continual-learning: write to user-scoped memory file by default

[yuranich]

Summary

Make the continual-learning plugin safe to enable inside team repositories.

Today the plugin defaults to writing learned bullets into the workspace
AGENTS.md and keeping its cadence + incremental-index state under
.cursor/hooks/state/. On a shared repo both of those paths are owned by
every contributor, so:

• personal preferences end up committed into the team's instructions, and
• the two state JSON files permanently show up in git status as untracked
  noise (or get committed by accident).

This PR moves the defaults to user-scoped locations under
~/.cursor/projects/<workspace-slug>/ — the same directory layout Cursor
already uses for agent-transcripts/ — and keeps any write inside the repo
strictly opt-in.

What changes

• New defaults
  • Memory file: ~/.cursor/projects/<slug>/AGENTS.local.md (personal,
    never committed). Cursor's stacked AGENTS.md loading picks it up
    automatically.
  • State dir: ~/.cursor/projects/<slug>/continual-learning/ with
    cadence.json (hook state) and index.json (incremental index).
• Opt-in shared file via CONTINUAL_LEARNING_WORKSPACE_FILE. If set,
  the hook runs git check-ignore against it and refuses to write when the
  file is inside a git repo and not gitignored. The check can be bypassed
  with CONTINUAL_LEARNING_ALLOW_SHARED=1.
• Explicit section routing
  • ## Learned User Preferences → user file only.
  • ## Learned Workspace Facts → workspace file when configured and not
    blocked.
  • ## Learned Workspace Facts (local) → user file fallback heading
    when no workspace file is configured or the workspace file is blocked.
• Hook embeds resolved paths in followup_message so the
  agents-memory-updater subagent gets concrete absolute paths and the
  block reason (if any) instead of having to recompute them.
• First-run migration copies any pre-existing
  .cursor/hooks/state/continual-learning{,-index}.json into the new
  user-scoped location, keeps a backup under the user state dir, and
  removes the originals so they stop appearing in git status.
• New env vars (all optional): CONTINUAL_LEARNING_USER_FILE,
  CONTINUAL_LEARNING_WORKSPACE_FILE, CONTINUAL_LEARNING_STATE_DIR,
  CONTINUAL_LEARNING_ALLOW_SHARED. Existing cadence vars and legacy
  CONTINUOUS_LEARNING_* aliases are unchanged.

Files

• continual-learning/hooks/continual-learning-stop.ts — use resolved
  targets for state paths; build a structured followup message; migrate
  legacy state once on startup.
• continual-learning/hooks/lib/memory-targets.ts (new) — slug
  computation, env-driven path resolution, git check-ignore safety check,
  legacy-state migration. ~180 lines, no new dependencies (node:fs,
  node:os, node:path, node:child_process only).
• continual-learning/agents/agents-memory-updater.md — describe the
  dual-target write model, the section-routing rules, the safety check, and
  the local-fallback heading. Prefers the embedded paths from the hook when
  present.
• continual-learning/skills/continual-learning/SKILL.md — header wording
  updated; orchestration unchanged.
• continual-learning/README.md — new "Safe defaults for team repos" and
  expanded configuration sections.
• continual-learning/CHANGELOG.md (new) — Unreleased entry following
  the format used by sibling plugins.

Backwards compatibility

• Default memory file path changes from workspace AGENTS.md to
  ~/.cursor/projects/<slug>/AGENTS.local.md. Existing users who relied on
  the old behavior can restore it by setting
  CONTINUAL_LEARNING_WORKSPACE_FILE=<absolute path to AGENTS.md> and
  CONTINUAL_LEARNING_ALLOW_SHARED=1.
• All existing env vars (CONTINUAL_LEARNING_MIN_TURNS, trial mode flags,
  the CONTINUOUS_LEARNING_* legacy aliases) keep their current meaning.
• Existing cadence/index files in .cursor/hooks/state/ are migrated
  automatically on the next hook run.

Test plan

• node scripts/validate-plugins.mjs — passes (no manifest changes).
• bun build continual-learning/hooks/continual-learning-stop.ts —
  bundles cleanly.
• Cold run (no transcript advance) — hook outputs {} and does not
  trigger the followup.
• Forced trigger with no workspace file configured — followup message
  contains the resolved user file path and routes workspace facts to
  the user file under ## Learned Workspace Facts (local).
• Forced trigger with CONTINUAL_LEARNING_WORKSPACE_FILE pointing at a
  tracked file — followup message marks the workspace target as
  BLOCKED with the git-leak reason and routes workspace facts to the
  user file under ## Learned Workspace Facts (local).
• Same scenario with CONTINUAL_LEARNING_ALLOW_SHARED=1 — followup
  message labels the workspace target as shared, may be committed and
  allows the write.
• First-run migration on a workspace that has the old
  .cursor/hooks/state/ files — files are moved to the user-scoped
  location, a backup is left there, and the originals are removed from
  git status.

Made with Cursor

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit 8039a82. Configure here.}