OMP vs looplane retrospective: what to borrow, what's over-engineered, what not to touch yet

OMP Internals Deep Dive series, post 15 — finale. The previous 14 dissected subsystems; this one zooms out, contrasting looplane (ex-rivumi) to answer: what to borrow, what's over-engineered, what not to touch yet.

Looplane referenced is the Python implementation under src/looplane; omp is the TypeScript + Rust monorepo. Citations use dual track: omp/package/file.ts#Symbol vs looplane/src/module.py#symbol.

TL;DR

Philosophy diff: omp bets "in-process batteries-included" — shoves grep/AST/PTY/isolation/voice into process via crates/pi-*; looplane bets "minimal harness + external runtime" — drives Codex / Claude Code as ExternalCodingRunner, focuses on thin prompt/permission/isolation layer.

Both work; difference is team size and maintenance cost.

Five Most Borrowable

1. Append-only context (post 2)

packages/agent/src/append-only-context.ts#AppendOnlyContextManager.syncMessages longestStablePrefix maximizes Anthropic/DeepSeek prompt cache hits. Old code log.clear() on any digest change forced ~40k token re-prefill on llama.cpp every turn (issue #3406). New code resends only divergence tail; provider KV cache stays warm.

Portability to looplane: Very high. Add fingerprint + digest memo + version tag to loop.py context assembly. Saves 30–60% prompt tokens per model call.

2. Four compaction strategies (post 3)

context-full (LLM summary) / snapcompact (PNG for vision) / branch summary (/tree nav) / shake (mechanical placeholder). Four strategies for four failure modes, auto-orchestrated by session-maintenance.ts.

Portability: High. Looplane has checkpoint but lacks strategy split — single compaction only. Prioritize shake (pure-local, deterministic, zero-failure) then branch summary (tree nav without amnesia).

3. Three-layer fail-closed approval (post 4)

packages/coding-agent/src/tools/approval.ts#resolveApproval: tool declaration → user override → mode threshold, defaults exec (fail-closed). Deny always wins; yolo still doesn't cross tool/user deny.

Portability: Very high. Looplane's permissions.py has deny/allow but lacks policyKey dynamic tier and CRITICAL_BASH_PATTERNS hardcoded floor. Adding those lifts security baseline immediately.

4. KDL rule tree (post 7)

packages/catalog/src/compat/rules/** layers 60+ providers' routing/compat/thinking/quota as taxonomy/classes/providers/runtime, compiled to rules.json, resolved by cascade.ts#resolveCascade with (exactness, dimensions, priority). Zero if (provider === "openai") in TS.

Portability: Medium. Looplane's provider_catalog.py is hardcoded table. For 10+ providers, KDL pays off; for 3–5, hardcoding is fine.

5. Session tree / fork (post 9)

packages/coding-agent/src/session/session-manager.ts#getTree + agent-session.ts#navigateTree, entries with id/parentId form tree, leaf switch optionally branch_summary, fork copies JSONL + artifacts.

Portability: High. Looplane's event journaling has similar base; add getTree + navigateTree to support /tree navigation.

Three Over-Engineered

1. Snapcompact pure-vision path

Rendering history as PNG for vision model saves LLM latency but assumes "vision model cheap and code-readable." Measured 11on16-bw vs 8on22-bw f1 delta 0.05 — ROI worse than shake + branch summary.

Looplane: Skip.

2. Metaharness self-built bench infra

packages/metaharness experiment→run→trace + Vibemon microVM + SQLite + REST dashboard vs promptfoo / evals / harbor. Self-build cost high.

Looplane: Use promptfoo + evals. Only build metaharness if "hardware isolation + auth gateway rewrite" required.

3. Bash tokenization extreme regex (post 5)

CRITICAL_BASH_PATTERNS 45 regexes + hasBashApprovalShellControl + glob-to-regex, triple layered. Good for security, but regex maintenance high; rm -rf -- / variants endless.

Looplane: Keep core 10–15 critical patterns + deny per-segment match; extreme strictness of allow whole-line can be simplified.

Three Don't-Touch-Yet

1. Full provider catalog (60+)

packages/catalog tracks 60+ providers' model lists, pricing, context windows, thinking ladders via KDL + bun run gen:compat. Needs dedicated maintainer.

Looplane: Keep 3–5 core providers (OpenAI/Anthropic/Google/OpenRouter), hardcode provider_catalog.py, add 6th only when users truly need it.

2. Full TUI self-build (post 11)

packages/tui differential rendering + composer multimode + vim + IME + virtual list ≈10k lines. Looplane's tui.py with textual / ink suffices.

Looplane: Don't self-build; focus on prompt assembly and approval UX.

3. Full collab wire protocol (post 13)

packages/wire AES-256-GCM + 4-byte envelope + snapshot-chunk + timing-safe token for asymmetric "host runs agent, guest observes" model — long-term relay + permission maintenance.

Looplane: For multi-user, use tmux + code-server / VS Code Live Share first; don't build wire.

Pragmatic Borrowing Roadmap (for looplane)

Phase 1 (1–2 weeks):

1. Add AppendOnlyContextManager longestStablePrefix + digest memo (post 2)
2. Add shake mechanical placeholder (post 3's DEFAULT_SHAKE_CONFIG)
3. Add core 15 CRITICAL_BASH_PATTERNS + deny per-segment (post 5)
4. Add deny-always-wins + policyKey to resolveApproval (post 4)

Phase 2 (1 month):

5. Add branch summary /tree nav (post 9)
6. Add minimal KDL or simplified rule tree (post 7's taxonomy + providers two layers)

Skip: snapcompact, metaharness, full TUI, full collab.

Lessons Learned

1. Borrow, don't rebuild — omp's 80k Rust lines + 60+ providers = 8 months, 18k commits. Looplane shouldn't copy scale, but copy "why behind the design"
2. Fail-closed is the only safe default — undeclared approval = exec, unknown provider = strictest tier — both should follow this
3. Cache-coherence is a cross-module contract — append-only-context.ts + message-cache.ts + tokenizer.ts share version tags; single-file change must notify others
4. Strategy split > single algorithm tuning — compaction isn't "one algorithm tuned", but "four strategies for four failure modes" — more impactful than any single optimization

References

• packages/agent/src/append-only-context.ts — longestStablePrefix
• packages/agent/src/compaction/compaction.ts / shake.ts / branch-summarization.ts — four strategies
• packages/coding-agent/src/tools/approval.ts — three-layer decision
• packages/coding-agent/src/tools/bash.ts — tokenized approval
• packages/catalog/src/compat/rules/** / cascade.ts — KDL rule tree
• packages/coding-agent/src/session/session-manager.ts — tree/fork
• packages/coding-agent/src/edit/hashline/execute.ts — hashline
• docs/compaction.md / docs/approval-mode.md / docs/tree.md — official docs

Part of OMP Internals Deep Dive series, post 15 (finale). Full 15-post series complete.