Repository navigation
Add vdt guide and vdt explain, and slim the agent skill - #28
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
@codex review |
|
Codex Review: Didn't find any major issues. You're on a roll. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Usage guidance now ships in the binary, so it always matches the installed version and any agent that can run a shell command can read it. - `vdt guide [topic]` prints the workflow and pages for adaptive, notification, drawables, diagnostics, and ci. - `vdt explain <code>` explains a diagnostic code in depth; `vdt explain` lists all 26. `vdt --help` points to both. - skills/vdt/SKILL.md shrinks from 226 lines to an entry point that sends the agent to `vdt guide` and keeps the rules about artwork and resources. - tests/skill.rs now also fails when a guide page names a command, flag, or topic vdt does not have, or a diagnostic code has no explanation. The old skill told agents to ask before deleting template launcher bitmaps, but `adaptive --legacy` replaces them itself; the guide describes what it actually does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/eval-guide sets up one Android project per agent and condition (long skill with the baseline binary, thin skill with vdt guide, or no skill), logs every vdt command, and grades the result. Four agents scored a mean of 14.9/16 with the long skill, 16.0 with the thin skill and guide, and 15.75 with no skill. Reports and command logs from that run are in results/. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
a743cb6 to
96d9d30
Compare
|
@codex review |
|
Codex Review: Didn't find any major issues. Breezy! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Moves the agent guidance out of a 226-line skill and into the binary, following cav's
guideconcept. Guidance built into the binary always matches the installed version, and any agent that can run a shell command can read it.Changes
vdt guide [topic]prints the workflow (inspect, fix, generate, wire up) and pages foradaptive,notification,drawables,diagnostics, andci. The source is markdown insrc/guide/, compiled in withinclude_str!.vdt explain <code>explains a diagnostic code such asVDT019in depth. With no code, it lists all 26. It acceptsVDT019,vdt019, or19.vdt --helpnow ends with a pointer to both.skills/vdt/SKILL.mdis now about 40 lines: install check, "runvdt guide", and the agent rules (inspect first, ask before changing artwork, never hand-edit output).tests/skill.rsnow fails when a guide page or the skill names a command, flag, orvdt guidetopic that doesn't exist, or when aDiagnosticCodehas noexplainentry.tools/eval-guide/is the eval below: a fixture project,setup.sh,grade.py, and the results.Eval
Four agents each ran under three conditions. The task was a launcher icon, a notification icon, and an illustration that can't be converted without changing the artwork. Scores are out of 16.
vdt guidethrough the--helphint and read it.vdtcalls (18–20 vs 11.5 on average). Token use wasn't measured.Found by the eval
The old skill told agents to ask before deleting template launcher bitmaps. In fact
adaptive --legacyreplaces them itself, on purpose, so they can't outrank the legacy icon on API 21–25. The guide now describes what actually happens and the--nameway to keep them. Separately, a logo on an opaque tile becomes a solid shape as a themed icon. The guide now says so; a vdt warning for it would be a worthwhile follow-up.Testing
cargo fmt --check,cargo clippy --all-targets -- -D warnings, andcargo testall pass.🤖 Generated with Claude Code