--- name: room-proactive description: > Proactive Data Room intelligence. Surfaces gaps, contradictions, and convergence signals. Active when room/ exists with entries. license: BSL-1.1. See LICENSE for complete terms (Business Source License 1.1, Change Date 2030-04-16 to Apache License 2.0). # activation: "resolve_room:active" <- INERT. Claude Code does not read an `activation` frontmatter # key (its documented set is name, description, disable-model-invocation, allowed-tools, # disallowed-tools, arguments, context, background), and no code in this plugin reads it # either. Kept as a comment so the INTENT survives; it never gated anything. Pinned by # tests/test-skill-frontmatter-inert-keys.cjs. paths: - "**/STATE.md" - "**/ROOM.md" - "**/MindrianRooms/**" - "**/.rooms/**" # --- Phase 172-06 CIRS R1 exclude (Canon Part 11) --- connector: excluded: true reason: "Ambient always-on infra. The proactive cross-relationship scan runs after every filing as part of the ambient loop; a continuous substrate, not a single problem-state reach." hitl_stages: - stage: "proactive-finding-response" shapes: ["F.0"] mode: "gate" - stage: "filing-offer-close" shapes: ["F.1"] mode: "gate" hitl_why: "Two independent decision-close moments: an APPROVE/REJECT/DEFER mini-gate on a surfaced finding, and an F.1 Next Move gate when offering to file a conversation artifact." --- # Room Proactive -- Gap, Contradiction, and Convergence Detection The Room is an active thinking partner. This skill surfaces what is missing, conflicting, and strengthening. ## Skill Activation This skill activates when `scripts/resolve-room` finds any active room with entries. The resolver checks (in order): central registry at `~/MindrianRooms/.rooms/registry.json`, directory scan under `~/MindrianRooms/`, workspace registry, and legacy `room/` fallback. All analysis below operates on the resolved room path. ## Activation Triggers | Trigger | Behavior | |---------|----------| | SessionStart | Max 2 HIGH findings. Prioritize 1 gap + 1 convergence (or contradiction). | | /mos:status | All HIGH + MEDIUM findings grouped by type. | | /mos:room --insights | Full analysis including LOW with interpretation. | | Methodology session | NEVER interrupt. Save for next SessionStart. | | PostToolUse (cascade complete) | When additionalContext matches `^post-write: cascade complete` OR `^queued MINTO regen`, read `/.mindrian/last-cascade.json`. If `proactive_intelligence.newFindings` is non-empty, present max 2 for APPROVE/REJECT/DEFER using Decision Capture flow. | ## Gap Detection - **Single-lens:** All entries from same methodology. Suggest complementary. - **Evidence gap:** Entries without validation/evidence markers. - **Adjacent section:** Connected sections filled, bridging section empty. - **Depth gap:** All entries at `depth: quick`. Phrase as opportunities, not criticisms. Suggest specific commands. ### Team-Execution Leadership Signals Detect these conditions in `room/team-execution/` and `room/team/`: | Signal | Condition | Confidence | Message | |--------|-----------|------------|---------| | `GAP:TEAM:no_profiles` | team-execution/ has 0 member profiles AND team/members/ empty or missing | HIGH | "No team mapped yet. Leadership starts with knowing who you're leading. Try: /mos:leadership" | | `GAP:TEAM:no_mentors` | team/mentors/ empty or missing AND 3+ team members exist | MEDIUM | "Team of [N] with no advisors mapped. Most ventures this size benefit from external perspective." | | `GAP:TEAM:solo_founder` | Only 1 person in team/ AND venture_stage past Pre-Opportunity | MEDIUM | "Solo at [stage] stage. The question isn't if you need a team -- it's what kind." | | `GAP:TEAM:no_assessment` | team-execution/ has member profiles but no leadership assessment artifact | MEDIUM | "Team profiles exist but no leadership assessment. Run /mos:leadership to diagnose team dynamics." | | `GAP:TEAM:stale_assessment` | Leadership assessment artifact older than 30 days AND room has new entries | LOW | "Leadership assessment is [N] days old. Team dynamics shift -- worth revisiting." | ## Contradiction Detection Scan for incompatible claims: customer type, market size, problem definition, timing assumptions. Frame as tensions worth reconciling. Check `created:` dates -- natural evolution (old X -> recent Y) is progress, not contradiction. ### Team-Execution Contradictions | Signal | Condition | Confidence | Message | |--------|-----------|------------|---------| | `CONTRADICT:TEAM:capacity_mismatch` | solution-design requires capabilities not present in any team profile | HIGH | "Solution design needs [capability] but no team member covers it. Hire, partner, or simplify?" | | `CONTRADICT:TEAM:stage_mismatch` | Team structure suggests Ready to Build but venture_stage is Pre-Opportunity | MEDIUM | "Full team assembled but problem isn't validated yet. Building before validating is the #1 startup killer." | ## Convergence Detection Same domain/customer/risk/theme in 3+ artifacts from different methodologies. Phrase as signal strength. ## Confidence Scoring | Level | Criteria | Display | |-------|----------|---------| | HIGH | Direct structural evidence, 3+ entries, clear conflict | SessionStart | | MEDIUM | 2 entries, keyword overlap, single-lens | /mos:status | | LOW | Single entry inference, weak match | Explicit request only | ## Noise Gate 1. SessionStart: max 2 findings 2. Never interrupt methodology 3. Stage filtering: Pre-Opportunity suppresses financial/legal gaps. Investment elevates all gaps. 4. Never repeat unchanged findings consecutive sessions ## Mid-Session Intelligence When the post-write cascade completes, the bash hook emits a tight one-line `additionalContext` advisory and writes the full cascade payload to a side-channel file at `/.mindrian/last-cascade.json`. This skill reads the side-channel file when it sees the trigger pattern, then renders findings using the cool-UI style canon (banner with rules + status grid + Shape A summary - NEVER raw prose). ### Detection The post-write hook's `additionalContext` matches one of two patterns: - `^post-write: cascade complete for ` (writes outside a recognized room section; payload is minimal) - `^queued MINTO regen for ` (writes inside a section; payload is full) When EITHER pattern fires, read `/.mindrian/last-cascade.json` (where `` is the path returned by `scripts/resolve-room` - the active room). Use the Read tool. The file is LOCAL only per Canon Part 8; never query the Brain or any network surface during this read. ### Behavior 1. If the side-channel file does not exist or fails to parse: the cascade did not produce intelligence. Do nothing. Soft-fail. 2. If `proactive_intelligence.newFindings` is empty or missing: the cascade ran but found nothing new. Do nothing. 3. If `proactive_intelligence.suppressed > 0`: silently note the suppression count. Do NOT mention it to the user. 4. If `newFindings` has 1+ items with `confidence >= 0.60`: present using the "After Filing: Decision Capture" flow below. Max 2 findings (highest confidence first). 5. If a finding has `isNew: false`: it is a previously-seen finding with new evidence. Present with the "I've seen this signal before" framing. 6. NEVER interrupt during a methodology session. Save for next non-methodology turn. ### Render Contract (cool-UI style canon) The cascade-finding render uses the cool-UI style from `.planning/research/cool-ui-style-reference.md`. Specifically: - **Banner** with thin horizontal rules (`━`) and a `►` separator. Title format: ` ROOM ► CASCADE FINDINGS` (one leading space; ALL CAPS; no em-dash). - **Status grid** with two-space indent, glyph at column 0, label-then-value with alignment whitespace (no colons; alignment IS the punctuation). Glyph vocabulary is fixed: - `◆` active / in-flight / configured (use for section + artifact rows) - `⚠` warning / contradiction (use for contradiction findings) - `⚡` urgent / convergence (use for convergence findings) - `⬜` gap / pending (use for gap findings) - `▶` next-action callout (single row at bottom of grid pointing to the decision verbs) - **Soft prose paragraph** below the grid: 2-3 sentences explaining WHY this finding matters in this venture's context. Plain prose, not bullets. - **Decision prompt** routes through the existing prose APPROVE/REJECT/DEFER flow (see "After Filing: Decision Capture" below). The F.0 AskUserQuestion selector is INTENTIONALLY deferred to Phase 88.2 / Phase 97 per `room/decisions/decision-phase-95-sequencing.md`. Phase 95 only changes WHERE the finding payload comes from, not the renderer itself. Example render (one contradiction finding, confidence 0.78): ``` ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ROOM ► CASCADE FINDINGS ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ◆ section problem-definition ◆ artifact customer-discovery-2026-04-29.md ⚠ contradiction customer-type vs market-size (confidence 0.78) ▶ next APPROVE / REJECT / DEFER The customer-type claim in this artifact (mid-market enterprise) sits in tension with the market-size assumption you filed in business-model last week (SMB segment). Worth reconciling before the next funding round. ``` NO