fix(session): recover stale encrypted reasoning

[1837620622]

Issue for this PR

Fixes #48915

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Resuming a Muse Spark / OpenAI Responses session after idle fails with:

reasoning `encrypted_content` was not issued to this caller

OpenCode stored that caller-bound blob and treated HTTP 400 as fatal, so the session stayed bricked.

This follows the existing context-overflow recovery path:

1. Classify those errors as stale-reasoning
2. If no assistant output has started, drop reasoningEncryptedContent / itemId from stored reasoning
3. Retry the turn once

Readable reasoning text stays. Live tool-loop encrypted state is left alone unless the provider rejects it.

Related: #48741, #48805

How did you verify your code works?

• packages/llm: bun test test/provider-error.test.ts
• packages/opencode: bun test test/provider/error.test.ts and the fromError case in test/session/message-v2.test.ts
• packages/core: bun test test/session-runner.test.ts (90 pass), including three new recovery tests plus existing overflow and encrypted-reasoning tests

Screenshots / recordings

n/a (non-UI)

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Duplicate Found:

PR #48908 - fix(session): recover from stale encrypted reasoning on provider rejection

• fix(session): recover from stale encrypted reasoning on provider rejection #48908

Why it might be related:
This PR addresses the same issue with stale encrypted reasoning recovery. Both PRs appear to handle the recovery path when encrypted reasoning content becomes stale during session resumption. They likely overlap in implementation or address the same bug from different angles.

Note: PR #43595 fix(opencode): recover stale encrypted reasoning state is an older related PR on this topic, but may be superseded.

[1837620622]

This bricks Muse Spark 1.3 on 1.18.30 when resuming an old session. Tests in packages/core session-runner pass. Please review/merge when you can.

[SynacNipo]

hopefully

[kungfooman]

I got that error message too now, what helps is using /models to e.g. switch from 1.2 to 1.3 or vice versa. So it's not a "full" brick.

[yakovexplorer]

still persists on 1.18.31. I hope this gets merged soon.

[kvnloo]

Why this matters

Resuming after idle still bricks sessions with stale encrypted_content / invalid_encrypted_content on 1.18.31 (see thread + #48915). Two open PRs both try to recover — maintainers need one shape on dev, not a third.

Evidence

Two open recoveries for stale encrypted reasoning (encrypted_content not issued to this caller / invalid_encrypted_content):

	#48918	#48908
Base	dev	v2 (currently conflicting)
Tip	87dd316	73e8e76
Error class	packages/llm + opencode provider error	packages/ai
Durability	strip + retry (also touches message-v2 / processor)	strip + SessionEvent.MessageContentUpdated so projection replay keeps sanitized state

Community still sees the brick on 1.18.31 (thread on #48918).

Architecture (proposal)

Maintainer call: which shape should land on dev? Happy to help rebase/tests on the chosen one — not opening a third PR.