Initiative: Pilot LSP (language-server) code intelligence in the agent harness — pr-review Shell first, measured go/no-go

[github-actions]

Why

Our dev-lead and pr-review agents navigate code by grep + file-reading, which returns textual not semantic matches and gives a sliced, incomplete view of large repos. Language Server Protocol (LSP) tooling — definition, references, hover, diagnostics — exposed to the agent as MCP tools gives real semantic code intelligence. The discussion's strongest, lowest-risk pilot is pr-review finding-verification on Shell in this repo (a bad verification only downgrades a finding; a bad code write breaks a build).

This is strictly opt-in plumbing on top of the MCP mechanism that already shipped — scripts/engine.sh threads --mcp-config/--strict-mcp-config into the deep + rubber-duck Claude tiers only (REVIEW_MCP_CONFIG / REVIEW_MCP_ALLOWED_TOOLS, epic #676, Context7 landed in #826). The triage tier stays fast/restricted and untouched. LSP is one more MCP server behind that same knob — we are not rebuilding the gating, the graceful-degradation path, or the token observatory.

Scope (confirmed from the discussion)

• Agent surface: pr-review deep/audit tiers first; dev-lead follows only after a positive go/no-go (out of scope here).
• Pilot language/repo: Shell in .github-private (this repo — the agent's own substrate; bash-language-server is lightweight, no cross-repo coordination).
• Server selection is an experiment, not a pre-decision. Per @don-petry, candidate LSP-MCP servers (agent-lsp, Serena, lsp-mcp) are compared on speed, quality, and cost across a corpus of PRs with and without LSP; selection is decided from that data, not the vendor narrative.

Initiative success metric

Go = on the frozen pilot PR corpus, LSP-on deep-tier review delivers a measurable navigation-token reduction (target >=2x fewer navigation tool-call tokens vs the LSP-off control) AND no regression in review quality (false-positive / precision no worse than the frozen LSP-off baseline), achieved within the cold-start SLA (<=30s P95). A win on tokens that costs precision is a no-go, not a win.

Cost cap (explicit)

• LSP is gated to deep/audit tiers only (triage untouched); cold-start is auto-skipped above the 30s P95 SLA (graceful degradation, never a workflow failure).
• The pilot is run-count bounded: <=20 PRs in the corpus x <=2 candidate servers x <=3 runs each ~= <=120 deep-tier review runs total, under the existing 60-min job cap. No fleet-wide rollout, and no dev-lead (code-writing) LSP, until the go/no-go decision record lands.

Shape

Phase 1 scopes the pilot and freezes the measurement corpus/baseline; Phase 2 wires the candidate server(s) into the runtime and adds the finding-verification step; Phase 3 runs the comparison and records a human go/no-go.

Untracked prerequisites

• Candidate LSP-MCP servers (agent-lsp / Serena / lsp-mcp) are third-party, non-org-owned dependencies — each must be pinned to a specific version for a reproducible pilot.
• bash-language-server (npm) must be installable in the pr-review runner environment for the Shell pilot.
• Companion Ideas discussions filed off Idea: add LSP (language-server) tooling to the dev-lead / pr-review agent harness #578 (LSP-Grounded Finding Verification; Per-Tier MCP Tool Gating; Token Cost Observatory LSP Metrics) are tracked separately and are not blockers — per-tier MCP tool gating already exists in engine.sh (deep/duck only).

---

Planned from idea discussion #578 by the BMAD Scrum Master initiative-planner. Inert until a maintainer adds initiative:auto.

[don-petry]

Epic complete. Shell LSP pilot = NO-GO (#845), pilot retired from the repo via #1110 (merged). Door left open per-language via #1124 → #1125 (Python) / #1126 (TS-JS).