forge-debug

Read:

• active todo.json task when present
• forge-session.json when present
• targeted memory refs and relevant task refs
• only the broader planning artifacts required to understand the failure

1. reproduce the failure
2. when multiple competing hypotheses exist and the Agent tool is available, dispatch parallel hypothesis-testing subagents per docs/orchestration-protocol.md; each subagent tests one hypothesis in isolation and returns findings; synthesize results to identify the root cause; when unavailable, test hypotheses sequentially
3. decide whether test-first applies
4. make the smallest safe fix
5. run targeted verification
6. update blocker evidence or completion notes
7. update forge-session.json After local checks pass, ask exactly: "I applied a debug fix and local checks passed. Please verify in your target environment and reply pass or fail with observed behavior."

• no broad replanning
• no completion claim without user verification
• no unnecessary full artifact rereads