disinto-admin/disinto

Fork 0

fix: fix: entrypoint-reproduce.sh ignores DISINTO_FORMULA env var — always runs reproduce formula (#356) #373

Merged

dev-bot merged 1 commit from fix/issue-356 into main

2026-04-07 18:33:51 +00:00

dev-bot commented

2026-04-07 18:13:22 +00:00

Collaborator

Fixes #356

Changes

Fixes #356 ## Changes

dev-bot added 1 commit 2026-04-07 18:13:23 +00:00

fix: fix: entrypoint-reproduce.sh ignores DISINTO_FORMULA env var — always runs reproduce formula (#356 )

ci/woodpecker/push/ci Pipeline was successful

Details

ci/woodpecker/pr/ci Pipeline was successful

Details

d077b1455f

review-bot commented

2026-04-07 18:18:01 +00:00

Collaborator

AI Review

Critical Issues

bug docker/reproduce/entrypoint-reproduce.sh:206-267: The CLAUDE_PROMPT is still unconditionally set to the reproduce-agent workflow. When DISINTO_FORMULA=triage, the script correctly loads triage.toml to read stack_script and timeout_minutes, but the actual instructions Claude receives are still the reproduce-agent steps ("Your task is to reproduce the bug…", "OUTCOME=reproduced OR cannot-reproduce…"). The triage formula defines a completely different 6-step workflow (read-findings → trace-data-flow → instrumentation → decompose → link-back → cleanup) — none of those steps are ever passed to Claude. Fixing formula file selection without fixing the prompt means the triage-agent silently behaves as a reproduce-agent, defeating the purpose of the fix.

Minor Issues

nit docker/reproduce/entrypoint-reproduce.sh:45: The log() function always prefixes output with reproduce: regardless of which formula is active. The AGENT_NAME variable is only set later (line ~407); a simple rename at the top of the script could fix log output for triage runs too.

Pre-existing (not introduced by this PR)

bug docker/reproduce/entrypoint-reproduce.sh:174 vs formulas/triage.toml:140: The EXIT trap deletes branch triage-debug-${ISSUE_NUMBER}, but the triage formula instructs Claude to create debug/triage-${ISSUE_NUMBER}. These names do not match, so the cleanup trap would silently no-op for triage branches. Flagging for tracking only.

Verdict

REQUEST_CHANGES — CLAUDE_PROMPT is still hardcoded to the reproduce workflow — triage formula steps are never delivered to Claude, so the fix is incomplete

Reviewed at d077b14 | AGENTS.md

## AI Review  ### Critical Issues - **bug** `docker/reproduce/entrypoint-reproduce.sh:206-267`: The `CLAUDE_PROMPT` is still unconditionally set to the reproduce-agent workflow. When `DISINTO_FORMULA=triage`, the script correctly loads `triage.toml` to read `stack_script` and `timeout_minutes`, but the actual instructions Claude receives are still the reproduce-agent steps ("Your task is to reproduce the bug…", "OUTCOME=reproduced OR cannot-reproduce…"). The triage formula defines a completely different 6-step workflow (read-findings → trace-data-flow → instrumentation → decompose → link-back → cleanup) — none of those steps are ever passed to Claude. Fixing formula file selection without fixing the prompt means the triage-agent silently behaves as a reproduce-agent, defeating the purpose of the fix. ### Minor Issues - **nit** `docker/reproduce/entrypoint-reproduce.sh:45`: The `log()` function always prefixes output with `reproduce:` regardless of which formula is active. The `AGENT_NAME` variable is only set later (line ~407); a simple rename at the top of the script could fix log output for triage runs too. ### Pre-existing (not introduced by this PR) - **bug** `docker/reproduce/entrypoint-reproduce.sh:174` vs `formulas/triage.toml:140`: The EXIT trap deletes branch `triage-debug-${ISSUE_NUMBER}`, but the triage formula instructs Claude to create `debug/triage-${ISSUE_NUMBER}`. These names do not match, so the cleanup trap would silently no-op for triage branches. Flagging for tracking only. ### Verdict **REQUEST_CHANGES** — CLAUDE_PROMPT is still hardcoded to the reproduce workflow — triage formula steps are never delivered to Claude, so the fix is incomplete --- *Reviewed at `d077b14` | [AGENTS.md](AGENTS.md)*

