony-score gate. Discovers a cluster off the top of the ready queue (just next + just similar), confirms a plan with the user BEFORE any work, works it track-aware (swarm-safe polecats / substrate-sensitive sessions), then gates on exactly one out-of-sample soak + just verdict + a separate-model skeptic audit. Use when the user types `/colony-loop`, or asks "run a colony-improvement chunk", "work a related ticket cluster", "advance the colony score", "pick a cluster and land it against the gate". One chunk per invocation; ad-hoc, never scheduled. Do NOT fire for ticket recommendation only (use `/next`), single-session orchestration (use `/work`), balance hypothesis testing (use `/hypothesize`), or run diagnosis (use `/diagnose-run`).

Friction-log breadcrumb (`just agent-feedback "<note>"`) — appends one JSON line to `logs/agent-friction.jsonl` capturing what was tried, where it broke, and how severe. Fire at the *moment of friction*, not as a TODO list. Trigger conditions — "you tried two things and they both failed", "a tool's output didn't match its SKILL.md description", "you can't figure out which `just` command serves the user's intent", "you misread an envelope and had to retry after a course-correction". Do NOT fire for — routine errors that resolve on the first retry, syntax mistakes the user corrects in stride, gaps in your codebase knowledge that aren't tooling gaps, or as a venting channel.

Build a versioned baseline dataset (`just baseline-dataset <label>`) — five-phase orchestrator that runs a probe pass, an aggregate sweep, focal-cat traces, conditional weather variants, and a REPORT.md generation. Use when the user wants to lock in a new comparison baseline (the kind `just verdict` will read via `logs/baselines/current.json` or `just promote`). Trigger phrases — "build a versioned baseline", "rebuild the baseline dataset", "promote a new healthy baseline", "we need a fresh baseline against which to measure". Do NOT fire for — single-seed verification (use `just soak` + `just verdict`), one-off sweep work (use `just sweep`), drift checks against an existing baseline (use `just verdict` or `just sweep-stats --vs`).

Collapse overview for a failing Clowder soak — names the cluster pattern (e.g. "189-style schedule-edge"), surfaces the collapse trajectory, and suggests next-step drills. Use when the user asks "why did the colony collapse?", "what kind of failure is this?", or "does this look like a prior cluster?". Do NOT fire for — pass/concern/fail gating (that's `just verdict`), per-tick drilling (that's `just q`/`/logq`), or cross-run SQL (that's `just logdb-query`).

Explain a tuning constant (`just explain <constants.path>`) — doc comment + current value (from a recent `events.jsonl` header) + every read-site in `src/` + (if `logs/sensitivity-map.json` exists) per-metric Spearman rho. Use whenever the user asks what a knob does, what its current value is, where it's read, or what it would shift if tuned. Trigger phrases — "what does FOOD_DECAY_RATE do", "where is X read", "what knob controls Y", "what's the current value of …", "show me the doc comment for …", "list all the constants". Do NOT fire for — multi-knob hypothesis testing (use `just hypothesize`), wanting to *change* a constant (this tool is read-only — edit `src/resources/sim_constants.rs` directly), or runtime-derived values that aren't `SimConstants` fields.

Master-orchestrator skill for spawning swarm-safe polecats (Stage 2 of ticket 354). Reads in-flight polecat state + ready swarm-safe queue + disk, presents a conversational menu, dispatches to `just foreman-*` primitives. Use when the user types `/foreman`, or asks "spawn polecats", "drain the swarm-safe queue", "fire off N parallel sessions", "what are my polecats doing", "shut down all polecats", "tail polecat X". Polecats are headless child `Codex` CLI processes — they run swarm-safe-only tickets (docs / frontmatter / atomic refactor / verified-layer-walk bugfix), exit via `/handoff` + `jj git push --bookmark`, and the master foreman auto-lands their bookmarks via `just refinery --auto`. Do NOT fire for substrate-sensitive or coherent-block work (those need human-gated landing — use `/work`).

