Skip to content

Make local Ollama models work out of the box, and stop a retired model id from disabling your LLM - #234

Merged
handstandsam merged 1 commit into
mainfrom
upstream-2026.08.10-2
Aug 11, 2026
Merged

Make local Ollama models work out of the box, and stop a retired model id from disabling your LLM#234
handstandsam merged 1 commit into
mainfrom
upstream-2026.08.10-2

Conversation

@handstandsam

Copy link
Copy Markdown
Collaborator

Summary

Syncs the latest upstream changes through 2026.08.10.

Local Ollama models now work without per-machine setup. Trailblaze requests a 64K context window (num_ctx: 65536) on every Ollama call, clamped down to the model's declared context_length when that is lower. Previously Trailblaze sent no num_ctx at all, so Ollama sized the window to whatever memory it happened to have — large on a workstation, very small on a laptop — and a single agent turn (~20K tokens on a content-heavy screen) failed with exceed_context_size_error. Override with TRAILBLAZE_OLLAMA_NUM_CTX.

A retired model id no longer silently disables your LLM. If you had a model selected that a later release dropped from the catalog, the lookup missed and fell through to a global default that is NONE in the open-source build — so Trailblaze quietly stopped calling an LLM. A saved id the provider no longer offers now resolves to that provider's own default_model instead.

gemma-4-12b-it removed from the Google catalog. Google's Gemini API serves exactly two Gemma 4 variants, gemma-4-31b-it and gemma-4-26b-a4b-it (supported models). The 12B is download-only, so selecting it under the google provider returned a 404 rather than a model. It is still available where it actually runs — ollama.yaml ships gemma4:12b.

Two new recorded tools: rememberTextBySelector and rememberNumberBySelector capture a value off the screen into agent memory by selector.

MCP tool surface: a stale target-scoped tool name that collides with a live host tool name no longer un-registers the host tool during a re-registration pass.

Docs corrections to adding_a_model.md and configuration.md: declaring an llm.providers block replaces the built-in catalog rather than adding to it, Ollama models are not auto-discovered at runtime, llm.defaults.model is read by the Android on-device runner rather than the desktop app or CLI, and workspace target auto-discovery scans trails/config/trailmaps/.

Also included: tapRoute override for recorded taps, recording tool rewrites, AXe clearState mode, and report/log viewer updates.

Note on #233

#233's Playwright driver-download hardening landed directly in this repo and was never mirrored back, so a naive sync would have reverted it. It has been backported to the mirror source first — this PR leaves PlaywrightDriverManager.kt untouched.

@handstandsam
handstandsam merged commit 4eb20b6 into main Aug 11, 2026
10 of 12 checks passed
@handstandsam
handstandsam deleted the upstream-2026.08.10-2 branch August 11, 2026 00:41
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