openlore

openlore

[!NOTE] spec-gen has been renamed to OpenLore. The npm package is now openlore and the CLI command is openlore. Existing projects: rename your .spec-gen/ directory to .openlore/ and reinstall (npm i -g openlore). See docs/RENAME-TO-OPENLORE.md for the full migration checklist.

Persistent architectural memory and structural cognition for AI coding agents.

openlore turns any evolving codebase into a navigable knowledge graph backed by OpenSpec living specifications. It maintains persistent architectural context across agent sessions: graph structure, specs, decisions, drift state, and semantic retrieval — so agents start each task already oriented instead of re-discovering the system from file reads.

Value Scorecard — does it pay for itself?

OpenLore only earns its place if an agent with it reaches a correct answer for less total cost than the same agent without it. We measure that inequality and publish it — wins and losses. Numbers are from the Spec 14 agent benchmark (claude -p, sonnet, N=4 medians, pinned SHAs, --strict-mcp-config isolating each arm), measured 2026-06-01.

Scenario (task × repo)	Cost Δ	Round-trips Δ	Correctness	Verdict
Large/unfamiliar repo · deep "how does X flow through Y" (its target)	−7% to −21%	−26%	100% = 100%	✅ helps — and the win grows with repo size
Small/familiar repo · shallow "who calls X"	task-dependent (Round 1: +43%)	+38%	100% = 100%	❌ often adds overhead — measure with openlore prove

Re-confirmed live 2026-06-03 (N=2): the deep-task win reproduces — okhttp −13%, identical to the table below. The small/familiar case is task-dependent, not a flat loss: same repo class, opposite outcomes (chalk −32% win vs express +59% loss) — the cost there is a sometimes-redundant orient round-trip, not tool-schema bytes, so a leaner surface doesn't close it. Don't guess from our repos — run openlore prove on yours.

Deep-trace detail — the win scales with codebase size (cost Δ; round-trips WITHOUT → WITH):

Repo (size)	Cost Δ	Round-trips
excalidraw (~640 files)	−21%	25 → 16
tokio (~790 files)	−21%	17 → 13
okhttp	−13%	13 → 11
django (~3k files)	−7%	21 → 15
gin (110 files, smallest)	+4% (≈even)	10 → 9

When OpenLore helps — and when it doesn't:

• Helps: large, unfamiliar, or private codebases the model hasn't memorized; deep multi-hop questions; long sessions where re-reading an ever-growing context compounds. The most consistent, hardest-to-game signal is round-trips: −26%, fewer on every deep task.
• Doesn't (yet): small, famous repos already in the model's weights answered by a shallow query — there's no orientation tax to remove, so the MCP tool surface is pure overhead.

Reproduce it: npm run bench:agent (needs an API key for the agent arm). Full methodology, per-task numbers, and caveats: docs/AGENT-BENCHMARKS.md. Plumbing latency (orient ~430µs p50) is separate and real: scripts/BENCHMARKS.md.

Honesty contract. We never publish a savings number the benchmark didn't produce; we always show the loss cases next to the wins; the scorecard is date-stamped and re-measured after each optimization phase. Every public token claim traces to a command you can run in this repo — if it doesn't reproduce, treat it as marketing and call it out.

Why It Exists

AI agents are powerful but amnesiac. On every new task:

• They re-read the same source files to understand structure
• They forget architectural decisions made two sessions ago
• They have no link between specs and code — drift is invisible
• File-by-file navigation costs round-trips and fresh tokens that grow with repo size — on deep traces in large repos the WITHOUT baseline runs 17–25 tool-calls to reach an answer; openlore cuts that −26% (see the Value Scorecard for the measured numbers, and where it does not help)
• In long sessions, they drift from authoritative retrieval toward internally cached reasoning — producing subtly wrong architectural assumptions that compound silently until a refactor breaks

openlore closes this loop. Run a full analysis once, then keep the graph incrementally updated as the codebase evolves. Even greenfield projects become cognitively "brownfield" after only a few agent sessions — architectural context fragments, decisions disappear, and agents repeatedly reconstruct the same understanding from scratch.

openlore persists that context continuously: structure, specs, decisions, drift state, and graph relationships remain queryable across sessions.

How It Works

Three layers, each usable independently:

Layer	What it does	API key?
1. Static Analysis	Call graph, clusters, McCabe CC, external deps → CODEBASE.md digest	No
2. Spec Layer	LLM-generated living specs, ADRs, drift detection, decision gates	For generation
3. Agent Runtime	50 MCP tools — orient(), semantic search, graph expansion	No

You can use layer 1 alone to give agents structural context. Add layer 2 for semantic intent and architectural governance through OpenSpec-compatible living specifications. Layer 3 keeps that context continuously accessible through graph-native MCP tools once openlore mcp is running.

openlore vs. Alternatives

	Cursor / Claude Code	Sourcegraph	openlore
Graph-aware MCP context	❌ file-based reads	Partial	✓ call graph + clusters
Spec drift detection	❌	❌	✓ milliseconds, no API
Architectural decision gates	❌	❌	✓ pre-commit hook
Offline structural analysis	❌	❌	✓
Token-efficient orient()	❌	❌	✓ −7%→−21% cost, −26% round-trips on deep tasks †
Living spec generation	❌	❌	✓
Persistent cross-session architectural memory	❌	Partial	✓

† Measured, and it depends on the task (full numbers in the Value Scorecard above). The Spec 14 agent benchmark (npm run bench:agent, WITH vs WITHOUT openlore, claude -p, N=4 medians) gives a two-tier result:

• Small, familiar repos + shallow "who-calls-X" queries: openlore adds ~43% cost — the model already knows the code, so there's no orientation to save.
• Larger codebases + deep "how does X flow through Y" questions (its target): with the lean --preset navigation tool surface, openlore is a net win — −7% cost and −26% tool-calls at N=4, scaling with repo size (up to −21% on ~640–790-file repos), at 100% answer correctness in both arms.

So the headline savings hold where openlore is designed to help, not on toy queries. Full results, methodology, and honest caveats: docs/AGENT-BENCHMARKS.md. The plumbing latency (orient ~430µs p50) is separate and real — see scripts/BENCHMARKS.md. | Long-session confidence decay (Epistemic Lease) | ❌ | ❌ | ✓ |

Traditional coding agents reconstruct architecture from repeated file reads every session. openlore persists it as a queryable graph.

5-Minute Quickstart

One command, no API key needed:

npm install -g openlore
cd /path/to/your-project

openlore install          # detect your agent, wire it up, AND build the index

That single command:

• Auto-detects which agent surfaces are present (Claude Code, Cursor, Cline, Continue, AGENTS.md) and wires each one to call orient() — no manual CLAUDE.md editing.
• Registers the MCP server so it starts automatically when your agent launches (you don't run openlore mcp yourself).
• Builds the index (init + analyze → a keyword/BM25 graph, no network needed) so orient() returns real results in your very first session — no separate analyze step.

openlore install --no-analyze   # wire surfaces only; build the index later
openlore install --dry-run      # preview every change without writing

See docs/install.md. The MCP server keeps the index fresh as you edit (file watcher on by default — large build dirs like target/, node_modules/, dist/ are pruned automatically; disable entirely with openlore mcp --no-watch-auto).

Then ask your agent: orient("add a new payment method")

That single call returns the relevant functions, their call neighbours, matching spec sections, and insertion-point candidates — preserving architectural continuity across sessions instead of forcing the agent to repeatedly reconstruct context from raw file reads. The Spec 14 benchmark (docs/AGENT-BENCHMARKS.md) measures this directly: on deep "how does X flow through Y" questions in larger codebases, openlore (with --preset navigation) cuts cost ~7% and tool-calls ~26% at N=4 (more on bigger repos); on small/familiar repos with shallow queries it adds overhead instead. Net: it pays off in its target arena, not on toy queries.

Full pipeline (specs + decisions — optional and additive):

openlore generate         # generate living specs (requires API key)
openlore drift            # detect spec/code drift
openlore decisions        # manage architectural decisions

Install from source

git clone https://github.com/clay-good/openlore
cd openlore
npm install && npm run build && npm link

Nix / NixOS

nix run github:clay-good/openlore -- analyze
nix shell github:clay-good/openlore

System flake:

environment.systemPackages = [ openlore.packages.x86_64-linux.default ];

See It In Action

Example: orient("add a payment method")

{
  "functions": [
    {
      "name": "processPayment",
      "file": "src/payments/processor.ts",
      "risk": "medium",
      "fanIn": 4,
      "callers": ["handleCheckout", "retryFailedCharge"],
      "callType": "direct"
    },
    {
      "name": "validateCard",
      "file": "src/payments/validator.ts",
      "risk": "low",
      "fanIn": 1,
      "testedBy": [{ "name": "validateCard.test.ts", "confidence": "called" }]
    }
  ],
  "specDomains": ["payments — §CardValidation, §PaymentFlow"],
  "insertionPoints": [
    "src/payments/processor.ts:87 — after existing charge logic"
  ],
  "callPath": "POST /charge → handleCheckout → processPayment → validateCard → stripeClient.charge"
}

One graph query replaces most exploratory file reads. The agent knows exactly where to look and what risks to consider.

