--- name: explore-opportunity description: Run the explored-stage analysis chain on a qualified opportunity help_jtbd: "Turn a qualified opportunity into a deep-researched, Minto-shaped, evidence-backed analysis." argument-hint: "[opportunity]" body_shape: E (Action Report) layer: "graph" layer_why: "Explicitly runs the analysis chain (deep dive across multiple methodology steps) for a qualified opportunity -- a command that dispatches a chain, the rubric's own step 1 signal." hitl_shape: "F.1" hitl_why: "Exploration spends navigator-controlled research cost and crosses material gates; explicit per-opportunity trigger only - never auto-fired on qualify." # Phase 265-04 reward-before-investment declaration (blocking-gate auto-fix, not # the dedicated ~85-command backfill phase named in docs/reward-before-investment-rule.md). # Grounded in the shipped chain below: the four analysis legs (web evidence, timing, # analogs, demand validation) run and their findings are visible at the material filing # gate as a structural preview of the Minto-shaped explored artifact, before the # navigator invests in the approve verb that actually files it. interactive_first_reward: schema_preview serves_jtbd: ["explore"] teaching: "A qualified opportunity is still conceptual: a connection without a defined problem. /mos:explore-opportunity runs the analysis chain (deep research, diffusion timing, analogies, web validation) and files a Minto-shaped explored artifact - a governing thought backed by cited sources and typed graph evidence. Opportunities must be researched and explored to become well-defined problems; this is that step, and its cost stays in your hands." allowed-tools: - Read - Bash - WebSearch - WebFetch - AskUserQuestion # --- Phase 219-05 connector frontmatter (born-wired, Canon Part 11 CIRS R1 / D-10) --- connector: connects_to_spine: true sensor_triggers: [] reach_id: deep_research sub_mode: explore-opportunity framework: null posture: hold hierarchy_rank: 5 filing: none plan_gated: true # rides the sanctioned deep_research exception (D-07) web_scope: green surface: F.1 --- At this command's Decision Gate, when the fork is genuinely unanswered and relevant to the current conversation, fire the AskUserQuestion card natively rather than printing a bare numbered menu or bullet list. Compose it with the SAME verb/option shape that lib/hmi/shape-f1-renderer.cjs (renderShapeF1) produces and that lib/hmi/selector-dispatcher.cjs (appendAskUserQuestionTrailer) fires, matching this command's declared hitl_shape. Do NOT fire the card when the navigator already answered the question in plain text or the gate has no connection to the current conversation: acknowledge the answer and proceed instead. Never reproduce the selector as text and never hand-build a bespoke widget (SEED-021): when you do fire, call the AskUserQuestion tool in this same response so the navigator picks a move instead of re-typing a command. Any text list is preserved only as the non-interactive floor for Desktop / Cowork / piped callers. # /mos:explore-opportunity -- The Explored-Stage Analysis Chain > The EXPLICIT [Explore] action on a QUALIFIED opportunity (REQ-4). It NEVER auto-fires on > Qualify (navigator cost control, T-219-18): you trigger it per opportunity, and material > steps halt at real gates (Canon Part 3). The output must read as ANALYZED - a governing > thought with cited web sources and typed graph evidence, never a candidate one-liner. ## How It Works The chain composes through the SHIPPED machinery only (D-06): framework names resolve to registry commands via `lib/workflow/command-resolver.cjs::composeWorkflow` (zero hardcoded slugs; a framework with no command degrades honestly to a manual step), and `lib/core/chain-executor.cjs::runChain` is the one gated loop. The entry point is `lib/core/eureka/explore-chain.cjs::exploreOpportunity(roomDir, opportunityNodeId, opts)`. The four analysis legs, in order: | Leg | Framework (registry-resolved) | What it contributes | |-----|-------------------------------|---------------------| | deep_research | Hypothesis-Driven Problem Solving | web evidence via the FROZEN deep_research reach (generic handles only, Part 8; the framework-runner's existing egress chokepoint - never a raw fetch) | | diffusion_timing | Adoption-Capacity Theory | is the window open? (current Track-1-unrewired form, D-07) | | analogies | Four Lenses of Innovation | analog domains that solved this shape of problem | | web_validation | Jobs to Be Done (JTBD) | demand validation against the customer segment | ### How the four legs dispatch, and what bounds the cost The four legs are analytically INDEPENDENT: each is a separate framing of the same fixed opportunity node (web evidence, timing, analogs, demand validation). No leg consumes another leg's output for its own analysis -- the `prev` parameter threaded through `onStep` carries provenance plus the offline-degrade and engine_mode stamping, not analytical input. Because of that independence, three of the four legs now run **in parallel**: `diffusion_timing`, `analogies` and `web_validation` fire concurrently, because there was never a data dependency between them to begin with (`lib/core/eureka/explore-chain.cjs::runAnalysisLegsParallel`, Phase 265-18). `deep_research` still runs FIRST, alone, as a probe. **The cost guard -- read this before assuming parallel means "spend everything".