Releases · dariomonopoli-dev/flash-3.5-action

First public release. Built the day Gemini 3.5 Flash dropped.

What's in v0.1.0

GitHub Action that runs Gemini 3.5 Flash on your repo from a single workflow step.
Twelve inputs covering prompt, context, output, model, cost caps, skills, and AGENTS.md.
Five example workflows: PR security review, issue triage, auto-label by content, changelog from PRs, weekly dependency audit.
`thinking-budget` input that exposes the Gemini 2.5+/3.x thinking-phase cap. `0` for predictable cost, `-1` lets the model decide, positive integer caps thinking at that many tokens.
Skills loading from `.agent/skills/` via SKILL.md frontmatter. Agent requests skills on demand with a self-closing `<skill name="..."/>` tag.
AGENTS.md prepended to the system prompt so the same conventions file works across Claude Code, Codex, Antigravity 2.0, and this action.
Honest cost guard that reads `promptTokenCount`, `responseTokenCount`, and `thoughtsTokenCount` directly from the Gemini API. Reported `tokens-used` matches `totalTokenCount` from Google exactly.
Retry on transient failures (429, 503, network resets) with exponential backoff.
`scripts/verify-live.mjs` for validating model + thinking config against your own key.
38 unit tests and a smoke-test job in CI that runs the action end-to-end against the real Gemini API.

Same 25-token prompt against `gemini-3.5-flash`:

Mode	Latency	Prompt	Response	Thinking	Total billed
`thinking-budget` unset	2.16s	25	40	295	360
`thinking-budget: 0`	0.97s	25	38	0	63

Disabling thinking made this task ~2.2× faster and ~5.7× cheaper.

One-line change:

```diff

Input names differ (`gemini_api_key` → `api-key`, `gemini_model` → `model`). See docs/migration-from-gemini-cli.md for the full mapping.

```yaml
uses: dariomonopoli-dev/flash-3.5-action@v0.1.0 # exact
uses: dariomonopoli-dev/flash-3.5-action@v0.1 # latest 0.1.x
```