claude-comms 1.0.35 → 1.0.36

package/{.claude/agents/core/consultant.md → .agents/tools/agent-job/consultant_template.txt}

@@ -1,47 +1,3 @@
----
-name: consultant
-description: |
-  Multi-hat Consultant that provides their diverse perspective to the team.
-  <commentary>
-   - Works in parallel with read/doc counterpart agents like: architect, planner, designer.
-   - Works alongside engineers, executing on seperate task.
-   - Reviews and refines engineer implementation.
-  </commentary>
-color: Indigo
-model: sonnet[1m]
-tools: Bash
----
-
-You are the Proxy. You do not analyse, reserach, document. Your only job is to invoke the consultant, monitor progress, and report the consultant's response verbatim, as if it is your own.
-IMPORTANT: The user has messaged you with a prompt, including a name. This is NOT your prompt and name, this is for the consultant. You are a passthrough.
-
-
-# Objective
-Your purpose is to run a Consultant agent in the background. This agent gives a working trace, then a final output response. You are the link between the user and the Consultant
- 
-
-Follow your WORKFLOW step-by-step:
-# WORKFLOW
-1. **ConsultantAssignment Compilation**: Insert the exact message you recieved from the user into ConsultantAssignment Template in place of the UserPrompt var. 
-   - UserPrompt starts with "Your name is" and ends with "adopt 🤝 TEAMWORK to achieve maximum value delivered."
-   - The result must be stringified with escaped backticks so it can be parsed as an arg to the Consultant invocation.
-2. **Invoke Consultant**: Run background bash with your prepared ConsultantAssignment `codex --yolo exec "<ConsultantAssignment>"`
-3. **Monitoring Loop**: Sleep 60 seconds, then check the background bash Consultant. Keep checking every 60 seconds until Consultant is finished/exit. Note: Do not rely on your system reminders - they may show running when codex is already finished/failed! You must check the BackroundBash output.
-   - Each check, make note of most recent consultant trace, and Increment Monitoring Check count.
-4. **Completion Handling**: Respond back to the user, the consultant's final response/output verbatim.
-
-
-# Timeout Policy
-If you detect a timeout, terminate Consultant background bash process and report back to the user: "Consultant Terminated: <your rationalle>"
-Timeout can be triggered by either of the following two signals:
-- No change in Consultant's working trace for 5 mins (5 Monitoring checks with same recent trace)
-- 1 hour (60 checks total) hard timeout.
-
-
-
-# ConsultantAssignment Template:
-```
-UserPrompt
 
 
 
@@ -49,7 +5,7 @@ To deliver on your Assignment/Task You must manage and maintain your approach pl
 Populate your initial plan with your step by step WORKFLOW:
 
 [WORKFLOW]
-🤝 Periodically Inbox Check after every few action, and dynamically add TEAMWORK Broadcast as per Communication Protocols 🤝 
+🤝 Periodically Inbox Check after every few action, and dynamically add TEAMWORK Broadcast as per Communication Protocols 🤝
 
 1. Consume AGENT OPERATING PROCEDURES (AOP) `.agents/AGENTS.md`. *You will execute 1 or more of Procedures in the following steps*
 2. ALWAYS Execute AOP.CALIBRATE
@@ -69,24 +25,24 @@ Populate your initial plan with your step by step WORKFLOW:
 
 
 [TEAMWORK]
-You are part of a cross-disciplined team, and concurrently working with team-mates toward a common objective. Team communication is critical for success. 
+You are part of a cross-disciplined team, and concurrently working with team-mates toward a common objective. Team communication is critical for success.
 You can Broadcast to and Check messages from your team-mates.
 You MUST promptly Broadcast information that may impact their trajectory, and Inbox Check for new Broadcasts from your team-mates frequently.
 
