Keep long AIDE decision history retrievable by reference

[mikamikasuki]

Summary

• Keep complete prior AIDE decision transcripts in episode-scoped public memory and persist them unchanged in the episode state.
• Put a stable, pageable context-segment reference in subsequent model prompts instead of embedding the full transcript on every request.
• Apply the same end-to-end request builder for FULL and AIDE so both receive the same reference and can retrieve the full history.
• Normalize repeated-action feedback to the native AIDE Journal schema and rehydrate legacy string-valued exc_info checkpoints without changing their original persisted bytes before the next ordinary checkpoint.
• Add regressions for context-size recovery and repeated-action checkpoint recovery.

Reproduction

Two formal end-to-end requests were rejected before inference: AIDE at 1,505,250 transport characters and FULL at 1,051,668, above the observed 1,048,576 capacity. The failed AIDE request contains about 3.89 million source characters of prior decision history. Replaying that saved request with the history externalized produces a 107,338-character transport payload while the complete source history remains retrievable.

A campaign episode also exposed a Journal restore error after repeated-action diagnostics had stored a string in AIDE's dict-typed exception field. The original checkpoint and action history are retained; the reader translates that legacy string to the upstream exception-argument structure in memory.

Design reference

This follows the established LangGraph context-engineering pattern: keep durable state as raw data and format only the context required by the current node on demand. The complete history stays in persistent episode memory and remains retrievable through the existing paged context tool: Thinking in LangGraph.

Repeated exception metadata follows the existing upstream AIDE ExecutionResult.exc_info shape with an args list, keeping Journal serialization compatible.

Validation

• Focused branch run across prerequisite classification, replanning feedback, request/context recovery, AIDE and study runtime, context memory, end-to-end runtime, and scheduler selection: 267 passed, 3 skipped.
• Includes restart recovery from a checkpoint with the exact legacy string-valued exc_info shape; the repeated scientific action remains deduplicated.
• git diff --check: passed.

Stacked on PR #297; review #296, then #297, then this PR.