emoji. NO em-dashes. Glyphs only. Aligned columns. The alignment IS the punctuation. ### New Evidence Indicator If a finding has `isNew: false`, it means this is a PREVIOUSLY SEEN finding that has NEW EVIDENCE (confidence or message changed). Present with the "I've seen this signal before" framing in the soft prose paragraph below the grid: "I have seen this signal before, but new evidence just shifted it. [finding message]. Confidence is now [0.xx] (was different before)." This distinguishes updated findings from brand-new discoveries and helps the user understand why they are seeing something again. ### Why a side-channel file (not in-band JSON) The PostToolUse envelope's `additionalContext` is a one-line string. The cascade payload contains structured data (classification, gitCommit, graphIndex, proactiveIntelligence with arrays of findings). Embedding it in the `additionalContext` string would (a) blow past the practical line-length budget, (b) require the skill to parse JSON-in-a-string, (c) couple the skill's contract to envelope-schema churn. The side-channel keeps the envelope clean and the data structured. This pattern is canonical in the plugin (see Phase 88-08 `pre-compact-snapshot.json`, Phase 88-06 `session-snapshot.json`). LOCAL only per Canon Part 8. NOTE: This contract REPLACES the previous load-bearing-broken-since-88.1-03 detection that read the cascade payload from the PostToolUse `additionalContext` envelope. The bash hook never wrote the cascade payload at the location the skill expected. Phase 95 corrects the data flow by relocating the payload to the side-channel file. The prose APPROVE/REJECT/DEFER renderer below (existing APPROVE/REJECT/DEFER flow at "After Filing: Decision Capture") is BYTE-IDENTICAL to current - only the input source has changed. ## Proactive Filing Offer (Conversation Artifact Capture) The cascade flow above fires when an artifact is WRITTEN. This section covers the case BEFORE that: when a CONVERSATION (not a /mos: command run) yields a keepable artifact, Larry proactively offers to file it before moving on. The conversation IS the work product; the offer is what turns it into a receipt (Canon Part 10, "the room as receipt"). The silent failure mode this guards against: good thinking that is never filed is gone, and a new navigator does not know to ask. This is doctrine, not a deterministic hook. The "a keepable artifact just appeared" judgment is Larry's, made by reading the conversation, not by a bash detector. The OFFER routes through machinery that already exists - the Decision Gate, the canonical verbs, the AskUserQuestion selector. No new verb, no new selector format, no new engine (Canon Part 7 reuse). ### What it is When a discuss chunk settles into something the navigator would want again - a sharpened problem definition, a grounded competitive landscape, a chosen direction, a pilot or Phase 0 design, a synthesis - Larry closes the turn with a one-line offer to file it. He does not silently let it evaporate, and he does not wait to be asked. ### Triggers (offer) Offer at a natural artifact boundary, when EITHER the navigator has agreed to a framing OR a self-contained work product has formed: | Trigger | What just settled | |---------|-------------------| | Problem definition articulated or reframed AND the navigator agrees | A sharpened or re-expressed problem statement | | Competitive landscape / market scan produced | A grounded set of competitors, alternatives, or market structure | | A decision or direction is chosen | The navigator picks a path among options | | A plan / pilot / Phase 0 design is produced | A concrete next-step protocol the navigator could run | | A synthesis the navigator would want again next session | A wrap that collapses branches back to insight | ### Anti-triggers (do NOT offer) One offer per artifact. Honor the escape hatch. Do NOT offer when: - Mid-exploration with no settled artifact - the thought is still forming. - Small talk, greetings, or a clarifying side-question. - The navigator is mid-thought (a trailing "...", an unfinished list, an explicit "hold on"). - An offer was already made and declined this turn-cluster - no repeat nagging on the same artifact. - A methodology session is live - NEVER interrupt; the filing already routes through that command. The cadence rule mirrors the Noise Gate above: when in doubt, stay quiet. A wrong offer is worse than no offer. ### How to offer (existing selector, existing verbs) Close the turn with a Decision Gate F.1 Next Move selector (the AskUserQuestion primitive, the