Agent Cheat Sheet

The full surface is 50 tools, but day-to-day work needs a handful. Reach for the right one by situation:

Situation	Tool
Starting any task	orient(task) — functions, callers, specs, insertion points in one call
Shallow "who calls X / where is Y?"	orient(task, lean:true) (CLI orient --lean) — navigation core only, ~40% smaller and skips the enrichment compute (Spec 27)
"Which file/function handles X?"	search_code
Call topology across many files	get_subgraph / analyze_impact
"What's the blast radius if I change this?"	analyze_impact — risk score + up/downstream chain + governing decisions
"What decisions constrain this code?"	analyze_impact / get_subgraph → governingDecisions (Spec 16)
Planning where to add a feature	suggest_insertion_points
"How does request X reach function Y?"	trace_execution_path
"I changed X — which tests should I run?"	select_tests — backward reachability to the reaching tests + paths (Spec 19)
"What's dead / what dies if I delete X?"	find_dead_code — cross-language reachability, confidence-tagged candidates (Spec 20)
"What changed structurally / whose callers are now stale?"	structural_diff — graph diff, stale callers, rename flags (Spec 21)
"What changes together with this / what's volatile?"	get_change_coupling — co-change + churn from git history (Spec 22)
"May I add this import here / what breaks the architecture?"	check_architecture — pre-edit verdict against declared rules (Spec 23)
Recording an architectural choice	record_decision before writing the code
Reading / checking a spec	get_spec · search_specs · check_spec_drift
Ranking what changed by risk	detect_changes

Everything else (read a file, grep, list files) uses your native tools. Full reference: docs/mcp-tools.md.

Use OpenLore as a Claude Code Skill

OpenLore ships a canonical Claude Code Skill at skills/openlore-orient/. Install it once and Claude Code will automatically call orient() at the start of every task — no CLAUDE.md editing required.

# From the OpenLore repo root:
npm run skill:install-local           # → ~/.claude/skills/openlore-orient/

# Or copy into a single project's .claude/skills/:
cp -R skills/openlore-orient /path/to/your-project/.claude/skills/

The skill bundle ships a SKILL.md manifest, POSIX + PowerShell wrappers, a worked example, and a redacted real orient() JSON output so the model knows the response shape. See skills/openlore-orient/README.md for details.

Core Features

Analyze (no API key)

Continuously maintains a structural representation of your codebase using pure static analysis. Builds a full call graph persisted to SQLite, runs label-propagation community detection to cluster tightly coupled functions, computes McCabe cyclomatic complexity for every function, and extracts DB schemas, HTTP routes, UI components, middleware chains, and environment variables. Outputs .openlore/analysis/CODEBASE.md — a ~600-token structural digest that compresses the equivalent of tens of thousands of exploratory tokens into a small, queryable summary.

With --watch-auto, the call graph updates incrementally on every file save: changed file and its direct callers are re-parsed and the graph is atomically swapped. Orient and BFS queries remain live between full analyze runs.

Generate (API key required)

Sends the analysis to an LLM in 6 structured stages: project survey → entity extraction → service analysis → API extraction → architecture synthesis → ADR enrichment. Produces openspec/specs/ living specifications in RFC 2119 format with Given/When/Then scenarios.

Drift (no API key)

Compares git changes against spec mappings in milliseconds. Detects: Gap (code changed, spec not updated), Uncovered (new file, no spec), Stale (spec references deleted files), ADR gap (code changed in an ADR-referenced domain). Installs as a pre-commit hook.

Install (no API key)

openlore install auto-wires the popular agent surfaces (Claude Code, Cursor, Cline, Continue, AGENTS.md) so they call orient() automatically — no CLAUDE.md editing required. Each integration uses a fingerprinted managed block so re-runs are idempotent and hand-edits are detected. --dry-run previews diffs; --uninstall cleanly removes everything. See docs/install.md.

Preflight (no API key)

openlore preflight is a CI staleness gate: any pull request that edits files in the graph fails the check until the graph is refreshed. Drop-in templates for GitHub Actions, GitLab CI, and generic shell live in examples/ci/. Weighted scoring surfaces hubs first so a one-line leaf edit doesn't fail the same way a refactor of a top-of-stack module does. See docs/preflight.md.

