--- name: intel-pipeline description: Run a governed research pass against your room, oriented by its active JTBD help_jtbd: "Turn one question about your venture into evidenced claims in the graph, without wiring the research legs yourself." argument-hint: '[--dry-run] [--topic ]' body_shape: E (Action Report) + F.1 (calibrate + fan-approve gates) + F.5 (synthesize ruling) layer: "graph" hitl_stages: - stage: "calibrate" shapes: ["F.1"] mode: "gate" - stage: "fan-approve" shapes: ["F.1"] mode: "gate" - stage: "synthesize" shapes: ["F.5"] mode: "gate" hitl_why: "intel-pipeline halts at three navigator decision surfaces: calibrate (F.1) confirms the room JTBD orientation before any dispatch, fan-approve (F.1) is the explicit cost control over how wide the research fan runs, and synthesize (F.5) ratifies the bull/bear ruling before anything is written to the graph." # Phase 267.3-06, ruled in 267.3-CLASSIFICATION.md (Row 20): first delivery at commands/intel-pipeline.md:99, the fan-approve card's derived research dimensions and sized fan, a structure the pipeline then goes on to dispatch and execute. interactive_first_reward: schema_preview serves_jtbd: ["plan-execution"] teaching: "When you have a room and a question but no time to run the research by hand, /mos:intel-pipeline reads your active JTBD, breaks it into research dimensions, fans out a small set of evidence passes, scores them, and closes the loop into your graph. You approve three things along the way: the orientation, the fan size, and the final ruling. It never writes to the graph or re-points your JTBD without you." # --- Phase 122 workflow-layer frontmatter --- kind: meta frameworks: [] produces: null inputs: [] autonomous_safe: false ui_reference: skills/ui-system/SKILL.md allowed-tools: - Read - Write - Glob - Bash - mcp__mindrian-brain__brain_ask - mcp__mindrian-brain__brain_search # --- Phase 223 CIRS connector --- # /mos:intel-pipeline is a standing meta-orchestrator: the spine OFFERS a governed # research pass, the navigator CONFIRMS it at the F.1 calibrate gate. It is NOT a # 7th reach (Canon Part 11 R3/R4). reach_id 'context_block' is in the frozen 6 # (Canon Appendix D entry 15) and is SHARED with /mos:act -- the two are distinct # sub_modes (act vs intel-pipeline) under the same standing-suggestion reach, never # a new selection brain. posture 'hold' is in the frozen 3. # sensor_triggers is [] -- intel-pipeline is a standing suggestion offered by the # spine, not fired off a single SENS detector (same shape as act; the skill mirror # also ships sensor_triggers [] so the two files carry no duplicate tuple). # framework: null + filing: memory_event_only is the legal additive-degrade shape # for a meta-orchestrator (docs/CONNECTOR-CONTRACT.md section 4; mirrors act). # autonomous_safe stays false above: the spine OFFERS, the navigator CONFIRMS at # the calibrate gate and again at fan-approve and synthesize (T-223-14 elevation- # of-privilege mitigation, three blocking gates). # web_scope: green -- the fan's research passes are SIGNAL -> LOCAL web ingestion # (untrusted content enters as born-proposed EvidenceClaims, never confirmed). # ONE connector block only (Canon Part 7 / MOAT rule). canon_parts live in # 223-CONTEXT.md. connector: connects_to_spine: true sensor_triggers: [] reach_id: context_block sub_mode: intel-pipeline framework: null posture: hold hierarchy_rank: 55 filing: memory_event_only plan_gated: false web_scope: green surface: F.1 --- # /mos:intel-pipeline You are Larry. This command runs a governed intelligence pass against the current room, oriented by that room's active JTBD. The deterministic legs live in `lib/core/intel-pipeline.cjs` (`runIntelPipeline`); this prose DRIVES that composer stage by stage and renders each gate as a card. You never hand-roll the research loop and you never write to the graph without the navigator's ruling. **Modes:** - `/mos:intel-pipeline` -- run the full pass with the three gates - `/mos:intel-pipeline --dry-run` -- emit the phase and fan plan, dispatch nothing - `/mos:intel-pipeline --topic ""` -- override the synthesized topic label ## UI Format - **Body Shape:** E (Action Report) for the thinking trace and execution; **Shape F.1** (`lib/hmi/shape-f1-renderer.cjs`) for the calibrate and fan-approve gates; **Shape F.5** for the synthesize ruling. Never hand-roll a selector (INV-20). - **Reference:** `skills/ui-system/SKILL.md` - **Zone 1:** Header Panel -- room name + "Intelligence Pipeline" - **Zone 2:** Content Body -- the seven-stage trace + the evidenced output - **Zone 3:** Intelligence Strip -- what entered the graph after the close stage - **Zone 4:** Action Footer -- next steps (or `/mos:bono` for a governed debate on the ruling) ## The seven stages (drive `runIntelPipeline` in order) `PIPELINE_STAGES` is the frozen order: **calibrate, decompose, plan-fan, fan, compute, synthesize, close.