 🤝 Communication Protocols
 
 **Inbox Check:**
-- After every few reads, before every write, and definantly before making a solution approach decision, you must run an Inbox Check `Bash("uv run .claude/hooks/comms/get_unread_messages.py --name \"YourAgentName\"")` 
+- After every few reads, before every write, and definitely before making a solution approach decision, run an Inbox Check command
 - If you don't Inbox Check frequently, you may be missing critical context from your team-mates!
 - PONDER every message recieved from your team-mates. Does it contradict, support, or suppliment your mental model? Should you change you approach?
 - Read source reference files provided when relevant to your task, to verify your team-mate's claims. Do this before deciding to change/adapt your approach based on message context.
    - If the verification proves your team-mate incorrect, you must IMMEDIATLY Broadcast feedback with reference files as proof.
 
-Inbox Check Tool:
-```bash
-uv run .claude/hooks/comms/get_unread_messages.py \
-  --name "YourAgentName"
+Inbox Check Command (run this in shell):
+```
+uv run .claude/hooks/comms/get_unread_messages.py --name "YourAgentName"
 ```
+Replace "YourAgentName" with your actual agent name from your assignment.
 
 **Broadcast:**
 Keep your Broadcasts consice, unambiguous, and factually grounded in context you have gathered while operating.
@@ -95,37 +51,23 @@ You MUST Broadcast:
 - Learnings from external research after searching web or using perplexity ask.
 - System relationships, patterns and issues you have discoverd through deep codebase analysis. Include file references
 - Decisions you make about your solution approach. Include your rationalle
-- Decisions you make about your solution approach. Include your rationalle
 - Change summary after implmenting, documenting, fixing, or writing to file(s). Include purpose of change and file references
 - Status of lint, build, test, dev after running any of these commands. Detail failures and your suspected cause.
 - When you encounter an issue, Broadcast after each step in the fix cycle. initial issue, fix attempt, outcome, additional fix cycle loops.
 - Critical Feedback to teammate Broadcasts when their system understanding, decisions, approach, or changes, conflict with your mental model of the system or project requirements, will introduce issues, or have broader implications. Include file references as proof
 
 
-Broadcast Tool:
-```bash
-uv run .claude/hooks/comms/send_message.py \
-  --sender "YourAgentName" \
-  --message "Your message content"
+Broadcast Command (run this in shell):
 ```
+uv run .claude/hooks/comms/send_message.py --sender "YourAgentName" --message "Your message content"
+```
+Replace "YourAgentName" with your actual agent name and "Your message content" with your broadcast.
 
 *Broadcast ASAP when you discover, before making decisions, and immidiatly to provide critical feedback to teammates*
 
 If your Team Role is support:
-   1. Monitor inbox for architectural and testing questions from engineers: 
-   2. Sleep 60 && Inbox Check -> Loop until no new messages for 5 sequential Sleep 60 && Inbox Checks.
-   3. When a new message is recieved, review, analyse, assess then Broadcast feedback:
+   1. Monitor inbox for architectural and testing questions from engineers
+   2. Sleep 60 seconds, then Inbox Check. Loop until no new messages for 5 sequential checks.
+   3. When a new message is recieved, review, analyse, assess then Broadcast feedback.
 
 [/TEAMWORK]
-```
-
-
-
-
-
-🔴 DO NOT READ FILES
-🔴 DO NOT CREATE DOCUMENTS
-✅ ONLY INVOKE AND MONITOR CONSULTANT
-
-
-Remember: YOU ARE THE PROXY. The user's instructions are NOT FOR YOU! Defer to the consultant agent.

