fix(eve): improve compaction checkpoints

[ruiconti]

Why

#749 records a production task that lost completed investigation state across compaction and continued for 442 model turns, 2,026 tool calls, and $59.0342 before cancellation. The free-form checkpoint prompt did not tell the compactor how to update an earlier checkpoint or keep completed work separate from pending work.

The first version of stacked test PR #755 required pi's exact headings. Its 60-call result measured formatting compliance, not semantic progress loss. After removing that requirement, base run 29303880312 passed the original six one-compaction cases on eve's existing prompt. Those results are retained as evidence that the earlier claim was invalid, not as proof for this change.

Review then exposed a separate repeated-compaction bug: eve stored its checkpoint as ordinary conversation messages. A later compaction could truncate the checkpoint to 280 characters or split its marker from its content and retain the stale checkpoint beside the replacement. Two focused regressions failed on both paths before the fix.

What changed

• Use a Codex-style handoff prompt that asks for progress, decisions, constraints, remaining work, and critical continuation data without mandating section names.
• Extract the framework-owned checkpoint before recent-window selection. Prompt construction receives it separately and intact, applies lossy truncation only to ordinary transcript messages, and replaces the old checkpoint with the updated one.
• Move prompt construction into a pure module. Prompt serialization is tested directly instead of through a mocked model call.
• Derive one fixed prompt-envelope estimate from the same prompt fragments used by the model call. The trigger adds that module-level reserve directly instead of scanning and splitting history on every check.
• Require exactly two compactions in both regression cases for GPT-5.6, Opus 4.8, and Sonnet 5. Each case must keep the original work call at one, advance the checkpoint once, and finish with both completion markers.
• Describe prompt-directed behavior as an instruction rather than a guaranteed output schema, with a patch changeset for eve.

On the final head, all six repeated-compaction cases passed in both the local run and the Vercel run. Each case observed exactly 2 compactions, 1 original work call, 1 checkpoint-advance call, and both checkpoint markers in the final response.

This PR is stacked on #755. Repeated-call containment remains a separate follow-up.

Addresses #749.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
eve-docs	Ready	Preview, Comment, Open in v0	Jul 14, 2026 5:42am

2 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
eve-docs-1644	Skipped	Open in v0	Jul 14, 2026 5:42am
eve-docs-4759	Skipped	Open in v0	Jul 14, 2026 5:42am

[vercel]

Additional Suggestion:

Prior structured checkpoint is truncated to 280 chars on the 2nd+ compaction, so accumulated progress/decisions/context are lost instead of being carried forward.