Skip to content

feat(brain): experimental model and thinking cadence foundation - #772

Draft
axelpey wants to merge 4 commits into
mainfrom
codex/agent-model-cadence
Draft

feat(brain): experimental model and thinking cadence foundation#772
axelpey wants to merge 4 commits into
mainfrom
codex/agent-model-cadence

Conversation

@axelpey

@axelpey axelpey commented Sep 6, 2026

Copy link
Copy Markdown
Contributor

Allows the core brain to opt into OpenAI Responses with gpt-6-astra / low reasoning and lets each directive vary its idle or running-skill thought pause. Gemini 3.6 Flash/minimal and the 3s/5s global defaults remain unchanged. Provider/model selection takes effect on node restart.

The foundation belongs below the Household Orders solution/skills/agent PR. Generic cadence is reused from #694 commit 3894f29fa; no Household skills or routing changes are included. “Astra light” is interpreted as Astra at its lowest documented effort (low), not a different model ID. The separate key-configuration PR owns secure local setup.

The adapter preserves native tool-call IDs and encrypted reasoning in bounded local history, uses store=false, and reuses the existing turn dispatch/cancellation path. Stop now waits for child-turn cleanup before returning. Failed/incomplete streams cannot commit tools or history. The managed proxy's wrapped SSE event-name lines are supported; errors are sanitized and direct keys are blocked in public demo mode.

Validation:

  • 121 tests passed in the cached ROS Humble image: brain, loader, OpenAI context, HTTP transport, settings guards; no external network. Plugin autoload disabled for the image's pre-existing anyio/pytest mismatch.
  • 101 local tests passed; all applicable pre-commit hooks passed. Added focused host CI for the brain/context/transport tests.
  • 12 real managed-provider probes: six per model, all basic behavioral outcomes correct. Median completion 1.072s Gemini/minimal versus 1.898s Astra/low; estimated token cost $0.01446 versus $0.05096 for six calls. Small sample with cache/startup differences, not a Household success-rate study.
  • Separate three-turn live Astra replay passed (tool call → outcome → next turn). No robot motion, new VM campaign or deployment.

Configuration, reproducible probe, raw results and limitations.

Full ROS node + live simulator UI behavior has not been exercised in this foundation; Household integration remains above it. As in the current worker-thread architecture, cancelled HTTP requests can finish in the background but cannot commit or act. Keep this PR DRAFT until Axel explicitly approves it. Do not merge or deploy it as a substitute for approval.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant