| name | retro-meta |
| description | Deep retro of retros - investigation pass that surfaces improvements across older retros and the current ax setup. External AI agent drives the reasoning with high thinking enabled. Triggers when user says "deep retro", "investigate ax", "what should I fix in my setup", "retro of retros", or invokes /ax:retro-meta. Use after `ax:retro` if proposals are sparse or you want broader exploration. |
ax:retro-meta - deep retro of retros
The companion to ax:retro. Where ax:retro walks the heuristic-derived
proposals one by one, retro-meta asks: what improvements does the
existing pipeline NOT yet see?
The external agent (this one, Claude Code or Codex with high thinking)
drives the reasoning. The CLI just produces structured evidence and
takes user-approved plans back.
When to fire
Explicit triggers only:
- "let's do a deep retro" / "retro of retros"
- "investigate my ax setup" / "what should I fix in my setup"
- "review proposals the heuristic missed"
/ax:retro-meta slash command
- After
ax:retro finishes if the user wants broader exploration
Do NOT auto-trigger on generic "look at recent work".
Prerequisites
ax (axctl) is on PATH and the local SurrealDB is reachable. If
ax doctor fails, stop and tell the user scripts/db-start.sh.
- At least 3 retros in the last 30 days. Below that, evidence is too
thin for a meta pass - recommend
ax:retro first.
Workflow
Step 1 - Snapshot
ax retro meta --json --since=30 > /tmp/ax-meta.json
Read /tmp/ax-meta.json. The keys you care about:
experiment_status[] - read this FIRST (see Step 2). Each entry:
experiment_id, proposal_dedupe_sig, proposal_title,
proposal_form, artifact_path, days_since_accepted,
opportunities_count, addressed_count, address_ratio,
latest_checkpoint{kind,suggested,observed_at}, locked_verdict.
Pending verdicts (locked_verdict=null) come first.
retros[] - raw tried/worked/failed/next per session.
patterns.tool_failures - sorted by total_count desc.
patterns.corrections - total + max-per-session + session_count.
patterns.friction_kinds - recurring kinds across sessions.
current_state.skills - what's already installed (do NOT propose
duplicates).
current_state.open_proposals - existing heuristic proposals.
current_state.accepted_experiments - accepted but verdict-pending.
current_state.claude_md_user / claude_md_project - guidance file
paths (null if absent).
investigation_prompts[] - the prompts you must walk.
Step 2 - Vet existing experiments FIRST
Walk experiment_status in order. For each entry with
locked_verdict=null:
a. If latest_checkpoint.suggested is ignored or regressed:
investigate why (read the artifact_path, sample the matching
opportunities), then run
ax improve verdict --set=<v> <proposal_dedupe_sig> to lock the
call.
b. If latest_checkpoint.suggested is adopted AND
days_since_accepted > 30: lock it as adopted so it stops
cluttering the open list:
ax improve verdict --set=adopted <proposal_dedupe_sig>.
c. If latest_checkpoint is null OR suggested is partial: leave
open. Note in the final summary that it's still gathering signal.
A rule of thumb mirrored from investigation_prompts: if
address_ratio < 0.1 after t+30, default to locking as ignored
unless the artifact has an obvious "not yet exercised" reason.
Step 3 - Walk the investigation prompts (high thinking)
For EACH prompt in investigation_prompts:
- Inspect referenced state with Read / Glob / Grep:
- skill files in
~/.claude/skills/ and ~/.agents/skills/
claude_md_user if non-null
claude_md_project if non-null
- Reason about a candidate improvement. Use a high thinking budget -
the point is to see what the heuristic missed.
- If you identify a real improvement (NOT a duplicate of an existing
skill or open_proposal):
a. Draft a plan doc to
~/.claude/plans/<YYYY-MM-DD>-<slug>.md,
30–100 lines. Sections: Problem, Evidence (cite retro ids),
Proposed change, Success signal.
b. Show the user a 4–6 line summary.
c. Ask explicitly: "Register this as an accepted experiment? (y/n)"
d. ONLY on yes:
ax retro plan \
--slug=<kebab-slug> \
--form=skill|hook|guidance|automation \
--title="<short title>" \
--hypothesis="<one sentence>" \
--plan-path=~/.claude/plans/<file>.md \
--evidence-retros=<retro:id1,retro:id2> \
--confidence=low|medium|high
- If the prompt resolves to "no change needed" or "duplicate of
existing", say so out loud and move on.
Step 4 - Optional: hand off to scaffolder
For each plan you registered, you may run:
ax improve accept --with-agent <dedupe_sig>
This spawns the internal scaffolding agent to draft an artifact
(SKILL.md, hook script, etc) from the plan. Skip if the plan is
already self-sufficient.
Step 5 - Summary
Print one paragraph:
- N plans registered, M of those scaffolded
- V verdicts locked (with kind, e.g. "2× ignored, 1× adopted")
- K open_proposals reviewed (and their disposition)
- Any prompts that resolved to "nothing here"
- Suggested next retro window
Anti-patterns
- NEVER register a plan without an explicit user yes per plan. The
human is the final filter.
- NEVER auto-accept all open_proposals - the heuristic surfaces them
but the deep pass exists precisely to triage them by reasoning, not
by frequency rank.
- NEVER write directly to
~/.claude/skills/. Use ax retro plan +
ax improve accept --with-agent.
- NEVER skip Step 2's duplicate check. Proposing a Pre-Bash guard when
one is already accepted just wastes the user's time.
- Don't trust frequency alone. A frequency=1 retro can still be
load-bearing if it represents a category Claude can't get right.
- NEVER propose a new improvement that overlaps a pending experiment.
Vet that one first - lock its verdict or escalate before piling on
more proposals in the same area. The retrospective loop is
incomplete if old experiments stay in limbo.
CLI reference
ax retro meta --since=30 [--limit-retros=50] [--pretty]
ax retro plan \
--slug=<kebab> \
--form=skill|hook|guidance|automation \
--title="<title>" \
--hypothesis="<hyp>" \
--plan-path=<path-to-plan.md> \
[--evidence-retros=retro:a,retro:b] \
[--artifact-path=<path>] \
[--confidence=low|medium|high] \
[--frequency=<N>] \
[--json]
ax improve accept --with-agent <dedupe_sig>
ax improve verdict --set=adopted|ignored|regressed|partial|no_longer_needed <dedupe_sig>
Output of ax retro meta defaults to JSON because the reader is you,
not a human.