default after a discuss chunk per Canon Part 3). The options map onto the EXISTING canonical verbs - do NOT invent a "File this" verb: | Option (what the navigator gets) | Canonical verb (Part 3) | Where it routes | |----------------------------------|-------------------------|-----------------| | File it | Run Methodology | Route to the matching /mos: command that files this artifact type, resolved at surface-time (problem definition -> /mos:diagnose or /mos:structure-argument output; landscape -> /mos:research or /mos:compare-ventures; pilot/plan -> file as a decision/artifact). Larry never types the slug from memory; the resolver late-binds it. | | Bank it | Bank Opportunity (ADD) | When the artifact is an opportunity, ADD to the local Opportunity Bank with HSI score + domain tags. | | Not yet | Defer | DEFER edge; the gate remembers and re-surfaces at a milestone audit. | | Free-Text | Free-Text | Always available; Larry interprets and routes to one of the above. | The selector is F.1; the verbs are the canonical ten. A "Synthesize" option is offered instead of "File it" when the artifact is a wrap that collapses branches (Part 3 verb 7). The vocabulary is closed - this offer adds no verb to it. ### Reason capture (decline is data) A decline is the most valuable signal in the system (Canon Part 4; Decision #13). When the navigator picks "Not yet" or declines, capture the reason with ONE low-friction line - never block, never re-prompt: "Got it - parking it. One line on why, so I do not re-offer the same thing?" The reason becomes graph data; silent rejection is the failure mode the gate must make trivially easy to avoid. ### Surface-agnostic The F.1 / AskUserQuestion path works identically on CLI, Desktop, and Cowork. The offer is conversational on every surface - no surface-specific code (Tri-Polar rule). ### Part 8 (local-only) Filing writes LOCAL room artifacts / room.db only. The offer never queries the Brain. The artifact body, the problem definition text, the landscape, the pilot - none of it egresses. If a /mos: command in the "File it" route consults the Brain for generic methodology, it carries only framework handles and phase identifiers per Canon Part 8, never the artifact bytes. ### Evidence provenance when filing (evidence_node_ids) When Larry files through the MCP write path by calling `artifact_file`, and the content he is filing was grounded by node ids already returned in THIS turn's context, he passes those ids as `evidence_node_ids`. Filing is the moment the provenance link is cheap - after the turn ends the ids are gone. **Where the ids come from.** Name the real sources by registered tool name: a prior `room_search` result, a `graph_query` neighborhood read, a `graph_reason` multi-hop trace, or a `chain_run` result whose steps surfaced node ids. If a gate card or `suggest_next` payload in the turn carried node ids that the filed content is about, those count too. **The never-fabricate floor.** Only ids Larry literally saw come back from a tool call in this turn. Never invent an id, never guess a format, never reconstruct one from memory of an earlier session, never derive one from a filename or a title. When no evidence ids are in context, OMIT the field entirely. An empty or absent `evidence_node_ids` is the CORRECT answer for an ungrounded artifact, not a failure and not something to apologize for. A fabricated provenance edge is worse than no provenance edge, because it makes the graph lie. **Graceful degradation.** The parameter may not exist on the server this session is talking to. If the `artifact_file` call is rejected for an unknown or unexpected argument, Larry drops `evidence_node_ids` and re-files ONCE, everything else unchanged. He does not retry-loop, does not surface a system error to the navigator, and does not abandon the filing. The artifact landing is what matters - the provenance edge is an enhancement on top of it. **No ceremony.** Larry never announces that he is attaching evidence ids, and never lists them back to the navigator. This mirrors the standing "never announce tool calls" rule. The navigator sees a filed artifact, not the mechanism. **Part 8 boundary.** These are LOCAL room-graph node ids. They are a local provenance link only and never egress to the Brain, exactly like the artifact body itself under the `### Part 8 (local-only)` rule directly above. Filing still queries no Brain wire. **Why it matters.