fix(cli): model-catalog cache — key by exact API key, shorten TTL, skip Ollama

[joaomdmoura]

Follow-ups to the model-catalog cache from #6462. Two correctness bugs + a freshness tweak.

1. Cache key didn't distinguish which API key

The catalog cache key only encoded key-present vs absent (openai#key / openai#nokey). After a live fetch was cached, swapping to a different key for the same provider still hit the old entry and showed the previous account's model list (for up to the TTL).

Fix: key the cache by the exact API key via a short, non-reversible sha256 digest — openai#<digest> (never the raw key). A different key → different entry → refetch. Absent key stays #nokey.

2. Local (Ollama) lists were cached

Successful Ollama lists were written with the full dynamic TTL, and the cache was read before hitting /api/tags. The key only reflected the base URL, not installed models — so after you deleted a model locally, the picker could keep offering it until the entry expired.

Fix: never cache local providers. _is_cacheable() gates both cache read and write (Ollama excluded); /api/tags is fast, so the picker re-probes every call and always reflects what's installed.

3. Dynamic TTL shortened 6h → 5m

A stale list (new/removed models, account changes) is worse than a ~1s refetch; the cache only needs to spare repeated fetches within a wizard session.

Tests

• distinct-key → separate cache entries (switching keys refetches)
• cache key is a digest and never contains the raw key
• Ollama is not cached: reflects a deletion between calls; never written to the cache file
• dynamic entry older than the TTL is not served

ruff format --check + ruff check + mypy clean; test_model_catalog.py 39 passed.

Note

These are the two items you flagged from #6462. The sha256-digest approach keeps the raw API key out of the on-disk cache (~/.crewai/model_catalog_cache.json) — a truncated digest is non-reversible, so no secret is persisted.

🤖 Generated with Claude Code

---

Note

Low Risk
CLI-only wizard model picker caching; changes improve correctness and freshness without touching runtime crew execution or secrets beyond hashed cache keys.

Overview
Fixes model catalog caching in the crew-creation wizard so lists stay correct per account and for local Ollama.

Cache keys now use a short SHA-256 digest of the API key (provider#<digest>), not just key-present vs #nokey. Switching keys for the same provider refetches instead of reusing another account’s cached models. The raw key is never written to ~/.crewai/model_catalog_cache.json.

Ollama is excluded via _is_cacheable(): no read or write of the catalog cache, so every picker call hits /api/tags and reflects install/uninstall changes immediately.

Dynamic catalog TTL drops from 6h to 5m so stale vendor lists don’t linger; negative fallback TTL stays 5m.

Tests cover per-key cache separation, digest-only keys, Ollama non-caching, and expired dynamic entries.

^{Reviewed by Cursor Bugbot for commit 0690a7f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 0ea3c42d-b8a2-44bc-9dca-704a11aeb276

📥 Commits

Reviewing files that changed from the base of the PR and between 799ab0f and 0690a7f.

📒 Files selected for processing (2)

• lib/cli/src/crewai_cli/model_catalog.py
• lib/cli/tests/test_model_catalog.py

---

📝 Walkthrough

Walkthrough

The dynamic model catalog's TTL was reduced to 300 seconds, and cache keying was reworked to hash provider API keys (truncated SHA-256) instead of special-casing Ollama. A new _is_cacheable helper gates cache reads/writes, and tests were updated to cover Ollama exclusion, key isolation, digest-only keys, and TTL expiry.

Changes

Model Catalog Cache Rework

Layer / File(s)	Summary
TTL constant and hashing setup lib/cli/src/crewai_cli/model_catalog.py	Adds hashlib import and reduces _CATALOG_TTL from 6 hours to 300 seconds.
Cacheability gating and hashed cache key lib/cli/src/crewai_cli/model_catalog.py	Removes the inline ollama check in get_provider_models, adds _is_cacheable, rewrites _cache_key to hash API keys, and updates _read_catalog_cache/_write_catalog_cache to early-return for non-cacheable providers.
Cache behavior tests lib/cli/tests/test_model_catalog.py	Adds tests for Ollama not being served stale or written to cache, per-API-key cache isolation, digest-only cache keys, and TTL-based expiry of stale entries.

Suggested reviewers: lorenzejay

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the cache key, TTL, and Ollama caching changes.
Description check	✅ Passed	The description clearly matches the cache key, Ollama, and TTL updates in the changeset.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch joaomdmoura/model-catalog-cache-fixes

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}