feat(selflearn): L1 per-turn background reviewer implementation (#52)

[yanmxa]

L1 self-learning implementation for #52, the per-turn background reviewer
designed in PR #58. Wires the feature end-to-end: trigger core →
restricted fork → memory + skill stores → action-permission gates →
runtime UI surfaces.

What's in this PR

Core (internal/selflearn/)

• Reviewer (per-turn trigger) — two arms with independent enable +
  cadence, single-flight drop-don't-reset, defensive snapshot copy before
  the goroutine, hydration on session resume.
• MemoryStore (memory_write tool surface) — project-partitioned
  store under ~/.gen/projects/<encoded-cwd>/memory/, atomic temp+rename
  writes, per-file 25 KB cap matching the injection cap (§4.2 invariant),
  injection scanner rejecting prompt-injection / exfiltration payloads.
• SkillManager (skill_manage tool surface) — six actions (create,
  edit, patch, write_file, remove_file, delete) gated by Deny-encoded
  permission flags + one advanced opt-in for patching user-authored
  skills; injection scanner on bodies + descriptions + patch results +
  support files.
• RunReview (fork mechanics) — fresh core.Agent inheriting the
  parent's cached system prompt verbatim for prefix-cache parity (§6
  invariant push #2), restricted toolset whitelist + action-level allowlist,
  5-min wall-clock deadline, best-effort + recover, single-flight.
• Three review prompts (memory / skill / combined) implementing the
  full §5.1 trigger conditions — UPDATE > DELETE > CREATE preference,
  CREATE four required criteria, the "user voiced a correction" UPDATE
  path, eviction-first memory framing with the cap-near retire loop.
• ResolveSettings bridges setting.SelfLearn → runtime values
  (Config, ActionPermissions, MemoryMaxChars) with Validate() up front
  so an illegal §3.1 combination is rejected at startup.

Config (internal/setting/)

• SelfLearnSettings + SelfLearnMemory + SelfLearnSkills types
  with Deny-encoded permission booleans (zero ⇒ allow, Go idiom).
• Validate() enforces the three illegal combinations from §3.1
  (no-update-with-create, no-delete-when-create+update,
  advanced-opt-in-without-base).
• UpdateSelfLearnAt(cfg, userLevel) persists the panel's edits
  through the existing atomic merge path.

Wire-up (internal/app/)

• wireSelfLearn at session start; reads setting.SelfLearn, applies
  the GEN_DISABLE_SELF_LEARN env override (§3.1 escape hatch),
  builds stores + manager + reviewer, hooks write observers, captures
  build params for cache-parity-preserving fork construction.
• OnTurnEnd forwards every Result to the reviewer; cancelled /
  interrupted turns are silently skipped inside Observe.
• StopAgentSession clears the reviewer reference; in-flight goroutines
  finish on the 5-min deadline.
• memory-auto reminder provider registered alongside memory-user
  and memory-project so agent-written entries are injected as their
  own scope and never mixed with user-authored GEN.md / CLAUDE.md.

User-visible surface (internal/app/selflearn_ui.go + UI hooks)

• SelfLearnUIState four-phase state machine (idle / reviewing / done
  / failed) with target debounce and auto-decay holds (2s / 3s).
• Status-bar evolving <braille> <target> indicator using the shared
  kit.BrailleSpinnerFrames table; hidden in both idle and disabled
  states (status line is pixel-identical to a feature-off session).
• In-stream recap block built from the action log (the actual tool
  calls captured via observers), not the model's narration — the recap
  reads as structurally accurate data, not LLM self-report.
• /config slash command opens a Self-Learning panel with editable
  rows for both arms, inline validation, scope toggle between user-wide
  and project-local saves.

Tests

• -race clean across internal/selflearn, internal/setting,
  internal/reminder, internal/app, internal/app/conv,
  internal/app/input, internal/app/kit.
• Coverage includes: trigger cadence (turns / iters / combined),
  interrupted-turn skip, concurrency cap behaviour, restricted-toolset
  enforcement, permission gate (every legal §5.5 combination +
  startup-rejection of the three illegal ones), allowUpdateUserCreated
  scope extension, injection scan on every memory + skill write path,
  recap-block formatting + verb mapping, state-machine transitions
  including done-hold decay survival across DrainActions, snapshot
  isolation under concurrent Observe.

What's NOT in this PR (per the design)

• §8 L2 curator — explicitly deferred to its own issue
• §10 user-level auto-memory (cross-project persona) — open question
• §10 topic-file count cap — L2's job
• §10 cache parity verification on non-Anthropic providers — needs
  real-provider testing
• §10 usage telemetry — Phase 2 prereq

Manual e2e

Not yet run against a real LLM. Reviewers / users who pull this branch
can exercise by setting in settings.json:

{
  "selfLearn": {
    "memory": { "enabled": true, "everyTurns": 3, "maxKB": 25 },
    "skills": { "enabled": true, "everyToolIters": 5 }
  }
}

Then watch for evolving in the bottom bar after 3 user turns / 5
cumulative tool iters, and the recap block appearing in the
conversation flow on completion. Disk artifacts will land in
~/.gen/projects/<encoded-cwd>/memory/ and ~/.gen/skills/<name>/ (or
./.gen/skills/ for project-level entries).

Design alignment

Design note (#58, now in main) tracked at every step; cross-references
in code comments use the stable §-anchors. Highlights of where this PR
deviates from the merged design and why:

• Permission flags renamed from allowCreate / allowUpdate / allowDelete
  (*bool) to denyCreate / denyUpdate / denyDelete (bool) — zero
  value should be the safe default. The design note was updated to
  match in the same PR that landed the rename (commit e2780e2).
• Inventory() reads SKILL.md directly from disk instead of via
  skill.Registry. The registry is loaded once at startup and only
  refreshed via ReloadPluginBackedState; a registry-backed Inventory
  would not see skills the reviewer created earlier in the same
  session, causing it to re-create duplicates. The disk scan is the
  correct freshness for this caller. Documented inline.

Closes #52 (implementation).

🤖 Generated with Claude Code