[BUG] Agent SDK: auto-compact completely broken on resumed sessions — reactive compaction handler is null

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

What's Wrong?

When using the Agent SDK with resumeSessionId, auto-compact never triggers and "Prompt is too long" is silently emitted as assistant text through the success path. Two independent bugs combine:

1. Proactive auto-compact uses local token counts, which are empty on resume.

The auto-compact check (RLq in the de-minified source) calls C_z, which computes qG(q) - z on the local messages array. For a resumed session, this array only contains the new prompt being sent — the full conversation lives server-side. The local count is always small, so auto-compact never triggers regardless of how large the server-side context has grown.

The pre-flight blocking limit check has the same issue — it uses qG(F) - K6, which is the local count.

2. Reactive compaction handler is permanently null.

The reactive compaction handler (wj6 in the de-minified bundle) is declared as var wj6 = null and never assigned anywhere in the 638k-line de-minified bundle. Every reactive compaction code path is gated behind wj6?. optional chaining and silently no-ops:

• wj6?.isWithheldPromptTooLong(y6) → always undefined, prompt-too-long messages are never withheld
• wj6?.isReactiveCompactEnabled() → always undefined
• if ((y6 || c6) && wj6) → always false, tryReactiveCompact is never called

This is the safety net for when the API rejects a request as too long. It's fully architected (withholding logic, compaction, retry loop) but completely inert.

Result: The API error is converted to an assistant message with content: "Prompt is too long" and emitted through subtype=success. The consumer sees it as regular model output, not as an error.

What Should Happen?

1. For resumed sessions, the SDK should track server-reported context size (from message_start stream events) and use that for auto-compact decisions, not just local message array size.
2. The reactive compaction handler should actually be instantiated.
3. At minimum, "Prompt is too long" from the API should surface as an error, not as subtype=success with assistant text content.

Steps to Reproduce

1. Use the Agent SDK to create a session
2. Run multiple queries accumulating context over time
3. Between queries, the session is evicted from memory and later resumed via resumeSessionId
4. Eventually the server-side context exceeds the model's limit
5. The SDK emits "Prompt is too long" as assistant text with subtype=success

This manifests in any long-lived Agent SDK application that resumes sessions across process restarts. In my case, an automation tool that runs periodic analysis sessions — each cycle resumes the same session ID. After significant accumulated context, the session silently died.

Error Messages/Logs

No error in logs. The SDK reported:

type=system, subtype=init
type=system, subtype=status
type=rate_limit_event
type=assistant  (1 content block: "Prompt is too long")
type=result, subtype=success, cost=$0.00

Claude Model

Opus 4.6 (1M context)

Is this a regression?

Unknown — reactive compaction may have never been wired up.

Claude Code Version

2.1.85

Platform

Anthropic API (Agent SDK)

Operating System

Linux (Debian)

Additional Information

Related: #36751 (same symptom from CLI resume path, no root cause identified). Also related: #35358 (session resume drops context).