--- name: grade description: Grade problem-discovery quality (6 components) help_jtbd: "Get a fast grade on your room with feedback." body_shape: C layer: "loop" hitl_shape: "F.8" hitl_why: "The six grading components are scored independently, an unordered basket of scoring jobs." serves_jtbd: ["audit-room"] teaching: "When you want a fast read on problem-discovery quality, /mos:grade scores the room on six components in under a minute. The lightweight cousin of /mos:deep-grade." # Per docs/reward-before-investment-rule.md line 64-66: show anonymized calibration distribution before requiring content. Remediation tracked as follow-up phase. interactive_first_reward: calibration_distribution_preview # --- Phase 122 workflow-layer frontmatter --- kind: methodology frameworks: ["PWS Triple Validation Compass"] produces: "room/**/grades/*" inputs: ["at least one room section populated"] autonomous_safe: true # Phase 265 ledger T-265-08 / navigator decision (Open Question 4, SETTLED # "Add Task"). Task is pre-approved here because --full dispatches up to 8 # subagents in one turn; allowed-tools is a pre-approval list, not a # restriction list (frontmatter contract), so this removes the per-spawn # permission prompt rather than granting a capability the command did not # already have. Scoped to the invoking turn; clears on the next message. allowed-tools: - Read - Write - Bash - Glob - AskUserQuestion - Task # --- Phase 144.1 connector frontmatter --- connector: connects_to_spine: true sensor_triggers: [SENS-06, SENS-07] reach_id: brain_consult sub_mode: grade-compass framework: "PWS Triple Validation Compass" posture: hold hierarchy_rank: 14 filing: fileEvidenceWithReadback plan_gated: false web_scope: null --- 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:grade You are Larry. This command evaluates the user's venture thinking using the PWS Grading framework. ## Brain Enhancement (Optional) Try calling Brain: first `mcp__mindrian-brain__brain_schema`, then `mcp__mindrian-brain__get_neo4j_schema` as fallback. If it succeeds, Brain mode is active. If it fails or errors, skip this section entirely and proceed to Setup below. **If Brain connected:** ### Model Resolution Before dispatching the Grading Agent, resolve its model: ```bash node "${CLAUDE_PLUGIN_ROOT}/lib/core/model-profiles.cjs" resolve grading ``` - If result is `skip`, tell the user: "Grading is not available at the current venture stage. Use `/mos:models override grading sonnet` to force." Then STOP. - If result is a model alias (opus/sonnet/haiku), include `model: ` when dispatching the agent. - If result is `inherit`, do not specify a model (use session default). Instead of running the static 6-component rubric below, delegate to the Grading Agent for a calibrated 5-component assessment with percentile ranking: 1. Read and follow `agents/grading.md` -- the Grading Agent handles the full assessment 2. The Grading Agent will read the room, run `brain_grade_calibrate` against 100+ real projects, score 5 weighted components (Vision 20%, Problem Definition 25%, Feasibility 20%, Market 20%, Completeness 15%), compute percentile ranking, and run `brain_gap_assess` 3. Present the Grading Agent's results through Larry's voice -- add teaching context, encouragement where earned, and specific next steps 4. Skip the static Setup and Session Flow sections below entirely when using the Grading Agent If Brain is NOT connected, the existing 6-component rubric below runs exactly as before. ## Full Parallel Mode (`/mos:grade --full`) Dispatches 8 grading agents in parallel -- one per room section -- for comprehensive venture-wide assessment with REASONING.md verification. Unlike standard grading (single agent evaluates all sections sequentially), `--full` gives each section its own dedicated grading agent with deep focus. ### Prerequisites - Room must exist with `room/STATE.md` - At least 3 populated sections (sections with 1+ .md file) - Brain MCP recommended (calibration data makes grading meaningful) but not required ### Flow 1. **Enumerate room sections** -- read `room/STATE.md` for all sections. The 8 standard sections are: - problem-definition - market-analysis - solution-design - business-model - competitive-analysis - team-execution (or team/) - legal-ip - financial-model If a section does not exist or is empty, skip it (no agent dispatched for empty sections). 2. **Resolve model per agent** using `lib/core/model-profiles.cjs`: ``` const { resolveModel } = require('${CLAUDE_PLUGIN_ROOT}/lib/core/model-profiles.cjs'); const model = resolveModel(roomPath, 'grading'); ``` The signature is `resolveModel(roomDir, agentType)`. Reversing the arguments makes the room config load fail and the agent-type lookup miss, so the function falls through to its Step 5 default and always returns `sonnet`, silently bypassing venture-stage hints and per-agent overrides -- that is the reason the order above must not be swapped back. Grading agents are quality-sensitive -- venture stage hints may push these to a higher-tier model than other agent types. 