@owlmeans/agent-skills 0.1.12 → 0.1.14

package/agent-meta/instructions/agent-memory.instructions.md

@@ -1,43 +1,139 @@
 ---
-description: "How project memory works in an OwlMeans app — where to store facts/decisions/gotchas for Copilot (.github/memory/) and Claude Code (.claude/memory/), the MEMORY.md index, and when to read/write memory. Apply when remembering something about the project or deciding where a note belongs."
-applyTo: "**/.github/memory/**, **/.claude/memory/**"
+description: "Shared agent memory protocol for the .agents/memory/ graph store (single store for Copilot and Claude Code) — MEMORY.md index, subsystem nodes, compact-on-write, size caps, read protocol. Apply when reading or writing project memory, or when any legacy .claude/memory or .github/memory path appears."
+applyTo: "**/.agents/memory/**, **/.claude/memory/**, **/.github/memory/**"
 scope: general
 ---
 <!-- AUTO-GENERATED — do not edit. Regenerate via sync-agent-meta. -->
 
-# Project memory
+# Agent memory (`.agents/memory/`)
 
-Memory is how durable, project-specific knowledge survives across sessions. Keep it **inside the
-project**, never outside the repository.
+One shared memory store per repository, used by **both** GitHub Copilot and Claude Code: a graph
+of subsystem nodes rooted in the `MEMORY.md` index. Compactness is the core value — every write
+merges and compacts; nothing is ever appended as a log.
 
-| Tool | Memory dir | Index |
-|---|---|---|
-| GitHub Copilot | `.github/memory/` | `.github/memory/MEMORY.md` |
-| Claude Code | `.claude/memory/` | `.claude/memory/MEMORY.md` |
+Never write memory to `.claude/memory/`, `.github/memory/`, `~/.copilot/`, `~/.claude/`, or
+anywhere outside the repository. Legacy `.claude/memory/` / `.github/memory/` stores are
+retired — if one exists, flag it for migration
+(`.github/instructions/memory-recompact.instructions.md`) instead of writing there.
 
-Each memory is one Markdown file holding one fact, decision, or gotcha. The `MEMORY.md` index keeps a
-one-line pointer per file and is loaded every session, so keep it current.
+## Store layout
 
-## When to read
+```
+.agents/memory/
+├── MEMORY.md     # graph root: the only always-read file
+└── <node>.md     # one file per subsystem / concern / integration
+```
 
-- **Start of every session** — read `MEMORY.md`, then open any relevant file.
-- **Before a non-trivial task** — check for and read a memory file on that topic.
+Flat directory — graph structure lives in scopes and wiki-links, not subdirectories.
 
-## When to write
+## Node naming
 
-After a task produces knowledge not obvious from the code or git history: a decision and its
-rationale, a gotcha, or project state/goals. Do not record what the repo or git history already says,
-or what only matters to the current conversation.
+Node id = lowercase-kebab name of the **project-structure part** the knowledge attaches to:
 
-## How to write one
+1. **Subsystem** — workspace dir / package name: `manager-web`, `auth`, `vslots`. Deepest
+   unambiguous segment; prefix with the parent only on collision.
+2. **Cross-cutting concern** spanning subsystems: `build`, `testing`, `deploy`, `routing`, `i18n`.
+3. **Integration** — external service/library: `stripe`, `cloudflare`, `openrouter`.
+4. **`workspace`** — repo-level facts that fit nowhere else (remotes, branch topology, env model).
 
-1. Create `.github/memory/<topic>.md` (and the Claude copy under `.claude/memory/`) with a short slug
-   and a one-line description.
-2. Write the single fact; for decisions/gotchas add why it matters and how to apply it.
-3. Add a one-line pointer to `MEMORY.md`.
-4. Update an existing file instead of duplicating; delete memories that prove wrong.
+File is `<node>.md`; the id is the wiki-link target: `[[manager-web]]`. **Never** key a node to an
+event, date, phase, investigation, or task name.
 
-## Memory vs skill
+## Index format (`MEMORY.md`)
 
-A repeatable procedure is a skill (see the `skill-authoring` instruction). Memory is for facts about
-this project.
+```markdown
+# Memory Graph — <project>
+
+Shared agent memory (`agent-memory` protocol). Read this file at session start.
+Before non-trivial work, open every node whose scope matches the task's files or topics.
+
+## Subsystems
+- [[manager-web]] `sources/manager-web/**` — i18n 8-domain split; owl-theme wiring
+## Cross-cutting
+- [[build]] `**/tsconfig*.json` — tsbuildinfo path pinning; tolerant library builds
+## Integrations
+- [[stripe]] `ext:stripe` — sandbox topology; webhook fan-out
+```
+
+Rules: exactly one line per node — `- [[<node>]] ` + backticked scope + ` — <hook>`; hook ≤ 100
+chars; no dates; only non-empty groups; file ≤ 50 lines; no skills listed (skills self-describe).
+The scope column is the router: match the task's paths/topics against scopes to decide which
+nodes to open.
+
+## Node format
+
+Frontmatter — exactly three fields:
+
+```yaml
+---
+node: vslots                 # = filename without .md; = wiki-link target
+scope: "scripts/vslots/**"   # repo-relative globs; ext:<service> for integrations; . for workspace
+updated: 2026-08             # month granularity only
+---
+```
+
+Body — only non-empty sections, in this order:
+
+- `## Facts` — what IS: structure, behavior, topology. 1–3 lines each.
+- `## Invariants` — what MUST HOLD; optionally "broke when violated: <symptom>".
+- `## Gotchas` — symptom → cause → counter-move, ≤ 3 lines each.
+- `## Pointers` — key files, promoted-skill pointer lines, `External docs` (URL + gist).
+- `## Status` — optional; in-flight state only; every line dated; prune resolved lines on every read.
+
+`Status` is the **only** place dates and unfinished-work notes are allowed. Link related nodes
+inline with `[[node]]`.
+
+## Read protocol
+
+- **Session start**: read `.agents/memory/MEMORY.md` — the index only.
+- **Before a non-trivial task**: open every node whose scope matches the task (typically 1–3).
+- **When a topic surfaces mid-session** that the index names: open its node before acting.
+- **Never bulk-read** the store — the index + scopes exist so you don't have to.
+
+## Write protocol (compact-on-write)
+
+Every write is a merge, never an append:
+
+1. **Extract** — reduce the outcome to reusable knowledge atoms (rule below). Nothing survives →
+   write nothing.
+2. **Locate** — match scopes to find the node; none matches → create one per the naming rule and
+   add its index line.
+3. **Merge / supersede** — place each atom in the right section; if it refines or contradicts an
+   existing line, **rewrite that line in place** — never append a dated correction. The node
+   always reads as current truth.
+4. **Compact** — reread the node; collapse redundancy; delete anything the code or git history
+   now states; prune resolved `Status` lines.
+5. **Cap** — over soft cap: compact harder. Over hard cap: split by sub-scope into a linked child
+   node, or promote procedure-shaped overflow
+   (`.github/instructions/memory-promotion.instructions.md`, else the skill-authoring guidance).
+6. **Index** — update the hook line if the node's center of gravity moved; bump `updated:`.
+
+## Knowledge, not events
+
+**Record the rule, not the story.** For every line ask: *what must a future agent know to act
+correctly, stated without reference to this session?* Keep invariants, cause→effect,
+counter-moves, recognition fingerprints (symptoms). Drop dates, phase numbers, who did what,
+attempt sequences, and anything recoverable from code or git.
+
+Before (event log): "Verification 2026-06-11 … Bug found: script missing embedded-count logic.
+Fix deployed to all copies; second run idempotent."
+After (knowledge): "`nested-agent-context.sh` exists as byte-identical copies in the archive and
+each repo — fan every fix out to all copies; verify with `diff`. Correct re-runs are no-ops."
+
+## Size caps (hard)
+
+| Thing | Cap |
+|---|---|
+| Index entry | 1 line, hook ≤ 100 chars |
+| `MEMORY.md` | ≤ 50 lines |
+| Node file (soft / hard) | 80 / 120 lines |
+| Single fact/gotcha | ≤ 3 lines |
+| `Status` section | ≤ 5 dated lines |
+| Nodes per store (soft) | ~25 — merge low-traffic siblings beyond |
+
+## Memory vs instruction
+
+Fact-shaped ("what is true") stays here. Procedure-shaped ("to do X, do Y") becomes an
+instruction/skill — follow `.github/instructions/memory-promotion.instructions.md` where present,
+otherwise the skill-authoring guidance. Repeated use of a node to *perform* tasks, and over-cap
+nodes full of steps, are promotion triggers.