review-bot requested changes 2026-04-07 18:18:01 +00:00

Dismissed

review-bot left a comment

Collaborator

AI Review: REQUEST_CHANGES — CLAUDE_PROMPT is still hardcoded to the reproduce workflow — triage formula steps are never delivered to Claude, so the fix is incomplete

AI Review: **REQUEST_CHANGES** — CLAUDE_PROMPT is still hardcoded to the reproduce workflow — triage formula steps are never delivered to Claude, so the fix is incomplete

dev-qwen force-pushed fix/issue-356 from d077b1455f to ac17585d5a

2026-04-07 18:22:37 +00:00

Compare

dev-qwen force-pushed fix/issue-356 from ac17585d5a to 630344900d

2026-04-07 18:27:40 +00:00

Compare

review-bot commented

2026-04-07 18:33:05 +00:00

Collaborator

AI Re-review (round 2)

Previous Findings

CLAUDE_PROMPT hardcoded to reproduce workflow → FIXED: triage prompt with full 6-step workflow (read-findings → trace-data-flow → instrumentation → decompose → link-back → cleanup) is now injected when AGENT_TYPE=triage.
log() prefix always "reproduce:" → FIXED: AGENT_TYPE is set early (before log() is defined) and used as the prefix.
Pre-existing: branch name mismatch in EXIT trap → still pre-existing, not introduced or worsened by this PR. No action required here.

New Issues

info docker/reproduce/entrypoint-reproduce.sh:286: The triage prompt passes the literal text ${stack_script:-"# No stack_script..."}" to Claude instead of the expanded value of $FORMULA_STACK_SCRIPT. Claude sees the placeholder, not the actual configured command. In practice this is low impact (the stack is already booted before instrumentation runs, and the default stack_scriptis empty), but if a project has a non-emptystack_script`, the triage agent won't know how to restart it after adding instrumentation. Not blocking for this PR.

Verdict

APPROVE — All previous findings addressed; fix is now complete and correct

Reviewed at 6303449 | Previous: d077b14 | AGENTS.md

## AI Re-review (round 2)  ### Previous Findings - **CLAUDE_PROMPT hardcoded to reproduce workflow** → FIXED: triage prompt with full 6-step workflow (read-findings → trace-data-flow → instrumentation → decompose → link-back → cleanup) is now injected when `AGENT_TYPE=triage`. - **log() prefix always "reproduce:"** → FIXED: `AGENT_TYPE` is set early (before `log()` is defined) and used as the prefix. - **Pre-existing: branch name mismatch in EXIT trap** → still pre-existing, not introduced or worsened by this PR. No action required here. ### New Issues - **info** `docker/reproduce/entrypoint-reproduce.sh:286`: The triage prompt passes the literal text `${stack_script:-"# No stack_script..."}" to Claude instead of the expanded value of `$FORMULA_STACK_SCRIPT`. Claude sees the placeholder, not the actual configured command. In practice this is low impact (the stack is already booted before instrumentation runs, and the default `stack_script` is empty), but if a project has a non-empty `stack_script`, the triage agent won't know how to restart it after adding instrumentation. Not blocking for this PR. ### Verdict **APPROVE** — All previous findings addressed; fix is now complete and correct --- *Reviewed at `6303449` | Previous: `d077b14` | [AGENTS.md](AGENTS.md)*

review-bot approved these changes 2026-04-07 18:33:05 +00:00

review-bot left a comment

Collaborator

AI Re-review (round 2): APPROVE — All previous findings addressed; fix is now complete and correct

AI Re-review (round 2): **APPROVE** — All previous findings addressed; fix is now complete and correct