Root cause observed in a real Codex session: the agent burned ~3min
on `make venv` + 4 sequential `make diagram-lookup` queries trying
to find a non-existent ref-arch (OCI PostgreSQL accessed FROM GCP
via Cross-Cloud Interconnect). The catalog has 123 entries; if two
honest queries can't surface a topology, none exists — but nothing
in the skill said "stop." Three coordinated fixes:
1) Makefile: `make venv` is now idempotent.
Marker file `.venv/.deps-installed` keyed off `requirements.txt`
mtime — the second `make venv` in a row is a 26ms no-op (was
1m53s on a Codex sandbox, paid every turn). Force a rebuild with
`rm .venv/.deps-installed`.
2) kb/architecture-center/catalog.yaml: new `known_gaps` block.
Surfaced by the lookup tool as a top-of-output banner whenever
the (synonym-expanded) query tokens match any one of a gap's
`triggers` token sets. Two gaps seeded from real engagements:
- gcp-to-oci-native-services — OCI PostgreSQL/OKE/Cache/etc.
reached FROM GCP via Cross-Cloud Interconnect (the
Database@Google Cloud entries are the OPPOSITE direction).
- newer-oci-services-without-icon — OCI Cache (Redis/Valkey)
and OCI PostgreSQL post-v24.2-toolkit; no ref-arch, no
toolkit icon, fall back to generic stencil + explicit label.
Each gap's `notice:` is a copy-paste recommended composition so
the agent doesn't have to invent a strategy.
3) tools/archcenter_pattern_lookup.py: detects + surfaces gaps.
`lookup()` now returns `{"matches": ..., "gaps": ...}`. The
text printer emits a "⚠ KNOWN GAP — <id>" banner with the
notice block ABOVE the score-based top-K, so the agent sees
the stop-and-compose signal before ever scrolling to results.
YAML output mode mirrors the same shape under `gaps:`.
4) SKILL.md (option 2 step 1) + Codex copy via sync-skill.
New explicit rule: "Lookup budget: max 2 queries — never loop."
Documents the gap banner, the closest-3-then-ask fallback, and
the prohibition against grasping at a 3rd refinement.
Verified manually:
- GCP+postgresql query → gap fires
- "redis ha multi-az" → newer-OCI gap fires
- "exadata cross region data guard" → no false-positive gap
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
7f8cba7f30