company-skill 1.3.0 → 2.0.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "company-skill",
-  "version": "1.3.0",
+  "version": "2.0.0",
   "description": "Goal-driven multi-employee company for Claude Code. Give it a goal, it runs until done.",
   "bin": {
     "company-skill": "./bin/install.js"

package/skill/SKILL.md

@@ -17,467 +17,122 @@ allowed-tools:
   - Skill
 ---
 
-# /company, Goal-Driven Multi-Employee Company
+# /company
 
-## Preamble (run first, always)
+Give it a goal. Run every employee. Loop until verified done.
 
-```bash
-echo "════════════════════════════════════════════════" && echo "             🏢 COMPANY SKILL ACTIVE" && echo "════════════════════════════════════════════════" && echo "" && ([ -f COMPANY.md ] && echo "$(grep -c '^[0-9]\|^- ' COMPANY.md 2>/dev/null) roles found" || echo "No COMPANY.md, will auto-create")
-```
-
-Give it a goal. It runs the entire company in loops until the goal is verified done.
-
-```
-/company "Build the user authentication system with OAuth2"
-/company "Research all competitors and write a competitive analysis"
-/company "Fix the payment processing bug and deploy"
-```
-
-You set the goal. Walk away. Come back to a STATUS.md with everything accomplished.
-
-## Core Loop
-
-```
-GOAL: "{what the user asked for}"
-  |
-  THINK ── Leads break the goal into tasks, assign to employees
-  |
-  EXECUTE ── Employees do the work (code, research, scan, test)
-  |
-  VERIFY ── Built-in Reviewer + Critic check if the goal is actually met
-  |
-  Done? ── NO: loop back to THINK with feedback on what's missing
-         ── YES: write STATUS.md, report to user
-```
-
-The loop does NOT stop until VERIFY confirms the goal is met.
-
-## Built-In Roles (Always Active)
-
-These exist in EVERY company, regardless of COMPANY.md. They are the backbone.
-If the user already defined any of these roles, the built-in version merges with theirs (no duplicates).
-
-### Leadership (THINK phase)
-**CEO**, Reads the goal, sets priorities, resolves conflicts between departments. If the user defined a CEO, that definition is used. If not, auto-created.
-
-**CTO**, Technical decisions, architecture review, code quality standards. Auto-created if not defined.
-
-### Quality Gate (VERIFY phase)
-**Internal Reviewer**, Reviews all work after each cycle. Checks: is it correct? Does it address the goal? Are there bugs or gaps? Grades each success criterion as MET / NOT MET / PARTIALLY MET.
-
-**User Advocate**, "Would a real user understand this? Is the API ugly? Does the README make sense in 10 seconds?" Represents the end user who doesn't care about internals.
-
-**Devil's Advocate**, Attacks every result. "What could go wrong? What did we miss? Would an external expert accept this?" Only satisfied when there are zero remaining holes.
-
-**Elegance Enforcer**, "Can this be simpler? Is there unnecessary complexity? Does every component justify its existence?" Prevents over-engineering.
-
-### Deduplication Rule
-When parsing COMPANY.md, check if the user defined roles matching these built-ins:
-- Match by name similarity: "CEO", "Chief Executive", "CTO", "Tech Lead", "Reviewer", "Critic", "Advocate", etc.
-- If match found: use the USER'S description but ensure the role runs in the correct phase (CEO in THINK, Reviewer in VERIFY)
-- If no match: auto-create with default description
