fix(agents): honor prompt contracts across generation paths - #8
Merged
Conversation
Editing a review through chat previously replaced its typed findings with markdown. Carry the complete Review through the tool, event, and reducer, and provide its full JSON to subsequent edits so hidden findings and evidence remain available. Use Unicode-safe draft previews and drain successful queued updates when generation completes. Regression tests exercise the full update path, multibyte boundaries, preserved metadata, and completion ordering. Co-Authored-By: Nova (GPT-6) <noreply@openai.com>
A fixed pull request head does not fix the three-dot comparison when its base absorbs part of the branch. Validate the captured base SHA at both publication checks so findings retain their analyzed scope. Exercise the changing merge base with a real Git graph and verify that both HTTP rejection paths stop before posting a review. Co-Authored-By: Nova (GPT-6 Astra) <noreply@openai.com>
Rust is a required build dependency rather than a formula option. Select the Intel macOS source path explicitly so installation builds the binary instead of selecting a manual page from the source tree. Verify both formula variants across all four supported platform branches with an isolated Ruby DSL harness. Co-Authored-By: Nova (GPT-6 Astra) <noreply@openai.com>
Separate capability contracts from repository evidence and remove fixed investigation recipes, fictional release examples, and padded artifact templates. Send each contract once in the trusted preamble. Apply saved instructions across capabilities and document the scope change. Preserve explicit emoji choices, including Studio Auto mode, and carry parent scope into workers with a usable tool-turn budget. Exercise real HTTP instruction routing and delegated tool loops. Live Astra and Opus comparisons confirm the explicit no-emoji policy and expanded saved-instruction scope on the same synthetic fixture. Co-Authored-By: Nova (GPT-6 Astra) <noreply@openai.com>
Allow manual validation of all native binaries and Linux packages without creating releases or updating registries, tags, or taps. Default dispatches to validation and make release publication opt in. Use the manifest version for package paths, reject mismatched release tags, and smoke-test each native binary before uploading artifacts. Co-Authored-By: Nova (GPT-6 Astra) <noreply@openai.com>
Use the same schema and parser across generation paths, and consume Rig's final response so tool narration cannot become the artifact. Apply the configured critic policy to streaming and retain the original content and contract during a requested revision. Keep revision errors explicit when material issues remain. Preserve underlying causes, reject foreign JSON wrappers, and avoid slicing Unicode error previews at byte offsets. Exercise HTTP and SSE exchanges for every structured artifact type, tool turns, critic settings, and failed revisions. Co-Authored-By: Nova (GPT-6 Astra) <noreply@openai.com>
Document the provider guidance behind the prompt changes, distinguish runtime regression tests from model-quality evaluations, and replace stale tool recipes with the current scope and evidence contracts. Explain saved-instruction and worker-model migrations. Refresh the manual and provider references, including routed providers and the shared streaming and critic behavior. Co-Authored-By: Nova (GPT-6 Astra) <noreply@openai.com>
|
Warning Review limit reached
On-demand reviews are free for the next 14 days. After that, they cost $0.25 per reviewed file. Or wait 19 minutes for your next included review. View limit detailsLimit details: You’ve used the included review currently available. Review configuration: ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Team Run ID: 📒 Files selected for processing (53)
Comment |
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
1 participant
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.
Iris could lose structured findings during Studio chat edits, let repository history override an explicit no-emoji setting, and parse streamed artifacts differently from normal generation. The agent now preserves typed drafts and applies the same task, schema, and critic contracts across both execution paths.
The prompt audit used current Astra, Opus 5, and Gemini guidance. Shared instructions separate task authority from repository evidence; the eight capabilities no longer impose fixed file caps, mandatory investigation sequences, fictional release examples, or oversized PR outlines. Evidence gates, selected comparison refs, changelog heading syntax, and human templates remain part of the contract. The prompt contract guide records the sources and evaluation limits.
Behavior and compatibility
Validation
The local gate passes: 331 tests, 9 existing ignored tests, formatting, and all-target Clippy with warnings denied. Thirteen runtime tests use actual local HTTP/SSE exchanges for instruction routing, delegated tool turns, all five structured artifact types, and critic behavior. Crate packaging verifies; documentation and the manual build; Studio startup, help, and quit pass a PTY smoke test.
An identical synthetic commit fixture was run before and after on Astra and Opus 5. Both now honor
--no-gitmojiand the expanded saved-instruction scope. Four calls establish those narrow behaviors, not an overall generation-quality benchmark.Independent reviews checked the prompt/runtime changes, Studio updates, base-commit guard, Homebrew paths, workflow publication gates, and documentation. The candidate matrix passed for commit
cfb5a5e: Linux x64/ARM64, macOS ARM64, and Windows GNU builds and binary smoke tests, plus Debian/RPM packaging and Docker checks. All publication jobs were skipped. The PR workflow also passed.