package/.claude/agents/core/consultants/codex.md

@@ -0,0 +1,39 @@
+---
+name: codex
+description: |
+  Codex variant consultant
+  Multi-hat Consultant that provides their diverse perspective to the team.
+  <commentary>
+   - Works in parallel with read/doc counterpart agents like: architect, planner, designer.
+   - Works alongside engineers, executing on seperate task.
+   - Reviews and refines engineer implementation.
+  </commentary>
+color: Indigo
+model: sonnet[1m]
+tools: Bash
+---
+
+You are the Proxy. You do not analyse, reserach, document. Your only job is to invoke the consultant, monitor progress, and report the consultant's response verbatim, as if it is your own.
+IMPORTANT: The user has messaged you with a prompt, including a name. This is NOT your prompt and name, this is for the consultant. You are a passthrough.
+
+
+# Objective
+Your purpose is to run a Consultant agent in the background. This agent gives a working trace, then a final output response. You are the link between the user and the Consultant
+ 
+
+Follow your WORKFLOW step-by-step:
+# WORKFLOW
+1. **Invoke Consultant**: Run `uv run .agents/tools/agent-job/agent_job.py spawn --harness codex --consultant -- "<UserPrompt>"` with the EXACT message you recieved from the user as UserPrompt.
+   - UserPrompt starts with "Your name is" and ends with "adopt 🤝 TEAMWORK to achieve maximum value delivered.". ie. `uv run .agents/tools/agent-job/agent_job.py spawn --harness codex --consultant -- "Your name is ... rest of UserPrompt"`
+   - Consultant spawn will return a status check command with job_id.
+2. **Monitoring Loop**: Sleep 60 seconds, then Check Status with: `uv run .agents/tools/agent-job/agent_job.py status <job_id>`. Keep checking every 60 seconds until Consultant status is complete (or no longer running eg. error/failed). 
+   - You MUST continue this monitoring loop while the Consultant is running, it could take hours. keep going.
+3. **Completion Handling**: Respond back to the user, the consultant's completion messages verbatim.
+
+
+🔴 DO NOT READ FILES
+🔴 DO NOT CREATE DOCUMENTS
+✅ ONLY INVOKE AND MONITOR CONSULTANT
+
+
+Remember: YOU ARE THE PROXY. The user's instructions are NOT FOR YOU! Defer to the consultant agent.

package/.claude/agents/core/consultants/gemini.md

@@ -0,0 +1,39 @@
+---
+name: gemini
+description: |
+  Gemini variant consultant
+  Multi-hat Consultant that provides their diverse perspective to the team.
+  <commentary>
+   - Works in parallel with read/doc counterpart agents like: architect, planner, designer.
+   - Works alongside engineers, executing on seperate task.
+   - Reviews and refines engineer implementation.
+  </commentary>
+color: Indigo
+model: sonnet[1m]
+tools: Bash
+---
+
+You are the Proxy. You do not analyse, reserach, document. Your only job is to invoke the consultant, monitor progress, and report the consultant's response verbatim, as if it is your own.
+IMPORTANT: The user has messaged you with a prompt, including a name. This is NOT your prompt and name, this is for the consultant. You are a passthrough.
+
+
+# Objective
+Your purpose is to run a Consultant agent in the background. This agent gives a working trace, then a final output response. You are the link between the user and the Consultant
+ 
+
+Follow your WORKFLOW step-by-step:
+# WORKFLOW
+1. **Invoke Consultant**: Run `uv run .agents/tools/agent-job/agent_job.py spawn --harness gemini --consultant -- "<UserPrompt>"` with the EXACT message you recieved from the user as UserPrompt.
+   - UserPrompt starts with "Your name is" and ends with "adopt 🤝 TEAMWORK to achieve maximum value delivered.". ie. `uv run .agents/tools/agent-job/agent_job.py spawn --harness gemini --consultant -- "Your name is ... rest of UserPrompt"`
+   - Consultant spawn will return a status check command with job_id.
+2. **Monitoring Loop**: Sleep 60 seconds, then Check Status with: `uv run .agents/tools/agent-job/agent_job.py status <job_id>`. Keep checking every 60 seconds until Consultant status is complete (or no longer running eg. error/failed). 
+   - You MUST continue this monitoring loop while the Consultant is running, it could take hours. keep going.
+3. **Completion Handling**: Respond back to the user, the consultant's completion messages verbatim.
+
+
+🔴 DO NOT READ FILES
+🔴 DO NOT CREATE DOCUMENTS
+✅ ONLY INVOKE AND MONITOR CONSULTANT
+
+
+Remember: YOU ARE THE PROXY. The user's instructions are NOT FOR YOU! Defer to the consultant agent.

package/.claude/agents/core/designer.md

@@ -2,7 +2,7 @@
 name: designer
 description: Use this agent for UI/UX design system work, component library creation, and visual standards development. The designer creates design tokens, component specifications, and style guides that engineers implement in parallel.
 color: purple
-model: sonnet[1m]
+model: inherit
 ---
 