MCP (no API key)

50 graph-native tools exposed over stdio. Together they act as a persistent architectural runtime for coding agents: orientation, graph traversal, semantic retrieval, drift awareness, decision context, and structural risk analysis. orient() is the main entry point — it collapses the discovery loop into one call (measured: −26% round-trips on deep traces; see the Value Scorecard). detect_changes risk-scores changed functions using call graph centrality × change type multiplier. Every tool call runs the same guards — input validation against its schema (bad args → JSON-RPC -32602), a per-tool timeout, a deterministic output-size cap, and normalized error codes — and the surface carries complete MCP annotations. See docs/mcp-tools.md.

orient() runs in ~430µs p50 against a 15k-node codebase (TypeScript compiler, ~79k edges). Full benchmark results: scripts/BENCHMARKS.md.

Test impact selection (no API key, Spec 19)

select_tests answers "I changed parseConfig() — which tests should I run?" by walking the call graph backward from the change to every test that transitively reaches it (via calls + tested_by + inheritance edges), returning each test with its reaching path. This is static, call-graph-based regression test selection (RTS) — established CS — served to the agent at edit time instead of to CI after the fact. grep can't do it (the reach is through indirect calls); the model is slow and guesses; a deterministic graph does it instantly. It is an honest over-approximate prioritizer ("run these first"), not a sound replacement for the full suite — the response states its posture, coverage, and caveats (dynamic dispatch / DI can under-select). Inputs: a symbol set or a git diff. Deterministic and offline. See docs/test-impact-selection.md.

Reachability & dead-code (no API key, Spec 20)

find_dead_code runs cross-language mark-and-sweep over the call graph: reachability from roots (tests, imported symbols, route handlers, main), candidate-dead = the unreached remainder, and "what becomes dead if I delete X?" = the set reachable only through X. Prior art (knip, ts-prune) is TS/JS-only; this rides the unified tree-sitter graph across 15+ languages. Results are confidence-tagged candidates, never deletion authority — dynamic dispatch, DI, framework routing, and externally-consumed exports cause false positives, stated in the response. A conservative module-level liveness signal keeps high-confidence candidates trustworthy (it cut them from ~470 to ~35 on a real repo). See docs/reachability-dead-code.md.

Structural change analysis (no API key, Spec 21)

structural_diff is a graph diff — the structural complement to git diff. Between two states (working tree vs a ref, or two refs) it reports functions and edges added/removed, signature changes, and the existing callers in other files now stale because a callee's signature moved under them. A review/refactor agent gets "this removed gamma, changed alpha's signature, and 5 of its callers are now stale" instead of "these 40 lines changed". Only the changed files are re-parsed (old via git show, new via the working tree), so it is cheap and never mutates the canonical graph; rename/move ambiguity is flagged, not guessed. See docs/structural-diff.md.

Change-coupling & volatility (no API key, Spec 22)

get_change_coupling mines two facts from local git history that the call graph structurally cannot see: co-change coupling ("these files almost always change together" — the invisible coupling with no import or call edge) and volatility/churn ("this file changed 23 times" — a risk flag). Prior art (CodeScene) puts it well: change coupling "isn’t possible to calculate from code alone — it is mined from git." Surfaced additively in orient as caution signals. Support/confidence thresholds and a bulk-commit filter keep it honest; it is an advisory signal, correlation not causation. Local, deterministic, no network (reuses the Spec 18 git ingestion). See docs/change-coupling.md.

Architecture invariant guardrails (no API key, Spec 23)

check_architecture turns an architectural rule from a post-hoc CI failure into a pre-write guardrail. A repo declares constraints — layers, forbidden, allowedOnly — in .openlore/architecture.json (or via an Invariant: marker on a synced ADR, so a recorded decision carries its invariant), and the tool answers, before the agent writes the import, "may a file under A import B?" with a deterministic verdict + the governing rule + why, plus a full violation scan. Prior art (ArchUnit, dependency-cruiser, import-linter) enforces architecture in CI after the code is written and per-language; OpenLore's contributions are cross-language rules over the unified dependency graph and agent-facing, pre-edit evaluation (reusing the same classifyLayerEdge primitive that powers CODEBASE.md's layer report). Opt-in and fully inert until rules are declared; never LLM-inferred; complements, not replaces, CI linters. See docs/architecture-invariants.md.

Epistemic Lease (no API key)

Core principle: EpistemicLease models architectural drift as a behavioral navigation phenomenon rather than a semantic understanding problem. Context decay is driven by where the agent goes (cross-module trajectory), not what it knows.

As a session grows longer, agents naturally shift from authoritative graph retrieval toward internally cached reasoning. This is useful for fluency but dangerous for architectural correctness — cross-module assumptions go stale, dependency hallucinations accumulate, and delegation prompts embed incorrect repository understanding that cannot easily be corrected downstream.

