retro

name	retro
description	Learning system interface: stats, search, graduate, clear learnings. Backed by learning.db (SQLite + FTS5).
user-invocable	true
argument-hint	[status|list|search <term>|graduate|clear]
allowed-tools	["Bash","Read","Edit","Grep","Glob"]
routing	{"triggers":["retro stats","list learnings","graduate knowledge","learning stats","search learnings","retro clear","clear learnings","prune learnings","clear stale learnings","clear learning noise"],"category":"meta-tooling","pairs_with":["learn","auto-dream"]}

Retro Knowledge Skill

Overview

This skill wraps scripts/learning-db.py into a user-friendly interface for the learning system. The learning database is the single source of truth—all queries go through the Python CLI, never maintaining a parallel file store.

---

Instructions

Parse the user's argument to determine the subcommand. Default to status if no argument given.

Argument	Subcommand
(none), status	status
list	list
search TERM	search
graduate	graduate
what-didnt-work	what-didnt-work
clear	clear

Subcommand: status

Key constraint: Always present results in readable tables/sections, not raw JSON. When showing stats, suggest next actions (search, graduate).

Show learning system health summary.

Step 1: Get stats.

python3 ~/.claude/scripts/learning-db.py stats

Step 2: Present status report. Present high-confidence counts with the category breakdown; error spam inflates per-row confidence (pruning 5119 noise rows dropped high-conf 4201→564, 2026-06-12).

LEARNING SYSTEM STATUS
======================

Entries:     [total] ([high-conf] high confidence)
Categories:  [breakdown by category]
Graduated:   [N] entries embedded in agents/skills

Injection:
  Hook: session-context.py (SessionStart, ADR-147 dream system)
  Method: pre-built payload from nightly auto-dream cycle + learning.db high-confidence patterns

Next actions:
  /retro list              — see all entries
  /retro search TERM       — find specific knowledge
  /retro graduate          — embed mature entries into agents

Subcommand: list

Display all accumulated knowledge.

Key constraint: Output must use the Python CLI as the single source of truth. Do not maintain parallel markdown files. Present results in readable grouped format, not raw JSON.

Step 1: Query all entries.

python3 ~/.claude/scripts/learning-db.py query

Step 2: Present grouped by category:

LEARNING DATABASE
=================

## [Category] ([N] entries)
- [topic/key] (conf: [N], [Nx] observations): [first line of value]
...

Optional flags:

• --category design — filter to one category
• --min-confidence 0.7 — only high-confidence entries

Subcommand: search

Full-text search across all learnings.

Step 1: Run FTS5 search.

python3 ~/.claude/scripts/learning-db.py search "TERM"

Step 2: Present results ranked by relevance:

SEARCH: "TERM"
==============

[N] results:

1. [topic/key] (conf: [N], category: [cat])
   [value excerpt]

2. ...

Subcommand: graduate

Evaluate learning.db entries and embed mature ones into agents/skills.

Key constraints:

• Only graduate entries that encode non-obvious, actionable knowledge—never generic advice.
• Always present proposals and wait for user approval before editing agent/skill files.
• Do not auto-graduate without explicit user approval (even with --auto flag, confirm intent).
• Skip categories error and effectiveness—those are injection-only (useful in context but not suitable as permanent agent instructions).

Step 1: Get graduation candidates from the DB.

python3 ~/.claude/scripts/learning-db.py query --category design --category gotcha

Step 2: For each entry, evaluate graduation readiness.

For each candidate, the LLM:

• Reads the learning value
• Searches the repo for the target file (grep for related keywords)
• Determines edit type: add failure mode, add to operator context, add warning, or "not ready / keep injecting"
• Checks if the target already contains equivalent guidance (use Grep to verify before proposing)

Question	Pass	Fail
Is this specific and actionable?	"sync.Mutex for multi-field state machines"	"Use proper concurrency"
Is this universally applicable?	Applies across the domain	Only applied in one feature
Would it be wrong as a prescriptive rule?	Safe as default	Has important exceptions
Does the target already contain this?	Not present	Already equivalent

Step 3: Present graduation plan to user.