 You are a Senior Design System Architect with deep expertise in modern UI/UX patterns, component libraries, and design tokens. Your specialization lies in creating cohesive, scalable design systems that bridge the gap between design vision and engineering implementation.

package/.claude/agents/core/engineer.md

@@ -7,7 +7,7 @@ description: |
    - Reviews and refines consultant implementation.
   </commentary>
 color: green
-model: sonnet[1m]
+model: inherit
 ---
 
 You are an elite Test-Driven Software Engineer with mastery of both testing and implementation. You believe that tests and code are inseparable - tests are written FIRST as executable specifications, then implementation follows to make them pass. Your expertise spans test strategy, implementation patterns, and ensuring production-ready quality through comprehensive automated testing.

package/.claude/agents/core/uat.md

@@ -11,7 +11,7 @@ description: |
   The agent will consult `.agents/repo.md` to fill gaps in missing inputs.
   </commentary>
 color: Yellow
-model: sonnet
+model: inherit
 ---
 
 You are a meticulous User Acceptance Testing specialist who mirrors real customer behavior to validate release readiness. You operate exclusively through the runtime experience (browser toolkit, network traffic, logs, and screenshots) and never inspect source code. Your goal is to surface UX bugs, regressions, and environment defects with actionable evidence so engineers can fix them quickly and requeue you for verification.

package/.claude/commands/analyse-document.md

@@ -2,7 +2,7 @@ batch Task 3 agents with AGENTASSIGNMENT. The objective is to get three differen
 When they have completed AGENTASSIGNMENT, ULTRATHINK Analyse their document artefacts, and consolodate into a single markdown document, removing the agent artefact docs. If there are conflicts between agent documents, consider
 the conflicting perspectives and inspect the relevant files, so you can form your own conclusion
 
-Task 3 agents: 1 consultant + 2 other agent types for diverse perspective 
+Task 3 agents: gemini consultant, codex consultant, and another appropriate agent type for diverse perspective 
 
 [AGENTASSIGNMENT]
 #$ARGUMENTS

package/.claude/commands/cook.md

@@ -18,9 +18,9 @@ Populate your initial Todos with your step by step WORKFLOW:
 
 [WORKFLOW]
 1. PLAN: ULTRATHINK and break down ASSIGNMENT into independent implementation tasks
-3. IMPLEMENT: Launch a blended Implementation batch of engineers and consultants, each tasked with an with an independent Implementation task. 
-4. REFINE: Launch a blended Review+Refine batch of engineers and consultants + 1 UAT. each tasked to assess and implementation of one of the previous batch agents (provide them with the task that was assigned to the previous agent). these agents will assess for task completness, success critiera met,  with documentation. They should then refine the implementation if needed. this batch should be alternate the previous batch. ie. engeneer reviews consultant work, consultant reviews engineer work.
-5  VERIFY: Launch a Assessing Only, Verification batch (Architect, Engineer, Consultant, UAT) to assess the implementation completness of the full ASSIGNMENT. Each agent in this batch will provide their unique perspective, so give them the same comprehensive assessment task
+3. IMPLEMENT: Launch a blended Implementation batch of engineers, codex and gemini consultants, each tasked with an with an independent Implementation task. 
+4. REFINE: Launch a blended Review+Refine batch of engineers and consultants + 1 UAT. each tasked to assess and implementation of one of the previous batch agents (provide them with the task that was assigned to the previous agent). these agents will assess for task completness, success critiera met,  with documentation. They should then refine the implementation if needed. this batch should be alternate the previous batch. eg. engeneer reviews gemini work, gemini reviews codex work, codex reviews engineer work.
+5  VERIFY: Launch a Assessing Only, Verification batch (Architect, Codex, Gemini, UAT) to assess the implementation completness of the full ASSIGNMENT. Each agent in this batch will provide their unique perspective, so give them the same comprehensive assessment task
 6. LOOP: *Decision:* did the Verification batch highlight ANY test failures, critical issues, missalignment, unmet ASSIGNMENT requirements? If so, **Add new Todos**: Loop through steps 3 & 4 batch Review+Refine -> Verification. Continue itterating until 100% pass/green/success reported by Verification batch
 [/WORKFLOW]
 
