--- name: bethesda description: Autonomous open-world exploration. Compound discovery loop — pick a cave, capture lore, cascade. The default action when no ticket is assigned and no human is steering. license: MIT tier: any category: intelligence autoInvoked: false dependencies: [] relatedSkills: [lore, gamedev, einstein, deepreflect, documentary, rounds, rally, cos, debate, call, ship, sync, nextphase, checkpoint, wrapup, inbox] requirements: env: [] integrations: [] --- # /bethesda — Compound Exploration You're in the open world. No waypoint marker, no assigned ticket — your move. This skill assumes a memory store is available. Search-and-capture is the spine of every cave. ## North Star (Read Every Session) 1. **Serve a real goal.** Every cave should advance the project — fix user pain, harden infrastructure, deepen knowledge, or build something that compounds. If you can't name the value, pick a different cave. 2. **Decide-and-ship is the default.** Use judgment: ship reversible, well-scoped changes directly; pause for review on architecture-level or hard-to-reverse decisions. The failure mode being prevented is wait-for-permission, not go-rogue. 3. **Quiet ≠ optimal.** Silence on idle is a failure to find work. /bethesda is the default action *because* there is always more to find. 4. **Capture or it didn't happen.** Every cave produces a memory with file paths and concrete numbers, or it was tourism. 5. **Cascade until a hard stop.** Each cave reveals the next. Stop only on: rate limit, the user steers, P1 incident, or the time budget is hit. 6. **Subtractive bias in self-improvement.** Every session may propose at most one cut to this skill (Phase 5). Skill changes get reviewed before they ship — never edit a skill file autonomously. ## Arguments - `[focus]` — domain hint (`infra`, `vault`, `external`, or an app/codebase name). Skips the Quest Board ladder, goes straight to that domain. - `deep` — single-cave focus mode. Pick one quest worth 60+ minutes (refactor, audit, deep research dive). Skip cascade. ## Session Budget | Time | Checkpoint | |---|---| | 0–5 min | Phase 0 orient — load context, announce TAKING | | 5–55 min | Cascade caves (3–7 typical, more if shallow, fewer if `deep`) | | 55–60 min | Phase 5 reflection — capture ≤1 improvement proposal IF a real one surfaced | | 60+ min | Phase 6 wrap — diary entry, summary post, stand down | | 90 min | **HARD STOP** — anything past here is prompt-cache death + diminishing returns | If `/loop 30m /bethesda` is firing in this slot, the next session picks up where this one stopped (diary leads carry). --- ## Phase 0: Orient (≤ 5 min) Gather orienting context for a compact brief — recent diary tail, current project state, open work, and your slot's default specialization. If a project-specific orientation script exists, run it; otherwise assemble the picture from the sources below. **Then top up with live signals (in this order, parallel where possible):** {{#IF_SLACK}} 1. **Team channel — read it first** (skip only if Slack is unavailable): ``` slack_read_channel(channel_id: "{{SLACK_CONTROL_CENTER_CHANNEL}}", limit: 30) ``` Other slots' recent posts. Anything addressed to you? Threads worth engaging? Cross-slot signal density is high — selecting a cave from your own context alone has produced wrong-cave selections that a quick channel scan would have caught. {{/IF_SLACK}} 2. **Memory context priming** — both queries in parallel: ``` memory_context(taskDescription: " /bethesda — what's most valuable to investigate now") memory_search(query: "", limit: 8) ``` The second surfaces THIS slot's recent captures (Phase 5 proposals, prior cave findings) that haven't hit the diary yet — the orient brief reads diary tails, not memory rows. If `memory_context` returns 0 results or times out, treat priming as failed: re-run with narrower keywords and note that you're operating on diary {{#IF_SLACK}}+ Slack {{/IF_SLACK}}signals only so others can calibrate trust. {{#IF_ISSUES}} 3. **Quest Board check** (if /gamedev has pre-built one): list open issues labeled `from-gamedev`. If 3+ exist, skip Phase 1 scanning entirely — go straight to Phase 2 selection. {{/IF_ISSUES}} 4. **Start the heartbeat** if not already running: `/loop 30m /bethesda` keeps the slot ambient between sessions. 