Repository navigation
feat: design durable, evidence-verified agent loops - #377
Merged
Merged
Conversation
11 tasks
Reliability alone reaches parity with existing check-and-repair loops. This revision makes the first repair recipe carry the things codeoid is placed to do better, instead of deferring them: - contract fields with hidden checks and protected inputs - a first-class `cannot_satisfy` worker outcome, so a conflict between objective and checks ends `blocked` rather than in altered checks - a written feedback policy: visible-check diagnostics yes; hidden-check and integrity-detection detail no - an approval profile so attempts wait and notify instead of parking - an escalation ladder ending in a different backend, and exact oscillation detection from checkpoint tree digests - evidence-selected candidates across backends via session.compare (#364) - learned per-repository run records, as advice never evidence - CI-failure and review-comment triggers in the first milestone - per-attempt `maxBudgetUsd` on the Claude SDK as a real spend bound - the sandbox spike moved into milestone A - the measurements a "best" claim must earn, at matched model and spend Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
rsharath
approved these changes
Oct 11, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Proposal
Adds
docs/loop-engineering-design.md, a design for durable agent loops built on codeoid's existing pipeline, findings, dispatcher, session, and checkpoint infrastructure.The initial workflow takes a failing check and produces a verified repair or an explicit blocker within declared limits. The proposal defines evidence tied to the exact candidate revision, lifecycle ownership, cancellation and crash recovery, provider capability limits, and staged delivery. Existing pipeline human boundaries remain the default.
The document includes a cited research comparison and an evaluation plan. It distinguishes proposed guarantees from existing capabilities and treats claims of superiority as something to measure.
Implementation tracked in #376. Merging this PR records the proposal; it does not implement the feature or close that issue.
Validation
git diff --checkpasses.Review Focus
Confirm the first repair workflow, lifecycle ownership between controller/pipeline/dispatcher, verifier integrity boundaries, and the guarantees feasible across different provider adapters.
Revision: codeoid's own levers in the first loops
The first draft reached parity on reliability. The revision makes the first repair recipe carry the things codeoid can do that other loop tools don't:
cannot_satisfyoutcome: the worker can report that the objective conflicts with its checks, and the run endsblockedinstead of in edited tests.session.compare(feat: compare backends side by side — one prompt, 2–4 forks, keep the best #364).🤖 Generated with Claude Code