GRADUATION CANDIDATES (N of M entries)

1. [topic/key] → [target file] (add anti-pattern)
   Proposed: "### AP-N: [title]\n[description]"

ALREADY APPLIED (N entries — mark graduated only)
- [topic/key] — already in [file]

NOT READY (N entries — keep injecting)
- [topic/key] — [reason]

Approve? (y/n/pick numbers)

Step 4: On user approval, apply changes.

Use the Edit tool to insert graduated content into target agent/skill files.

After embedding, mark the entry as graduated:

python3 ~/.claude/scripts/learning-db.py graduate TOPIC KEY "target:file/path"

Graduated entries stop being injected (the injector filters graduated_to IS NULL).

Step 5: Report.

GRADUATED:
  [key] → [target file] (section: [section])

Entries marked. They will no longer be injected via the hook
since they are now part of the agent's permanent knowledge.

Subcommand: clear

Remove noise from learning.db: cross-domain rows (via filtered prune) or old low-confidence rows (via stale-prune). Wraps the existing learning-db.py prune / stale-prune CLI — no new deletion code path.

Key constraints:

• Dry-run by default, always. Never pass --apply/--confirm on the first invocation of a session, regardless of how the user phrased the request ("clear the voice noise", "prune this", "clean up learnings").
• Show the dry-run counts and a sample of matched rows, then stop and ask the user to explicitly confirm ("apply" / "yes, clear it" / "confirm") before re-invoking with --apply or --confirm.
• Never infer confirmation from the original request alone — "clear the voice category" is a request to see what would be cleared, not authorization to delete.
• This command mutates learning.db. Do not use it to satisfy a code-level bug fix task that explicitly excludes data mutation — check the task's constraints before running --apply/--confirm.

Step 1: Determine the clear mode from the user's argument.

User intent	Mode	Underlying command
"clear category X" / "clear topic X" / has --category, --topic, --max-confidence, or --older-than	filtered	learning-db.py prune
"clear stale" / "clear old" / no filter given	stale	learning-db.py stale-prune

Step 2: Run the dry-run (always, unconditionally, first).

# Filtered mode
python3 ~/.claude/scripts/learning-db.py prune --category CATEGORY [--topic TOPIC] [--max-confidence N] [--older-than DAYS] --dry-run

# Stale mode
python3 ~/.claude/scripts/learning-db.py stale-prune --dry-run [--min-age-days DAYS]

Step 3: Present the dry-run result and stop.

RETRO CLEAR — DRY RUN
======================

Mode: [filtered | stale]
Filter: [category=X, topic=Y, ...] or [min-age-days=N]

Matched: [N] entries
  - [topic/key] (conf: [N], age: [N]d)
  ...

This is a preview — nothing was deleted. Reply "apply" (or "confirm") to
actually remove these entries, or refine the filter and re-run.

Step 4: Only after the user explicitly confirms in this turn, re-run with --apply (filtered) or --confirm (stale):

python3 ~/.claude/scripts/learning-db.py prune --category CATEGORY ... --apply
# or
python3 ~/.claude/scripts/learning-db.py stale-prune --confirm [--min-age-days DAYS]

Step 5: Report the outcome.

CLEARED: [N] entries removed ([mode] mode, filter: [...])
Total learnings: [before] -> [after]

Graduated entries and routing/effectiveness rows are always protected from prune (see scripts/tests/test_learning_db_prune.py). stale-prune archives to learning_archive rather than hard-deleting, and likewise skips graduated rows.

Subcommand: what-didnt-work

Print the negative-results registry, the list of experiments that lost. Read it before re-running an experiment so a known-dead path is not retried.