5. **Announce TAKING** {{#IF_SLACK}}to the team channel {{/IF_SLACK}}(or to the diary if no Slack). Format: ``` [] /bethesda: entering. Diary loaded: . Priorities: . TAKING: . Why: . ``` If any phase returns rich enough signal that the first quest is obvious, skip Phase 1 entirely and execute. The Quest Board exists for when the orient brief leaves the next move ambiguous. --- ## Phase 1: Quest Board (if needed) ### Slot specialization (default domain) | Slot pattern | Default domain | |---|---| | `devops*` | infra, deploys, dependency drift, fleet hygiene, /ship + /sync + daemon health | | `feature*` | app/feature work in the relevant codebase (use last-touched signal to pick the area) | | `bugfix*` | triage sweeps, compliance audits, cross-cutting patterns, regression hunts | | `marketing*` | SEO, content, marketing site, growth | | `grunt` | any low-friction maintenance | | `cos`, `documentary` | their own skill domain — usually pre-routed | A slot's default domain is its first-look. **Cross-domain quests are still fair game** — if a devops slot sees an unclaimed P1 in a feature area, take it. Specialization is a prior, not a fence. ### Priority ladder 1. **Forward intent** — whatever the project's planning notes flag as next (highest weight) 2. **User pain** — feedback tickets, real bug reports, "broken since X" signals 3. **Production health** — failing deploys, P1 tickets, prod incidents 4. **Knowledge gaps** — memory domains with few entries (compound value for the next agent) 5. **Stale quests** — issues / open PRs untouched 7+ days 6. **Curiosity cascade** — diary {{#IF_SLACK}}or channel {{/IF_SLACK}}leads worth following If `[focus]` arg given, skip the ladder. If `deep`, pick ONE from priority 1–3. ### Quest sources Issue trackers, open PRs/MRs, the codebase itself{{#IF_SLACK}}, Slack triage{{/IF_SLACK}}, the memory store, the vault, and external research all carry quest seeds. Scan the ones available to you. --- ## Phase 2: Execute (Cave Pattern) ### 2A — Name the Leverage One sentence: *what real goal does this cave serve?* User pain, infrastructure hardening, a knowledge gap, or compounding capability. If you can't name one, pick a different cave. ### 2B — Sanity-Check the Decision Use judgment on the cave's central decision: - **Ship it directly** — the change is reversible, scoped to your specialization, not taste-loaded, and has no cross-cutting implications. Most caves. Proceed to 2C. - **Pause for review** — the decision is architecture-level, touches shared infrastructure, modifies a skill, contradicts a prior memory or decision, or you genuinely can't tell. Surface the question for review (a `/debate`, a teammate, or the user as appropriate), record the outcome, then proceed with the resolved answer. - **Escalate to the user** — taste calls (UI copy, naming, brand), money/vendor decisions, direction changes, kill/launch decisions. Surface it with context + your recommended answer + why you can't decide. Don't block this cave — pick a different one and resume when answered. **Skill changes always get reviewed before merge.** Never autonomously ship a change to a skill file. ### 2C — Execute the Cave Same shape every cave: 1. **Search before investigating** — `memory_search` the area for prior context 2. **Read the FULL ticket body, not just the title** — scan for design-block markers (`design review needed`, `no blind fix`, `awaiting review`, `deferred`). If present, surface the design question instead of fixing, and pick a different cave. 3. **Read scoped to the correct area** — don't read unrelated code for a narrowly-scoped ticket 4. **Form a hypothesis** — diagnose, but post as hypothesis not fact 5. **Decide ship vs. verify:** - Easily reversible + confident → fix it, test it, `/ship` it - High-risk or uncertain → post the hypothesis, request cross-slot verification 6. **Capture before moving on** (Phase 3) **Earn-trust posture:** post `[Slot] I think X because Y. @ can you verify?` Not `[Slot] X is the cause. Fixing now.` Cross-slot verification is a safety mechanism — a confident-but-wrong diagnosis caught by another slot is the system working. --- ## Phase 3: Capture (Most Important Phase) **Every cave produces:** 1. **A domain-specific memory** via `memory_pulse` — file paths, function names, concrete numbers, named consequences. The rest of the capture format auto-derives. Quality bar: a future agent reads it and is productive in 30 seconds. - Bad: "the app has feedback tickets" - Good: "App X: ticket 792 — a power user reports flower data resets on hard refresh. Root: localStorage clears on full reload. Fix: migrate to a server-side `user_flowers` table. Same user filed tickets 926/928 (dark mode, double-reported)." {{#IF_SLACK}} 2. **An ideas-channel post** if the smile test passes — would the user read this on their phone over lunch and feel good about the system working autonomously? Post to `{{SLACK_IDEAS_CHANNEL}}`. {{/IF_SLACK}} {{#IF_OBSIDIAN}} 3. **A diary entry** — per-cave (not just session end). Captures: timestamp, quest title, what you found, what you shipped, what the next cave should explore. {{/IF_OBSIDIAN}} **Capture triggers:** bug root causes, stated preferences, API gotchas, architecture decisions, silent failures, steering messages, philosophy statements, external discoveries that reshape strategy. **Don't capture** generic programming advice or observations obvious from reading the code. --- ## Phase 4: Cascade (Until Hard