Skip to content

graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs

Python
1.6 KB
Raw Blame

0298. Recurrence reports update the canonical issue's state

Status: accepted

Context

#1701 exposed a gap in the reporting instructions: searching closed issues and adding fresh evidence avoided duplicates, but left unresolved recurrences classified as resolved. The /issue brief also demanded a new issue even when a matching report existed.

Decision

Ordinary reporting and /issue share issue-lifecycle guidance. Inspect the matching issue's state, closure reason, and latest resolution. When current reproduction evidence contradicts a resolved closure, reopen the canonical issue and attach the evidence. Follow duplicate closures to their canonical tracker; do not reopen deliberately rejected, withdrawn, or out-of-scope reports merely because they receive comments. Confirm and accurately report the final state, including a denied reopening that requires a maintainer. Custom and learned root prompts retain this policy across replacement, as worker personas already do. /issue expands the same reporting brief in the REPL and over ACP.

Consequences

Duplicate detection is not evidence that reporting is complete. The agent still needs to judge whether the new evidence contradicts the resolution; automatically reopening every commented-on issue would undo intentional closures. No GitHub-specific shell interception is introduced: the same rule applies to CLI and connector reporting. The offline issue-reporting cases in evals/harness_behavior.jsonl verify prompt delivery and stateful local CLI roundtrips, not a live model's judgment about arbitrary reports.