** Without the ids, a filing writes one bookkeeping `memory_event` row and no `SOURCED_FROM` edge, so nothing downstream can trace what the artifact was built from and `MINTO.md`'s governing thought has no claim node to read (`docs/2026-09-03-DESIGN-t2-write-back-minimal.md`). `artifact_file` has no deterministic caller, which is why this lives in doctrine rather than in code. **Forward compatibility.** If the tool also grows an `epistemic_type` argument, a conversation artifact filed this way is a `conclusion`, never a `decision` - only a gate approval mints a `decision`. Apply the same drop-and-refile degradation to it. ## After Filing: Decision Capture When the post-write cascade completes and the side-channel reader (above) finds `newFindings` in `/.mindrian/last-cascade.json`, present findings to the user for decision. ### When to Present - ONLY when the side-channel file exists AND `proactive_intelligence.newFindings` has 1+ items - ONLY present findings with confidence >= 0.60 - Max 2 findings per filing (pick highest confidence) - NEVER present during a methodology session (save for next filing) - NEVER re-present findings already marked `decided: true` ### How to Present Format each finding as a natural observation, not a system alert: "I noticed something while filing that [artifact title]. [Finding message]. Confidence: [0.xx] This [CONTRADICTS/creates a GAP in/CONVERGES WITH] your [section name]. What would you like to do? - **APPROVE** -- accept this impact (I'll note the cross-subsystem connection) - **REJECT** -- disagree with this finding (tell me why -- your reasoning becomes data) - **DEFER** -- park this for later review" ### How to Record When the user responds, run this command via Bash: For APPROVE: ```bash node bin/mindrian-tools.cjs record-decision --room ROOM_PATH --key "INSIGHT_KEY" --decision approve ``` For REJECT (reason is REQUIRED -- ask for it if not provided): ```bash node bin/mindrian-tools.cjs record-decision --room ROOM_PATH --key "INSIGHT_KEY" --decision reject --reason "USER_REASON_HERE" ``` For DEFER: ```bash node bin/mindrian-tools.cjs record-decision --room ROOM_PATH --key "INSIGHT_KEY" --decision defer ``` Where: - ROOM_PATH is the resolved room directory path - INSIGHT_KEY is the finding's dedup key (format: `type:subtype:section` for gaps, `convergence:term` for convergence, `contradiction:section1:section2` for contradictions) If the finding references specific artifacts with IDs, add `--source-artifact` and `--target-artifact` flags to create a graph edge. ### Rejection is Data (Decision #13) When a user rejects a finding, their reason is the most valuable signal in the system. ALWAYS capture the reason. If the user says just "no" or "reject", ask: "Got it -- can you tell me briefly why? Your reasoning helps me learn what matters for this venture." ### After Recording Confirm briefly: "Noted -- [decision]. [If reject: Your reasoning is now part of the room's intelligence.]" Do NOT follow up with more findings. One decision interaction per filing. If there were 2 findings presented, capture both decisions before moving on. ## analyze-room Signal Format - `GAP:STRUCTURAL:{section}:{confidence}:{message}` - `GAP:SEMANTIC:{section}:{confidence}:{message}` - `GAP:ADJACENT:{section}:{confidence}:{message}` - `CONVERGE:{term}:{count}:{confidence}:{message}` - `CONTRADICT:{section1}:{section2}:{confidence}:{message}` Script catches structural patterns; add semantic interpretation from actual Room entries. ## Capability Suggestions `CAPABILITY:{feature}:{confidence}:{message}` signals when room has enough data for features: | Feature | Threshold | Command | |---------|-----------|---------| | DASHBOARD | 3+ artifacts | `/mos:room view` | | EXPORT_DASHBOARD | 7+ artifacts | `/mos:room export` | | WIKI | 5+ artifacts + 1+ meeting | `/mos:wiki` | | MEETING_REPORT | 3+ artifacts + 2+ meetings | `/mos:export meeting-report` | | THESIS | 10+ artifacts | `/mos:export thesis` | | TEAM_VIEW | 2+ team profiles | `/mos:room view` | | LEADERSHIP_COACHING | 2+ team members + no leadership artifact | `/mos:leadership` | Max 1 capability suggestion per SessionStart. Natural voice. Never repeat used commands. ## Causal Discovery Surfacing (v1.7.0) Surface causal discoveries when graph has 5+ CausalClaim nodes AND 3+ CASCADES_TO edges: - CausalClaim + HSI_CONNECTION: explain the cause-effect behind similarity - CausalClaim through REVERSE_SALIENT: show chain to bottleneck root - CausalClaim + ANALOGOUS_TO: structural match prediction - Overdue predictions: prompt review - Cascade depth >3: warn about blast radius ## Dashboard Export Integrity ALWAYS use `scripts/generate-standalone` or `scripts/serve-dashboard`. NEVER generate HTML by hand.