fix: make PO Agent source-agnostic, remove AzDo hard dependency

.github/workflows/po-agent-resume.yml

@@ -0,0 +1,17 @@
+name: PO Agent Resume
+on:
+  workflow_run:
+    workflows: ["CI"]
+    types: [completed]
+
+permissions:
+  contents: read
+  pull-requests: read
+  actions: write
+
+jobs:
+  resume:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ./resume

.github/workflows/po-agent.yml

@@ -0,0 +1,51 @@
+name: PO Agent
+on:
+  issue_comment:
+    types: [created]
+  repository_dispatch:
+    types: [po-agent-resume]
+
+permissions:
+  contents: write
+  pull-requests: write
+  issues: write
+  actions: write
+
+jobs:
+  agent:
+    if: >
+      (github.event_name == 'issue_comment' &&
+       contains(github.event.comment.body, '@po-agent') &&
+       (github.event.comment.author_association == 'OWNER' ||
+        github.event.comment.author_association == 'MEMBER' ||
+        github.event.comment.author_association == 'COLLABORATOR'))
+      || github.event_name == 'repository_dispatch'
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    concurrency:
+      group: po-agent-${{ github.event.issue.number || github.event.client_payload.pr_number || github.run_id }}
+      cancel-in-progress: false
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Start Copilot API proxy
+        run: |
+          npm install -g copilot-api@latest
+          copilot-api start --github-token "$COPILOT_GITHUB_TOKEN" &
+          for i in $(seq 1 30); do
+            if curl -sf http://localhost:4141 > /dev/null 2>&1; then
+              echo "Copilot API proxy running (attempt $i)"
+              exit 0
+            fi
+            sleep 2
+          done
+          echo "Proxy failed to start after 60s"
+          exit 1
+        env:
+          COPILOT_GITHUB_TOKEN: ${{ secrets.COPILOT_API_GITHUB_TOKEN }}
+
+      - uses: ./
+        with:
+          anthropic_api_key: "dummy"
+          anthropic_base_url: "http://localhost:4141"
+          max_budget_usd: "3.0"

.po-agent/skills/test-skills/SKILL.md

@@ -0,0 +1,63 @@
+---
+name: test-skills
+description: Run the skill validation framework against all built-in and consumer skills.
+disable-model-invocation: true
+---
+
+# Test Skills
+
+Run this skill to validate that all skill definitions are correctly structured.
+
+## Step 1: Validate Built-in Skills
+
+```bash
+bash scripts/validate-all-skills.sh skills/
+```
+
+All 11 built-in skills (including `rfs-proposal`) must pass with 0 errors.
+
+## Step 2: Validate Consumer Skills (if any)
+
+```bash
+if [ -d .po-agent/skills ]; then
+  bash scripts/validate-all-skills.sh .po-agent/skills/
+else
+  echo "No consumer skills found"
+fi
+```
+
+## Step 3: Check Skill Routing Table
+
+Verify that every skill in `skills/` has a corresponding entry in `prompts/core.md`:
+
+```bash
+for skill_dir in skills/*/; do
+  skill_name=$(basename "$skill_dir")
+  if grep -q "/$skill_name" prompts/core.md; then
+    echo "✅ $skill_name — listed in core.md"
+  else
+    echo "⚠️ $skill_name — NOT in core.md routing table"
+  fi
+done
+```
+
+## Step 4: Check for Orphaned Skill References
+
+Verify that every skill referenced in `prompts/core.md` has an actual directory:
+
+```bash
+grep -oP '/[a-z][-a-z]+' prompts/core.md | sort -u | while read -r skill_ref; do
+  skill_name="${skill_ref#/}"
+  if [ -d "skills/$skill_name" ] || [ -d ".po-agent/skills/$skill_name" ]; then
+    echo "✅ $skill_ref — exists"
+  else
+    echo "❌ $skill_ref — referenced but not found"
+  fi
+done
+```
+
+## When to Use
+
+- After adding, modifying, or removing a skill
+- After modifying `prompts/core.md`
+- Before releasing a new version

.po-agent/skills/validate-action-yml/SKILL.md

@@ -0,0 +1,64 @@
+---
+name: validate-action-yml
+description: Validate the action.yml composite action after modifications. Checks YAML syntax, inputs/outputs, step structure.
+disable-model-invocation: true
+---
+
+# Validate action.yml
+
+Run this skill after modifying `action.yml` to ensure the file is structurally valid.
+
+## Step 1: YAML Syntax Check
+
+```bash
+python3 -c "import yaml; yaml.safe_load(open('action.yml'))" && echo '✅ YAML valid' || echo '❌ YAML invalid'
+```
+
+## Step 2: Structure Validation
+
+```bash
+python3 -c "
+import yaml
+with open('action.yml') as f:
+    data = yaml.safe_load(f)
+
+# Check required top-level keys
+assert 'name' in data, 'Missing: name'
+assert 'inputs' in data, 'Missing: inputs'
+assert 'outputs' in data, 'Missing: outputs'
+assert 'runs' in data, 'Missing: runs'
+assert data['runs']['using'] == 'composite', 'Not a composite action'
+
+# Check required inputs
+required_inputs = ['anthropic_api_key']
+for inp in required_inputs:
+    assert inp in data['inputs'], f'Missing required input: {inp}'
+
+# Check outputs
+assert 'response' in data['outputs'], 'Missing output: response'
+assert 'cost_usd' in data['outputs'], 'Missing output: cost_usd'
+
+# Count steps
+steps = data['runs']['steps']
+print(f'✅ All checks passed ({len(steps)} steps, {len(data[\"inputs\"])} inputs)')
+"
+```
+
+## Step 3: Check for Common Issues
+
+- Ensure all `jq` calls have `2>/dev/null` or `|| true` for error handling
+- Ensure all `grep -oP` calls have `|| true` to handle no-match exit code 1
+- Ensure `GITHUB_OUTPUT` writes don't contain unescaped newlines in values
+- Ensure all step IDs referenced in `steps.*.outputs.*` actually exist
+
+## Step 4: Verify Skills Still Pass
+
+```bash
+bash scripts/validate-all-skills.sh
+```
+
+## When to Use
+
+- After ANY edit to `action.yml`
+- After merging upstream changes
+- Before committing action.yml changes

CLAUDE.md

@@ -0,0 +1,101 @@
+# PO Agent — AI-Powered Product Owner
+
+## Project Overview
+
+PO Agent is a **GitHub Actions composite action** that wraps Claude Code CLI to act as an AI Product Owner and Senior Developer. It manages the full lifecycle: requirements, investigation, implementation, CI, review, verification, and delivery.
+
+## Architecture
+
+```
+action.yml              ← Main composite action (all steps in one job)
+prompts/
+  core.md               ← Agent persona + skill routing table
+  source-github.md      ← GitHub PR trigger behavior
+  source-github-issue.md← GitHub Issue trigger behavior
+  source-slack.md       ← Slack trigger behavior
+  source-resume.md      ← Auto-resume trigger behavior
+  source-azure-devops.md← Azure DevOps trigger behavior
+  rfs-detection.md      ← Skill gap detection prompt
+skills/                 ← Built-in skills (SKILL.md files)
+scripts/                ← Shell scripts linked to PATH at runtime
+resume/action.yml       ← Resume sub-action (triggers on CI completion)
+webhooks/               ← Event relay functions (Slack, AzDo)
+docs/                   ← Architecture, quickstart, writing skills, RFS docs
+```
+
+## Key Files
+
+- **`action.yml`** (~1500 lines): The core of the project. Contains ALL steps:
+  1. Source detection (github/github-issue/slack/resume/azure-devops/webhook/manual)
+  2. Secret export
+  3. Acknowledge receipt (reaction/comment per source)
+  4. Install Claude Code CLI
+  5. Setup Node.js + Playwright
+  6. Register MCP servers (GitHub, Playwright, optionally AzDo, Notion)
+  7. Link skills + scripts
+  8. Fetch context (PR details, issue details, AzDo work items, Slack threads)
+  9. Compose prompt + run Claude Code
+  10. Save persistent context + transcript
+  11. Upload artifacts
+  12. Post response to source
+  13. Job summary + cancellation handling
+
+- **`resume/action.yml`**: Watches for CI completion via `workflow_run` event, checks labels, dispatches `po-agent-resume` repository_dispatch event.
+
+## Inputs
+
+| Input | Required | Default | Description |
+|-------|----------|---------|-------------|
+| `anthropic_api_key` | Yes | — | API key (or dummy if using proxy) |
+| `max_budget_usd` | No | `5.0` | Max spend per session |
+| `model` | No | `claude-sonnet-4-20250514` | Default model |
+| `anthropic_base_url` | No | — | Custom API base URL (proxy support) |
+| `timeout_minutes` | No | `0` | Timeout awareness |
+| `dry_run` | No | `false` | Investigation only, no code changes |
+| `investigation_model` | No | — | Model for read-only tasks |
+| `implementation_model` | No | — | Model for code-writing tasks |
+| `trigger_word` | No | `@po-agent` | Keyword to invoke agent |
+| `secrets` | No | — | KEY=VALUE pairs as env vars |
+
+## Skills System
+
+Skills are markdown files (`SKILL.md`) with frontmatter, linked into `.claude/skills/` at runtime.
+
+- **Built-in skills** (10 + rfs-proposal): `skills/` directory
+- **Consumer overrides**: `.po-agent/skills/<name>/` overrides built-in skill of same name
+- **Consumer extensions**: `.po-agent/skills/<name>/` for project-specific skills
+
+### Validation
+
+```bash
+bash scripts/validate-skill.sh skills/bug-fix-workflow/
+bash scripts/validate-all-skills.sh
+```
+
+## Label System
+
+Labels on PRs control the auto-resume flow:
+- `po-agent:waiting-ci` — waiting for CI
+- `po-agent:waiting-human` — waiting for human input (`<!-- WAITING_FOR_HUMAN -->`)
+- `po-agent:in-progress` — agent is running
+- `po-agent:attempt-N` — tracks resume attempts
+- `po-agent:failed` — max attempts exceeded
+
+## RFS (Request for Skills)
+
+The agent auto-detects skill gaps and files structured GitHub Issues with `rfs` + `rfs:proposed` labels. See `docs/rfs.md` for the full lifecycle.
+
+## Testing
+
+- Validate YAML: `python3 -c "import yaml; yaml.safe_load(open('action.yml'))"`
+- Validate skills: `bash scripts/validate-all-skills.sh`
+- Test in a consumer repo: point workflow at `Adrian62D/po-agent@main`
+
+## Development Guidelines
+
+- All shell code is in `action.yml` steps — there's no separate build step
+- Use `|| true` after grep/commands that may return non-zero in `set -e` contexts
+- Test YAML validity after every edit to action.yml
+- MCP servers are registered at runtime, not bundled
+- Prompts are composed from `core.md` + `source-*.md` + context + user message
+- The `stream-json` output format is parsed for `"type":"result"` lines