Skip to content

feat: auto hill-climbing - #189

Open
wangauone wants to merge 21 commits into
mainfrom
feat/decouple-phase-skills
Open

feat: auto hill-climbing#189
wangauone wants to merge 21 commits into
mainfrom
feat/decouple-phase-skills

Conversation

@wangauone

@wangauone wangauone commented Jul 15, 2026

Copy link
Copy Markdown
Collaborator

Summary

Introduces autonomous hill-climbing and decouples the context-engineering lifecycle into standalone phase skills. Each phase (init, bootstrap, evaluate, hill-climb) now runs on its own and accepts a context set from a local file or a remote Context Store resource — no need to run the whole pipeline in order.

Key changes

  • Autonomous hill-climbing: loops evaluate → analyze → mutate → re-upload until convergence with no per-iteration approval; resumable via state.md. The agent uploads/mutates context sets directly through MCP tools (no manual browser step).
  • Decoupled phase skills: rewrote init, bootstrap, evaluate, hillclimb, generation-guide, and workflow (now a router/guide) so phases are independent with flexible local/remote inputs.
  • Workspace convention: autoctx/.context-engineering/ (holds tools.yaml and experiments/).
  • Context Store + QueryData → prod: switched to prod endpoints; quota project falls back to the ADC default.
  • Evals overhaul: restructured core-cujs fixtures/configs; bumped evalbench to 1.10.0.
  • CI hardening: 45-min step timeouts; pinned GitHub Actions to SHAs and fixed template-injection findings (zizmor) in release.yml.

@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

@gemini-code-assist gemini-code-assist Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request restructures the context engineering framework by renaming skills, updating the configuration path for tools.yaml to .context-engineering/tools.yaml, exposing context upload/download MCP tools, and adding a CI output validation script. It also introduces new evaluation datasets and workspaces for testing. The review feedback highlights a critical foreign key constraint mismatch in the petstore test database setup, along with recommendations to address deprecated SQLAlchemy and Python datetime methods in the Flask application. Additionally, the feedback suggests specifying file encoding and avoiding variable shadowing in the CI script, as well as correcting minor typos in the documentation.

Comment thread .ci/check_eval_outputs.py
Comment thread .ci/check_eval_outputs.py
Comment thread evals/eval-dataset-gen/README.md
Comment thread evals/eval-dataset-gen/README.md
Comment thread evals/eval-dataset-gen/README.md
Comment thread evals/eval-dataset-gen/README.md
@wangauone
wangauone force-pushed the feat/decouple-phase-skills branch from 0fcc336 to 64ad2dd Compare July 19, 2026 23:48
@wangauone
wangauone force-pushed the feat/decouple-phase-skills branch from 64ad2dd to 1cf64e3 Compare July 20, 2026 00:35
@wangauone wangauone changed the title feat: golden dataset generation and expansion feat: auto hill-climbing Jul 20, 2026
@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

2 similar comments
@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

…en release.yml (pin actions, fix template-injection)
@wangauone
wangauone force-pushed the feat/decouple-phase-skills branch from b8503d0 to 6545c4e Compare July 24, 2026 20:50
@wangauone
wangauone requested review from g-lynnzee and helloeve July 24, 2026 21:21
@wangauone

Copy link
Copy Markdown
Collaborator Author

/gcbrun

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

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant