fix(router): Sync resolved model's context window onto the Auto session - #1137
Conversation
After the router resolves a concrete model for a session (or subagent delegation), advertise the routed model's contextWindow and maxTokens on the active kimchi-dev/auto model. This makes Pi's context-usage accounting (getContextUsage) reflect the real window (e.g. deepseek ~1M) instead of the fixed pre-routing <=128k floor, so a fresh deepseek turn shows ~4% used rather than ~34%. - autoModelForTarget now copies contextWindow/maxTokens from the resolved target - hasTargetCapabilities compares all four fields to skip redundant setModel - syncAutoReasoningCapabilities -> syncAutoCapabilities (renamed for accuracy) - regression tests for the fresh-routing and session-restore sync paths Co-Authored-By: Kimchi <noreply@kimchi.dev>
|
📚 No documentation changes were needed for this PR. |
Kimchi Code Review
Summary📊 Review Score: 78/100 (overall code quality — 0 lowest, 100 highest) 🧪 Tests: yes — Two new tests cover context window and max token syncing for routed and restored sessions. However, the restore test's expectation of a single 📝 Found 1 issue(s). See inline comments for details. What to expectKimchi will analyze the changes in this pull request and post:
The review typically completes within a few minutes. This comment will be updated once the review is ready. Interact with Kimchi
ConfigurationReviews are configured by your organization admin. Powered by Kimchi — AI-powered code review by CAST AI |
![kimchi-review[bot]](https://avatars.githubusercontent.com/in/3223788?s=60&v=4){kind=small}
There was a problem hiding this comment.
📊 Review Score: 78/100 (overall code quality — 0 lowest, 100 highest)
⏱️ Estimated effort to review: 2/5 (1 = trivial, 5 = very complex)
🧪 Tests: yes — Two new tests cover context window and max token syncing for routed and restored sessions. However, the restore test's expectation of a single setModel call appears inconsistent with the visible handler logic, which may call setModel twice.
📝 Found 1 issue(s). See inline comments for details.
…target Addresses PR #1137 review comment: the set_model tool and the model_select handler validated the context and vision guards against the raw target/event.model descriptor, which for Auto reports Auto's 128K catalog floor instead of the resolved concrete model. After Auto routed to a 1M target, a ~200K conversation would fit the real window but be rejected (and model_select would revert the switch) purely because Auto's descriptor stuck at 128K. - set_model: resolve the effective target (resolveEffectiveModel) for the context and vision guards; keep pi.setModel(target) so Auto stays selected. - model_select: resolve the effective target for the context and vision guards and the overflow-message numbers; event.model still drives revert/compact/ new-session identity. - When Auto is unresolved, resolveEffectiveModel falls back to Auto's descriptor, so switching into an unresolved Auto with an oversized context is still conservatively rejected. Regression test: ~200K context, resolved 1M Auto target, Auto descriptor at 128K -> the Auto switch is not reverted. Co-Authored-By: Kimchi <noreply@kimchi.dev>
2 participants
Linked issue
Closes #0
What does this PR do?
After the router resolves a concrete model for a session (or subagent delegation), advertise the routed model's contextWindow and maxTokens on the active kimchi-dev/auto model. This makes Pi's context-usage accounting (getContextUsage) reflect the real window (e.g. deepseek ~1M) instead of the fixed pre-routing <=128k floor, so a fresh deepseek turn shows ~4% used rather than ~34%.
Checklist
pnpm run test)pnpm run check)