Skip to content

fix: strengthen orchestrator discipline — dispatch, branch verification, todo discipline, property patterns#163

Merged
nullhack merged 1 commit into
mainfrom
fix/agentic-improvements-post-session
May 19, 2026
Merged

fix: strengthen orchestrator discipline — dispatch, branch verification, todo discipline, property patterns#163
nullhack merged 1 commit into
mainfrom
fix/agentic-improvements-post-session

Conversation

@nullhack
Copy link
Copy Markdown
Owner

Summary

Post-session analysis of cex-mm project revealed systemic failures in orchestrator discipline. Five fixes across 5 files:

  1. Dispatch enforcement (AGENTS.md): Added explicit Dispatch step (Including a small change in README to test GH Actions #2) in todo template with "MUST NOT do the work itself" constraint and owner mapping table.
  2. Branch verification (AGENTS.md): Preparation step now verifies current branch matches attrs.git. Golden rule 7 strengthened. New golden rule 8: feature branches must be merged back to dev.
  3. Todo discipline (AGENTS.md): Added paragraph requiring todo update after every step and regeneration if missing.
  4. Review-gate smell detection (.opencode/skills/review-gate/SKILL.md): Step 5 now loads full docs for detection, not #key-takeaways fragment.
  5. Property patterns for BDD examples (new file + skill update): Added property-patterns knowledge file with seven Wlaschin (2014) patterns and a decision tree for Example vs Scenario Outline. Updated write-bdd-features skill step 4.

Files Changed

File Change
AGENTS.md Golden rules 7+8, dispatch step, skill-derived work items, branch verification, todo discipline
.opencode/skills/review-gate/SKILL.md Full docs for smell/ObjCal/SOLID detection
.opencode/skills/write-bdd-features/SKILL.md Property patterns for Example selection
.opencode/knowledge/requirements/property-patterns.md New: 7 property patterns with decision tree
docs/research/software-engineering/quality/wlaschin_2014.md New: source research

Testing

Applied these changes during a live cex-mm session (configurable_rate_limit_buckets feature). The dispatch step, branch verification, and todo discipline prevented the three failure modes that triggered the post-mortem.

…h verification, todo discipline, property patterns for BDD examples

Post-session analysis of cex-mm project revealed three systemic failures:

1. Orchestrator routinely bypasses owner dispatch and does work directly.
   The todo template had no Dispatch step — it jumped from Preparation
   to Load Skills, so the orchestrator never saw the instruction to dispatch.
   Fix: added explicit Dispatch step (#2) in todo template with MUST NOT
   do the work itself constraint and owner mapping table.

2. Branch discipline not enforced at state entry. Agents entered states
   declaring git:dev while on feature branches and vice versa.
   Fix: Preparation step now verifies branch matches attrs.git. Golden
   rule 7 now says 'Verify before starting'. New golden rule 8: feature
   branches must be merged back to dev before new work starts.

3. Todo list goes stale or disappears mid-state as agents focus on work.
   Fix: added Todo discipline paragraph requiring update after every step
   and regeneration if missing.

4. Review-gate skill loaded smell-catalogue at #key-takeaways but
   detecting violations needs the full document (per progressive
   knowledge loading rules in AGENTS.md).
   Fix: step 5 now loads full docs for detection, #key-takeaways only
   for recall.

5. No guidance for choosing Example vs Scenario Outline during BDD
   example creation. Agents either over-used Scenario Outlines or
   under-used them.
   Fix: added property-patterns knowledge file (Wlaschin, 2014) with
   seven patterns and a decision tree. Updated write-bdd-features skill
   step 4 to apply patterns systematically. Added research reference.
@nullhack nullhack merged commit b2b6493 into main May 19, 2026
10 checks passed
@nullhack nullhack deleted the fix/agentic-improvements-post-session branch May 19, 2026 08:30
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant