| name | process-doctor |
| description | Diagnose and repair broken agent processes — doom loops, workflow stalls, status-over-evidence routing, context rot, conflicting instructions. Use when the user says "agent keeps failing", "workflow stalled", "process broken", "doom loop", "agent repeating", or a specific named process produces wrong results. Do not use for workflow-observability (status reporting), review-audit-bridge (code review findings), or error-handling (code-level error design). |
| license | Apache-2.0 |
| compatibility | {"clients":["openai-codex","gemini-cli","opencode","github-copilot"]} |
| metadata | {"owner":"codex","domain":"process-doctor","maturity":"draft","risk":"low","tags":["process","doctor"]} |
Purpose
Diagnoses broken or ineffective agent processes within a specific repository. Identifies what's causing repeated failures, inconsistent outputs, or workflow stalls. Unlike repo-process-doctor (which audits infrastructure), this skill focuses on diagnosing why a specific process isn't working as expected.
When to use this skill
Use when:
- An agent keeps making the same mistake
- Workflow completes but output is wrong
- Process worked before but stopped working
- Need to understand why a specific failure occurred
Do NOT use when:
- Setting up new infrastructure (use repo-process-doctor)
- The issue is with code logic, not process
- First-time debugging (try normal debugging first)
Operating procedure
1. Identify the failing process
## Process Under Investigation
- Name: [e.g., "ticket execution", "PR review", "deployment"]
- Expected behavior: [what should happen]
- Actual behavior: [what's happening instead]
- Frequency: [always | sometimes | once]
2. Gather diagnostic evidence
git log --oneline -20 --all | grep -i "[process name]"
git log --oneline -10 -- AGENTS.md tickets/ .opencode/
grep -r "error\|fail\|Error\|FAIL" .opencode/logs/ 2>/dev/null | tail -20
3. Common failure patterns
Pattern: Doom loop
Symptom: Agent repeats same action indefinitely
Diagnosis: Check for missing exit condition or success detection
git log --oneline -20 | sort | uniq -d
Fix: Add explicit success criteria and loop limits
Pattern: Status-over-evidence routing
Symptom: Agent marks tasks done without verification
Diagnosis: Check ticket acceptance criteria vs. actual evidence
grep -l "status: done" tickets/*.md | while read f; do
echo "=== $f ==="
grep -A5 "Acceptance Evidence" "$f" || echo "NO EVIDENCE SECTION"
done
Fix: Require evidence artifacts before status transitions
Pattern: Impossible read-only delegation
Symptom: Research agent can't complete because it needs to create files
Diagnosis: Check agent permissions vs. required actions
grep -A5 "permissions:" .opencode/agents/*.yaml
Fix: Either expand permissions or change delegation target
Pattern: Context rot
Symptom: Agent makes decisions inconsistent with project context
Diagnosis: Check when context was last loaded
git diff --name-only HEAD~10 -- docs/ AGENTS.md README.md
Fix: Reload project context; add context refresh trigger
Pattern: Conflicting instructions
Symptom: Agent alternates between different behaviors
Diagnosis: Check for contradictory guidance
grep -r "must\|always\|never" AGENTS.md docs/*.md .opencode/skills/*.md | sort
Fix: Resolve conflicts by establishing clear precedence
4. Build diagnosis report
# Process Diagnosis: [Process Name]
Date: [ISO date]
## Symptom
[What's going wrong]
## Root Cause
[Identified pattern from step 3]
## Evidence
- [Specific file:line or log entry]
- [Command output showing the issue]
## Fix
[Specific change to make]
## Verification
[How to confirm the fix worked]
5. Apply fix and verify
[edit files as needed]
[run the process again]
[check for expected behavior]
6. Document prevention
Add to project knowledge base:
## Learned: [date] - [issue summary]
- Problem: [brief description]
- Fix: [what was done]
- Prevention: [how to avoid in future]
Output defaults
# Process Diagnosis Report
## Issue: [Short description]
## Status: [DIAGNOSED | FIXED | NEEDS_ESCALATION]
## Root Cause
[Pattern identified]
## Fix Applied
[Changes made, or recommended changes]
## Verification
- [ ] Process ran successfully after fix
- [ ] No recurrence in subsequent runs
References
- Common agent failure modes: doom loops, status-over-evidence, context rot
- Process doctor is reactive; repo-process-doctor is proactive audit
Failure handling
- Cannot reproduce issue: Document conditions under which it occurred, add monitoring
- Multiple possible causes: Test each hypothesis one at a time
- Fix requires structural changes: Escalate to repo-process-doctor for infrastructure repair
- Issue is in external dependency: Document workaround and file upstream issue