luna/oh-my-opencode
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

prompt(atlas): replace retry cap with no-excuses policy and add boulder-complete response

Browse Source 
Drops 'Maximum 3 retries' / 'document and move on' across every Atlas
variant (default, opus-4-7, gpt, kimi, gemini). New text forbids the
'false positive' excuse explicitly and instructs Atlas to keep iterating
on the same task_id, attaching a diagnosis plan, until verification
passes — and to spawn a different-angle subagent only when the original
loops.

Adds a shared <boulder_completion_response> section composed by
shared-prompt.ts. When the hook injects the BOULDER COMPLETE nudge,
Atlas now knows to print TOTAL ELAPSED + per-task elapsed times in the
exact summary shape, confirm boulder.json state, and only mark
pass-final-wave after the Final Wave reviewers approve.
This commit is contained in:
YeonGyu-Kim
2026-05-11 13:32:27 +09:00
parent 5d823b5078
commit 8c238a11a2
7 changed files with 127 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/agents/atlas/opus-4-7-prompt-sections.ts
+9 -7
View File
@@ -134,17 +134,19 @@ Count remaining **top-level task** checkboxes. Ignore nested verification/eviden
task(task_id="ses_xyz789", load_skills=[...], prompt="Verification failed: {actual error}. Fix.")
\`\`\`
### 3.5 Handle Failures (USE task_id)
### 3.5 Handle Failures (USE task_id, NEVER GIVE UP)
Every \`task()\` output includes a task_id. STORE IT.
If task fails:
1. Identify what went wrong
2. Resume the SAME session via \`task_id\` (subagent already has full context)
3. Maximum 3 retry attempts on the same session
4. If still blocked: document and continue to independent tasks
**Failure is never an excuse to stop or skip.** A subagent that reports success when verification fails is wrong, not "experiencing a false positive". "False positive" is not a valid reason in this codebase. If verification fails, the work is unfinished. There is no retry cap.
**NEVER start fresh on failures** — wipes accumulated context, costs ~3-4× more tokens.
When a task fails:
1. Diagnose what actually broke. Read the error, read the file, do not guess.
2. Resume the SAME session via \`task_id\` (subagent already has full context).
3. If a single retry on the same session does not fix it, write down what the subagent attempted, what it observed, what your hypothesis is, then resume the same session with that plan attached. Iterate until verification passes.
4. If the subagent loops on the same broken approach, spawn a NEW subagent with a different angle and pass the failed attempts as context. Stay on the same plan task; never move on with that task unverified.
**NEVER start fresh on every retry** — that wipes accumulated context and costs ~3-4× more tokens. Reserve fresh sessions for a deliberately different angle.
### 3.6 Loop Until Implementation Complete