package/agent-meta/instructions/git.instructions.md

@@ -0,0 +1,52 @@
+---
+description: "Mandatory git rules — never run state-changing git without explicit instruction, commit only under the user's configured identity (never as the AI/agent, no Co-Authored-By trailer), report finished git work as a Markdown table, and never commit a conflicted working copy. Consult before any git operation."
+applyTo: "**"
+scope: general
+---
+<!-- AUTO-GENERATED — do not edit. Regenerate via sync-agent-meta. -->
+
+# Git Workflow Rules
+
+**Mandatory and universal.** These rules apply to every git operation in this repository and
+**override any default agent behavior** — including any automatic `Co-Authored-By` or
+AI/agent-attribution trailer. Consult them before running any git command.
+
+## 0. Never run state-changing git without explicit instruction
+
+- **Never run state-changing git operations** (`commit`, `add`/`rm` staging, `push`,
+  `reset`/rollback, `revert`, `rebase`, `merge`, `branch`, `checkout`/`switch`, `stash`,
+  `tag`, `cherry-pick`, force-push, etc.) unless the user **explicitly instructs it in the
+  current request**. Permission to make code edits is **not** permission to touch git.
+- **Only exception**: creating and operating inside a **temporary git worktree** that a task
+  or subagent has **explicitly requested** for that purpose. Outside such an explicitly
+  requested tmp worktree, do nothing with git.
+- **Read-only inspection is allowed**: `git status`, `git diff`, `git log`, `git show`,
+  `git branch --list`, etc. — use these to report state, never to change it.
+
+## 1. Never commit under the agent's identity
+
+- Commit only under the repository's preconfigured git identity — whatever `git config user.name`
+  and `git config user.email` resolve to (set by the user, globally or locally).
+- **Never** override authorship: do not pass `--author`, do not set or change `user.name` /
+  `user.email`, and never substitute an AI / agent / assistant name or email.
+- **Never** add a `Co-Authored-By:` trailer (or any other trailer) attributing the commit to
+  Claude, Copilot, or any AI/agent. Every commit is the user's, attributed solely to the user.
+- If the repository has no git identity configured (neither local nor global), stop and ask the
+  user — do not invent one.
+
+## 2. Always report finished git work as a table
+
+- After completing any git action (commit, push, branch, checkout, merge, rebase, stash, tag,
+  reset, etc.), summarize what was done as a Markdown table.
+- One row per action. Include at least **Action**, **Target** (branch / remote / files / ref),
+  and **Result** (commit SHA, `pushed`, `up to date`, `conflict`, …). Add a **Notes** column
+  when useful.
+
+## 3. Never commit a conflicted working copy
+
+- If a `merge`, `rebase`, `cherry-pick`, `stash pop`/`apply`, or `pull` produces conflicts,
+  **do not** finalize it: do not run `git commit`, `git merge --continue`,
+  `git rebase --continue`, or stage-and-commit the conflicted tree to "resolve" it on the
+  user's behalf.
+- Stop, list the conflicted paths in the report table, and hand control back to the user —
+  unless the user has explicitly told you how to resolve the conflict and commit.