** The sequential chain's `quality_early_stop` (`lib/core/chain-executor.cjs::runChain`) used to end the run the moment a step returned LOW quality, which meant a cold `deep_research` leg returning `insufficient_evidence` never spent the other three legs, honoring this command's own `hitl_why` commitment that "Exploration spends navigator-controlled research cost." Parallel dispatch has no "early" left to stop at once three legs are already firing together -- so the guard did NOT get dropped, it moved one step earlier: `resolveExploreCostPolicy` decides the mode and `probeClears` evaluates the `deep_research` probe leg with the EXACT SAME quality test `quality_early_stop` applied, and the remaining three legs are dispatched only when the probe clears. Plainly: a cold corpus still costs one leg, not four. **The override.** `EXPLORE_COST_MODE` picks the mode: `cost-conscious` (the default) runs the probe-first guard above; `all-legs` skips the probe and fans all four legs concurrently unconditionally. An unrecognized value falls back to `cost-conscious` -- a typo can never select the more expensive path. `EXPLORE_PARALLEL=0` is the operator escape hatch: it forces the original sequential `runChain` path regardless of cost mode. **The fallback, never silent.** If the parallel pre-pass throws, if the engine is absent (the D-20 OFFER below), or if `EXPLORE_PARALLEL=0`, the run falls back to the existing sequential `runChain` path and reports exactly which path ran and why in the `dispatch` key of `exploreOpportunity`'s return (`{mode: 'sequential'|'parallel', reason, ...}`). **What did not change.** The material filing step still runs through `runChain` with its unchanged `gateFn`, so filing still halts for your approve verb and nothing files without it. `lib/core/chain-executor.cjs` has zero diff, and the Phase 264 zero-diff gate in `tests/run-all-264.sh` (base SHA `c7c33eea449f6f227c4cfbb86f220acaac9b5ab8`) still passes. The navigator settled parallelizing these legs as `build-now-in-265` (265-04 Task 3); the build landed in Phase 265 plan 265-18. This is an in-process `Promise.all` fan-out inside `lib/core/eureka/explore-chain.cjs`, not a subagent dispatch, so this command needs no `Task` / `Agent` grant in its `allowed-tools` for it. Then the MATERIAL filing step halts at the gate. On your approve verb: 1. **Research corpus artifact (D-16 item 1):** files nested as `research//.md` with frontmatter (date, url-cited sources with accessed dates, topic handles) and registers as a `memory_artifact` graph citizen. 2. **Minto explored artifact (D-08):** governing thought + SCQA + MECE sections + citations, validated by `feynman-minto-invariants` BEFORE filing, nested per Decision 16 at `opportunity-bank/
//.md` (real domain-slug section, 216 contract); re-exploration updates in place via the `problem_hash` dedup. 3. **Graph wiring:** >=2 typed evidence edges (SUPPORTS/INFORMS), lifecycle + stage advance to `explored` through `advanceOpportunityStage` (D-17 append-only stage_history). 4. **Post-filing extraction (D-16 item 2):** the 218 extractor runs scoped to exactly the new artifact paths (`opts.paths` seam - never a full-room re-extract); extracted entities land proposed with DERIVED_FROM edges to the artifact node. The research is graph-visible to ALL engines at filing time. ## Offline Degrade (D-16 item 3) When the web leg is unavailable, the chain queries the room's OWN research corpus (tri-modal retrieval scoped to `research/` artifacts) with the step provenance `web: absent (room-corpus degrade)` and the run typed `web_degraded_local_fallback`. A COLD corpus returns `insufficient_evidence` with zero fabricated results and files nothing - never a crash, never a silent skip (D-19). ## D-20: The Engine-Unavailable Offer When a chain step's engine cannot run (probe failures beyond the graceful rungs, crash, missing substrate), the chain HALTS at a gate OFFERING **[LLM manual run (high effort)]**: the model performs the analysis directly via the frozen deep_research reach + native web tools (same Part 8 handles). NEVER the default, NEVER a silent substitution (the corepower lesson). Manual output carries `engine_mode: llm_manual_baseline` in the run trace and every filed artifact's frontmatter, and is EXCLUDED from calibration sets by that marker. Test seam: `MINDRIAN_FORCE_ENGINE_ABSENT=1`. ## What This Command Does NOT Do - It does NOT fire on Qualify. Qualification (`/mos:qualify-opportunity`) and exploration are SEPARATE material decisions; a qualified opportunity waits until you explicitly explore it. - It does NOT explore candidates. A non-qualified opportunity is refused with `not_qualified` - the human gate always precedes the spend. - It does NOT mint sub-rooms. A promoted opportunity graduating to its own workspace is a follow-on (D-21: sub-room minting is out of v1 scope). ## Canon Compliance - **Part 3**: the autonomous_safe prefix runs; filing is material and halts at the gate. - **Part 8**: the web legs carry generic handles through the existing audited egress chokepoint; all graph writes are LOCAL room.db via navigation.cjs. - **Part 9**: extracted entities land `proposed`; only human confirms promote. - **Part 11 (CIRS)**: born WIRED - explicit surface, hitl_shape F.1 declared above. - **Part 12**: the explored artifact teaches the analysis, not just the verdict.