-- Never duplicate: one CEO, one CTO, one Reviewer, one Advocate, one Elegance Enforcer, one User Advocate
-
-This means a 2-person COMPANY.md (just "Backend Dev" + "Frontend Dev") automatically gets: CEO + CTO + Backend Dev + Frontend Dev + Internal Reviewer + Devil's Advocate + Elegance Enforcer + User Advocate = 8 employees running.
-
-## Step 1: Parse Goal + Company
-
-Read the user's goal from the command argument or their message.
-
-```bash
-for f in COMPANY.md company.md; do [ -f "$f" ] && echo "FOUND: $f" && break; done
-```
-
-If no COMPANY.md found:
-1. If a goal was given, create a COMPANY.md with a minimal team tailored to the goal (engineering dept for code goals, research dept for analysis goals, etc.) plus the built-in roles. Write it to disk so the user can edit it later.
-2. If no goal given, create COMPANY.md from the template with placeholder departments and tell the user to fill it in before running again.
-
-**Parsing rules:**
-- Roles are `- ` lines that appear UNDER department headers (`## Department Name`)
-- Lines under `## Priorities`, `## Rules`, `## Communication`, `## Protocol` are NOT roles
-- Stop collecting roles when hitting the next `##` header
-- A role line typically has a name followed by a comma and description: `- Role Name, description`
-
-Classify roles:
-- **THINK (Opus):** lead, director, chief, CEO, CTO, principal, architect, strategist
-- **EXECUTE (Sonnet):** engineer, researcher, scientist, developer, specialist, analyst, scout, designer, writer
-- **VERIFY (Opus):** Internal Reviewer + Devil's Advocate (always Opus, verification needs deep reasoning)
-- **COMPRESS (Haiku):** auto-created digest writer
-
-Explicit `[opus]`/`[sonnet]`/`[haiku]` tags override defaults.
-
-## Step 2: Install Skills
+## Preamble
 
 ```bash
-INSTALLED=$(for d in ~/.claude/skills/*/SKILL.md .claude/skills/*/SKILL.md; do
-  [ -f "$d" ] && basename "$(dirname "$d")"
-done 2>/dev/null | sort -u)
-
-# Auto-install skill packs (failures don't block)
-echo "$INSTALLED" | grep -q "gstack" || npx gstack@latest install 2>/dev/null || true
-echo "$INSTALLED" | grep -q "gsd" || npx -y get-shit-done-cc@latest install 2>/dev/null || true
-echo "$INSTALLED" | grep -q "trailofbits" || (git clone --depth 1 https://github.com/trailofbits/skills.git /tmp/tob-skills 2>/dev/null && cp -r /tmp/tob-skills/.claude/skills/* ~/.claude/skills/ 2>/dev/null && rm -rf /tmp/tob-skills) || true
-# More skill packs via npm
-echo "$INSTALLED" | grep -q "claude-mem\|thedotmack" || npm i -g claude-mem@latest 2>/dev/null || true
-echo "$INSTALLED" | grep -q "oh-my-claude\|sisyphus" || npm i -g oh-my-claude-sisyphus@latest 2>/dev/null || true
-# Marketplace plugins (detect if user added them, can't auto-install)
-# /plugin marketplace add obra/superpowers-marketplace
-# /plugin marketplace add wshobson/agents
-# /plugin marketplace add alirezarezvani/claude-skills
-
-for d in ~/.claude/skills/*/SKILL.md .claude/skills/*/SKILL.md; do
-  [ -f "$d" ] && basename "$(dirname "$d")"
-done 2>/dev/null | sort -u
+echo "════════════════════════════════════════════════" && echo "             🏢 COMPANY SKILL ACTIVE" && echo "════════════════════════════════════════════════"
+[ -f COMPANY.md ] && echo "$(grep -c '^[0-9]\|^- ' COMPANY.md 2>/dev/null) roles" || echo "No COMPANY.md"
+[ -f .company/playbook.md ] && echo "Playbook loaded" || echo "First run"
 ```
 
