learn

Learn

Capture behavior-changing knowledge in Mnemosyne. This workflow always creates a branch, signed commit, push, and pull request. It never writes directly to a default branch and never treats a local-only edit as success.

Prepare the knowledge repository

Prepare Mnemosyne at $HOME/.agent_brain/knowledge by following the canonical dependency-resolution contract exactly. Do not restate its owner precedence, trust gates, checkout rules, or revalidation requirements in this skill. Report the exact repository, commit SHA, and trust basis. Any resolution, authentication, checkout, update, or revalidation failure is blocking.

Before writing

1. Run advise with the proposed lesson.
2. Search flat skills/*.md, excluding optional notes, for semantic overlap.
3. Search Git history for prior consolidation and provenance.
4. Query open pull requests in the resolved Mnemosyne repository by candidate title and changed skills/<name>.md path. If one already changes the canonical entry, inspect it and either stack on its branch with explicit authority or stop for user direction; never create a conflicting PR.
5. Amend the canonical entry when its intent matches; create a new entry only for a distinct search intent. If the proposed lesson contains no material knowledge or verification change, fail with no learnable change before mutating anything. Do not report learn as completed.

Repository audits belong in repo-review; PR audits belong in pr-review; review depth is a mode, not another skill.

Learn records verified knowledge; it does not embed executable Athena behavior in Mnemosyne. When a lesson requires an Athena implementation, first make that change through normal Athena development: put each Bash or Python helper in skills/<name>/scripts/, reference it from the owning SKILL.md, and add executable behavior tests under tests/unit/. Never paste an inline Bash or Python program into skill Markdown. Run the complete Athena gates before learning the verified result through the mandatory Mnemosyne PR.

Athena skill guidance must follow ../../docs/policies/development.md. Do not teach agents to create prose-string tests, documentation snapshots, manually maintained changelogs, generated documentation, duplicated registries/catalogs/inventories, or unrelated files. Tests must exercise computable behavior or executable artifact contracts and fail for the defect they claim to detect. Apply KISS, YAGNI, TDD, DRY, SOLID, modularity, and least astonishment when deciding whether a lesson should cause repository work at all.

External-write authority checkpoint

Before creating a branch or worktree, editing Mnemosyne, committing, pushing, or opening the mandatory pull request, establish explicit user authority for the resolved repository and the complete branch, commit, push, and PR workflow. A direct user request to invoke learn supplies that authority. An indirect recommendation or invocation by another skill does not: show the repository, trust basis, base revision, proposed branch, intended files, and PR target, then obtain explicit user approval before mutation.

Read-only resolution, search, and planning do not authorize later mutation. If authority is absent, stop before creating mutable state and report that Learn has not run successfully. Once authorized, the workflow may not substitute a local-only edit for its mandatory PR outcome.

Delegation and integration

When the host supports subagents, partition independent discovery, overlap analysis, drafting, and verification into bounded work items. Run dependency-independent items concurrently in the background, up to the host's safe limit. If delegation or background execution is unavailable, run the same items sequentially without weakening their evidence requirements.

Every writing subagent receives an isolated worktree based on the same resolved Mnemosyne default- branch revision and an explicit, non-overlapping file ownership set. Each canonical knowledge entry belongs to the coordinator or one designated integration item. Read-only agents may inspect shared evidence but must not edit it. Stop concurrent work on any ownership overlap, changed base revision, or unexpected scope. When the host does not provide native isolation, retain the resolved Mnemosyne checkout as the current working directory and invoke Athena's tested ../git-worktrees/scripts/prepare_worktree.py by its absolute path resolved from this skill directory, with the exact skill/<slug> branch, --path $HOME/.agent_brain/worktrees/knowledge-<slug>, --path-root $HOME/.agent_brain/worktrees, and --start-point <resolved-default-SHA>.

The coordinator reviews each result and diff, rejects unrelated edits, and integrates accepted work sequentially into the single delivery worktree described below. Run focused validation after each integration and the resolved repository's complete relevant validation after the combined result. Only the coordinator performs the authorized commit, push, and PR creation.

Isolated write contract

Never modify the shared checkout's active worktree. From its fetched default branch:

1. Derive slug and name using only lowercase ASCII letters, digits, and single hyphens. Require the pattern [a-z0-9][a-z0-9-]*; reject /, .., leading -, control characters, and empty values. Add a short collision-resistant suffix when the branch or worktree already exists.
2. Create an isolated worktree under $HOME/.agent_brain/worktrees/knowledge-<slug> on branch skill/<slug>. Resolve the path before creating it and require it to remain directly beneath $HOME/.agent_brain/worktrees; reject symlinked parents or destinations.
3. Write skills/<name>.md only after resolving the destination and proving it remains directly beneath the worktree's skills/ directory. Include name, searchable description, category, date, semantic version, verification level, tags, when-to-use, verified workflow, failed attempts, results, parameters, and evidence.
4. For an amendment, update only the canonical entry. Git and pull-request history provide provenance; optional raw evidence may be added to .notes.md only when a current consumer needs it.
5. Run the resolved Mnemosyne repository's own validation and tests.
6. Verify no duplicate intent or stale consolidated name was introduced.
7. Commit with a cryptographic signature and DCO sign-off, push the feature branch, and open a PR against the resolved repository's default branch. The PR body must contain Closes #N when a tracking issue exists.
8. Report the PR URL and exact validation evidence. Do not auto-merge.

If a push or PR cannot be created, preserve the isolated worktree and report the blocker. A Learn run is successful only when it returns a PR URL. Never fall back to writing inside Athena or a different repository.

Preserve all delegated and delivery worktrees until their unique work is integrated or explicitly rejected. Cleanup is a separate mutation: remove only worktrees created by this Learn invocation, only with user authority, and only after rechecking that no uncommitted or unintegrated state remains. Otherwise report each worktree's path, owner, revision, cleanliness, and integration state and leave it intact. Never delete branches, discard changes, force removal, or touch a pre-existing worktree.