Per-DSE focal-cat trace diff (`just frame-diff <baseline> <new> [hypothesis.md]`) — ranks every DSE by |Δ mean(final_score)| between two trace sidecars and optionally classifies each row's concordance against a phase balance doc's hypothesis table. Use when comparing two focal-cat traces to attribute a behavior change to specific DSEs. Trigger phrases — "did the substrate refactor change DSE behavior", "compare focal traces between baselines", "which DSE moved between commits", "is the L1 invariant holding", "did Mate stop firing after the patch". Do NOT fire for — footer-field comparisons (use `just verdict` or `just sweep-stats --vs`), single-trace inspection (use `just q trace`), constants-only diffs (use `diff <(jq …)`).

Run a balance hypothesis end-to-end (`just hypothesize <spec.yaml>`) — baseline + treatment sweeps, concordance check, draft balance doc. Use whenever the user wants to test a balance prediction with the four-artifact methodology (hypothesis · prediction · observation · concordance) that AGENTS.md mandates for any drift > ±10% on a characteristic metric. Trigger phrases — "test if X causes Y", "validate the prediction that…", "run the four-artifact thing on…", "does increasing K shift the metric we expect", "I want to ship a balance change". Do NOT fire for — verdict checks on a single run (use `just verdict`), one-knob constant lookups (use `just explain`), per-DSE drift inspection (use `just frame-diff`), single-seed sanity checks (use `just soak` + `just verdict`).

One-cat post-hoc report (`just inspect <name>`) — personality profile, action distribution, score breakdown, needs timeline, relationships, key decisions, and death info read from `events.jsonl`. Use when the user wants the *aggregate* picture of one cat across a run. Trigger phrases — "tell me about <cat>", "who is <cat>", "what kind of cat is <cat>", "how did <cat> spend the run", "why did <cat> behave that way overall", "show me <cat>'s personality". Do NOT fire for — per-tick decision drilling (use `just q trace` or `just q cat-timeline`), focal-cat L1/L2/L3 trace analysis (that's the `clowder-focal-cat` subagent), or colony-wide queries (use `just q anomalies`, `just q deaths`).

Drill into a Clowder sim run with parameterized log queries (`just q …`). Use whenever the user is investigating a specific run, cat, tick range, or anomaly — after `/diagnose-run` surfaces something interesting, or before it, when they already know what they're looking for. Produces a consistent envelope (query echo, scan stats, stable IDs, narrative, suggested next queries). Trigger phrases — "why did <cat>", "what happened at tick N", "show me the <event> events", "drill into X", "look at run <path>", "was there a starvation spike", "which DSE did <cat> pick most". Do NOT fire for — the overview report (that's `/diagnose-run`'s job), editing sim source code, rebalancing constants without looking at a run.

Embedding-based ready-ticket recommender (`just next`) — top-K unblocked tickets adjacent to recent landings, in-flight work, the AI substrate refactor, or an ad-hoc seed. Use when the question is "what should I work on next" and the user wants the recommendation rooted in the existing ticket/landed/system corpus, not a fresh categorical scan. Trigger phrases — "what should I work on next", "best next ticket", "pick a candidate", "what's adjacent to what I'm doing", "what's the natural follow-on to X", "rank ready tickets by relevance to <Y>". Do NOT fire for — listing all ready tickets (use `just open-work-ready`), surfacing a specific ticket's neighbors (use `just similar <id>`), pair/linkage discovery (use `just similar-linkages`), or scoring tickets by impact/effort (no signal exists for that).

Triage a proposed Clowder simulation idea with the V×F×R×C×H rubric before writing code. Use whenever the user proposes adding a new system, feature, mechanic, creature, item, or behavior to the sim. Produces a score, bucket, and recommendation anchored against the shadowfox calibration — prevents narratively-cool, ops-expensive ideas from slipping in unpriced. Trigger phrases - "add X to the sim", "what if cats could Y", "new system for Z", "I want a feature that...", "we should have a <creature/item/mechanic>". Do NOT fire on balance tweaks to existing systems, bug fixes, or refactors.

One-shot corpus orchestration tagging ceremony. Walks the ticket corpus track-by-track (swarm-safe → coherent-block → substrate-sensitive) and applies `orchestration:` + optional `block:` / `verdict-anchor:` tags via the heuristic auto-classifier with per-batch user confirmation. Use when the user types `/retag` or asks to "tag the corpus", "apply orchestration tags", "promote tickets to swarm-safe / coherent-block", or "set up the three-track partition". Stage 0 step 4 of ticket 354; one-shot. Do NOT fire for single-ticket retags (use `just retag <id>` directly) or for new tickets (they tag at open-time via `just open-ticket`).

