Reflect — Brain Routing Correction Recorder

Record routing corrections with confidence scoring. Called by the brain orchestrator when Nathan signals a wrong routing decision. Uses the Claude-Reflect pattern: corrections must occur 3+ times before being promoted to active learnings.

Input

Receive correction context as args:

input="<original request>" routed_to="<wrong skill>" correct_skill="<right skill>" reason="<Nathan's words>"

Parse these 4 fields from $ARGUMENTS.

Workflow

Step 1 — Read Current Learnings

para_read({ file: "plugins/para-obsidian/skills/brain/references/learnings.md" })

Parse the Raw Correction Log table to find existing entries.

Step 2 — Check for Existing Pattern

Search the Raw Correction Log for a row where Routed To matches routed_to AND Correct Skill matches correct_skill AND the Input Pattern is semantically similar to the new input.

• If match found: Increment the count for that row. Update the Input Pattern to be a generalized pattern if the inputs differ slightly (e.g., "process inbox" and "handle my inbox" → "inbox processing requests").
• If no match: Add a new row with count = 1 and today's date.

Step 3 — Update Raw Correction Log

Use para_replace_section to update the "Raw Correction Log" section with the modified table:

para_replace_section({
  file: "plugins/para-obsidian/skills/brain/references/learnings.md",
  heading: "Raw Correction Log",
  content: "<updated table with all rows>",
  response_format: "json"
})

Step 4 — Promote to Active Learnings (if count >= 3)

If the updated count reaches 3 or higher, also update the "Active Learnings" section by adding a human-readable routing rule:

para_replace_section({
  file: "plugins/para-obsidian/skills/brain/references/learnings.md",
  heading: "Active Learnings (count >= 3)",
  content: "<existing rules>\n- When input matches \"<generalized pattern>\", route to `<correct_skill>` (not `<routed_to>`). Reason: <Nathan's reason>.",
  response_format: "json"
})

Step 5 — Confirm

Output a single confirmation line:

SKILL_RESULT:{"status":"ok","skill":"reflect","summary":"Recorded correction: <routed_to> → <correct_skill> (count: <N>)"}

If count just reached 3, add: "promoted": true to the JSON.

Rules

1. Never modify anything outside learnings.md — this skill has a single responsibility.
2. Semantic matching, not exact — "process inbox" and "triage my inbox" are the same pattern.
3. Generalize patterns — After 2+ similar inputs, generalize the Input Pattern column to a broader description.
4. Preserve existing entries — Always read before writing. Never lose existing corrections.
5. One correction per invocation — Process exactly one correction, then exit.