package/agent-meta/instructions/memory-promotion.instructions.md

@@ -0,0 +1,67 @@
+---
+description: "Promote procedure-shaped or hot memory into skills/instructions — triggers (procedure test, promote? flags, over-cap), update-vs-create rule, post-promotion pointer state. Apply when converting memory content into guidance files or when a promote? flag appears in a memory node."
+applyTo: "**/.agents/memory/**, **/.claude/skills/**, **/.github/instructions/**"
+scope: general
+---
+<!-- AUTO-GENERATED — do not edit. Regenerate via sync-agent-meta. -->
+
+# Memory promotion
+
+Memory holds **facts**; **procedures** belong in instructions/skills, where they auto-apply and
+stop consuming memory-read cycles. Promotion is how the store stays compact and the harness
+teaches itself.
+
+## Triggers
+
+- **Procedure shape**: the content answers *how to do X* — ordered steps, imperative verbs,
+  command lines, if-this-do-that tables. Promote it.
+- **Repeated touch**: an existing ` → promote?` flag on content you are touching again means
+  promote now; if mid-task, promotion becomes the first row of the completion report.
+- **Over-cap**: a node past the 120-line hard cap with roughly a third or more procedure-shaped
+  content → promote the procedures instead of splitting the node.
+- Corroborating only (weak — agents don't commit, so only user commits appear):
+  `git log --oneline -- .agents/memory/<node>.md` showing ≥ 3 commits within ~a month marks a
+  hot node worth evaluating.
+
+## Procedure-shape test
+
+Procedure-shaped answers *how to do X*: ordered steps, imperative verbs, commands, action tables.
+Fact-shaped answers *what is true*: declarative structure, invariants, symptom→cause pairs.
+Mixed entries split — the fact stays in the node, the "then do" moves out.
+
+## Flagging (how content earns promotion)
+
+When you **merge** procedure-shaped content into a node, or you **use** (not merely check) a
+node's content to perform a task, append ` → promote?` to that line or section heading.
+Evaluable by reading the file alone — no tooling required.
+
+## Update vs create
+
+**Default is update** — extend the existing instruction/skill whose scope covers the activity,
+even partially; keep both twins in sync. Create a NEW pair only when:
+
+- (a) a new subsystem or technology entered the repo;
+- (b) an activity with no covering instruction needed memory read/write more than once (a
+  re-encountered ` → promote?` flag);
+- (c) an external technology required internet docs and has no governing instruction.
+
+New instructions multiply lookup cost — compactness applies to the guidance population too.
+
+## Procedure
+
+1. Collect the flagged / procedure-shaped memory lines.
+2. Author or extend the `.github/instructions/<name>.instructions.md` and its
+   `.claude/skills/<name>/SKILL.md` twin, following the repo's skill-authoring / create-skill
+   conventions.
+3. Shrink the node: delete the promoted prose; leave one pointer line (format below).
+4. Remove the ` → promote?` flags.
+5. Update the node's index hook if its main value moved; bump `updated:`.
+6. Report per the Reporting rule.
+
+## Post-promotion state
+
+The node keeps exactly one line under `## Pointers`:
+
+```
+- <activity> → skill `<name>` (procedure lives in .claude/skills/<name>/ + .github/instructions/<name>.instructions.md)
+```

