---
digest: >
  v0.36.3 is a patch release of one Change Request, CR160, with CR159's docs.
  On Pi, a skill picked from the command menu, or typed as `/skill:<name>`,
  reached Mirror after Pi had expanded it into the skill's whole body, and
  Mirror stored that body as something the person said. Pi's turns now go
  through the rule every other runtime has: a skill invocation and its reply
  are not recorded, and a question asked with `/mm-mirror` is recorded as the
  question. `mirror conversations skill-turns` repairs what earlier versions
  stored, only on `--apply` and only after a verified backup. No migration.
---

# v0.36.3 — A Skill You Invoke Is Not Something You Said

**Date:** 2026-10-23

## Highlights

- **On Pi, a skill you invoke is not recorded as your words**
  ([CR160](../project/refinement/rs010-cv22-oracle-and-port-hygiene/cr160-a-pi-skill-invoked-as-skill-is-logged-as-its-body.md)),
  whether you pick it from the menu or type it, and neither is its answer. A question
  you ask with `/mm-mirror` is recorded as the question, with one reply. A prompt has
  one reply row.
- **`mirror conversations skill-turns` repairs what an earlier version stored.** It
  shows, by skill, what it would change, and each conversation it would delete because
  no user turn would be left; on `--apply` it takes a dated backup, verifies it, and
  repairs in one transaction. Mirror questions keep their question; other skill turns
  go; a conversation left with no user turn goes, its memories kept, unlinked; a title
  that came from a skill's body is rebuilt from the first remaining user turn.
- **Pi joins the rule the other runtimes already had.** Mirror's Pi extension sends
  what you typed, read before Pi expands a skill, to the same prompt and reply handlers
  Claude Code, Codex, and Grok Build use.
- **The docs say how Pi invokes a skill**
  ([CR159](../project/refinement/rs010-cv22-oracle-and-port-hygiene/cr159-pi-skills-are-spelled-by-their-bare-name-and-step-7a-lost-two-dashes.md)):
  typing `/mm-` opens Pi's command menu, which lists each skill as `skill:mm-<skill>`;
  pick it there.

## Where We Started

Writing down how extensions reach each runtime (CV21.DS5.TS4), the docs called Pi's
skill commands `/mm-<skill>`, as if Pi registered them. CR159 measured Pi 1.4.0: Pi
registers each skill as `/skill:<name>`, its menu offers it under that name, and it
expands the skill into its whole body before any extension sees the prompt. Mirror's
Pi extension skipped a prompt only when it opened with `/`. An expanded skill opens
with `<skill`, so it was stored as the person's turn. On the maintainer's two homes,
225 and 185 such turns, from April to October: conversations titled
`<skill name="mm-build" location=…`, and extraction reading a skill's instructions as
something the person said.

## What Changed

**One rule, in one place.** The core already held the rule: what you say to the mirror
is remembered, a skill you invoke is not. Claude Code, Codex, and Grok Build reach it
through the hooks' prompt and reply handlers; Pi's extension kept a check of its own,
and Pi's expansion slipped past it. Rather than teach the extension Pi's wrapper, Pi
now uses the same handlers: the extension reads what you typed from Pi's documented
`input` event, gives each run a turn id, and sends the prompt's whole reply once, when
Pi settles. The payload travels as an argument, because Pi's extension API cannot
write a process's input; Pi's prompts already travelled that way.

**Measured before and after.** A route committed with the Change Request runs the same
four prompts on Pi 1.4.0 against the tree before the fix and after it. Before, a skill
picked from the menu or typed as `/skill:` was stored as the person's turn, a question
typed with `/mm-mirror` was lost, and every skill's reply was kept. After, a skill
invocation leaves nothing, and a Mirror question leaves the question and one reply.

**A repair, careful by construction.** Shaped as v0.36.2's `remove`: a dry run first,
with the exact command that applies it; on `--apply`, the set decided read-only, then a
dated archive, verified, then one transaction that decides again and refuses if
anything moved. It matches only Pi's whole skill block, and the one the old logger cut
at its 50,000-character limit, and only in Pi conversations: a person may paste
`<skill name=` anywhere. On copies of the maintainer's two homes it kept 20 Mirror
questions in each, removed 205 and 165 skill turns, deleted 31 and 24 conversations
left with no user turn, and a second dry run found nothing.

**A review before any code.** Five personas reviewed the plan. The engineer found that
the hooks already had the rule and a tested handler for each side of a turn, so Pi
joined them rather than the core learning a second copy. The devops engineer found that
a running Pi session keeps the extension it loaded, so the old logging commands stay
for one more release, and the repair is run after every session has restarted.

### What you will notice

- In Pi, a skill you pick or type no longer appears in your conversations as something
  you said, and neither does its answer.
- A question you ask with `/mm-mirror` is recorded as the question; typed in full, it
  used to be lost.
- A new command, `mirror conversations skill-turns`.

## Upgrading

```bash
mirror runtime update          # or: npm install -g mirror-mind
```

There is no database migration, and no runtime wiring changes. Then, for every home
you use with Pi:

1. Restart every Pi session on that home. A session keeps the extension it loaded and,
   until it restarts, stores a skill's body as before.
2. See what the repair would change, then repair:

   ```bash
   mirror conversations skill-turns
   mirror conversations skill-turns --apply
   ```

The dry run lists each conversation it would delete, with its message count. Running it
again later repairs anything a session that had not restarted stored meanwhile. The
dated backup it takes, and the pre-write backup, keep copies until they age out (30
days for an archive).

## Conscious Exclusions

- **Past replies to skipped turns.** A reply to a skill typed as `/mm-<skill>` was
  recorded before this release, and cannot be told from a reply to an ordinary turn,
  since one prompt can record more than one reply. Those replies stay.
- **Summaries, their embeddings, and memories** of affected conversations stay: they
  describe a session in which a skill was invoked.
- **Pi's own session files** are not read or touched.
- **Older weak titles.** Startup maintenance improves only the titles of the most
  recently ended conversations; the rest wait for
  [CR162](../project/refinement/rs010-cv22-oracle-and-port-hygiene/cr162-maintenance-retitles-only-the-latest-conversations.md).
- **A tool for models.** No MCP tool and no skill runs the repair: a person does.

## What We Learned

A rule kept in two places drifts at the edges. The hooks' rule and the Pi extension's
check agreed for months, until what Pi hands an extension stopped opening with `/`.
The second copy kept the letter of the rule after its input changed shape; one copy,
reached by every runtime, cannot.

The data that says a thing is wrong is often already there. Four hundred rows titled
`<skill name=…` sat in two homes for six months, read by every title, summary, and
extraction pass, and named by none of them.

## Next Horizon

CV21.DS11, the contract and docs refresh with the parity matrix, then CV21's own
release. Captured and open: CR161, the Pi extension's types and the check it lacks;
CR162, the titles maintenance never reaches. `log-user` and `log-assistant`, kept here
for sessions that have not restarted, retire in the release after this one.