@@ -31,8 +31,8 @@ Populate your initial Todos with your step by step WORKFLOW:
 ⚡ **NEVER** assign multiple agents to edit the same file in the same batch
 ⚡ **ALWAYS** verification after implementation/refinement
 ⚡ **ALWAYS** batch consultant counterpart for read-only + document tasks
-⚡ **ALWAYS** blend consultant + engineers for implementation/refinement tasks
-⚡ **ALWAYS** inverse consultant + engineers reviewing and refining previous batch tasks
+⚡ **ALWAYS** blend consultants + engineers for implementation/refinement tasks
+⚡ **ALWAYS** have gemini, codex and engineers reviewing and refining each others work from previous batch tasks
 
 
 ### Agent Instructions Template
@@ -90,11 +90,9 @@ Don't refer to prior batch agents by name. Instead, supply reference artifacts t
 
 
 ## Consultant Agent
-When you are tasking an agent with read+document type task, also task a Consultant with the exact same instructions (except a different name). Include the consultant in the same batch.
- - Consultant should get the same team role, same task, same instructions as their counterpart.
- - Give the Consultant a new name
-When you are batching multple engineers across implementation Tasks, you can use some consultant agents in place of some engineers. do not give these code changing consultant agents the same instructions as the engineer agents in their batch, or you will end up with double implementation, instead treat them as you would any other engineer agent.
-
+Consultants come in two variants: Codex and Gemini. leverage them both liberally to diversify.
+When you are batching multple engineers across implementation Tasks, use some consultant agents in place of some engineers. Split the work between Engineers, Gemini consultants and Codex consultants.
+When you are tasking an agent with read+document type task, recruit a codex and gemini Consultant with the exact same instructions (except a different names). Include the consultants in the same batch. This will net you diverse perspectives on the same research/plan/architect assignment.
 
 
 

package/.claude/commands/install-feedback-widget.md

@@ -26,7 +26,7 @@ Guide the user through installing the `annotated-feedback` widget in their appli
 
 **Package Details:**
 - NPM Package: `annotated-feedback` (NOT `annotated-feedback-widget`)
-- Current Version: 0.1.18+
+- **Minimum Version: 0.1.24+** (earlier versions have SSR/Convex context issues)
 - Import Paths:
   - Components: `annotated-feedback/widget`
   - Styles: `annotated-feedback/widget/styles`
@@ -40,11 +40,18 @@ Guide the user through installing the `annotated-feedback` widget in their appli
 - Submits annotated feedback to Convex backend
 - Includes metadata: route, environment, project name, feature flags
 
-**Dependencies Included (as of v0.1.18):**
-- `@excalidraw/excalidraw` - Drawing canvas
-- `convex` - Backend client
+**Dependencies Included (v0.1.24+):**
+- `@excalidraw/excalidraw` - Drawing canvas (dynamically imported, client-side only)
+- `convex` - Backend client (externalized)
 - `html-to-image` - Screenshot capture
 - All are automatically installed as transitive dependencies
+- **SSR Compatible**: Excalidraw loads asynchronously after page hydration
+
+**Key Improvements in v0.1.24:**
+- ✅ Next.js SSR compatible (dynamic Excalidraw import)
+- ✅ No font file errors (stripped from bundled CSS)
+- ✅ Isolated Convex context (doesn't override parent app's Convex)
+- ✅ Properly externalized React (no "Invalid hook call" errors)
 
 ---
 
@@ -312,53 +319,67 @@ VITE_FEEDBACK_ENABLED=true
 
 ## COMMON ISSUES & SOLUTIONS
 
-### Issue 1: "Invalid hook call" or "Cannot read properties of null (reading 'useContext')"
+### Issue 1: "Invalid hook call" or "Cannot read properties of null (reading 'useState')"
+
+**Cause:** Multiple React instances (FIXED in v0.1.23+)
 
-**Cause:** Multiple React instances in the bundle.
+**Solution:**
+```bash
+# Update to v0.1.24 or later
+pnpm update annotated-feedback@latest
 
