Repository navigation
Make evolution domain-aware and preserve behavioral families - #2
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d57b52f4fc
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| self.used += 1 | ||
| cfg = self.config | ||
| actions = p.case(actions, cfg.vocabulary, cfg.limits["max_actions"]) | ||
| actions = p.case(actions, cfg.vocabulary, cfg.limits["max_actions"], cfg.domain) |
There was a problem hiding this comment.
Preserve the required setup prefix during shrinking
With a nonempty domain.setup and sufficient shrink budget, a confirmed failure aborts the campaign because shrink() still deletes chunks starting at index zero, while this validation now rejects candidates missing the setup prefix. The resulting ValueError escapes before the finding and summary are written, losing the confirmation receipts. Reproduced with the payment example using setup [restart] and seed [restart, pay(A), refund(A)]. Restrict shrink deletions to the suffix after the required setup so both search and explicit minimization can preserve replayable findings.
AGENTS.md reference: AGENTS.md:L8-L9
Useful? React with 👍 / 👎.
EvoFuzz now implements the evolvability design: payment seeds can generate domain-valid identity, authorization, replay, reversal and recovery sequences, and the search retains different families instead of evicting by age alone.
Changes:
Validation at d57b52f:
Scope: family labels are supplied search descriptors, not semantic-equivalence proofs. Multiple descendants may explore the same behavior; duplicates are logged. Online scheduling is measured reward adaptation, not training the LLM or a benchmark claim of improved bug discovery. Existing oracles and process-isolation limits remain unchanged. The separate resonance-retrieval paper investigation adds no dependency or acceptance gate.