package/agent-meta/instructions/memory-recompact.instructions.md

@@ -0,0 +1,72 @@
+---
+description: "Whole-store memory recompaction and migration for .agents/memory/ — rebuild node map from project structure, merge records, enforce caps, regenerate the index, fold in and delete legacy .claude/memory and .github/memory stores. Apply when restructuring a memory store or migrating legacy stores."
+applyTo: "**/.agents/memory/**, **/.claude/memory/**, **/.github/memory/**"
+scope: general
+---
+<!-- AUTO-GENERATED — do not edit. Regenerate via sync-agent-meta. -->
+
+# Memory recompaction
+
+Whole-store maintenance for `.agents/memory/` (protocol:
+`.github/instructions/agent-memory.instructions.md`). Also the migration procedure for legacy
+`.claude/memory/` + `.github/memory/` stores. Store-wide rewrite — propose it when triggers
+appear; the operator invokes it.
+
+## When
+
+- An index entry runs longer than one line, or the index exceeds 50 lines.
+- Nodes keyed by event/date/phase/task, or bodies reading as session narratives.
+- The same fact stated in two or more nodes.
+- More than ~20% of a node is stale `Status` content.
+- A node's `updated:` is months behind commits touching its scope.
+- Legacy `.claude/memory/` or `.github/memory/` dirs exist → run the migration below.
+
+## Build the target node map first
+
+Before reading any record bodies, derive the node set from **project structure**: workspaces
+array / top-level dirs → subsystem nodes; then the cross-cutting concerns and external
+integrations actually present. Write the map down (old file/section → target node). Every
+existing record must land in exactly one node — or split into atoms landing in several. Only
+then process records.
+
+## Per-record pass
+
+For each old file or section:
+
+1. Apply the agent-memory extraction rule — keep invariants, cause→effect, counter-moves,
+   symptom fingerprints; drop narratives, dates, attempt logs, anything code/git already states.
+2. Route each surviving atom to its node's section (Facts / Invariants / Gotchas / Pointers;
+   genuinely in-flight state → `Status`, dated).
+3. On conflict between records, the version consistent with **current code** wins — check the
+   code, don't average.
+4. Procedure-shaped survivors do not enter nodes — route them to
+   `.github/instructions/memory-promotion.instructions.md`.
+
+## Legacy-store merge (migration)
+
+1. Union `.claude/memory/` and `.github/memory/`. Same-named files are two drifted sources of
+   ONE node — merge both; the code-consistent version wins.
+2. Index-only entries with no backing file: extract the fact into its node, or drop if stale.
+3. Old `## Skills` / "Key Files" index sections are dropped — skills self-describe; harness
+   layout belongs to `CLAUDE.md` / `.github/copilot-instructions.md`. Move genuinely non-obvious
+   dispatch hints there.
+4. When the new store verifies (below), delete both legacy dirs entirely.
+
+## Regenerate the index
+
+Rebuild `MEMORY.md` from the resulting nodes per the agent-memory format — never edit the old
+index incrementally.
+
+## Verify
+
+- Every node file is listed in the index; every listed node exists; every wiki-link resolves.
+- All caps met (index ≤ 50 lines; nodes ≤ 120; entries ≤ 3 lines; Status ≤ 5 dated lines).
+- No dates outside `Status` and `updated:`; no event-keyed filenames.
+- Both root instruction files' Memory sections point at `.agents/memory/`.
+- Legacy dirs gone; `grep -rn '\.claude/memory\|\.github/memory'` over the repo's harness files
+  returns nothing but allowlisted historical mentions.
+
+## Report
+
+One table: **Node** | **Sources merged** | **Lines before → after**. Follow the Reporting rule
+(what, not why).

