SkillOpt-Sleep for OMP

This is the OMP thin shell for Microsoft SkillOpt-Sleep. It follows the upstream Claude Code, Codex, and Devin pattern: harvest local agent sessions, mine recurring tasks, replay them offline, gate proposed memory/skill edits on a held-out score, and stage improvements for review before adoption.

When to use

Use this when the user asks OMP to:

• learn from prior OMP sessions or repeated feedback;
• run a nightly/offline sleep, dream, or self-improvement cycle;
• inspect SkillOpt-Sleep status, harvest, dry-run, run, adopt, schedule, or unschedule;
• refine reusable Agent Skills from real OMP usage with validation gates.

Mechanism

The upstream skillopt_sleep engine currently knows Claude Code and Codex transcript sources. The bundled wrapper mirrors OMP JSONL sessions into a Claude-compatible, sanitized mirror at ~/.skillopt-sleep/omp/claude-home, then runs python -m skillopt_sleep with --source claude --claude-home <mirror>.

• Source sessions: ~/.omp/agent/sessions/**/*.jsonl
• Mirror state: ~/.skillopt-sleep/omp/claude-home
• SkillOpt state/staging: ~/.skillopt-sleep/omp/.skillopt-sleep/state.json and ~/.skillopt-sleep/omp/staging/
• Live changes: none until adopt; every adoption is backed up by the engine

Tool outputs are not mirrored. The wrapper keeps user/assistant text and assistant tool names, enough for SkillOpt to mine recurring task patterns without copying raw tool results.

Commands

Set SKILLOPT_SLEEP_REPO when SkillOpt is cloned somewhere the wrapper cannot auto-detect:

export SKILLOPT_SLEEP_REPO=/path/to/SkillOpt

Run from the project whose memory/skills should evolve:

# Safe smoke check, no proposal staged
/skillopt-sleep dry-run \
  --backend mock --max-sessions 5 --max-tasks 3 --progress

# Full cycle; stages a proposal only
/skillopt-sleep run \
  --backend codex --max-sessions 10 --max-tasks 5 --progress

# Inspect staged proposals and history
/skillopt-sleep status

# Adopt only after explicit review/approval
/skillopt-sleep adopt

# Nightly wrapper schedule; this installs cron for the OMP wrapper, not the
# upstream bare `python -m skillopt_sleep run`.
/skillopt-sleep schedule \
  --hour 3 --minute 17 --backend codex --max-sessions 10 --max-tasks 5

/skillopt-sleep unschedule

Actions are status, harvest, dry-run, run, adopt, schedule, and unschedule. schedule / unschedule are handled by this OMP wrapper so the nightly job refreshes the OMP transcript mirror before invoking SkillOpt-Sleep. Other actions pass through to upstream python -m skillopt_sleep with --source claude and the mirrored OMP home injected by default. Upstream flags such as --target-skill-path, --tasks-file, --backend, --model, --edit-budget, --lookback-hours, --max-sessions, --max-tasks, --progress, and --json pass through unchanged.

Slash command usage is /skillopt-sleep <action> [flags...]; the extension also exposes the skillopt_sleep_omp tool for structured calls.

Wrapper-only flags:

Flag	Meaning
--omp-sessions DIR	Override the OMP session root. Defaults to ~/.omp/agent/sessions.
--omp-mirror-home DIR	Override the Claude-compatible mirror home. Defaults to ~/.skillopt-sleep/omp/claude-home.
--help-omp-wrapper	Show wrapper-specific help.
--self-test	Validate the OMP-to-Claude transcript translator without SkillOpt.

Recommended workflow

1. Start with dry-run --backend mock --json to validate harvesting and task mining without API spend.
2. For real improvement, use run --backend codex or another upstream-supported backend. Keep --max-sessions, --max-tasks, and budget flags bounded.
3. Read the staged report.md and proposed edits before summarizing.
4. Adopt only after explicit user approval. Do not hand-edit memory or skills as a substitute for adopt.

Safety rules

• OMP session harvest is read-only; never edit ~/.omp/agent/sessions.
• Keep raw secrets, private transcript contents, and raw tool outputs out of chat, logs, commits, and generated skill text.
• Treat generated edits as proposals until the held-out gate and human review approve them.
• Prefer the mock backend for plumbing checks; real backends spend the user's own agent budget.
• If using --tasks-file with a real backend, review and mark the task file as reviewed according to upstream SkillOpt-Sleep rules.