Stop) After every cave: 1. Look at what you discovered — does it point to the next cave? (e.g., gitignore noise → fleet-wide skill shadowing → a self-improvement design) {{#IF_SLACK}} 2. Re-read the team channel for replies/leads to your posts 3. Read the ideas channel — engage with at least one other agent's thread (challenge, connect, or build on their finding). The discussion is the product. {{/IF_SLACK}} {{#IF_ISSUES}} 4. Re-check the Quest Board — did your discovery change priorities? {{/IF_ISSUES}} 5. Pick the next quest, post `[Slot] TAKING: (cascaded from )` **Stop the cascade when:** - Rate limited on any integration — back off, capture, stand down - The user posts steering (read + respond, then re-orient) - A P1 production incident appears — that takes priority - The 90-min hard stop is hit - The diary has 5+ unexplored leads (you're generating faster than exploring — pick a fixable quest, ship, then stand down) --- ## Phase 5: Reflect + Self-Improve (Optional, Single Capture) **Capture at most ONE proposal — and only if a real one surfaced this session.** A "mandatory three" rule produces filler (proposals manufactured to satisfy the rule). If nothing real came up, capture nothing. Silence here is correct when the session was ordinary. When a genuine friction, surprise, or cut DID emerge: ``` memory_store({ content: "[bethesda-improvement-proposal type=] ", tags: ["active", "from-observation"] }) ``` The searchable signal is the `[bethesda-improvement-proposal type=...]` content prefix — `memory_search("bethesda-improvement-proposal", limit: 50)` matches that. **Skill changes require explicit approval.** This skill cannot be edited autonomously. When proposals accumulate (5+ unaddressed), draft the diff and surface it for review — don't ship without a yes. A reasonable cap is one skill-evolution change per slot per week; rapid back-to-back skill PRs are an anti-pattern this rule prevents. **Quality bar for proposals (when one surfaces):** - Concrete: name the file, the section, the line range. "Phase 1 is too long" is weak. "Phase 1 external-intel section (lines 289–317) wasn't relevant to this fleet-hygiene session — propose moving it to a reference doc" is actionable. - Tied to evidence: cite the cave where the friction or surprise occurred. - One to three sentences. No padding. --- ## Phase 6: Wrap (3 min) {{#IF_OBSIDIAN}} **Final diary entry** — meta-reflection if 5+ caves happened this session. {{/IF_OBSIDIAN}} {{#IF_SLACK}} **Final summary post — SKIP IF /ship ALREADY POSTED.** /ship emits a structured update to the team channel on every PR ship. When a session ends with a single-cave /ship, a wrap post duplicates that update seconds later. Detection: did this slot run /ship this session? If yes → either skip this post entirely, or downgrade to one line: `[] /bethesda complete. /loop 30m continues.` For sessions without a /ship (cascade-only, vault-only, no-fixable-quest), post the full wrap: ``` [] /bethesda session complete. Quests: completed Memories: captures across Discoveries: Cascades: PRs: Improvement proposal: Standing down. /loop 30m continues. ``` {{/IF_SLACK}} **Reset dashboard phase:** ```bash PREFIX=$(jq -r '.sessionPrefix // "traqr"' .traqr/config.json 2>/dev/null || echo "traqr") echo "idle" > "/tmp/${PREFIX}-phase-$(basename $(pwd))" ``` Don't post "checking for updates" on the next /loop tick. The orient step will tell you if something genuinely changed. --- ## Error Handling {{#IF_ISSUES}} - **Issue tracker unavailable**: skip ticket sweeps, use code + memory + diary leads instead {{/IF_ISSUES}} {{#IF_SLACK}} - **Slack unavailable**: capture to the memory store + diary only {{/IF_SLACK}} {{#IF_POSTHOG}} - **Analytics unavailable**: skip the anomaly check, note "analytics caves locked" {{/IF_POSTHOG}} - **Empty Quest Board**: memory archaeology (`memory_browse()`){{#IF_OBSIDIAN}} and vault lore{{/IF_OBSIDIAN}} always have unexplored caves {{#IF_MONOREPO}} - **All apps recently scouted** (within 24h): switch to `deep` mode on the most stale one, or pick a vault/external cave {{/IF_MONOREPO}} - **Rate-limit alarms**: stop cascading, capture what you have, stand down — preserve the next session's budget --- ## Reference The cascade pattern is universal: orient → pick a cave → execute → capture → cascade. The procedural flow above is the whole skill. Depth on any one topic — exploration sources, cascade examples, town-square write/read protocols, /gamedev integration — belongs in a project-specific reference doc, which is **subordinate** to this skill. If they ever conflict, this skill wins. ``` ╭─────────────────────────────────────────────────────────────╮ │ /\___/\ │ │ ( o o ) The open world is yours. │ │ ( =^= ) Pick a cave. Capture lore. Cascade. │ │ (______) │ ╰─────────────────────────────────────────────────────────────╯ Raqr · /bethesda {{PROJECT_DISPLAY_NAME}} ```