Status: 📋 Proposed Source: 5-lens adversarial architecture eval vs Serena (2026-07-03) × open backlog cross-map Verdict being answered: precision Serena 8.5 > CodeLens 6.5 (maturity gap, not ceiling); evolution Serena 8.0 > 6.5 (language marginal cost = structural); economics/ops/governance CodeLens 8.5 already superior — maintain, don't over-invest.

Strategy: invest where we LOSE (precision, language economics), finish where we WIN (governance/ops closure items), and never regress the axes already won (lean contract, MRR/perf gates stay mandatory).

P0 — Deploy hygiene (immediate, <1h)

P1 — Precision parity (highest leverage; the one axis we lose on product identity)

Goal: "delivered precision" ≥ Serena on the default path. Target score 6.5 → 8+.

1. LSP client protocol parity — crates/codelens-engine/src/lsp/session.rs - Answer server→client requests (workspace/configuration, client/registerCapability, window/workDoneProgress) instead of discarding (session.rs:535-578). - Per-server initializationOptions/settings table (start: rust-analyzer, pyright, ts-ls, gopls). - Quiescence readiness handshake: rust-analyzer experimental/serverStatus quiescent; equivalent signals per server; replace CODELENS_LSP_STARTUP_GRACE_MS blind sleep. - Acceptance: requests during indexing wait for quiescent (integration test with rust-analyzer); zero protocol-violation discards in a warm session trace.
2. Confidence calibration — closes #295 (11-day stale SCIP still 0.95) + panel condition 4 - SCIP staleness + LSP readiness feed confidence_basis; stale precise tier caps ≤0.6 with explicit degraded_reason. - Acceptance: regression test — stale index.scip vs newer source ⇒ confidence ≤0.6 + warning.
3. Python default-path precise routing — the one MEASURED recall gap (imports/type annotations) - Option A: opt-in pre-warm LS pool (daemon detects project languages at bind → warms pyright); Option B: scip-python auto-generation (see P2.1). Ship A first (smaller), B supersedes. - Acceptance: Python find_referencing_symbols returns import + annotation references on the DEFAULT path; add benchmark dataset asserting it (recall floor in CI).

P2 — SCIP multi-language + language-scaling economics (structural weakness)

1. SCIP auto-generation pipeline: language detect → run indexer (scip-python / scip-typescript / rust-analyzer scip) → partial publish → mtime freshness watch. Today: load-only, generation script Rust-only. Interlocks with #295/#298 (hot-reload session invalidation must not lock out clients).
2. Grammar ABI unblock: tree-sitter 0.25→0.26 unified upgrade releases the 5 blocked languages (make/dockerfile/vim/fsharp/perl — lang_config.rs:85); evaluate dynamic grammar loading (libloading) as the durable fix for the ABI lattice class.
3. New-language policy pivot: LSP/SCIP-first for semantics of any NEW language; freeze growth of per-language regex import extractors (import_graph/parsers.rs). Pipeline a dataset generator so each new language lands WITH an embedding-quality dataset + MRR floor.

P3 — Governance closure (win the axis fully; small)

1. Secure-by-default: mutation-capable daemon without principals.toml → startup warning; install-http-daemons-launchd.sh --principals-scaffold writes a starter file (planner=ReadOnly / builder=Refactor cross-principal split as the documented default).
2. #347 promotion: unbound-project calls on a shared daemon warn→block (advisory hint today).

P4 — Ops/scale closure

1. Watcher start failure: silent .ok() degrade → explicit watcher_unavailable warning in bootstrap + freshness path.
2. Direct 100K-file benchmark (39K measured today — confirm the extrapolated claim once).

3. 300/#301: session token survives daemon restart → graceful re-init instead of

   'Unknown session' lockout.

4. 342: analysis cache invalidation on file move + redeploy.

P5 — Mutation tier completion (extend the verified safety lead)

1. pending-D3 promotion decision: symbolic_edit_core 4 tools + refactor_substrate 5 tools out of allowlisted_dispatch_only — together with #287 (LSP code-action backends for replace_symbol_body / insert_*_symbol), i.e. mutation goes LSP-first too.
2. #341 hash-anchored semantic_edit — widens the defensive-mutation lead the eval already credited (old_text verify + overlap reject + dry-run).

Sequencing & effort

Quality gates per phase: existing infra (fmt/clippy matrix, nextest, regen/surface drift, MRR floor + benchmark CI). New: Python reference-recall dataset (P1.3), stale-SCIP confidence regression (P1.2), quiescence integration test (P1.1).

Non-goals (explicitly deferred)

• Tool-count growth: 93 tools is already past the "consolidate" guidance; new capability should land inside existing workflow entrypoints, not as new top-level tools.
• Re-litigating won axes: lean contract, budget/compression, RBAC design are done — only the closure items above remain.

P6 — Evidence discipline & context hygiene (LazyCodex-referenced, 2026-07-03)

Source: LazyCodex/omo 4.13.0 hook-layer audit (claims verified against ~/.codex/plugins/cache/sisyphuslabs/omo/4.13.0 source + CodeLens seams). Principle: adopt the verification habits and context hygiene, never the host-hook architecture (CodeLens is a cross-host MCP server, not a hook runner).

Also rejected (boundary violations): transcript parsing (breaks cross-host portability), .omo/boulder.json-style external plan ledger as core contract (typed session state owns this), embedding a hook runner.