RLM/log workbench: route large local logs and structured payloads out of the parent transcript

[Hmbown]

Problem

Large local logs and structured payloads are currently easy to inspect directly in the parent thread, even when an RLM-style bounded analysis pass would be safer. This pushes huge output into the model transcript, increases context pressure, and hides the distinction between "the model is confused" and "the parent transcript is overloaded."

Evidence from maintainer-private local CodeWhale session logs, scanned 2026-05-24:

• 56 turns contained 8 or more read/search calls and had no observed RLM-style delegation.
• 67 function-call outputs reported more than 50,000 original tokens.
• The largest observed tool output reported about 800,893 original tokens.
• Two parent turns used structured local analysis while also producing very large outputs.
• No RLM-style calls were observed in the inspected CodeWhale logs.

No prompts, raw tool outputs, secrets, absolute local paths, or user text are copied here. These are aggregate counters from local logs.

Desired Behavior

When CodeWhale sees local log mining, structured JSONL scanning, long command output, or repeated broad search, it should steer toward an RLM/log-analysis workbench:

• Keep the large source in a local handle.
• Run focused structured analysis near the data.
• Return aggregate metrics, examples, and line/source handles.
• Avoid dumping raw logs into the parent model context.
• Preserve a privacy-first export mode for public issues.

This should not require the user to know when RLM is appropriate. The system can infer it from data size and work shape.

Acceptance Criteria

• Large local files, JSONL session logs, and command outputs above a configurable threshold trigger an RLM/log-analysis suggestion.
• The analysis result is summary-first and includes bounded examples or handles, not raw full content.
• The parent can request follow-up slices by handle.
• Public issue export redacts prompts, secrets, local absolute paths, raw command output, and transcript text by default.
• Tests cover large JSONL logs, long stdout/stderr, repeated rg/sed/jq sweeps, and false positives for small file reads.

Related

• Session context: cap raw tool-output replay and keep details behind handles #2021 raw tool-output replay caps.
• Session logs: classify environment/tool failures before blaming the model #2022 environment/tool failure classification.
• Slash commands: PEEK-backed command receipts and continuity #1889 PEEK-backed command receipts and continuity.
• Feature: DeepSeek cache-aware prompt diagnostics and wire payload optimization #1253 cache-aware prompt diagnostics and wire payload optimization.