Cocowisdom

---
name: wisdom
description: CocoWisdom — displays project quality rejection history summary. Shows total rejection count, most frequent gate, most blocked dimension, last rejection, and quality trend. Invoked via $wisdom.
version: "1.1.0"
author: CocoPlus
tags:
  - cocowisdom
  - institutional-memory
  - quality-gate
user-invocable: true
blocking: false
---

## Objective

Display a CocoWisdom status summary for the current project.

Before proceeding, verify that `.cocoplus/` exists. If not, output: "CocoPlus is not initialized. Run `$pod init` first." Then stop.

## CocoPlus 2.0.2 Command Surface

- `$wisdom reject "<description>"` records a rejected approach in `.cocoplus/wisdom/do-not-use.md`.
- `$wisdom index [--harness <name>]` builds the session-log lexical recall index.
- `$wisdom recall "<query>"` searches curated wisdom plus the session-log index.
- `$wisdom learnings [--filter <text>]` lists curated learning and rejection records.
- `$wisdom learn "<text>" [--session <id>] [--reason <text>]` records an operator-curated learning.

Negative memory in `do-not-use.md` is universal. Always load it before positive topic-gated wisdom, never auto-prune it, and never merge rejected approaches into positive wisdom category files.

Positive wisdom uses manifest-first memory. Read `.cocoplus/wisdom/SCHEMA.md` for stage mappings before loading topic content, load only the mapped topics for the current stage role, and keep all other topic files in progressive disclosure mode until `$wisdom get <topic>` is invoked. Memory is branch-scoped Markdown: changes travel with the feature branch and are reviewed like code.

## Step 1 — Read Rejections Store

Read `.cocoplus/wisdom/rejections.jsonl`. If it does not exist or is empty, output:

```
CocoWisdom — No rejection records yet.
Quality gates have not produced any BLOCKED outcomes in this project.
Run $sentinel <file> or $secondeye to begin evaluating artifacts.
```
Then stop.

## Step 2 — Compute Summary Statistics

Parse all JSONL records. Compute:
- Total record count (excluding `type: "retraction"` records)
- Count by gate: secondeye, sentinel, da_critic
- Most frequent gate (highest count)
- Count by dimension across all gates
- Most blocked dimension (highest count)
- Most recent record by timestamp
- Sessions in last 8: count records from last 8 distinct `session_id` values — if count is 0, trend is "Improving"

## Step 3 — Output Summary

```
CocoWisdom — Project: <project name from .cocoplus/ or directory name>
Total rejection records: <count>
Most frequent gate: <gate> (<count> records)
Most blocked dimension: <dimension> (<count> records)
Last rejection: <date> | <gate> | <dimension>
Quality trend: <Improving (no rejection in last 8 sessions) | Active (<count> in last 8 sessions)>
```

## `$wisdom route` — Correction Routing (Feature 37 Enhancement)

Routes CocoCupper's auto-captured corrections (`.cocoplus/cupper/auto-captured.json`) to proposed skill-file edits, closing the loop between "the developer corrected the agent" and "the skill that caused it gets fixed."

1. Invoke `cocowisdom/wisdom-route`. It reads `.cocoplus/cupper/auto-captured.json` entries that carry a `skill_context` field, groups them by skill file, and applies a routing classification per group:
   - **incorrect-behavior** — the skill's instructions produced an action the developer reversed
   - **missing-variant** — the correction asks for an option or mode the skill doesn't support
   - **agent-misapplication** — the skill or persona was invoked in a context it doesn't fit

   This classification may invoke Haiku (Tier 3, async) since it requires judging correction intent — unlike CocoCupper's Tier 1 capture, which is pure regex.

2. For each group, `cocowisdom/wisdom-route` generates a proposed edit (the specific skill section and suggested change) but does not apply it.

3. Present each proposed edit to the developer for explicit confirmation:

   ```
   CocoWisdom Route — [skill-file].skill.md
   Routing: incorrect-behavior (3 corrections)
   Proposed edit: [section] — [specific change]

   Apply this edit? (yes/no)
   ```

4. **The routing workflow requires explicit developer confirmation before any file is modified.** No proposed edit is ever applied automatically, regardless of how many corrections support it.

## `$wisdom keep` and `$wisdom forget`

`$wisdom keep --id <id> --text "<rule>"` writes protected institutional memory to `.cocoplus/wisdom/must-keep.md`. Must-keep entries are not candidates for automatic consolidation.

`$wisdom forget --id <id> --rationale "<reason>"` records an explicit removal rationale in `.cocoplus/wisdom/consolidation-log.md`. Forgetting without rationale is rejected.

Use:

```text
invoke cocowisdom/wisdom-route --keep --id <id> --text "<rule>"
invoke cocowisdom/wisdom-route --forget --id <id> --rationale "<reason>"
```

## Evidence Gate and Denser-Not-Larger Rule

Stable CocoWisdom promotion requires at least 3 distinct confirmed sessions by default. The threshold is configurable in `cocoplus.toml` under `[wisdom].min_evidence_sessions`.

Before promoting a consolidated thesis, validate it with:

```text
invoke cocowisdom/wisdom-route --candidate <candidate.json>
```

The candidate is rejected if it increases word count without reducing entry count, unless a justified exception is recorded. Consolidation should make memory denser, not merely larger.

## Anti-Rationalization Table

| Shortcut / Temptation | Why It Fails |
|-----------------------|--------------|
| Show "Improving" when only a few sessions exist | Trend requires 8 distinct sessions — fewer sessions cannot establish a pattern |
| Skip reading rejections.jsonl when file is large | The quality trend depends on the last 8 sessions — truncating the read produces wrong counts |
| Auto-apply a routed edit when many corrections agree | Volume of corrections is evidence for review, not authorization to skip developer confirmation — the workflow is structurally confirm-first |
| Promote a wisdom pattern from one dramatic incident | Stable institutional memory requires repeated confirmed evidence across distinct sessions |
| Grow a thesis without reducing entries | Denser-not-larger keeps memory useful under context pressure |

## Exit Criteria

- Summary displays correctly even if only 1 record exists
- Trend shows "Improving" only when zero rejections in last 8 distinct sessions
- `$wisdom route` groups auto-captured corrections by skill file and applies one of the three routing classifications to each group
- No skill file is modified by `$wisdom route` without explicit developer confirmation of the specific proposed edit
- `$wisdom keep` writes protected entries to `must-keep.md`
- `$wisdom forget` records rationale in `consolidation-log.md`
- Promotion candidates pass evidence and density gates before becoming stable knowledge