[Jay Gordon (jaydestro)]

The first main-branch native Copilot run (40537110727) installed the CLI and provisioned the workflow environment but all three CLI processes rejected --max-ai-credits 10 with Use at least 30 AI credits before making inference requests. Raise the cap to the supported minimum of 30 per session and correct the README ceiling to 150 for three tasks plus at most two resolver sessions. No application, test, default PR behavior or publication/media changes. Actionlint passes with only the documented new Copilot permission-scope exception. The original task/isolation/stand-in jobs passed in that same run; the failed real-job summary and logs were retained, and its temporary Git resources were removed. The native run will be retried on main after this fix lands.

[Jay Gordon (jaydestro)]

Verified native Copilot on merged main at 3c413a9:
https://github.com/AzureCosmosDB/fleetcharge-agent-cosmos-ci/actions/runs/40539601322

All five jobs passed, including Real Copilot worktree swarm. Downloaded copilot-swarm-1 and independently checked its result, patches and JUnit files:

• Three real Copilot CLI sessions, three distinct emulators/endpoints, verified BAY-7 / BAY-8 / BAY-9 state.
• Per-task tests: 18 / 18 / 57. Agent A/B changed dispatch error diagnostics; agent C added a fleet-rule explanation comment. All patches and notes were retained.
• Genuine conflict in app/dispatch.py resolved by Copilot; both Git parents retained, combined patch retained all notes/source changes, and 75 combined tests passed (11 seconds).
• The negative control used deterministic stand-ins, reproduced one lost update, and was not mislabeled as real-agent inference.
• Owned containers, temporary worktrees and branches were removed. Runner duration: 140.802 seconds.
• Native GITHUB_TOKEN authentication worked without a PAT/repository secret. The corrected CLI minimum is 30 AI credits per session, up to 150 for five sessions. The earlier failed run made no inference requests because 10 was rejected during argument validation.

The 211-test local gate and all five regular PR checks (including real Ollama) also passed. Generated experimental application edits remain artifacts only; they were not pushed into main. Blog/WordPress/video/image work remains paused.