DRG profile routing not applied on first tasks generation

When I ran /spec-kitty.tasks for mission charter-ux-and-org-pack-vocabulary-01KSAF14, every Python-implementation work package was assigned the generic implementer-ivan profile. The HiC had to ask explicitly for "proper assignment based on dynamic DRG routing and task context" before the more-specific python-pedro (routing priority 80, with pydantic / pytest / pathlib / refactor / python in its action-domain set) was applied to WP01-WP08.

What happened

  • All eight Python-implementation WPs got agent_profile: implementer-ivan on first pass (priority 50, generic implementer).
  • WP09 correctly got curator-carla (documentation/glossary work).
  • WP10 correctly got reviewer-renata (acceptance/sign-off framing).
  • The remaining mismatch on WP01-WP08 was caught by a second-pass HiC nudge, not by the routing logic itself.

What should have happened

The /spec-kitty.tasks skill explicitly instructs the agent (Step 8a) to "review all available doctrine-provided and user-created agent profiles and assign the most relevant profile to each work package" using task_type, authoritative_surface, owned_files, and subtask content. With python-pedro.routing-priority = 80 vs implementer-ivan.routing-priority = 50, and with every owned-file path under src/**/*.py and every test under tests/specify_cli/**/*.py or tests/integration/**/*.py, Pedro is the strict overmatch from the DRG signal alone.

Why the gap

Two compounding causes:

  1. Anchoring on the most generic acceptable profile. I reached for implementer-ivan because it is the literal canonical "implementer" role and trivially correct. That is the default that satisfies the role-fit check but ignores the routing-priority + action-domain overlap. The skill's Step 8a is explicit about doing the routing match; I executed it as a one-line shortcut.
  2. Skill-instruction recency vs. task-pressure inversion. Step 8a sits late in a long prompt, after seven steps that all emphasised completing the planning artifacts. By the time it fired, the gravitational pull was "land the WPs and finalize-tasks" rather than "run a separate profile-matching pass". The skill's framing rewards completion over precision on this last step.

Neither cause is novel — both are classic LLM failure modes (default-to-generic, step-fatigue near the end of multi-step prompts). The interesting observation is that the DRG routing data needed to do this correctly was sitting one CLI call away (spec-kitty agent profile list --json), but I didn't think to invoke it on the first pass.

Follow-up

  • If this recurs, consider tightening Step 8a of the /spec-kitty.tasks skill: make the profile-matching call mandatory and require a 1-line justification per WP referencing the matched action-domain. That converts "pick a profile" into a small structured argument that's harder to discharge with a generic default.
  • Or expose a CLI helper spec-kitty agent profile match --wp <id> --json that resolves the best profile from owned_files + action domains, so the agent doesn't have to reproduce the matching logic in prose.
  • For now, the lesson is local to this mission: profile assignment is a separate step that benefits from doing the lookup explicitly, not a "default to ivan" decision.
Edit this page