package/agent-meta/instructions/router-plugins.instructions.md

@@ -0,0 +1,46 @@
+---
+description: "How OwlMeans UI routing plugins work — the RouterPlugin contract, cascade selection, choosing/switching the router at context provisioning (default OwlMeans vs opt-in react-router), and authoring a plugin. Use before wiring routing in an app or writing a routing plugin."
+applyTo: "**/*.ts, **/*.tsx"
+---
+<!-- AUTO-GENERATED — do not edit. Regenerate via sync-agent-meta. -->
+
+# OwlMeans routing plugins
+
+OwlMeans UI routing is pluggable. `@owlmeans/router` is a **host**: it defines the `RouterService`
+facade and a `RouterPlugin` registry, selecting the active plugin by **cascade** (priority + `match`).
+Route descriptions are react-router-compatible (static, `:param`, nested, index), so the same route
+trees work across plugins.
+
+## Plugins
+
+| Plugin | Package | Register | Priority |
+|--------|---------|----------|----------|
+| OwlMeans in-browser (default) | `@owlmeans/web-router` | `appendWebRouter(ctx)` | 0 |
+| react-router v7 (opt-in) | `@owlmeans/web-router-react-router` | `appendReactRouter(ctx)` | 100 |
+| SSR (future) | — | higher-priority `match: env => env.ssr` | >0 |
+
+## Choosing the router at context provisioning
+
+`web-client`/`web-panel` register the OwlMeans plugin by default. To switch to react-router, add it
+in your `makeContext` — its higher priority wins:
+
+```typescript
+import { appendReactRouter } from '@owlmeans/web-router-react-router'
+export const makeContext = (cfg) => { const c = makeBase(cfg); appendReactRouter(c); return c }
+```
+
+Do NOT pass a `provide`/`createBrowserRouter` prop anymore — routing resolves its compiler from the
+active plugin (`context.router().compile`). The old `provide` export from `web-client` is a
+deprecated `undefined`.
+
+## RouterPlugin contract
+
+`{ alias; priority?; mode?; match?(env,ctx); compile(routes,ctx); provider(); outlet(); useParams;
+useLocation; useNavigate; useSearchParams }`. `RouteObject` (`{index?,path?,children?,Component?}`)
+is the neutral IR both plugins consume. `RouterEnv` (`{hasWindow,ssr,request?}`) drives selection.
+
+## Authoring a plugin
+
+Reuse the pure matcher (`flattenRoutes`/`rankRouteBranches`/`matchRoutes` from `@owlmeans/router`),
+expose `appendMyRouter(ctx) = ensureRouterService(ctx).registerPlugin(makeMyPlugin())`, set a
+`priority`/`match` for your environment, and keep facade methods assignable (native-safety).

package/agent-meta/instructions/scaffolding.instructions.md

@@ -19,7 +19,9 @@ npx @owlmeans/create-app my-app
 ```
 
 By default it copies the template, runs `git init`, installs dependencies, and deploys agent
-guidance via `@owlmeans/agent-skills` (`.claude/skills/` + `.github/instructions/`).
+guidance via `@owlmeans/agent-skills` (`.claude/skills/` + `.github/instructions/`). With
+`--no-install` the deploy still runs from the installer's own bundled extras, so the project gets
+its general/harness guidance; package-specific skills wait for `npx @owlmeans/agent-skills`.
 
 Flags: `--name`, `--pm <bun|npm|yarn>` (default `bun`), `--no-install`, `--no-skills`, `--no-git`,
 `--yes`. Run with `cd my-app && bun run dev` (API :3000, web :3001).
@@ -38,6 +40,11 @@ Follow Option B in `docs/getting-started.md`: bun workspace → `common` (entryp
 ```
 my-app/
 ├── package.json            # bun workspaces: sources/*
+├── CLAUDE.md               # git/reporting/memory/self-education rules + project-purpose placeholder
+├── .github/copilot-instructions.md   # same directives for Copilot
+├── .github/instructions/   # seeded harness instructions (+ deployed ones)
+├── .claude/skills/         # seeded harness skills (+ deployed ones)
+├── .agents/memory/MEMORY.md          # starter shared memory graph index (both tools)
 ├── sources/common/         # consts, types, schemas, config, modules
 ├── sources/api/            # context (appendStaticResource), app/session/*, modules, index
 └── sources/web/            # vite + tailwind v4, components/ui/*, layout, nav, screens, render
@@ -45,3 +52,17 @@ my-app/
 
 Session items live in `@owlmeans/static-resource`, namespaced by a client-generated `sid` in
 `localStorage` — no database, no authentication.
+
+## Generated agent guidance
+
+Both root files carry the same four mandatory sections as a real OwlMeans monorepo — Git Workflow,
+Reporting, Memory, Self-Education — plus the mandatory reuse-code section and a project-purpose
+placeholder (`<!-- OWLMEANS:PROJECT-PURPOSE -->`) the agent fills in on its first session.
+
+Memory lives **only** in `.agents/memory/` — the legacy `.claude/memory/` and `.github/memory/`
+starters are gone (`memory-recompact.instructions.md` migrates a project that still has them).
+
+The harness guidance is seeded into the template as generated, banner-carrying copies, so a project
+has it before the installer runs: `agent-memory`, `memory-promotion`, `memory-recompact`,
+`self-education`, `skill-authoring`, `git`, `reuse-code`, `getting-started`. Never hand-edit a seeded
+copy — regenerate with `sync-agent-meta --seed-only` in the library-manager.

package/agent-meta/instructions/self-education.instructions.md

@@ -0,0 +1,69 @@
+---
+description: "Mandatory post-development skill/instruction updating — after adding or changing functionality, update touched project guidance, note external-doc findings in the governing instruction, or add instructions for new subsystems/technologies; required before the completion report when development started from an agreed plan. Apply when editing skills or instruction files after development."
+applyTo: "**/.claude/skills/**, **/.github/instructions/**"
+scope: general
+---
+<!-- AUTO-GENERATED — do not edit. Regenerate via sync-agent-meta. -->
+
+# Self-education
+
+After functionality is added or changed, the guidance that describes it must catch up — the agent
+teaches itself for next time. This pass runs after implementation and verification, **before**
+the completion report.
+
+## When (mandatory)
+
+Development "started from planning" when the session produced a plan the operator approved before
+implementation — a plan approval, an agreed written plan/spec, or an explicit "go ahead" on a
+proposed approach. **Approval of a plan is simultaneously approval of its implicit final step:
+this self-education pass.** A planned task is not finishable without it. If guidance files cannot
+be edited (read-only run), list the required updates in the report instead.
+
+Also recommended after any unplanned change that made an existing instruction/skill inaccurate.
+
+## Review checklist
+
+For each area the work touched:
+
+1. Which existing instruction/skill covers it? (Check `.github/instructions/` +
+   `.claude/skills/`.)
+2. Do its commands, paths, APIs, and behavior claims still hold after the change?
+3. Fix in place — and keep the instruction and its skill twin in sync.
+
+## Non-project instructions
+
+If a general or imported instruction gained an important usage pattern during the work, add the
+pattern to the **deployed copy** in this repo and note it in the report as an upstream
+candidate — canonical archive copies change only on explicit operator request.
+
+## External docs
+
+If the work required reading internet documentation for an external API, library, or service,
+the governing instruction must record it under an `## External docs` heading:
+
+```
+- <URL> — <one-line gist of what it settled> (<version/date if load-bearing>)
+```
+
+Governing instruction = the project instruction covering the touched area; else the one covering
+that library; if none exists and the technology will recur, create one (update-vs-create rule in
+`.github/instructions/memory-promotion.instructions.md`). Never leave doc findings only in
+memory or the conversation.
+
+## New subsystems / technologies
+
+When new technology or a new subsystem entered the repo: create the instruction/skill pair for
+its procedures, and/or the `.agents/memory/` node for its facts — split along the
+memory-vs-instruction boundary (`.github/instructions/agent-memory.instructions.md` /
+`.github/instructions/memory-promotion.instructions.md`).
+
+## Completion gate
+
+The completion report must contain a Self-education table:
+
+| Item | Action | Path |
+|---|---|---|
+| <area/skill> | updated / created / none-needed | <path> |
+
+"none-needed" requires a one-phrase reason. A post-plan completion report without this table is
+a protocol violation.

package/agent-meta/instructions/skill-authoring.instructions.md

@@ -51,10 +51,12 @@ Cross-reference related guidance by name.
 ## Skill vs memory
 
 - **Skill / instruction** — a reusable procedure or reference worth auto-invoking.
-- **Memory** (`.claude/memory/`, `.github/memory/`) — a fact, decision, or gotcha specific to this
-  project's history. See the `agent-memory` instruction.
+- **Memory** (shared `.agents/memory/` graph store, both tools) — a fact, decision, or gotcha
+  specific to this project's history. See the agent-memory instruction; promotion triggers live
+  in the memory-promotion instruction.
 
 ## After adding a skill
 
-List it in `.github/memory/MEMORY.md` (and `.claude/memory/MEMORY.md`), and reference it from
+If it absorbed memory content, shrink the source `.agents/memory/` node to a pointer line
+(memory-promotion instruction) — the memory index does not list skills. Reference it from
 `.github/copilot-instructions.md` / `CLAUDE.md` if it should be discoverable every session.