Semantic retrieval over Clowder prose (`just similar <id-or-path-or-text>`) — top-K adjacent chunks across tickets / landed / balance / system docs / DSE doc-comments. Use whenever a layer-walk audit, bugfix triage, or balance investigation needs to find prior work that "rings the same bell" but doesn't share keywords. Trigger phrases — "any tickets like this", "what's similar to ticket N", "have we seen this before", "find related balance threads", "which DSEs mention X", "this collapse looks like…". Do NOT fire for — symbol lookup (use `rg` / LSP), cross-run SQL (`just logdb-query`), per-tick drilling (`just q`), tuning-knob explanation (`just explain`), or full-cat reports (`just inspect`).

Statistical summary of a sweep (`just sweep-stats <dir> [--vs <other>]`) — per-metric mean / stdev / 95% CI / sample size from every `_footer` in a sweep directory, plus Welch's t / Cohen's d / effect-size band when comparing two sweeps. Use when the user has a sweep result and wants a structured numeric summary, or when comparing two sweeps quantitatively. Trigger phrases — "is this sweep different from baseline", "Welch's t between these two sweeps", "what's the variance on metric X across the sweep", "summarize this sweep", "sweep-stats this". Do NOT fire for — single-run drift checks (use `just verdict`), per-DSE focal-trace diffs (use `just frame-diff`), hypothesis-driven runs that compose sweeps automatically (use `just hypothesize`).

One-call run validation for a Clowder soak. Use after running `just soak` or `just soak-trace`, or whenever the user asks "is this run OK?", "did I break the colony?", "verify this run", "check this directory". Composes survival canaries + continuity canaries + constants drift + footer-vs-baseline drift into a single JSON envelope with a pass/concern/fail verdict and `next_steps` hints. Replaces the retired `just autoloop`. Do NOT fire for drill-down queries (that's `just q`/`/logq`), overview reports (that's `/diagnose-run`), or balance hypothesis testing (that's `just hypothesize`).

Daily-driver parallel-session orchestration. Reads in-flight sessions + ready queue + disk, presents an interactive menu, dispatches to the right `just` primitive (session-new / refinery / session-done / block-info). Use when the user types `/work`, or asks "what should I do today", "let's start a session", "land what's ready", "show me what's in flight", "spawn a foreman against the swarm-safe queue". Composes the three-track partition (substrate-sensitive / coherent-block / swarm-safe) from ticket 354. Do NOT fire for ticket recommendation only (use `/next`), specific log queries (use `/logq`), or balance investigation (use `/diagnose-collapse`).

Master-orchestrator skill for spawning swarm-safe polecats (Stage 2 of ticket 354). Reads in-flight polecat state + ready swarm-safe queue + disk, presents a conversational menu, dispatches to `just foreman-*` primitives. Use when the user types `/foreman`, or asks "spawn polecats", "drain the swarm-safe queue", "fire off N parallel sessions", "what are my polecats doing", "shut down all polecats", "tail polecat X". Polecats are headless child `claude` CLI processes — they run swarm-safe-only tickets (docs / frontmatter / atomic refactor / verified-layer-walk bugfix), exit via `/handoff` + `jj git push --bookmark`, and the master foreman auto-lands their bookmarks via `just refinery --auto`. Do NOT fire for substrate-sensitive or coherent-block work (those need human-gated landing — use `/work`).

One-shot corpus orchestration tagging ceremony. Walks the ticket corpus track-by-track (swarm-safe → coherent-block → substrate-sensitive) and applies `orchestration:` + optional `block:` / `verdict-anchor:` tags via the heuristic auto-classifier with per-batch user confirmation. Use when the user types `/retag` or asks to "tag the corpus", "apply orchestration tags", "promote tickets to swarm-safe / coherent-block", or "set up the three-track partition". Stage 0 step 4 of ticket 354; one-shot. Do NOT fire for single-ticket retags (use `just retag <id>` directly) or for new tickets (they tag at open-time via `just open-ticket`).

