Skip to content

fix: read Ollama context length from flat model_info keys - #1622

Merged
avoidwork merged 3 commits into
mainfrom
fix/ollama-context-probe-fix
Nov 1, 2026
Merged

fix: read Ollama context length from flat model_info keys#1622
avoidwork merged 3 commits into
mainfrom
fix/ollama-context-probe-fix

Conversation

@avoidwork

@avoidwork avoidwork commented Nov 1, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

Description

The Ollama /api/show probe in src/provider/modelInfo.js iterated Object.values(model_info) looking for nested family.context_length, but Ollama returns model_info as a flat object of dotted keys (e.g. "gemma4.context_length": 131072). The branch was dead, so the probe returned undefined for stock Ollama models and the summarization trigger fell back to the 128000-token default — wrong for small local models.

This change rewrites the Ollama branch to scan Object.entries(model_info) for keys ending in .effective_context_length (the runtime cap, min(native, OLLAMA_CONTEXT_LENGTH), highest precedence), then .context_length (native), preserving the num_ctx fallback. It corrects the masking unit test to the flat shape and adds precedence and neither-field cases. The change is delivered as an OpenSpec change (ollama-context-probe-fix), now archived with its main-spec sync.

Type of Change

  • Bugfix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to not work as expected)
  • Documentation update
  • Refactor (no functional changes)
  • Performance improvement
  • CI / build / tooling

Testing

Corrected the Ollama /api/show unit test to the flat model_info shape and added: a precedence case (both effective_context_length and context_length present → effective wins), and a neither-field case asserting undefined. All 26 modelInfo unit tests pass. modelInfo.js coverage: 100% line / 97% branch / 100% func. (PDF/Chrome and MCP-server integration tests fail in this environment for unrelated reasons — no Chrome binary, no live MCP server.)

Coverage

  • Line coverage maintained

Checklist

  • npm run lint passes
  • Tests pass with maintained line coverage
  • No forbidden patterns used
  • Conventional Commit style applied

Closes #1621

@avoidwork
chore: add OpenSpec change for Ollama context-window probe fix
a8bae30 
Add proposal, design, spec delta, and tasks for the ollama-context-probe-fix
change. The Ollama /api/show probe never matches the real flat model_info shape,
so it returns undefined for stock models and the summarization trigger falls back
to the 128000 default.
@avoidwork avoidwork added the bug Something isn't working label Nov 1, 2026
@avoidwork avoidwork self-assigned this Nov 1, 2026
@avoidwork
fix: read Ollama context length from flat dotted model_info keys
ae7d8ee 
The Ollama /api/show endpoint returns model_info as a flat object of
dotted keys (e.g. "gemma4.context_length": 131072), not nested family
objects. The previous branch iterated Object.values checking
family.context_length, which never matched, so the probe always fell
through to the 128000-token default and summarization triggered at
102400 tokens — wrong for small local models.

Scan Object.entries for keys ending in .effective_context_length
(runtime cap, min(native, OLLAMA_CONTEXT_LENGTH)) with highest
precedence, then .context_length (native), preserving the num_ctx
fallback. Correct the unit test to the flat shape and add precedence
and neither-field cases.
@avoidwork
chore: archive ollama-context-probe-fix change and sync main spec
0d32990 
Sync the two ADDED requirements (flat model_info parsing,
effective_context_length precedence) into the main
model-context-length spec, then archive the completed change.
@avoidwork avoidwork changed the title chore: add OpenSpec change for Ollama context-window probe fix fix: read Ollama context length from flat model_info keys Nov 1, 2026
@avoidwork

avoidwork commented Nov 1, 2026

Copy link
Copy Markdown
Owner Author

Implementation complete

The Ollama context-window probe fix is implemented, tested, and archived.

Root cause: src/provider/modelInfo.js iterated Object.values(model_info) looking for nested family.context_length, but Ollama's /api/show returns model_info as a flat object of dotted keys (e.g. "gemma4.context_length": 131072). The branch was dead, so the probe returned undefined for stock Ollama models and the summarization trigger fell back to the 128000-token default — wrong for small local models.

Fix: Scan Object.entries(model_info) for keys ending in .effective_context_length (runtime cap, min(native, OLLAMA_CONTEXT_LENGTH), highest precedence), then .context_length (native), preserving the num_ctx fallback.

Tests: Corrected the masking unit test to the flat shape; added a precedence case (effective wins over native) and a neither-field case. All 26 modelInfo unit tests pass. modelInfo.js coverage: 100% line / 97% branch / 100% func.

Note: PDF/Chrome and MCP-server integration tests fail in this environment for unrelated reasons (no Chrome binary, no live MCP server) — not regressions from this change.

Closes #1621

@avoidwork
avoidwork merged commit c1312a6 into main Nov 1, 2026
2 checks passed
@avoidwork
avoidwork deleted the fix/ollama-context-probe-fix branch November 1, 2026 01:17
@avoidwork avoidwork mentioned this pull request Nov 1, 2026
Merged
6 of 12 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@avoidwork avoidwork

Labels

bug Something isn't working

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

fix: Ollama context-window probe never matches the real /api/show shape

1 participant

@avoidwork