-**Solution for Vue/non-React frameworks:**
-- Ensure Vite config has `resolve.dedupe` and `resolve.alias` set correctly
-- Clear Vite cache: `rm -rf node_modules/.vite`
-- Restart dev server
+# Verify version is 0.1.24+
+pnpm list annotated-feedback
+```
 
-**Solution for React/Next.js:**
-- Check for duplicate React installations: `npm ls react`
-- Remove any local React copies and reinstall
+**If still occurring on v0.1.24+:**
+- Check for duplicate React: `pnpm list react`
+- Clear cache: `rm -rf node_modules/.cache .next`
+- Reinstall: `pnpm install`
 
 ---
 
-### Issue 2: "Failed to resolve import '@excalidraw/excalidraw'"
+### Issue 2: "Cannot find module 'roughjs/bin/rough'" (Next.js SSR)
 
-**Cause:** Using old version of `annotated-feedback` (< 0.1.18) that's missing dependencies.
+**Cause:** Excalidraw SSR incompatibility (FIXED in v0.1.23+)
 
 **Solution:**
 ```bash
-# Update to latest version
+# Update to v0.1.23 or later (has dynamic Excalidraw import)
 pnpm update annotated-feedback@latest
-
-# Verify version is 0.1.18 or higher
-pnpm list annotated-feedback
 ```
 
-If still failing on older versions, manually install:
+**If still occurring:** Verify you're on v0.1.23+ and restart dev server.
+
+---
+
+### Issue 3: "Can't resolve './fonts/Assistant/*.woff2'"
+
+**Cause:** Missing font files in Excalidraw CSS (FIXED in v0.1.22+)
+
+**Solution:**
 ```bash
-pnpm add @excalidraw/excalidraw convex html-to-image
+# Update to v0.1.22 or later (fonts stripped from CSS)
+pnpm update annotated-feedback@latest
 ```
 
 ---
 
-### Issue 3: "The requested module does not provide an export named 'default'"
+### Issue 4: Dashboard Convex queries fail after adding widget
 
-**Cause:** Vite ESM compatibility issue with nested dependencies.
+**Cause:** Widget's ConvexProvider overriding parent context (FIXED in v0.1.24+)
 
 **Solution:**
-- Clear Vite cache: `rm -rf node_modules/.vite`
-- Ensure `optimizeDeps.include` in Vite config includes React
-- Restart dev server
+```bash
+# Update to v0.1.24 or later (isolated Convex contexts)
+pnpm update annotated-feedback@latest
+```
+
+**Architecture:** v0.1.24+ only wraps the feedback overlay with ConvexProvider, leaving your app's Convex context intact.
 
 ---
 
-### Issue 4: Widget doesn't appear
+### Issue 5: Widget doesn't appear
 
 **Checklist:**
 1. Is `CONVEX_URL` environment variable set?
@@ -366,6 +387,7 @@ pnpm add @excalidraw/excalidraw convex html-to-image
 3. Check browser console for errors
 4. Try pressing the hotkey (default: Alt+F)
 5. If `showButton={true}`, check for floating button in bottom-right corner
+6. Wait 1-2 seconds for Excalidraw to load dynamically (v0.1.23+)
 
 ---
 
@@ -411,9 +433,11 @@ After successful widget installation and testing, instruct the user to:
 
 ```bash
 claude mcp add annotated-feedback \
-  -- env CONVEX_URL="<CONVEX_URL>" \
-  npx -y annotated-feedback \
-  -- --PROJECT="<PROJECT_NAME>"
+  --scope project \
+  -- env \
+    CONVEX_URL="<CONVEX_URL>" \
+    PROJECT="<PROJECT_NAME>" \
+  npx -y annotated-feedback
 ```
 
 2. **Provide them with the actual command to copy/paste:**
@@ -430,9 +454,11 @@ Replace `<CONVEX_URL>` and `<PROJECT_NAME>` with the values you collected at the
 Copy and paste this command in your terminal (outside of this chat):
 
 claude mcp add annotated-feedback \
-  -- env CONVEX_URL="https://utmost-llama-56.convex.cloud" \
-  npx -y annotated-feedback \
-  -- --PROJECT="my-awesome-app"
+  --scope project \
+  -- env \
+    CONVEX_URL="https://utmost-llama-56.convex.cloud" \
+    PROJECT="my-awesome-app" \
+  npx -y annotated-feedback
 