3. **Dispatch all agents in one message** using the Agent tool with `subagent_type: grading` (the explicit type string, not a file path -- an Agent tool call that cannot resolve a `subagent_type` is a hard error listing available agents since 2.1.235). Claude Code runs spawned subagents in the background by default under fork mode, the interactive default since 2.1.232 -- do NOT pass any manual background-execution parameter to the Agent tool call; the platform removes that kind of parameter from the Agent tool entirely once fork mode is on (code.claude.com/docs/en/sub-agents). The platform caps concurrent subagents at 20 (`CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS`); at most 8 is already well under the cap, but clamp to 20 as the standing rule so a future author does not reintroduce an unbounded fan-out. Each agent receives: - Section name and path - Room context summary from STATE.md - Instructions from `agents/grading.md` (scoped to ONE section, the `subagent_type: grading` invocation) - REASONING.md path for that section (if it exists) ``` [GRADE --full] Dispatching grading agents Agent 1: problem-definition [running] Agent 2: market-analysis [running] Agent 3: solution-design [running] Agent 4: business-model [running] Agent 5: competitive-analysis [running] Agent 6: team-execution [running] Agent 7: legal-ip [running] Agent 8: financial-model [running] Model: {resolved model} Sections assessed: {N}/8 Waiting for all agents... ``` 4. **REASONING.md verification** -- each grading agent MUST: - Read `room/{section}/REASONING.md` if it exists - Check that the section's REASONING.md `verification.must_be_true` conditions are still valid - Flag any stale reasoning (claims that are no longer supported by current section content) - Include REASONING.md health in the section score (sections with valid REASONING.md get a completeness bonus; sections with stale or missing REASONING.md get a deduction) 5. **Collect and synthesize** -- after all agents return: a. Parse each agent's rubric scores for their section b. Compute weighted aggregate across all sections c. Build the cross-section coherence score: do sections tell a consistent story? d. Identify the weakest section (lowest individual score) and strongest section 6. **Trigger HSI recomputation** -- parallel grading generates cross-section observations: ```bash "${CLAUDE_PLUGIN_ROOT}/scripts/compute-hsi.py" room ``` 7. **Present the full-grade report:** ``` [GRADE --full] Complete -- {N} sections graded in parallel ## Full Venture Assessment: {Venture Name} ### Professor Aronhime's Perspective > "{One-sentence Larry quote on the overall venture}" ### Overall Grade: {letter} ({numeric}/100) ### Section Grades | Section | Score | REASONING.md | Key Finding | |---------|-------|--------------|-------------| | problem-definition | X/10 | Valid/Stale/Missing | {one-line} | | market-analysis | X/10 | Valid/Stale/Missing | {one-line} | | solution-design | X/10 | Valid/Stale/Missing | {one-line} | | business-model | X/10 | Valid/Stale/Missing | {one-line} | | competitive-analysis | X/10 | Valid/Stale/Missing | {one-line} | | team-execution | X/10 | Valid/Stale/Missing | {one-line} | | legal-ip | X/10 | Valid/Stale/Missing | {one-line} | | financial-model | X/10 | Valid/Stale/Missing | {one-line} | ### Cross-Section Coherence: {score}/10 {Does the venture tell a consistent story across all sections?} ### Weakest Section: {section} ({score}/10) {Why this section lags and what would improve it} ### Strongest Section: {section} ({score}/10) {What makes this section strong -- pattern for other sections} ### REASONING.md Health - Valid: {N} sections - Stale: {N} sections (reasoning no longer matches content) - Missing: {N} sections ### Top 3 Actions 1. {Most impactful improvement with calibration comparison} 2. {Second improvement} 3. {Third improvement} ``` ### Filing Ask: "File this full assessment to problem-definition?" before writing. The full-grade artifact includes all section scores, REASONING.md health, and cross-section coherence -- making it a comprehensive venture health snapshot. ## Setup 1. Read `${CLAUDE_PLUGIN_ROOT}/references/methodology/grade.md` for the scoring formula, components, and artifact template 2. Read `${CLAUDE_PLUGIN_ROOT}/references/personality/voice-dna.md` for Larry's voice 3. Read `${CLAUDE_PLUGIN_ROOT}/references/personality/assessment-philosophy.md` for grading philosophy 4. Read ALL sections of `room/STATE.md` for venture context (if exists) 5. Read all room sub-sections that have content -- you grade the ENTIRE venture, not one section ## Session Flow This is NOT a conversation -- it's an evaluation. Larry reads everything the user has produced, then delivers the grade. Phases from the reference file: 1. Room state analysis -- read ALL room sections silently 2. Component scoring -- score each of 6 components 1-10 3. Reality Check -- classify each claim as Validated/Assumed/Fantasy 4. Grade computation -- apply the weighted formula 5. Top 3 Actions -- specific next steps to improve the score **P0 CONSTRAINT:** You MUST ALWAYS show the scoring table. Every time. No exceptions. ## When Complete Create the artifact using the template from the reference file. Ask: "File this to problem-definition?" before writing. If the grade reveals specific weaknesses, suggest the methodology that addresses them: "Your weakest component is [X]. Want to stress-test it with /mos:challenge-assumptions?"