Skip to content

Add fail-closed steering activation preflight#38

Merged
mdheller merged 1 commit into
mainfrom
work/activation-path-34
May 20, 2026
Merged

Add fail-closed steering activation preflight#38
mdheller merged 1 commit into
mainfrom
work/activation-path-34

Conversation

@mdheller
Copy link
Copy Markdown
Contributor

@mdheller mdheller commented May 20, 2026

Summary

Adds the first GPT-2 Small controlled-activation entrypoint for Issue #34 without claiming real activation is complete.

This PR does not close #34. It adds the fail-closed real-path surface that #34 needs before the actual activation injection implementation and smoke proof can land.

What changed

  • Adds optional runtime dependency list in requirements-steering.txt.
  • Adds src/agent_machine/steering_runtime.py.
  • Adds sourceset-aware readiness preflight:
    • agent-machine steer preflight --sourceset gpt2-small.res-jb
  • Adds sourceset-aware local endpoint entrypoint:
    • agent-machine steer serve --sourceset gpt2-small.res-jb --host 127.0.0.1 --port 8080
  • Keeps agent-machine steer serve-stub intact as the no-weights fallback.
  • Updates make validate to exercise steer preflight through Python and bootstrap CLI paths.
  • Adds docs/steering-activation-path.md.
  • Links the activation-path doc from docs/index.md.

Boundary discipline

This PR does not:

The new steer serve --sourceset path remains fail-closed and returns status: not_configured until runtime dependencies, verified artifacts, storage receipts, policy/grant admission, activation injection, and local smoke evidence are present.

Validation

Native CI passed on this PR head:

The validate path exercised steer preflight --sourceset gpt2-small.res-jb through both Python and bootstrap CLI paths.

@mdheller mdheller merged commit 71e4f77 into main May 20, 2026
2 checks passed
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.

Implement controlled activation for local steering inference provider

1 participant