技能 2026-10-10 00:45 UTC
agent-skillsv2.5
版本【2.5】由提交人 copybara-service 提交,主要更新类型为:功能特性新增与扩展。提交说明:[skills] Make the Firebase AI Logic skill platform-independent. (#225)。为项目引入了新的能力与模块支持。
Commit提交信息原文
[skills] Make the Firebase AI Logic skill platform-independent. (#225)
`SKILL.md` mixed Web SDK specifics (the `npm install` step, Chrome hybrid inference, reCAPTCHA Enterprise) into the shared setup flow, and its description mentioned only web apps even though the skill also covers Android, iOS, and Flutter.
`SKILL.md` now keeps only cross-platform guidance: provisioning the backend with the Firebase CLI via `npx`, which every platform needs, followed by a pointer to the platform reference for adding the SDK. The Web specifics move to the web reference, and the iOS reference now routes SDK installation to the `xcode-project-setup` skill. App Check debug-provider instructions, previously inline for Web, Android, and iOS only, now live in each platform reference; the iOS and new Flutter sections cover the required dependency, call order, debug-only gating, and CI token setup.
The same split applies to content from the latest GitHub import, cl/996819095. Its three per-platform debug token churn tips in `SKILL.md` become one platform-neutral tip, and their details merge into the debug token persistence steps of the Android, iOS, and Web references, which now sit in each reference's App Check section. Its new Text-to-Speech section in `SKILL.md` drops the Swift and JavaScript syntax and the platform audio APIs, which `references/tts_generation.md` already covers.
The frontmatter description now lists all supported platforms and says when to use the skill. The Node.js minimum moves from 16 to 20 to match `firebase-basics`, and the hybrid and App Check links point to the platform-neutral AI Logic docs.
Agents no longer install Node.js themselves; they ask the user to install it as `firebase-basics` describes. In an earlier eval run, an agent following the old "Install them if they aren't already available" line ran `sudo apt-get install nodejs npm` on the machine running the eval. The provisioning warning now also covers apps that already use Firebase, and agents that can't run `init ailogic` tell the user to run it, since leaving it out was the most common eval failure.
Evals after syncing past the import, baseline (`local_base`) vs. this CL, 3 runs per case with all 12 Firebase skills loaded:
- `EVAL.txtpb`: 5/18 → 4/18 passed,
https://evalin.corp.google.com/compare/5d831db67f0142cb/ff8e0caaf5754ce6
- `activationEval.txtpb`: 63/63 → 63/63 passed,
https://evalin.corp.google.com/compare/9c60d402ec5e4e9a/292460151a314091
Most of the `EVAL.txtpb` failures come from the eval setup, not the skill. In both arms, all nine runs of the three Flutter cases failed before the agent started, because `eval_fixtures/setup_flutter.sh` from cl/996169567 has a Python syntax error; pending cl/996505101 fixes it. All `integrate_ai_logic_complex` runs in both arms hit the case's 300-second limit, one CL run after meeting every expectation. The only difference is `basic_text_generation`, 2/3 → 1/3, and every failure of that case in both arms is the agent choosing `gemini-2.5-flash`.
Before the import, when the Flutter cases still ran, `EVAL.txtpb` results varied a lot between runs: across the four base-vs-CL runs on snapshots 3 to 6, the baseline passed 9 to 13 of 18 cases and this CL 12 to 15. The snapshot 6 run was 13/18 → 13/18: https://evalin.corp.google.com/compare/d7c79edc6c054110/6be69aba9a9341a2. The reviewer edits in snapshot 5 (numbered provisioning steps, section rules, and heading changes) scored the same as snapshot 4 in a head-to-head run, 13/18 → 13/18 with identical per-case results: https://evalin.corp.google.com/compare/299f1ff5d5ba4ac6/03d9c6874b3a406e.
PiperOrigin-RevId: 996079357
Co-authored-by: Copybara <no-reply@google.com>