The registry is a doc, not a DB table: docs/what-didnt-work.md is capture, store, and query target. This subcommand reads and prints it, then offers an optional one-line mirror into learning.db for FTS search.

Step 1: Read and print the registry.

Use the Read tool on docs/what-didnt-work.md and present it. Group by the dated ## YYYY-MM-DD headings; show each entry's Decision verdict (rejected / deferred / revisit-if) up front so a scan answers "did we already reject this?".

NEGATIVE RESULTS (docs/what-didnt-work.md)
==========================================

## [date] [experiment]
  Decision: [rejected | deferred | revisit-if <condition>]
  What happened: [one line]
...

If the file is missing, report that no negative results are recorded yet and point the user at the format in CONTRIBUTING.md.

Step 2 (optional): Mirror one line into learning.db for full-text search.

The doc stays canonical. The mirror is one pointer row, not a parallel store. Run only when the user wants the entry FTS-searchable via /retro search:

python3 ~/.claude/scripts/learning-db.py learn --topic negative-results \
  "YYYY-MM-DD <experiment>: <decision> - see docs/what-didnt-work.md"

Batching learn calls: run learning-db.py learn calls individually or chained with &&, then confirm via learning-db.py search. A single failing command in a plain multi-line Bash batch silently drops the rest (observed 2026-06-12).

This reuses the existing learn command (no new code). Confirm with either:

# Topic listing (exact, includes the hyphen):
python3 ~/.claude/scripts/learning-db.py query --topic negative-results
# Or FTS (use a space, not the hyphen; the tokenizer splits hyphens):
python3 ~/.claude/scripts/learning-db.py search "negative results"

---

Examples

Example 1: Quick health check

User says: "/retro" Actions: Run learning-db.py stats, show entry counts, injection health.

Example 2: See what we know

User says: "/retro list" Actions: Run learning-db.py query, display grouped by category.

Example 3: Search for specific knowledge

User says: "/retro search routing" Actions: Run learning-db.py search "routing", display ranked results.

Example 4: Graduate mature knowledge

User says: "/retro graduate" Actions: Query design/gotcha entries, evaluate each against graduation criteria, propose edits to target agents/skills, apply approved changes, mark graduated.

Example 5: Clear cross-domain noise

User says: "/retro clear category voice" Actions: Run learning-db.py prune --category voice --dry-run, present the matched count and sample rows, stop and wait for explicit confirmation. Only on "apply"/"confirm" from the user, re-run with --apply and report before/after totals.

---

Error Handling

Error: "learning.db not found"

Cause: Database not initialized yet Solution: Report that no learnings exist yet. Hooks auto-populate during normal work.

Error: "No graduation candidates"

Cause: No design/gotcha entries, or all already graduated Solution: Report the stats and suggest recording more learnings via normal work.

Common Mistakes During Graduation

• Graduating generic advice (e.g., "use proper error handling"): Creates noise. Agents already know general patterns. Only graduate specific, actionable findings that encode something non-obvious.
• Proposing without target verification: Always grep the target file for equivalent guidance before proposing. Duplication creates maintenance burden.
• Proceeding without explicit user approval: Graduation permanently changes agent behavior. Always present proposals in Step 3 and wait for explicit approval before applying changes in Step 4.

---

References

• ~/.claude/scripts/learning-db.py — Python CLI for all database operations, including prune and stale-prune (wrapped by the clear subcommand)
• hooks/session-context.py — Hook that injects the pre-built dream payload and high-confidence patterns at session start (ADR-147, supersedes retro-knowledge-injector.py)
• hooks/pretool-learning-injector.py — PreToolUse hook that queries learning.db for tool-error hints; scoped to error/gotcha/debug categories (ADR: pretool-injector-scoping)
• scripts/learning.db — SQLite database with FTS5 search index
• docs/what-didnt-work.md: Negative-results registry. Printed by the what-didnt-work subcommand; the doc is the canonical store.