Skip to content

Update egm-lite skill for automatic pre-implementation spec review - #446

Closed
eewing314 wants to merge 2 commits into
google:feat/egm-lite-corefrom
eewing314:feat/egm-lite-gate-alignment
Closed

eewing314 wants to merge 2 commits into
google:feat/egm-lite-corefrom
eewing314:feat/egm-lite-gate-alignment

Conversation

@eewing314

@eewing314 eewing314 commented Nov 15, 2026 •

Copy link
Copy Markdown
Collaborator

Summary of Key Changes

Updates the /egm-lite skill implementation (skills/egm-lite/), test suite, and Starlight documentation (docs/site/skills/egm-lite.md) in #438 (implementing the updated design in #445 / #419) so pre-implementation Goldfish specification review is automatic, post-implementation Goldfish diff re-review is optional, and Goldfish is reserved strictly for isolated fresh-context subagents across a 6-phase state machine:

  • Skill Protocol (skills/egm-lite/SKILL.md):
    • Assigns Section 1 (The Critique Pause) to The Elephant (Critic) (in-session deep-context critique) and flattens checklist generation and spec review into Section 2 (Execution Checklist — Elephant Planner) chaining automatically in the same turn into Section 3 (Automatic Specification Review — Goldfish Specification Reviewer via Subagent) (falling back to human approval without simulating a same-context review if the host exposes no subagent tool).
    • Adds State Transitions & Command Aliases and [State Advisory] guards preventing premature exec or diff review before checklist generation and specification review.
    • Records the starting review baseline in Section 4 (Exec Protocol) before the first task edit, finishes cleanly on passing tests while presenting post-implementation diff re-review (review / goldfish) as optional, and updates Section 5 (Optional Adversarial Diff Re-Review) to collect committed task changes, unstaged edits, and untracked files against the recorded baseline while keeping unrelated pre-existing edits separate.
    • Updates Section 6 (Unified Post-Review Resolution Gate) to discriminate pre-implementation spec reviews (stamping Verified [✅ LGTM] and gating on exec / Proceed without premature task closure) from post-implementation diff re-reviews.
  • Verification & Site Docs (skills/egm-lite/TEST.md, skills/egm-lite/test_scenarios.yaml, docs/site/skills/egm-lite.md):
    • Updates the interactive protocol checklist, Starlight skill page, and 19 declarative evaluation scenarios to the 6-phase model covering Elephant Critique Pause, automatic pre-implementation spec review, [State Advisory] guards, spec LGTM vs. findings triage, optional post-implementation re-review, and baseline-aware untracked/committed change collection.

Contribution Type

Track: Feature / Architectural Change -- Stage 2: Implementation PR

Verification

  • pytest tests/unit/test_skill_load_limits.py tests/unit/test_frontmatter.py tests/unit/test_link_integrity.py tests/unit/test_markdown_tables.py tests/unit/test_packaging.py tests/unit/test_harnesses.py tests/unit/test_opencode_adapter.py tests/unit/test_scenario_coverage.py — passed (skills/egm-lite/SKILL.md is 20,186 bytes total, within Antigravity's 24,000-byte skill body limit).
  • Validated all 19 scenario regex patterns in skills/egm-lite/test_scenarios.yaml.
  • Node-dependent workflow/site checks and live model behavioral probes (EGM_CLAUDE_LIVE, EGM_CODEX_LIVE) — not run locally (covered by GitHub Actions CI for offline site/workflow checks).

Human Accountability Affirmation

@eewing314 eewing314 added docs Improvements or additions to documentation site Changes to the project website labels Nov 15, 2026
@eewing314

Copy link
Copy Markdown
Collaborator Author

Closing in favor of a single chained PR stacked on #459.

@eewing314 eewing314 closed this Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

docs Improvements or additions to documentation site Changes to the project website

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant