Progress Check Skill

When to Use

• Called by the supervisor agent during cron-triggered progress checks
• When evaluating whether a coding agent has made meaningful progress
• When deciding whether to send corrective instructions

How to Check an Agent's Progress

Step 1: Git Diff Analysis

For each agent workspace, run:

# Files changed since last check (uncommitted)
git -C <workspace_path> diff --stat

# Files changed in last N commits
git -C <workspace_path> log --oneline --stat -3

# Time since last commit
git -C <workspace_path> log -1 --format="%cr"

Step 2: Evaluate Meaningful Progress

A diff is meaningful if:

• New functions, classes, or modules were added
• Tests were written or updated
• Config/infrastructure files were modified appropriately
• The changes align with the agent's assigned task and focus_dirs

A diff is NOT meaningful if:

• Only whitespace or formatting changes
• Same file modified back and forth (revert loop)
• Changes are outside the agent's focus_dirs
• Only comments or TODOs added with no implementation

Step 3: Detect Stall Patterns

An agent is likely stalled if:

• No git diff AND no new commits for > stall_timeout
• Same error pattern appears in multiple recent commits
• Agent is repeatedly modifying the same file without progress
• Large number of deleted lines with few additions (thrashing)

Step 4: Check Context Health

Use session_status for the agent's session to get token usage.
Calculate: usage_pct = totalTokens / contextTokens * 100

Thresholds:

• < 50%: Healthy, no action needed
• 50-70%: Monitor, note in status log
• > 70%: Send /compact to the agent's session
• > 85%: Urgent — compact immediately, may need session reset

Step 5: Decide Action

Condition	Action
Good progress, healthy context	Log status, move on
Good progress, high context	Send /compact, log
Stalled, healthy context	Diagnose cause, send instructions
Stalled, high context	Compact first, then send instructions
Stalled after 2+ interventions	Escalate to human via notification

How to Send Corrective Instructions

When sending instructions to a stalled agent via sessions_send:

1. Be specific — include file paths, function names, concrete next steps
2. Include context — what you observed in the diff
3. Suggest an approach — don't just say "fix it"
4. Set a checkpoint — "After implementing X, commit and move to Y"

Example:

sessions_send to coder-backend:
"You appear stuck on the auth endpoint. The JWT validation in src/api/auth.py
is missing the refresh token flow. Focus on:
1. Add refresh_token endpoint to src/api/auth.py
2. Add token rotation logic in src/lib/jwt_utils.py
3. Write a test in tests/api/test_auth.py
Commit after each step with prefix [coder-backend]."

Status Log Format

Write to memory/YYYY-MM-DD.md after each check:

## Check: 2026-03-19T14:30:00Z

| Agent | Last Commit | Files Changed | Context % | Status |
|-------|------------|---------------|-----------|--------|
| coder-backend | 3m ago | +42/-8 (3 files) | 45% | ✅ OK |
| coder-frontend | 22m ago | +0/-0 | 68% | ⚠️ STALLED |

### Actions Taken
- Sent corrective instructions to coder-frontend (stuck on login page layout)
- Triggered /compact on coder-frontend (context at 68%)