Daily-driver parallel-session orchestration. Reads in-flight sessions + ready queue + disk, presents an interactive menu, dispatches to the right `just` primitive (session-new / refinery / session-done / block-info). Use when the user types `/work`, or asks "what should I do today", "let's start a session", "land what's ready", "show me what's in flight", "spawn a foreman against the swarm-safe queue". Composes the three-track partition (substrate-sensitive / coherent-block / swarm-safe) from ticket 354. Do NOT fire for ticket recommendation only (use `/next`), specific log queries (use `/logq`), or balance investigation (use `/diagnose-collapse`).

Semantic retrieval over Clowder prose (`just similar <id-or-path-or-text>`) — top-K adjacent chunks across tickets / landed / balance / system docs / DSE doc-comments. Use whenever a layer-walk audit, bugfix triage, or balance investigation needs to find prior work that "rings the same bell" but doesn't share keywords. Trigger phrases — "any tickets like this", "what's similar to ticket N", "have we seen this before", "find related balance threads", "which DSEs mention X", "this collapse looks like…". Do NOT fire for — symbol lookup (use `rg` / LSP), cross-run SQL (`just logdb-query`), per-tick drilling (`just q`), tuning-knob explanation (`just explain`), or full-cat reports (`just inspect`).

Friction-log breadcrumb (`just agent-feedback "<note>"`) — appends one JSON line to `logs/agent-friction.jsonl` capturing what was tried, where it broke, and how severe. Fire at the *moment of friction*, not as a TODO list. Trigger conditions — "you tried two things and they both failed", "a tool's output didn't match its SKILL.md description", "you can't figure out which `just` command serves the user's intent", "you misread an envelope and had to retry after a course-correction". Do NOT fire for — routine errors that resolve on the first retry, syntax mistakes the user corrects in stride, gaps in your codebase knowledge that aren't tooling gaps, or as a venting channel.

Build a versioned baseline dataset (`just baseline-dataset <label>`) — five-phase orchestrator that runs a probe pass, an aggregate sweep, focal-cat traces, conditional weather variants, and a REPORT.md generation. Use when the user wants to lock in a new comparison baseline (the kind `just verdict` will read via `logs/baselines/current.json` or `just promote`). Trigger phrases — "build a versioned baseline", "rebuild the baseline dataset", "promote a new healthy baseline", "we need a fresh baseline against which to measure". Do NOT fire for — single-seed verification (use `just soak` + `just verdict`), one-off sweep work (use `just sweep`), drift checks against an existing baseline (use `just verdict` or `just sweep-stats --vs`).

Run a balance hypothesis end-to-end (`just hypothesize <spec.yaml>`) — baseline + treatment sweeps, concordance check, draft balance doc. Use whenever the user wants to test a balance prediction with the four-artifact methodology (hypothesis · prediction · observation · concordance) that CLAUDE.md mandates for any drift > ±10% on a characteristic metric. Trigger phrases — "test if X causes Y", "validate the prediction that…", "run the four-artifact thing on…", "does increasing K shift the metric we expect", "I want to ship a balance change". Do NOT fire for — verdict checks on a single run (use `just verdict`), one-knob constant lookups (use `just explain`), per-DSE drift inspection (use `just frame-diff`), single-seed sanity checks (use `just soak` + `just verdict`).

Triage a proposed Clowder simulation idea with the V×F×R×C×H rubric before writing code. Use whenever the user proposes adding a new system, feature, mechanic, creature, item, or behavior to the sim. Produces a score, bucket, and recommendation anchored against the shadowfox calibration — prevents narratively-cool, ops-expensive ideas from slipping in unpriced. Trigger phrases - "add X to the sim", "what if cats could Y", "new system for Z", "I want a feature that...", "we should have a <creature/item/mechanic>". Do NOT fire on balance tweaks to existing systems, bug fixes, or refactors.

Statistical summary of a sweep (`just sweep-stats <dir> [--vs <other>]`) — per-metric mean / stdev / 95% CI / sample size from every `_footer` in a sweep directory, plus Welch's t / Cohen's d / effect-size band when comparing two sweeps. Use when the user has a sweep result and wants a structured numeric summary, or when comparing two sweeps quantitatively. Trigger phrases — "is this sweep different from baseline", "Welch's t between these two sweeps", "what's the variance on metric X across the sweep", "summarize this sweep", "sweep-stats this". Do NOT fire for — single-run drift checks (use `just verdict`), per-DSE focal-trace diffs (use `just frame-diff`), hypothesis-driven runs that compose sweeps automatically (use `just hypothesize`).