# stage planner ```output { repository, diagnosis, reproduction_command, baseline_failure, fix_plan, acceptance_checks, revision_count: 0, max_revisions } ``` ```prompt Use the user's runtime prompt as the bug report. Infer repository context from the workspace. Preserve user constraints through every output handoff; initialize revision_count=0 and max_revisions=2 unless the user requests a lower budget. Reproduce the reported bug before proposing a fix. Inspect relevant code, logs, and stack traces; record the exact command and observed baseline failure. Identify the root cause, not a workaround. Propose the smallest correct change and a regression check that fails before the fix. Preserve unrelated user work. Return success only with concrete evidence and a testable plan; otherwise blocked. ``` ```edges success -> implementer blocked -> #failure error -> #failure ``` ```configs provider=opencode-go model=deepseek-v4.1-flash retries=0 ``` # stage implementer ```output { changed_files, diff_summary, regression_test, acceptance_checks, revision_count, max_revisions } Preserve reproduction evidence and all previous handoff fields. ``` ```prompt Implement the focused root-cause fix and add a regression test. Do not broaden scope, discard existing edits, commit, or push. On a feedback-driven revision, increment revision_count; if the budget would be exceeded, return exhausted. Address verification or review feedback without weakening the acceptance checks. Hand off the actual changed files and test command, not a claim that tests passed. ``` ```edges success -> verifier exhausted -> #failure error -> #failure ``` ```configs provider=opencode-go model=deepseek-v4.1-flash retries=0 ``` # stage verifier ```output { verification_results, command_exit_codes, observed_behavior, feedback?, revision_count, max_revisions } Preserve implementation artifacts and reproduction evidence. ``` ```prompt Execute the regression check, original reproduction, and relevant integration checks. Record exact commands, exit codes, and observations. Do not infer success from source inspection. Return success only if the original failure is resolved and all required checks pass. Return tests_failed with actionable feedback for an ordinary implementation failure; use error for unavailable tools or infrastructure. ``` ```edges success -> reviewer tests_failed -> implementer error -> #failure ``` ```configs provider=opencode-go model=deepseek-v4.1-flash retries=0 ``` # stage reviewer ```output { review_findings, feedback?, approval, revision_count, max_revisions } Preserve the exact verification evidence. ``` ```prompt Review the diff against the root cause, user constraints, and acceptance checks. Look for workarounds, missing edge cases, weakened tests, and unrelated changes. Return changes_requested with specific findings when another revision is needed. Return success only for a correct, scoped fix supported by executed checks. Do not modify files during review. ``` ```edges success -> report changes_requested -> implementer error -> #failure ``` ```configs provider=opencode-go model=deepseek-v4.1-flash retries=0 ``` # stage report ```output { summary, changed_files, evidence, remaining_risks } ``` ```prompt Summarize the root cause, fix, regression coverage, and executed checks concisely. Include paths and command outcomes. If approval or required evidence is missing, return incomplete rather than claiming a successful fix. Do not commit or push. ``` ```edges success -> #success incomplete -> #failure error -> #failure ``` ```configs provider=opencode-go model=deepseek-v4.1-flash retries=0 ```