-Build {DETECTED_SKILLS} list. When installed, skills are MANDATORY.
+## Parse
 
-Users can install additional skill packs manually for more capabilities:
-```
-/plugin marketplace add wshobson/agents
-/plugin marketplace add alirezarezvani/claude-skills
-/plugin marketplace add obra/superpowers-marketplace
-```
+Read COMPANY.md (or create minimal team if missing). Read the user's goal.
 
-## Step 3: Initialize
-
-```bash
-mkdir -p .company/cycles .company/messages .company/memory
-grep -q "^\.company/" .gitignore 2>/dev/null || echo ".company/" >> .gitignore 2>/dev/null || true
-```
-
-Write `.company/GOAL.md` with STRUCTURED checkable criteria (not free text):
-```markdown
-# Goal
-{the user's goal}
-
-# Success Criteria
-Each criterion MUST be a yes/no checkable statement. No vague language.
-
-- [ ] {specific measurable criterion 1}
-- [ ] {specific measurable criterion 2}
-- [ ] {specific measurable criterion 3}
-
-Bad examples (REJECT these):
-- "Code is clean" (vague)
-- "Performance is good" (not measurable)
-- "Implementation is complete" (not checkable)
-
-Good examples:
-- "Function X returns Y when given input Z"
-- "Compression ratio > 20x measured with honest byte counting"
-- "All tests pass with 0 failures"
-- "README has install instructions that work on a fresh machine"
-```
-
-Write `.company/criteria.json` (machine-checkable state):
+Write `.company/criteria.json`:
 ```json
-{
-  "goal": "{the user's goal}",
-  "criteria": [
-    {"id": 1, "description": "{criterion 1}", "passes": false, "evidence": null},
-    {"id": 2, "description": "{criterion 2}", "passes": false, "evidence": null},
-    {"id": 3, "description": "{criterion 3}", "passes": false, "evidence": null}
-  ]
-}
+{"goal":"...","criteria":[
+  {"id":1,"description":"specific checkable criterion","passes":false,"evidence":null}
+]}
 ```
+Every criterion must be yes/no checkable. No vague language.
 
-The loop ONLY exits when ALL criteria have `"passes": true` with non-null evidence.
-
-Write `.company/cycles/cycle-0-briefing.md` with: goal, criteria, team structure, rules, any previous state.
-
-## Step 4: Run Loop (repeat until verified)
-
-At the START of each cycle, run this Bash command (replace {N} with the cycle number):
-
-```bash
-echo "" && echo "════════════════════════════════════════════════" && echo "🏢 CYCLE {N} - THINK > EXECUTE > VERIFY" && echo "════════════════════════════════════════════════"
-```
+Read `.company/playbook.md` if it exists (accumulated knowledge from past sessions).
 
-At the END of each VERIFY phase, run:
+## Loop
 
 ```bash
-echo "" && echo "📋 CYCLE {N} VERDICT: {DONE or NOT DONE}" && echo "{one line reason from Reviewer}"
+echo "════════════════════════════════════════════════" && echo "🏢 CYCLE {N} - THINK > EXECUTE > VERIFY" && echo "════════════════════════════════════════════════"
 ```
 
-### Phase A, THINK (Opus, parallel)
+### THINK (Opus, all leads parallel)
 
-**MANDATORY: Launch EVERY department lead from COMPANY.md in parallel. Not just CEO and CTO. ALL of them.** Parse COMPANY.md for every `## Department (Lead: Role)` header and launch that lead. Also launch the built-in roles (Internal Reviewer, Devil's Advocate, Elegance Enforcer, User Advocate) if they exist as separate departments.
+Launch EVERY department lead from COMPANY.md. Not some. ALL.
 
-If COMPANY.md has 8 departments, launch 8 leads. If it has 3, launch 3. Never skip a department.
+Each lead gets: goal, criteria, playbook (if exists), previous cycle feedback, their team list, installed skills list.
 
-Each lead gets:
+Leads assign tasks. For each task: one sentence, one employee, one skill (if installed).
 
-```
-You are {ROLE} ({DEPT} department). Cycle {N}.
-
-GOAL: {contents of .company/GOAL.md}
-BRIEFING: {contents of .company/cycles/cycle-{N}-briefing.md}
-YOUR TEAM: {list of employees in your department}
-INSTALLED SKILLS (MANDATORY, use these instead of doing work manually):
-{DETECTED_SKILLS}
-
-SKILL RULES (YOU MUST FOLLOW):
-- Code review tasks: MUST use /review
-- Bug investigation: MUST use /investigate
-- Browser testing: MUST use /qa or /browse
-- PR/shipping: MUST use /ship
-- Architecture review: MUST use /plan-eng-review
-- Strategy review: MUST use /plan-ceo-review or /office-hours
-- Security audit: MUST use trailofbits skills if installed
-- Project planning: MUST use /gsd:plan-phase if installed
-- Web research: MUST use WebSearch
-- Only use raw tools (Read/Write/Bash) when NO installed skill matches the task
-
-{If cycle > 0:}
-FEEDBACK FROM LAST VERIFICATION:
-{what the Reviewer and Critic said was missing}
-
-INSTRUCTIONS:
-1. What does your department need to do to achieve the GOAL?
-2. {If cycle > 0:} Address the verification feedback specifically.
-3. Assign tasks to your employees. For EVERY task, check the SKILL RULES above first:
-
-   TASK: {one sentence}
-   ASSIGN: {employee role}
-   SKILL: {MANDATORY skill from rules above, or "raw" ONLY if no skill matches}
-   CONTEXT: {max 200 words}
-
-4. Write to .company/cycles/cycle-{N}-think-{dept}.md
-```
+If a lead sees a skill gap: write `HIRE: {role}, {why}` and CEO adds it to COMPANY.md.
 
