fix(runtime): side-conversation forks must not diverge from the parent's prompt-cache prefix

[Astro-Han]

Problem

A committed Side Conversation inherits the parent Session's transcript so the user can ask an auxiliary question against a hot context without paying for it twice. On the wire, provider prompt caching (Anthropic's request-level cache_control; OpenAI's automatic prefix caching) is what makes that inheritance nearly free: the fork's first request should match the parent's cached prefix and hit.

It cannot. The side-conversation boundary instructions are joined into the single system prompt string, so the fork's system block differs from the parent's at the first divergence point — before the inherited history. Prefix matching stops at the system block and the entire inherited transcript is re-read at full input price on the fork's first request. The feature's own economic premise (cheap context reuse) is defeated by its own implementation.

Mechanism

1. Fork creation copies the parent transcript and pins the same connection + model (session-revision-coordinator.ts, #createInput), branching at branchOfTurnId. The replay path is deterministic and shared with the parent's own per-request replay, so the message segment can be wire-identical.
2. Every run on the fork composes the system prompt with a fork-scoped fragment: interactive-run-composer.ts passes input.sideConversation ? buildSideConversationSystemPromptFragment() : undefined into assembleMainSessionSystemPrompt.
3. assembleMainSessionSystemPrompt (system-prompt/main-session-prompt.ts) join('\n\n')s all fragments into one string. ai@7.0.70 maps a string instructions to a single { role: 'system' } message, so the composed text is one system block on the wire.
4. The parent's requests do build full-prefix cache entries: model-factory.ts sets request-level providerOptions.anthropic.cacheControl, and @ai-sdk/anthropic@4.0.40 forwards it as a top-level cache_control body field. The fork just cannot match past the system block.
5. Net effect: the fork's first request is a full-price read of tools + system + entire inherited history, then it builds its own cache from its second request onward.

On the OpenAI Responses wire there is an additional blocker: promptCacheKey is maka:${sessionId} (openai-responses-continuation.ts), so the fork routes under a different cache key than the parent even where prefixes would match.

The cost lands exactly where the feature is most used: the user forks while the parent session is hot (within the cache TTL), and the re-read scales with session length — precisely the sessions where a side conversation is worth opening. On Anthropic, a cached read is a fraction of full input price, so the gap is proportional to the whole inherited context.

Invariant (authoritative fix direction)

Fork-scoped semantics must not modify anything the provider caches ahead of the inherited transcript. Concretely:

• The side-conversation flag must not alter the system prompt. buildSideConversationSystemPromptFragment must stop flowing through assembleMainSessionSystemPrompt.
• The fork's first request must be a strict prefix-extension of the parent's request at the fork point: identical tools, identical system prompt, identical message history up to branchOfTurnId.
• The boundary instructions ride after the inherited history — recommended placement is prepended to the fork's first user turn (or a synthetic first user message the UI renders as a boundary notice), which also keeps the fork's own history stable for its subsequent requests.
• No other mechanism may mutate the pre-history region for forks. This invariant should be enforced by a test, not documented in a comment.

Regression guard

Add an integration test in the runtime-host suite (where session.branch.create already runs real executions against fake providers) that captures the provider request bodies — ProviderRequestTracker already captures them for telemetry — for the parent's last request and the fork's first request, and asserts:

1. the tools arrays are identical;
2. the system content is identical byte-for-byte;
3. the fork's message array strictly extends the parent's (deep-equal prefix);
4. the first divergence is the fork-owned turn, and the boundary text occurs after the inherited prefix — never in the system prompt.

This guard is the point of the fix: any future mechanism that mutates pre-history content for forks fails here instead of silently re-pricing every fork.

Open question: OpenAI cache-key routing

Even with prefix equality, the fork's promptCacheKey differs from the parent's. Options:

• (a) carry the parent's cache key for the fork's requests — the fork's prefix is a superset of the parent's, so sharing the routing key stays coherent for the fork's lifetime;
• (b) accept best-effort on OpenAI and guarantee the hit on Anthropic.

Recommend (a); decide during implementation.

Acceptance criteria

1. The fork's first request carries a system prompt byte-identical to the parent's.
2. The boundary instructions appear after the inherited history on the fork's first request and remain stable in the fork's own history afterward.
3. The wire-level regression guard above exists and fails if any mechanism reintroduces pre-history divergence for forks.
4. Fork behavior is otherwise unchanged: the model still treats inherited history as reference-only, and the UI renders the boundary without leaking raw preamble text as user content.
5. Replay admission rules are untouched (thinking replay at fork time remains identity + model scoped).

Non-goals

• Making mid-session skill/memory/workspace-instruction drift cache-neutral; that applies to every session and predates this bug.
• Caching the final assistant response that postdates the parent's last request; the parent never sent it, so that delta is unavoidable full-price input.
• Changing feat(desktop): allow Side Conversations to select an independent model #4495. Independent model selection is orthogonal: fixing this bug makes the default same-model fork cheap, which is the path feat(desktop): allow Side Conversations to select an independent model #4495 leaves untouched.

Related: #4331, #4495.

[scs0209]

take