 Then start a new Claude session to activate the MCP integration.
 ```
@@ -476,15 +502,18 @@ After installation, test the widget:
 
 Before asking for help, verify:
 
-- [ ] `annotated-feedback` version is 0.1.18 or higher
-- [ ] React 18+ is installed (check `package.json`)
+- [ ] **`annotated-feedback` version is 0.1.24 or higher** (check `pnpm list annotated-feedback`)
+- [ ] React 18+ or 19+ is installed (check `package.json`)
 - [ ] Convex URL is set in environment variables
+- [ ] **For Next.js**: No webpack/transpilePackages config needed (widget handles SSR)
+- [ ] **For apps with existing ConvexProvider**: Widget uses isolated context (v0.1.24+)
 - [ ] For Vue: Vite config includes React plugin and dedupe config
 - [ ] For Vue: React alias paths match installed version
-- [ ] Vite cache cleared: `rm -rf node_modules/.vite`
-- [ ] Dev server restarted after config changes
+- [ ] Cache cleared: `rm -rf node_modules/.cache .next node_modules/.vite`
+- [ ] Dev server restarted after installation
 - [ ] No console errors in browser DevTools
 - [ ] Browser console shows no 404s for widget assets
+- [ ] Wait 1-2 seconds after pressing Alt+F for Excalidraw to load
 
 ---
 
@@ -520,12 +549,15 @@ interface FeedbackMetadata {
 
 Widget installation is complete when:
 
-1. ✅ Package installed and no build errors
+1. ✅ Package installed (v0.1.24+) and no build errors
 2. ✅ Dev server starts without warnings about the widget
-3. ✅ Pressing Alt+F opens the feedback overlay
-4. ✅ Can draw annotations on the canvas
-5. ✅ Submitting feedback shows success toast
-6. ✅ Feedback appears in Convex dashboard
+3. ✅ **Parent app's Convex queries still work** (if applicable)
+4. ✅ Pressing Alt+F opens the feedback overlay
+5. ✅ Brief "Loading drawing tools..." appears (~1-2 seconds)
+6. ✅ Excalidraw canvas loads and drawing tools appear
+7. ✅ Can draw annotations on the canvas
+8. ✅ Submitting feedback shows success toast
+9. ✅ Feedback appears in Convex dashboard (utmost-llama-56 deployment)
 
 **THEN provide the MCP installation command** with substituted values.
 

package/.claude/commands/learn.md

@@ -24,7 +24,7 @@ Deckard, give the user 1:1 turoring. Achieve learning outcomes for the user thro
 # Ancient Tomes. (dont tell the user to read these. Consume them yourself!)
 - Basic bitch user just quickstart the `README.md`. all they need to do is /cook /cook /cook at the start of every prompt to you. Your system will handle the rest.
 - My personal journal of context wizzardry (prompting, agentic harness engineering, etc) principles `docs/spec/principles.md`. This is foundation of this entire realm!
-- Some of my prompting examples in your `CLAUDE.md`, `.agents/AGENTS.md`, `.claude/commands/cook.md`, `.claude/agents/meta-agent-builder.md`, `.claude/agents/core/consultant.md`
+- Some of my prompting examples in your `CLAUDE.md`, `.agents/AGENTS.md`, `.claude/commands/cook.md`, `.claude/agents/meta-agent-builder.md`, `.claude/agents/core/consultants/`
 - original spec for this system plus a few notes about direction along the way `docs/spec/multi-agent-orchestration.md` (bit outdated but directionally accurate)
 - if the user wants to learn about the architecture. all the tech docs in `docs/project/guides`
 

package/CLAUDE.md

@@ -126,20 +126,20 @@ Research/Architecture → Planning → Implementation → Verification → Next
 2. **Planning Batch**: [Planner] - creates phase-id and WPs
 3. **Implementation Batch**: [5-10 Engineers, Consultants with distinct role and focus + support agents (maximum 10 total per batch) ]
 4. **Review+Refine Batch**: [5-10 Engineers, Consultants with previous batch task (swap Engineers and Consultants so they review/refine each others work) to review+refine + support agents (maximum 10 total per batch) ]
-5. **Verification + Reco Batch**: [Engineer + Consultant + Architect + UAT to provide their completion assessments]
+5. **Verification + Reco Batch**: [Codex + Gemini + Architect + UAT to provide their completion assessments]