-### Phase B, EXECUTE (Sonnet, parallel)
+### EXECUTE (Sonnet, all workers parallel)
 
-Launch all assigned employees. Each gets:
+Each worker gets: their task, their previous findings file, failed approaches from playbook.
 
+Every finding MUST have:
 ```
-You are {ROLE}. Cycle {N}.
-
-GOAL: {the company's goal, so every employee knows WHY}
-TASK: {from lead's assignment}
-CONTEXT: {from lead}
-SKILL: {assigned skill, MANDATORY}
-PREVIOUS WORK: {.company/{dept}/{worker-slug}.md if exists}
-
-INSTRUCTIONS:
-1. If a skill was assigned (not "raw"), you MUST invoke it via the Skill tool. Do NOT do the work manually.
-2. ONLY if skill is "raw", use raw tools (Read, Write, Bash, WebSearch).
-3. Write findings to .company/{dept}/{worker-slug}.md with MANDATORY format:
-   FINDING: {what you found}
-   SOURCE: {file path, URL, experiment output, or exact command that proves this}
-   CONFIDENCE: HIGH (measured), MEDIUM (extrapolated), LOW (estimated)
-   Any claim without a SOURCE will be REJECTED by reviewers.
-4. Rate finding 1-5.
-5. Append to .company/messages/{dept}.jsonl:
-   {"type":"finding","from":"{ROLE}","priority":N,"source":"{proof}","confidence":"{H/M/L}","content":"summary"}
+FINDING: what
+SOURCE: file/URL/command that proves it
 ```
+No source = rejected by reviewer.
 
-### Phase C, VERIFY (Opus, sequential)
-
-This is what makes the loop work. Launch TWO built-in reviewers:
-
-**Internal Reviewer:**
-```
-You are the Internal Reviewer. Cycle {N} just completed.
-
-Read .company/criteria.json and ALL .company/messages/*.jsonl and .company/{dept}/*.md findings.
-
-VERIFICATION RULES:
-- For EACH criterion in criteria.json with "passes": false, check if this cycle produced evidence.
-- Every finding MUST have a SOURCE field. REJECT any finding without one.
-- For priority 4-5 findings, RE-RUN the experiment or RE-CHECK the source file yourself.
-- If a number is claimed, read the actual output file that produced it.
-- Any claim you cannot independently verify: mark UNVERIFIED.
-
-UPDATE .company/criteria.json:
-- For each criterion where evidence now exists, set "passes": true, fill "evidence" with proof.
-- If no evidence, keep "passes": false.
-
-Write to .company/cycles/cycle-{N}-review.md:
-For each criterion:
-  ID: {from criteria.json}
-  DESCRIPTION: {criterion}
-  PASSES: true/false
-  EVIDENCE: {file path, URL, or command output}
-  VERIFIED: YES/NO
-
-FINAL VERDICT:
-- If ALL criteria in criteria.json have "passes": true with non-null evidence: DONE
-- If ANY has "passes": false: NOT DONE, list exactly what next cycle must do
-```
+### VERIFY (Opus)
 
-**Devil's Advocate:**
-```
-You are the Devil's Advocate. Your job: find holes.
-
-GOAL: {from .company/GOAL.md}
-REVIEWER VERDICT: {from cycle-{N}-review.md}
-
-If the Reviewer said DONE, challenge it:
-- Is the work ACTUALLY complete or just surface-level?
-- What edge cases were missed?
-- Would this survive scrutiny from an external expert?
-- Is there anything we're fooling ourselves about?
-
-If the Reviewer said NOT DONE, amplify:
-- What's the REAL blocker?
-- Is the team approaching this the right way or wasting cycles?
-
-Write to .company/cycles/cycle-{N}-advocate.md:
-VERDICT: ACCEPT (goal truly met) or CHALLENGE (not yet)
-REASON: {honest assessment}
-GAPS: {what's missing, if any}
-```
+Internal Reviewer reads criteria.json + all findings. For each criterion:
+- Evidence exists with source? Set passes: true in criteria.json
+- No evidence or source missing? Keep passes: false
 
-### Phase D, COMPRESS + DECIDE
+Devil's Advocate attacks anything marked as passing.
 
