wisdom-insights

Objective

Generate a structured insights report from CocoWisdom rejection history using a Haiku sub-agent synthesis.

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

Step 1 — Read Rejections and Prior Thesis

Read .cocoplus/wisdom/rejections.jsonl. If absent or fewer than 3 records, output: "Not enough rejection data for insights (minimum 3 records needed). Current count: ." Then stop.

Load prior thesis (carry-forward): Glob for insights-*.md files in .cocoplus/wisdom/. If any exist, read the most recent one (by filename date). Extract the ## Thesis section. Store as prior_thesis. If no prior insights file exists, prior_thesis is null (first-run case).

Step 2 — Prepare Analytics Data

Compute the following deterministically before invoking any sub-agent:

Total records by gate
Total records by dimension
Records by dimension for last 30 days vs prior 30 days (trend per dimension)
Sessions with rejections vs sessions without (from distinct session_ids in record set and recent sessions)
Most improved dimension (highest reduction in last 30 days)
Most persistent dimension (most records, no reduction)

Step 3 — Spawn Haiku Synthesis Sub-Agent

Pass the pre-computed analytics data AND prior_thesis to a Haiku sub-agent with this mandate:

"Read the provided rejection analytics data and produce a structured insights report. The report must:

Name the most frequently blocked dimension and explain the pattern from the evidence (do not speculate — cite actual rejection reasons)
Describe the quality trend (improving/degrading/stable) with specific session counts
Produce a per-dimension health table
Provide 2-3 actionable recommendations based on the patterns observed

Carry-forward thesis rule (mandatory):

If prior_thesis is provided: your ## Thesis section MUST begin with the prior thesis verbatim, then add a ### New Evidence subsection that extends or refines it based on the new records. Never replace the prior thesis — only extend it.
If prior_thesis is null (first run): write a fresh ## Thesis section.
The thesis is the project's accumulated quality narrative. It must grow incrementally across insight runs, not reset.

Do not produce vague generalities. Every statement must be supported by the analytics data provided."

Step 4 — Write Insights File

Write the sub-agent output to .cocoplus/wisdom/insights-<YYYY-MM-DD>.md with header:

# CocoWisdom Insights — <date>
Generated: <ISO8601 timestamp>
Records analyzed: <total count>
Prior thesis carried from: <prior insights filename, or "none — first run">

## Thesis
[Prior thesis verbatim (if any), followed by ### New Evidence subsection]

## Dimension Health
[per-dimension health table]

## Trend
[quality trend narrative with session counts]

## Recommendations
[2-3 actionable items]

Invariant: Never overwrite an existing insights-<YYYY-MM-DD>.md file if one already exists for today. If a same-day file exists, write to insights-<YYYY-MM-DD>-<HHmmss>.md instead.

Step 5 — Output Confirmation

Display: "CocoWisdom insights written to .cocoplus/wisdom/insights-<date>.md" Show the first 20 lines of the insights file as a preview.

Anti-Rationalization Table

Shortcut / Temptation	Why It Fails
Write insights to stdout only	Insights are a project artifact — the file is the persistent record, stdout is transient
Skip sub-agent for small datasets	Synthesis is only meaningful when patterns exist; minimum record count gate prevents meaningless output
Let LLM compute the statistics	Statistics are deterministic — compute them in skill, let LLM interpret them
Replace prior thesis with new narrative	Thesis is the accumulated quality story — replacing destroys institutional memory; always extend
Skip prior thesis load when file is large	Prior thesis is one section of one file — read cost is negligible vs. the memory lost by skipping

Exit Criteria

Insights file written to date-stamped path in .cocoplus/wisdom/
File includes pre-computed statistics and narrative interpretation
## Thesis section present: extends prior thesis if one exists, or fresh thesis on first run
Developer sees preview of insights file