8000 chore: migrate io-auto-mode classifier defaults to gemini-3.1-flash-lite by simon-inkie · Pull Request #1 · simon-inkie/inkie-auto-mode · GitHub
[go: up one dir, main page]

Skip to content

chore: migrate io-auto-mode classifier defaults to gemini-3.1-flash-lite - #1

Merged
simon-inkie merged 2 commits into
masterfrom
chore/gemini-3.1-flash-lite-sweep
Jul 9, 2026
Merged

simon-inkie merged 2 commits into
masterfrom
chore/gemini-3.1-flash-lite-sweep

Conversation

@simon-inkie
Copy link
Copy Markdown
Owner

Migrates all four classifier stage slots in DEFAULT_CONFIG (plus the matching README/INSTALL default tables, example config, and prose) from google/gemini-2.5-flash to google/gemini-3.1-flash-lite.

What changed

  • stage1Model, stage1Fallback, stage2Model, stage2Fallback -> google/gemini-3.1-flash-lite
  • README and INSTALL default tables + OpenClaw example config + requirements prose updated to match

Verification (both stages probe-verified against the live model-call.ts config, no thinkingConfig, GEMINI_API_KEY from env)

  • Stage 1 at its real call shape (maxOutputTokens 8, one-word prompt): flash-lite returns finishReason=STOP, non-empty, parseable ALLOW/BLOCK with correct decisions. The incumbent gemini-2.5-flash returns EMPTY at maxTokens 8 (finishReason=MAX_TOKENS, thinking consumes the budget), so this also fixes a latent empty-output failure on the classify path.
  • Stage 2 at its real call shape (maxOutputTokens 2048, JSON reasoning prompt, escalated-block inputs): flash-lite returns finishReason=STOP, non-empty, valid parseable JSON with a legal ALLOW/ASK/BLOCK decision and a populated thinking field, so the prompted chain-of-thought reasoning is retained. No MAX_TOKENS truncation.

Blast radius

Inert until config cutover: the live ~/.io-auto-mode/config.json hardcodes all four stage models and runs yolo mode, so this default change takes effect only when that config is updated.

Test suite: 211/211 pass (32 suites).

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com

Io and others added 2 commits July 9, 2026 22:35
Switch DEFAULT_CONFIG stage1Model and stage1Fallback from
google/gemini-2.5-flash to google/gemini-3.1-flash-lite, and update the
README and INSTALL default tables/prose to match.

Stage 2 (stage2Model / stage2Fallback) is left on gemini-2.5-flash: it is
the deliberate thinking classifier for escalated blocks, and flash-lite is
text-optimised. Not switched here.

Verified via a probe replicating the live model-call.ts config (no
thinkingConfig) at Stage 1's maxOutputTokens of 8: flash-lite returns a
clean one-word decision (finishReason STOP, non-empty, parseable ALLOW/
BLOCK). The incumbent gemini-2.5-flash returns empty at maxTokens 8
(finishReason MAX_TOKENS, thoughtsTokenCount consumes the budget), so this
also removes a latent empty-output failure mode on the classify path.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Simon's decision, Stage 2 now also defaults to gemini-3.1-flash-lite
(stage2Model + stage2Fallback in DEFAULT_CONFIG), completing the sweep so
all four stage slots use flash-lite. README and INSTALL default tables,
example config, and prose updated to match.

Verified against the live model-call.ts config (no thinkingConfig) at
Stage 2's real call shape: maxOutputTokens 2048, the JSON reasoning prompt
(STAGE2_SUFFIX), and escalated-block style inputs. flash-lite returns clean
non-empty output on primary and fallback prompts: finishReason STOP,
candidatesTokenCount ~115-148 (well under budget), parseable JSON with a
valid ALLOW/ASK/BLOCK decision and a populated thinking field, so the
prompted chain-of-thought reasoning is retained. No empty-output or
MAX_TOKENS truncation (the Sentinel failure mode) observed.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@simon-inkie
simon-inkie merged commit 6c6c658 into master Jul 9, 2026
1 check passed
@simon-inkie
simon-inkie deleted the chore/gemini-3.1-flash-lite-sweep branch July 9, 2026 22:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

0