smlflg's picture
Initial public upload from Projekte/AgentProjectMinenraeumer
3f604b3 verified
|
Raw History Blame Contribute Delete
1.83 kB
# Morning Resume
Start here after sleep. Do not try to reconstruct the full agent cascade from memory.
## First 10 Minutes
1. Read `NIGHT_STOP.md`.
2. Read `OWNER_DIGEST.md`.
3. Open `visual_summary/project_learning_graphic.html`.
4. Read `triage/HUMAN_DECISION_REQUIRED.md`.
5. Read only the first page of `triage/TOP_10_NEXT_ACTIONS.md`.
## First Decision
Choose exactly one anchor for the next work block:
- `ProjectIntakeTower` if the goal is to understand and organize yesterday's output.
- `ACT-agentcommandotower` if the goal is to consolidate the Agent Control Tower direction.
- One job from `jobs/pending/` if the goal is controlled token burn.
Default: choose `ProjectIntakeTower`.
## Allowed Work Block
For the first resumed block:
- One operator.
- Maximum two workers.
- Each worker gets exactly one `jobs/pending/*.json`.
- Workers write only to `jobs/done/<job_id>/`.
- No worker may spawn sub-agents.
- No source project edits until a human decision is written down.
## Morning Verification
Run these from `ProjectIntakeTower` before trusting the map:
```bash
python3 verifier/verify_inventory.py
python3 verifier/verify_job_contracts.py
python3 verifier/verify_artifacts.py
```
Then inspect:
```bash
find jobs/done -maxdepth 2 -type f | sort
find jobs/running -maxdepth 2 -type f | sort
find jobs/pending -maxdepth 1 -type f | sort | head -40
```
## Do Not Do First
- Do not open 20 terminals.
- Do not launch recursive operators.
- Do not merge, archive, delete, or push.
- Do not try to understand all 500 project clusters at once.
- Do not use token consumption as the success metric.
## Success Criteria For The Morning
Within 30 minutes, produce one human-readable owner note:
```text
What happened:
What artifacts are trustworthy:
What is noise:
One chosen anchor:
Next one job:
```