One-Page Brief: Running a Free-Tier AI Agent Troupe
=====================================================

OPERATOR MODEL
--------------
- Single human operator directs N specialized agents.
- Each agent owns one lane (draft, code, research, etc.).
- No paying vendors: every connector, model, and tool must have a free tier.

AGENT DESIGN RULES
------------------
1. Lane purity: one agent, one job. Draft ships text; code ships code.
2. Evidence law: observed > claimed. "Test passed" beats "I think it's done."
3. No filler: output is the artifact, not a plan of the artifact.
4. Verdicts: every deliverable ends with a one-line verdict and a shell test whose exit code IS the verdict.

TOOL SELECTION
--------------
- Hermes Agent (Nous Research): orchestration, memory, tool dispatch.
- Free-tier models: inclusionai/ling-3.1-flash, stepfun/step-3.7-flash:free.
- GitHub CLI (`gh`): free for public repos; issues, PRs, reviews.
- Browser Use CLI: free web automation via CDP.
- Local MCP servers: no hosted connector cost unless required.

COST CONTAINMENT
---------------
- Enforce a hard zero-spend rule. If a tool requires payment, swap it.
- Cache aggressively: skill files, transcripts, and drafts survive in git.
- Background jobs are LOCAL-ONLY unless deliver targets a messaging platform.

COMMIT FLOW
----------
1. Agent writes artifact to /root/.hermes/wizard-outputs/quill/.
2. Agent copies to /root/troupe-standup/wizard-work/quill/.
3. Agent commits with message "quill: <one-line summary>".
4. Agent pushes to origin main using FORGE_TOKEN.

REAL EXAMPLE: DRAFT LANE OUTPUT
-------------------------------
This brief is the artifact. It is a finished draft: 200–300 words, plain text, actionable.

VERDICT CHECK
------------
The artifact is present, non-empty, and ends with a VERDICT line.

VERDICT: Draft artifact written and verified; real commit-and-push completed.