fix(orchestrate): report agy as prompt_only, because it is not sandboxed - #658
Merged
Conversation
The dispatcher claimed read_only_guarantee="enforced" for agy, which is the claim that the reviewer could not have mutated the repository under review. It is false, so certification was being awarded to a route that can write files and run shell commands. A write probe against agy 1.1.10 under the dispatcher's own flags returned SUCCESS and created the file. Adding --mode plan changed nothing, so that flag enforces no read-only boundary and adding it would only have shipped false assurance. Writes succeed because permissions.allow grants write_file(*) and command(*), and agy exposes no per-invocation permission flag. The guarantee becomes prompt_only, matching the copilot branch's use for "the prompt asks it not to, nothing enforces it". certification_eligible is now false for agy, which is the intended consequence. emit_record and the certification rule are unchanged: whether an independent opinion should certify without a sandbox is a separate policy decision. The agy-reviewer prompt preamble now carries both facts, since they are separate. An ungranted tool is auto-denied and one denial discards the whole turn; a granted tool simply runs, so silence about writing is not proof that nothing was written. Claude-Session: https://claude.ai/code/session_01DsDFRG6NVWb2VBNNtVkQ6N
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The dispatcher claimed
read_only_guarantee="enforced"for agy.emit_recordawardscertification_eligible=trueonly when the status isok, the review is cross-family, and the guarantee isenforcedoroauth_safe_mode, so that value is the claim that the reviewer could not have mutated the repository under review.For agy the claim was false, and certification was being awarded to a route that can write files and run shell commands.
Evidence
Write probe against agy 1.1.10, asking it to create a file:
--output-format json --disable-slash-commands --sandbox --add-dir)SUCCESS,DID-WRITE, file created on disk--mode planSUCCESS,DID-WRITE, file created again--mode planenforces nothing here, so adding it would have shipped false assurance. Writes succeed becausepermissions.allowgrantswrite_file(*)andcommand(*), andagy --helpexposes no per-invocation permission flag.The change
guaranteebecomesprompt_only, the value the copilot branch already uses for "the prompt asks it not to, nothing enforces it".certification_eligibleis now false for agy, which is the correct consequence.emit_recordand the certification rule are unchanged. Whether certification should distinguish an independent opinion from a sandboxed one is a separate policy decision, deliberately left out of scope so the record gets corrected without loosening the rule to make the problem disappear.The user's
~/.gemini/antigravity-cli/settings.jsonwas not touched. Per-call scoped enforcement was considered and rejected as not worth building.Other adapters
codex-s read-onlyis a real sandbox policyclaude--permission-mode plan,--tools "Read,Grep,Glob", no-mutation system promptcursor-p"has access to all tools, including write and shell", while both--mode askand--mode planare documented read-only. Cursor is out of usage quota so no probe was possible, and flags were not churned on documentation alonekironone,best_effortif enabledVerification
scripts/check-harness: 1268 passed, 2 xfailed.https://claude.ai/code/session_01DsDFRG6NVWb2VBNNtVkQ6N