Skip to content

codex_antigravity_bridge.md

Latest commit

 

History

History
74 lines (54 loc) · 4.06 KB

codex_antigravity_bridge.md

File metadata and controls

74 lines (54 loc) · 4.06 KB
Raw

Codex And Antigravity Bridge Protocol

Purpose

This protocol lets TriageCore play to the strengths of both Codex and Antigravity while preserving local-first evidence.

  • Local backends draft, classify, benchmark, or attempt bounded work first.
  • Antigravity can act as the IDE-native supervisor for implementation flow and local pipeline coordination.
  • Codex can act as the repo-grounded supervisor for code review, patching, verification, and documentation.
  • TriageCore records the decision trail instead of silently treating any supervisor as magic.

Recommended Division Of Labor

Lane Best At TriageCore Evidence
Local LLM fast first drafts, small benchmark tasks, low-risk parsing and repair backend, model, tokens, duration, energy, validator result
Worker Council parallel local specialist review worker role, aggregate tokens, duration, escalation decision
Antigravity IDE-native orchestration, visual implementation flow, Gemini supervisor coordination supervisor_tool=antigravity, model/profile, artifact path, decision
Codex repo-grounded review, patching, tests, methodology/docs, Git hygiene supervisor_tool=codex, notes, tests/artifacts, decision

Credibility And Throughput Separation

TriageCore intentionally separates high-throughput implementation from publication-facing credibility work.

Antigravity is the preferred lane for moving quickly inside the IDE: exploring larger implementation slices, coordinating local model pipelines, sketching UI or workflow changes, and producing enough concrete code for later review. Antigravity outputs should be treated as productive drafts until they are reconciled against the ledger, tests, documentation, and evidence boundaries.

Codex is the preferred lane for credibility review: checking claims against the repo state, tightening methodology language, preserving APA7-ready citation habits, reconciling docs with implementation, running stability checks, and making sure paper-facing language stays cautious, reproducible, and falsifiable. Codex review does not make a result scientifically valid by itself; it helps label what is verified, what is estimated, and what still needs evidence.

This separation is an operating protocol, not a ranking of tools. The goal is to use Antigravity for momentum and Codex for stabilization, documentation credibility, and evidence-bound interpretation.

Workflow

  1. Run a local draft, worker council, benchmark, or auto pipeline task.

  2. Review the task in the ledger.

  3. If Antigravity or Codex supervises the result, record that review:

    triagecore record-supervisor-review <task_id> --tool codex --decision needs_revision --notes "Local draft was close but missed tests." --model gpt-5 --profile high --artifact-path triage_tasks\codex_task_example.md
    triagecore record-supervisor-review <task_id> --tool antigravity --decision accepted --notes "Gemini supervisor accepted the local draft after IDE review." --model gemini-3.1-pro-high --profile supervisor --artifact-path .agent_tasks\example\TASK.md

  4. Use accepted, rejected, needs_revision, or escalated consistently.

  5. Treat estimated supervisor token fields as estimates unless exact tool usage is available.

  6. When exact tool usage is available as JSON or JSONL, import it instead of retyping values:

    triagecore scan-supervisor-usage supervisor_logs\
triagecore import-supervisor-usage supervisor_usage.jsonl --tool codex --token-source imported_exact --dry-run
triagecore import-supervisor-usage supervisor_usage.jsonl --tool codex --token-source imported_exact

Interpretation Rule

Do not compare local-only runs against supervised runs without labeling the supervision lane. A Codex-reviewed patch and an Antigravity-supervised IDE implementation are different workflows, even if they start from the same local draft.

Next Improvements

  • Capture exact supervisor token usage when an API or tool surface exposes it.
  • Import supervisor usage from tool logs when exact values can be verified.