| name | advisor |
| description | Anthropic Claude Advisor pattern — consulted at executor inflection points (low-confidence extraction, ambiguous dimensions, conflicting governance, unexplained drift) for strategic reframe rather than parameter-tuning. |
| user-invocable | false |
| allowed-tools | ["Bash","Read"] |
advisor — strategic reframe at inflection points
Anthropic's Claude Advisor pattern applied to CogSig's pipeline. Executor agents (extract, review, historian) are fast/local/cheap; they do routine work. When they hit an inflection — a moment where tuning-within-frame won't help because the FRAME is wrong — they pause and consult the advisor. Advisor reframes strategically in a short call; executor continues execution in the new frame.
Inflection classes
| Class | Trigger | Where fired |
|---|
low-confidence | any dimensions[x].confidence < 0.5 after extract | extract.py post-extraction |
schema-soft-fail | jsonschema validation surfaces type-coercion / missing-subfield warnings | extract.py post-validate |
conflicting-governance | Brutus=WEAK + QA=PASS + Historian=EXPECTED simultaneously | managed-agents/review.py post-review |
drift-unexplained | Historian flags >2 dimensions drifted with no corpus-source change | historian analysis |
What advisor DOES
- Reads the full context (signature draft, samples, review outputs, history)
- Identifies whether the executor's PROBLEM FRAMING is wrong, not just whether parameters need tuning
- Returns a short strategic reframe + concrete next action
- Advisor responses are 400-700 tokens (short), so cost is bounded even on Opus
What advisor does NOT do
- Doesn't retry the extraction itself
- Doesn't modify any files
- Doesn't override governance decisions
- Doesn't block executor — executor decides whether to act on reframe
Deploy modes — advisor runs across all
- Standalone — advisor fires automatically on trigger, user sees reframe output
- In-session team — advisor + governance agents both run; advisor consulted when governance agents conflict
- Cloud (Managed Agents) — advisor can itself be a Managed Agent (future) or standalone inline (current)
Scaffold precedent
This pattern emerged from prior internal work documenting cases where surfacing a model-rated decision-quality verdict at inflection points changes downstream behavior. Three validated instance types: rule-level fixes substituted for verbal-discipline; benchmark reframings; clarifying ambiguous mode boundaries. The pattern maps operationally to Anthropic's Claude Advisor tool (Haiku-executor + Opus-advisor pairing).
