[fix] A cross-modality model switch keeps the warm sandbox

[mmabrouk]

Context

Cold/warm audit finding 2. Even with the router keying fixed (#6364), switching between a vision model and a text-only model still destroyed the warm sandbox. The request field modelCapabilities (the resolved model's input modalities) sits in the session fingerprint and the harnessSession facet, and it changes with the model. So a cross-modality switch moved two facets, the mixed plan refused the live setModel route, and the runner rebuilt: fresh Daytona Secrets, the substitution race of #6362, about 15 seconds. The Pi catalog splits 249 text-and-image models against 163 text-only ones, so this is ordinary playground use.

The field is per-turn data: only the attachment-delivery chain reads it, from each incoming request, every turn. Nothing bakes it into the sandbox, the daemon, or the harness session.

Changes

Before: a model switch from claude-sonnet-5 (text+image) to a text-only model planned {model: apply-live, harnessSession: reopen-session} and rebuilt.

After: it plans {model: apply-live} and applies live on the running session.

modelCapabilities leaves configFingerprint and the harnessSession facet, with the reason documented in both places, exactly the treatment workflowRevision and isDraft got in step 1. Attachment behavior is unchanged: the chain keeps reading the incoming request's value each turn.

Tests

• The per-turn-volatile test gains the field (it must move no facet); the old "changes the fingerprint" assertion is reversed, with the reason in the comment.
• Full runner suite green (157 files / 2579 tests), typecheck clean.

What to QA

• In the playground on a warm session, switch from a vision-capable model to a text-only model and send. The turn starts fast (no rebuild) and runs on the new model.
• Regression: attach an image, switch to a text-only model, send. The attachment handling still follows the new model's modalities.

Stacked on #6372.

https://claude.ai/code/session_014s6jqKCsVKsNMmnrJVJkZt

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview	Aug 31, 2026 3:51pm

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• release/.*

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository YAML (base), Organization UI (inherited)

Review profile: CHILL

Plan: Pro Plus

Run ID: 8f1a3b21-fe36-4a87-b700-0855cf8f636d

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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.}

[mmabrouk]

lgtm