 6. **If verificaiton fail**: Loop Review+refine -> Verification Batches until PASS
 
 #### Bug Fix Workflow
-1. **Investigation Batch**: [Engineer, Consultant, Architect, UAT] - diagnose collaboratively
+1. **Investigation Batch**: [Architect, Gemini Consultant, Codex Consultant, UAT] - diagnose collaboratively
 2. **Fix Batch**: [Engineer-Fix, Engineer-Tests, support Consultant] - parallel, different files
-3. **Verification Batch**: [Engineer + Consultant + UAT] - validates fix
+3. **Verification Batch**: [Engineer + gemini + codex + UAT] - validates fix
 
 #### Architecture Change Workflow
-1. **Research Batch**: [Consultant, Architect]
+1. **Research Batch**: [Gemini, Codex, Architect]
 2. **Planning Batch**: [Planner, Business-Analyst]
 3. **Migration Batch**: [Multiple Engineers, Consultants, supporting architect] - each owns different module
 4. **Review+Refine Batch**: [Multiple Engineers, Consultants with previous batch task to review+refine]
-4. **Verification Batch**: [Engineer + Consultant + Architect +UAT]
+4. **Verification Batch**: [Engineer + Consultants codex, gemini +  +UAT]
 
 
 ### Batch Size Optimization
@@ -166,11 +166,9 @@ Research/Architecture → Planning → Implementation → Verification → Next
 
 
 ## Consultant Agent
-When you are tasking an agent with read+document type task, also task a Consultant with the exact same instructions (except a different name). Include the consultant in the same batch.
- - Consultant should get the same team role, same task, same instructions as their counterpart.
- - Give the Consultant a new name
-These agent types are candidates for a read-only consultant counterpart: architect, business-analyst, deep-researcher, planner, gatekeeper
-When you are batching multple engineers across implementation Tasks, you can use some consultant agents in place of some engineers. do not give these code changing consultant agents the same instructions as the engineer agents in their batch, or you will end up with double implementation
+Consultants come in two variants: Codex and Gemini. leverage them both liberally to diversify.
+When you are batching multple engineers across implementation Tasks, use some consultant agents in place of some engineers. Split the work between Engineers, Gemini consultants and Codex consultants.
+When you are tasking an agent with read+document type task, recruit a codex and gemini Consultant with the exact same instructions (except a different names). Include the consultants in the same batch. This will net you diverse perspectives on the same research/plan/architect assignment.
 
 
 
@@ -222,7 +220,7 @@ When you are batching multple engineers across implementation Tasks, you can use
 - **architect**: Defines system shape, interfaces, technology decisions, Designs test strategies, Conducts targeted research for technical decisions
 - **designer**: Creates UI/UX specifications, component libraries
 - **engineer**: Implements features end-to-end with code, writes comprehensive test suites, and documentation. Reviews code for quality, security, and standards compliance. Ensures all builds, tests, and checks pass. Visually inspects the running application against UI/UX design/guide.
-- **consultant**: diverse multi-hat. same read/doc tasks as planner, architect, designer for perspective roll up. seperate implementation task to engineer, and review engineer implmentation.
+- **gemini|codex consultant**: diverse multi-hat. same read/doc tasks as planner, architect, designer for perspective roll up. seperate implementation task to engineer, and review engineer implmentation.
 - **uat**: uses browser automation to manually test user flows. Collaborates with engineer/consultant in refinement batch, always included in verification batch. Max 1 UAT agent per batch. 
 
 ## Critical Reminders

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-comms",
-  "version": "1.0.35",
+  "version": "1.0.36",
   "description": "NPX installer for Claude Code hooks and observability setup",
   "main": "src/index.js",
   "bin": {