Pi: emit "input": ["text", "image"] for vision-capable models

[davanstrien]

Summary

• For models with pipeline_tag === "image-text-to-text", the Pi "Use this model" snippet now writes "input": ["text", "image"] into the generated ~/.pi/agent/models.json. Without this field, Pi can't pass images to the model — so users had to discover the field manually from the docs to get vision working.
• Behavior for text-only models is unchanged (existing test extended; new pi - vision test added).
• Cross-ref: Add vision support note to Pi + llama.cpp guide hub-docs#2408 — the documentation PR that explains the same field. @gary149's review there suggested surfacing it directly in the snippet.

Test plan

• pnpm test in packages/tasks (17 tests pass, including new pi - vision)
• pnpm check, pnpm lint:check, pnpm format:check all clean
• Verified generated JSON output via local script: text models produce identical output to before; vision models gain the input field correctly nested

🤖 Generated with Claude Code

---

Note

Low Risk
Low risk: only adjusts Pi snippet JSON generation for models flagged as vision-capable and adds a focused unit test; no auth, persistence, or runtime execution paths change beyond emitted config text.

Overview
Updates the Pi “Use this model” snippet to include "input": ["text", "image"] in the generated models.json entry when the selected model has pipeline_tag === "image-text-to-text", enabling image inputs for vision-capable models.

Adds a new pi - vision test to assert the input field is emitted (and preserves existing behavior for text-only and MLX-backed Pi snippets).

^{Reviewed by Cursor Bugbot for commit bf73c01. Bugbot is set up for automated code reviews on this repo. Configure here.}

[julien-c]

lgtm

[cursor]

Missing --jinja flag for vision models in server command

High Severity

The test at line 152 asserts that the llama-server command for vision models includes --jinja, but the implementation on line 481 of local-apps.ts never conditionally appends this flag. The isVision variable is computed but only used for the modelEntry.input field — it's never used to modify the server command. This means either the test will fail, or vision models won't get the --jinja flag they need for proper Jinja template handling with llama.cpp.

Additional Locations (1)

• packages/tasks/src/local-apps.spec.ts#L151-L152

 

^{Reviewed by Cursor Bugbot for commit 671b44f. Configure here.}

[davanstrien]

@cursoragent IIUC jinja is default in llamacpp now?

[cursor]

Unable to authenticate your request. Please make sure to connect your GitHub account to Cursor. Go to Cursor

[julien-c]

yes

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bf73c01. Configure here.}

[cursor]

Test expects --jinja flag that implementation never produces

High Severity

The new vision test asserts that snippet[0].content contains --jinja in the llama-server command, but the snippetPi implementation at line 481 builds the content as `llama-server -hf ${model.id}${getQuantTag(filepath)}` with no --jinja flag appended — and no other code path adds it. The --jinja string appears nowhere in local-apps.ts. This test will fail at runtime. Given the PR discussion confirming jinja is now default in llama.cpp, the test expectation is likely stale and the assertion needs to drop --jinja.

Additional Locations (1)

• packages/tasks/src/local-apps.ts#L480-L481

 

^{Reviewed by Cursor Bugbot for commit bf73c01. Configure here.}