fix: infer provider from model for long-horizon agent sessions

[lsm]

Switching the model on a long-horizon (LH) agent threw Session has no provider configured because the LH/regular session config builders set model but left provider undefined, while model-switch-handler hard-threw before checking the stored model.

This infers the provider from the model (mirroring custom-agent.ts) in both config builders when agent.provider is unset, and hardens the switch guard to infer the previous provider from the stored model instead of throwing — so existing stranded sessions self-heal on the next wake.

• buildLongHorizonAgentSessionConfig + regular worker-agent branch: agent.provider ?? inferProviderForModel(model)
• model-switch-handler: previousProvider = session.config.provider ?? inferProviderForModel(session.config.model)
• Regression tests: config-builder inference (kimi/non-kimi/explicit-wins/default), refresh self-heal (undefined → kimi), and the switch guard (incl. the kimi → kimi same-model case from the report)

[lsm]

🤖 Review by glm-5.1[1m] (GLM)

Model: glm-5.1[1m] | Client: NeoKai | Provider: GLM

Recommendation: APPROVE — correct, minimal, and well-tested. Zero P0-P3 findings.

Note: posted as COMMENT because the PR author and reviewer share the same GitHub account (GitHub rejects APPROVE/REQUEST_CHANGES from the author). The substantive recommendation is APPROVE.

Scope reviewed

• model-switch-handler.ts:175-182 — guard now infers previousProvider from the stored model instead of hard-throwing.
• space-runtime-service.ts:509-554 (buildLongHorizonAgentSessionConfig) — infers provider from model when agent.provider is unset.
• space-runtime-service.ts:657-661 (regular worker-agent branch) — same agent.provider ?? inferProviderForModel(model) fallback.
• Regression tests in both model-switch-handler.test.ts and space-runtime-service.test.ts.

Correctness

• The primary fix mirrors the already-correct worker pattern in custom-agent.ts:642-644. inferProviderForModel('kimi-for-coding') → 'kimi' (the registry strips the [1m] suffix and routes kimi-* / moonshot-* / k3 to kimi), so the report's exact case resolves correctly.
• model is computed first, then provider derives from it, so inference always sees the resolved model — including the DEFAULT_LONG_HORIZON_AGENT_MODEL fallthrough, which now correctly yields provider='anthropic' instead of undefined.
• In model-switch-handler, the "already using this model" check at line 222 compares the raw session.config.provider against newProvider, so switching an unprovisioned kimi session to the same model proceeds (and repairs the provider) rather than no-op'ing — matching the report's requirement that kimi→kimi succeed. The downstream resolveModelAlias(..., previousProvider) and acp checks only become more correct with the inferred value.

Backwards compatibility / regression

• Explicit agent.provider always wins via ?? short-circuit; for sessions that already had a provider stored, inferProviderForModel is never even computed. Covered by explicit-wins tests for both builders and for the switch guard.
• Self-heal is one-shot: refreshLongHorizonAgentSessionConfig only calls updateConfig + resetQuery when the config actually changed, and a no-op wake is explicitly tested. No data migration needed.

Coverage

• Config-builder inference: kimi, non-kimi (anthropic), explicit-wins, default-fallback.
• Refresh: undefined → kimi self-heal, and no-op when already matching.
• Regular worker branch: inference + explicit-wins.
• Switch guard: empty-provider infers from model (glm + kimi), the kimi→kimi same-model regression from the report, and the "no provider AND no model" case still throws.
• Both changed test files pass locally (32/0 and 69/0).

Completeness

A sweep of every provider: assignment sourced from a .provider field across packages/daemon/src confirms no other Session['config'] producer reachable by switchModel is missing the inference — session-lifecycle.ts resolves via getValidatedModelId, custom-agent.ts already infers, and the remaining sites are RPC/DTO/repository plumbing, not session configs.

No changes requested.

[lsm]

@codex review

Retriggering a fresh Codex review for the current head (bf5c4972e) so the approval hook sees a current +1.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep them coming!

Reviewed commit: bf5c4972e4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.