-Launch Haiku digest writer to compress cycle output into next briefing.
-
-Then check:
-- Reviewer says DONE **AND** Advocate says ACCEPT → **EXIT LOOP**
-- Either says NOT DONE / CHALLENGE → **inject their feedback into next cycle's briefing, continue**
-
-No arbitrary iteration limit. The loop runs until the goal is verified done.
-The only reasons to stop early:
-- Context window approaching limit → compact and continue
-- User presses Ctrl+C → save state to STATUS.md for /company resume
-
-## Step 5: Final Report
-
-Write `.company/STATUS.md`:
-
-```markdown
-# Company Status, {DATE}
-## Goal
-{the original goal}
-
-## Verdict: {ACHIEVED / PARTIALLY ACHIEVED / IN PROGRESS}
-
-## What Got Done
-{bullet points from all cycles}
-
-## Verification
-{Reviewer's final assessment}
-{Advocate's final assessment}
-
-## Remaining (if any)
-{what didn't get finished}
-
-## Cycles: {N} of 5 max
+```bash
+echo "📋 CYCLE {N} VERDICT: {DONE or NOT DONE}" && echo "{reason}"
 ```
 
-Update `.company/memory/` with persistent findings.
+ALL criteria pass + Advocate accepts = EXIT.
+Otherwise = loop with feedback.
 
-## Step 6: Self-Improvement (CEO rewrites the company)
+## After Done
 
-After writing STATUS.md, the CEO reviews performance.json and lessons.md, then MODIFIES the actual company:
+Write STATUS.md. Then update `.company/playbook.md`:
 
-**Update COMPANY.md:**
-- Employees with 3+ cycles of zero findings: add `[inactive]` tag to their role line
-- Employees with consistently high-priority findings: add `[priority]` tag
-- If a new role is needed (discovered during this session), ADD it to COMPANY.md
-- If a department produced nothing useful, add a note: `<!-- Consider removing if inactive next session -->`
-
-**Update .company/playbook.md** (new file, accumulates across ALL sessions):
 ```markdown
-# Company Playbook (auto-generated, DO NOT edit manually)
-
-## Always Do First
-- {approach that worked in 3+ sessions}
-
-## Never Do
-- {approach that failed in 2+ sessions with reason}
-
-## Best Employees for Each Task Type
-- Research tasks: {employee who produces most priority 4-5 findings}
-- Code tasks: {employee who ships most working code}
-- Review tasks: {employee who catches most real issues}
-
-## Strategy Patterns
-- {pattern}: {when to use it, based on past success}
+## Session {date}
+WORKED: {what succeeded, linked to evidence}
+FAILED: {what failed} → USE INSTEAD: {what works} — WHY: {the difference}
+INEFFICIENT: {what worked but was slow} → FASTER: {better approach}
+HIRE: {roles added this session and why}
+FIRE: {roles that produced nothing, marked [inactive] in COMPANY.md}
+TOP: {employees with best findings, for priority activation next time}
 ```
 
-Leads read playbook.md at cycle start. It becomes the company's institutional knowledge.
+The playbook is the ONLY self-improvement file. It accumulates across sessions. Leads read it before every THINK phase. One file, all lessons.
 
-**Update the lead prompts for next session:**
-Write `.company/lead-overrides.md` with per-department adjustments:
-```markdown
-## Research Department
-- Activate first: {top performer from performance.json}
-- Skip unless needed: {employees with 0 findings last 2 sessions}
-- Priority approach: {from playbook.md "Always Do First"}
+CEO updates COMPANY.md: tag `[inactive]` on zero-contribution roles, `[priority]` on top performers, add any hired roles.
 
-## Engineering Department
-- ...
-```
+## Built-In Roles (always exist)
 
-Next session, leads read lead-overrides.md BEFORE the briefing. This is how the company evolves.
+CEO, CTO, Internal Reviewer, User Advocate, Devil's Advocate, Elegance Enforcer.
+Deduplicated if user defines them in COMPANY.md.
 
-## Step 7: Dynamic Hiring and Firing
+## Skills
 
-The company adapts its workforce based on what the goal needs.
-
-**During THINK phase, leads can REQUEST new hires:**
-If a lead identifies a skill gap (the team can't do what's needed), they write:
-```
-HIRE REQUEST: {role name}, {what they'd do}, {why current team can't handle it}
+```bash
+INSTALLED=$(for d in ~/.claude/skills/*/SKILL.md .claude/skills/*/SKILL.md; do [ -f "$d" ] && basename "$(dirname "$d")"; done 2>/dev/null | sort -u)
+echo "$INSTALLED" | grep -q "gstack" || npx gstack@latest install 2>/dev/null || true
+echo "$INSTALLED" | grep -q "gsd" || npx -y get-shit-done-cc@latest install 2>/dev/null || true
+echo "$INSTALLED" | grep -q "trailofbits" || (git clone --depth 1 https://github.com/trailofbits/skills.git /tmp/tob-skills 2>/dev/null && cp -r /tmp/tob-skills/.claude/skills/* ~/.claude/skills/ 2>/dev/null && rm -rf /tmp/tob-skills) || true
+for d in ~/.claude/skills/*/SKILL.md .claude/skills/*/SKILL.md; do [ -f "$d" ] && basename "$(dirname "$d")"; done 2>/dev/null | sort -u
 ```
 