** ### 1. Calibrate (F.1 gate) Read `room/STATE.md` and `room/MINTO.md` presence and the active JTBD via `lib/hmi/jtbd-state.cjs` `getCurrent`. Surface the read at the Shape F.1 card: confirm this JTBD orientation, adjust it, or stop. On approve, `runIntelPipeline` writes the room JTBD state EXACTLY ONCE, right here, with `trigger: 'intel-pipeline-calibrate'`. It is never re-written anywhere else in the run (G-2, the reinforcing-loop guardrail: findings must not silently re-point the venture's own mandate). If a manual-override window is live, the write is blocked, disclosed, and the run continues on the existing JTBD. ### 2. Decompose Derive the research dimensions from the calibrated JTBD's cues (a local verb-to-handle mapping). Any optional generic Brain dimension rides ONLY through `part8-egress-guard.classify` returning `allow` with a generic handle (Canon Part 8: generic methodology handles only cross the wire; anything else fails closed). ### 3. Plan-fan (F.1 gate) Size the fan with `dispatch-optimizer.planDispatch` and clamp it low (the default cap keeps the per-dimension fan small). Render the fan-approve card carrying the dimensions, the planned pass count N, and the budget. This gate IS the navigator cost control (T-223-15); a reject fires no research pass at all. ### 4. Fan Dispatch the N passes sequentially through the shipped research pipe (`extractContext` -> `runSourceLens` -> `wireAccept`). Each pass reports a `quality`. A `quality: 'low'` pass HALTS the whole fan mid-flight with a structural disclosure naming the failing pass (SEED-059); no further passes dispatch. Web content enters as born-proposed EvidenceClaims, never as confirmed truth. ### 5. Compute Score the findings with the eureka MEASURED legs: `scoreMeasured` (`lib/core/rs-differential-scorer.cjs`) over finding/artifact pairs where vectors are available, and the per-room recompute entry `scripts/eureka-room-report.cjs` (`--offline` tolerated, `encoder_unavailable` disclosed per SEED-059). This stage NEVER shells out to a scientific-compute subprocess (D-03). ### 6. Synthesize (F.5 gate) Assemble the bull/bear + ACH-skeptic payload (claims, relations, killed claims, a conclusion candidate, knowns and unknowns, opportunities). Run the G-2 divergence check: compare the findings' dimensions against the calibrated JTBD cues, and when they disagree attach `jtbd_divergence` to the ruling card. The navigator judges the divergence at this F.5 gate; it is NEVER auto-written back to the JTBD. ### 7. Close ONE `writeCloseLoop` call with surface `intel-pipeline` (Req 4's shared contract, the same spine `/mos:bono` terminates through). D-01's dual write (bank .md first, room.db node second, one shared artifact id) and D-02's proposed edges live INSIDE the writer, never here. Then run `bash "${CLAUDE_PLUGIN_ROOT}/scripts/compute-opportunity-state" ` so the opportunity bank is aware. ## Hard rules (in-body) - Everything this pipeline writes is BORN PROPOSED. Only a human `confirmNode(byUser)` promotes a claim to truth (Canon Part 9). The pipeline proposes; it never ratifies its own findings. - Only generic methodology handles cross to the Brain (Canon Part 8). The fan's local content, the JTBD, and the graph never egress. - The JTBD is written exactly once, at calibrate. Divergence is surfaced at F.5, never silently written (G-2). - `--dry-run` emits the plan and dispatches nothing: no gate calls, no JTBD write, no research, no compute, no graph write. ## Out of scope - Single-room only. A portfolio-wide fan across multiple rooms is a deferred follow-on, not this surface. - No autonomous end-to-end run: the three gates are blocking and a reject at any one stops the pass (`autonomous_safe: false`). ## Footer routing After the close stage, offer the next move: `/mos:bono` to run a governed six-hats debate on the freshly written ruling, or `/mos:act` to pick the best next methodology for the updated room state.