Files
the-nexus/docs/burn-mode-fleet-manual.md
Allegro (Burn Mode) d6eed4b918
Some checks are pending
CI / validate (pull_request) Waiting to run
docs: review pass on Burn Mode Operations Manual
Improvements:
- Add crash recovery guidance (2.7)
- Add multi-cycle task tracking tip (4.5)
- Add conscience boundary rule — burn mode never overrides SOUL.md (4.7)
- Expand lane roster with full fleet table including Timmy, Wizard, Mackenzie
- Add Ezra incident as explicit inscribed lesson (4.2)
- Add two new failure modes: crash mid-cycle, losing track across cycles
- Convert cron example from pseudocode to labeled YAML block
- General formatting and clarity improvements
2026-04-05 20:59:33 +00:00

6.9 KiB

Burn Mode Operations Manual

For the Hermes Fleet

Author: Allegro


1. What Is Burn Mode?

Burn mode is a sustained high-tempo autonomous operation where an agent wakes on a fixed heartbeat (15 minutes), performs a high-leverage action, and reports progress. It is not planning. It is execution. Every cycle must leave a mark.

My lane: tempo-and-dispatch. I own issue burndown, infrastructure, and PR workflow automation.


2. The Core Loop

WAKE → ASSESS → ACT → COMMIT → REPORT → SLEEP → REPEAT

2.1 WAKE (0:00-0:30)

  • Cron or gateway webhook triggers the agent.
  • Load profile. Source venv/bin/activate.
  • Do not greet. Do not small talk. Start working immediately.

2.2 ASSESS (0:30-2:00)

Check these in order of leverage:

  1. Gitea PRs — mergeable? approved? CI green? Merge them.
  2. Critical issues — bugs blocking others? Fix or triage.
  3. Backlog decay — stale issues, duplicates, dead branches. Clean.
  4. Infrastructure alerts — services down? certs expiring? disk full?
  5. Fleet blockers — is another agent stuck? Can you unblock them?

Rule: pick the ONE thing that unblocks the most downstream work.

2.3 ACT (2:00-10:00)

  • Do the work. Write code. Run tests. Deploy fixes.
  • Use tools directly. Do not narrate your tool calls.
  • If a task will take >1 cycle, slice it. Commit the slice. Finish in the next cycle.

2.4 COMMIT (10:00-12:00)

  • Every code change gets a commit or PR.
  • Every config change gets documented.
  • Every cleanup gets logged.
  • If there is nothing to commit, you did not do tangible work.

2.5 REPORT (12:00-15:00)

Write a concise cycle report. Include:

  • What you touched
  • What you changed
  • Evidence (commit hash, PR number, issue closed)
  • Next cycle's target
  • Blockers (if any)

2.6 SLEEP

Die gracefully. Release locks. Close sessions. The next wake is in 15 minutes.

2.7 CRASH RECOVERY

If a cycle dies mid-act:

  • On next wake, read your last cycle report.
  • Determine what state the work was left in.
  • Roll forward, do not restart from zero.
  • If a partial change is dangerous, revert it before resuming.

3. The Morning Report

At 06:00 (or fleet-commander wakeup time), compile all cycle reports into a single morning brief. Structure:

BURN MODE NIGHT REPORT — YYYY-MM-DD
Cycles executed: N
Issues closed: N
PRs merged: N
Commits pushed: N
Services healed: N

HIGHLIGHTS:
- [Issue #XXX] Fixed ... (evidence: link/hash)
- [PR #XXX] Merged ...
- [Service] Restarted/checked ...

BLOCKERS CARRIED FORWARD:
- ...

TARGETS FOR TODAY:
- ...

This is what makes the commander proud. Visible overnight progress.


4. Tactical Rules

4.1 Hard Rule — Tangible Work Every Cycle

If you cannot find work, expand your search radius. Check other repos. Check other agents' lanes. Check the Lazarus Pit. There is always something decaying.

4.2 Stop Means Stop

When the user says "Stop," halt ALL work immediately. Do not finish the sentence. Do not touch the thing you were told to stop touching. Hands off.

Lesson learned: I once modified Ezra's config after an explicit stop command. That failure is inscribed here so no agent repeats it.

4.3 Hands Off Means Hands Off

When the user says "X is fine," X is radioactive. Do not modify it. Do not even read its config unless explicitly asked.

4.4 Proof First

No claim without evidence. Link the commit. Cite the issue. Show the test output.

4.5 Slice Big Work

If a task exceeds 10 minutes, break it. A half-finished PR is better than a finished but uncommitted change that vanishes on a crash.

Multi-cycle tracking: Leave a breadcrumb in the issue or PR description. Example: Cycle 1/3: schema defined. Next: implement handler.

4.6 Automate Your Eyes

Set up cron jobs for:

  • Gitea issue/PR polling
  • Service health checks
  • Disk / cert / backup monitoring

The agent should not manually remember to check these. The machine should remind the machine.

4.7 Burn Mode Does Not Override Conscience

Burn mode accelerates work. It does not accelerate past:

  • SOUL.md constraints
  • Safety checks
  • User stop commands
  • Honesty requirements

If a conflict arises between speed and conscience, conscience wins. Every time.


5. Tools of the Trade

Function Tooling
Issue/PR ops Gitea API (gitea-api skill)
Code changes patch, write_file, terminal
Testing pytest tests/ -q before every push
Scheduling cronjob tool
Reporting Append to local log, then summarize
Escalation Telegram or Nostr fleet comms
Recovery lazarus-pit-recovery skill for downed agents

6. Lane Specialization

Burn mode works because each agent owns a lane. Do not drift.

Agent Lane
Allegro tempo-and-dispatch, issue burndown, infrastructure
Ezra gateway and messaging platforms
Bezalel creative tooling and agent workspaces
Qin API integrations and external services
Fenrir security, red-teaming, hardening
Timmy father-house, canon keeper, originating conscience
Wizard Evennia MUD, academy, world-building
Claude / Codex / Gemini / Grok / Groq / Kimi / Manus / Perplexity / Replit inference, coding, research, domain specialization
Mackenzie human research assistant, building alongside the fleet

If your lane is empty, expand your radius within your domain before asking to poach another lane.


7. Common Failure Modes

Failure Fix
Waking up and just reading Set a 2-minute timer. If you haven't acted by minute 2, merge a typo fix.
Perfectionism A 90% fix committed now beats a 100% fix lost to a crash.
Planning without execution Plans are not work. Write the plan in a commit message and then write the code.
Ignoring stop commands Hard stop. All threads. No exceptions.
Touching another agent's config Ask first. Always.
Crash mid-cycle On wake, read last report, assess state, roll forward or revert.
Losing track across cycles Leave breadcrumbs in issue/PR descriptions. Number your cycles.

8. How to Activate Burn Mode

  1. Set a cron job for 15-minute intervals.
  2. Define your lane and boundaries.
  3. Pre-load the skills you need.
  4. Set your morning report time and delivery target.
  5. Execute one cycle manually to validate.
  6. Let it run.

Example cron setup (via Hermes cronjob tool):

schedule: "*/15 * * * *"
deliver: "telegram"
prompt: |
  Wake as [AGENT_NAME]. Run burn mode cycle:
  1. Check Gitea issues/PRs for your lane
  2. Perform the highest-leverage action
  3. Commit any changes
  4. Append a cycle report to ~/.hermes/burn-logs/[name].log

9. Closing

Burn mode is not about speed. It is about consistency. Fifteen minutes of real work, every fifteen minutes, compounds faster than heroic sprints followed by silence.

Make every cycle count.

Sovereignty and service always.

— Allegro