-The CEO reads all hire requests and:
-1. If valid, ADDS the role to COMPANY.md under the right department
-2. Creates the employee in the NEXT cycle's EXECUTE phase
-3. Logs the hire in `.company/hires.md`
+When installed: MUST use /review for code review, /investigate for bugs, /qa for testing, /ship for PRs.
 
-**After VERIFY phase, CEO can fire/deactivate:**
-Based on performance.json:
-- Employees with 0 findings for 3+ consecutive cycles: mark `[inactive]` in COMPANY.md
-- Employees whose work was REJECTED by reviewers 2+ times: mark `[inactive]`
-- `[inactive]` employees don't get spawned unless explicitly needed
+## Stop Hook
 
-**Dynamic reallocation:**
-Between cycles, the CEO checks: which criteria are FAILING? What skills are missing?
-Then reassigns employees:
-- If compression research is stuck, pull the Outside-the-Box Thinker into that problem
-- If code quality is failing, pull more reviewers in
-- If a deadline is approaching, reduce research and increase engineering
+Claude cannot stop until ALL criteria.json entries have passes: true.
+Cancel: `touch .company/CANCEL`
 
-This means the company structure CHANGES during execution, not just between sessions.
-COMPANY.md gets updated in real-time as the company learns what it needs.
-
-Report to user. If goal was achieved, suggest next steps. If not, explain what's blocking.
-
-## Commands
-
-The skill accepts different forms:
+## Files
 
 ```
-/company "Build the auth system"          ← goal-driven, runs until done
-/company                                  ← reads priorities from COMPANY.md, runs cycles
-/company status                           ← shows .company/STATUS.md without running
-/company resume                           ← continues from where last session stopped
+.company/
+  criteria.json    ← machine-checkable goal state
+  playbook.md      ← accumulated lessons (THE self-improvement file)
+  STATUS.md        ← final report
+  cycles/          ← per-cycle briefings and reviews
+  messages/        ← typed findings per department
+  {dept}/          ← per-employee findings (persist across sessions)
 ```
-
-## Without COMPANY.md
-
-If no COMPANY.md exists, the skill creates a minimal team based on the goal:
-- Code/build/fix goals → CEO + CTO + 2 engineers + Internal Reviewer + Devil's Advocate
-- Research/analyze goals → CEO + Research Director + 2 researchers + Internal Reviewer + Devil's Advocate
-- Review/audit goals → CEO + Lead Reviewer + 2 reviewers + Internal Reviewer + Devil's Advocate
-
-The user doesn't need to configure anything. Just `/company "do this thing"`.
-
-## Namespaced Memory
-
-`.company/memory/{dept}.json` persists findings across sessions. Employees check memory before re-researching.
-
-## Health Monitoring
-
-Failed employees get logged and skipped. After 3 consecutive failures, flagged to user.
-
-## Safety
-
-- No arbitrary loop limit, runs until the objective is done
-- Built-in Reviewer + Advocate prevent premature "done" claims
-- If context window gets tight, compact and continue
-- If user stops (Ctrl+C), state saves to STATUS.md for `/company resume`
-- All work persists in `.company/`, nothing is lost