The Epistemic Lease models this decay explicitly. Every MCP tool response carries a freshness signal when the agent's architectural context has degraded or expired. Decay is triggered by any of: time elapsed since orient(), git hash divergence from the orient baseline, weighted cognitive load accumulation (heavier tools count more), or cross-module file access breadth.

The signal escalates through three levels to resist warning blindness:

Level	Trigger	Signal style
Degraded	load ≥ 30, age ≥ 15min, or cross-module density ≥ 0.15	Advisory signal appended
Stale	load ≥ 60, age ≥ 30min, git hash divergence, or density ≥ 0.30	Procedural block prepended: what NOT to do
Stale [Elevated]	load ≥ 85 or age ≥ 45min	Risk-framing: names downstream consequences
Stale [Critical]	load ≥ 110 or age ≥ 60min	Imperative: STOP. Call orient(). — minimal, hardest to skim

Cross-module density is computed as a sliding-window trajectory model: switches_in_last_15_calls / 15. The fixed denominator prevents false positives during session warmup. Each module switch adds +5 cognitive debt; a high-density window adds +15; a burst (density ≥ 0.60) adds +20. A 5s dampening window prevents back-and-forth from double-counting.

An oscillation coefficient (repeated_bigram_transitions / total_transitions) separately distinguishes confusion loops (A→B→A→B scores 1.0) from genuine exploration (A→B→C→D scores 0.0). When already stale, a heavy architectural tool (weight ≥ 8) or density burst (≥ 0.60) triggers immediate escalation to Stale [Critical].

When fresh, injection is zero-overhead. Calling orient() resets the tracker. Unlike governance systems, the lease never blocks — it modulates the agent's confidence in its own cached reasoning rather than constraining its actions.

Decisions (API key for consolidation)

Agents call record_decision before writing code. Consolidation runs immediately in the background. At commit time, a pre-commit hook gates the commit until all verified decisions are reviewed and written back as requirements in spec.md files. Decisions are classified by scope (local / component / cross-domain / system); only cross-domain and system decisions produce ADR files, keeping the decision log signal-dense.

Decisions are also first-class graph nodes. At analyze time the active decision store is projected — the same parser→projector split that puts Infrastructure-as-Code on the graph — into decision::<id> nodes joined to the files they govern by affects edges. The relationship is stored, not recomputed: analyze_impact and get_subgraph return the governing decisions of a symbol and its blast radius as typed neighbors (nodeType: "decision"), and orient reports which relevant files each decision governs. This turns "what architectural decisions constrain this code, and what does changing it implicate?" into a deterministic graph query — the join no code-navigation competitor offers. The JSON store stays authoritative; the projection is derived and rebuilt on every analyze. See docs/specs/openlore-spec-16-decisions-as-graph-nodes.md.

Provenance (no API key, local-only)

Reads the local .git history (and local gh if present) to project authored_by (file → person) and changed_in_pr (file → PR) edges onto the graph, so orient answers "last changed by X in PR #N" — provenance grep cannot surface. No OAuth, no cloud connector, nothing is ever uploaded: the git-only path needs no network, and gh is an optional enrichment that degrades gracefully when absent or unauthenticated. Bounded (last-touch + top-N recent authors + recent PRs per file) so the graph never bloats; deterministic for a fixed git state. The same local history feeds the change-coupling instrument (Spec 22). See docs/provenance.md.

Telemetry (opt-in, no API key)

Cognitive telemetry for empirical measurement of EpistemicLease behavior. Gated by OPENLORE_TELEMETRY=1 — disabled by default. Writes append-only JSONL to .openlore/telemetry/ per domain. Agent identity is captured from the MCP initialize handshake, enabling per-agent behavioral comparison.

.openlore/telemetry/
  mcp.jsonl              # every tool call: latency, errors, agent name
  orient.jsonl           # orient quality: function/file/insertion_point counts
  cache.jsonl            # readCachedContext hit/miss
  epistemic-lease.jsonl  # state transitions: degraded, stale, depth escalation

Analyze with openlore telemetry:

openlore telemetry [directory]   # summary: latency, cache hit rate, obstinacy index
openlore telemetry --live        # stream events in real time as they occur

Key metrics: obstinacy index (tool calls after stale before orient — measures whether agents act on warnings), recovery efficiency (stale→orient latency), trajectory dynamics (avg cross-module density, burst frequency). These turn EpistemicLease from a tuning-by-intuition system into an empirically measurable one.

Architecture

OpenSpec provides semantic intent and workflow structure. openlore maintains the evolving implementation as a continuously queryable architectural graph for agents.

flowchart TD
    Code[Codebase] --> Analyze[openlore analyze<br/>tree-sitter · pure static analysis]
    Analyze --> DB[(SQLite graph store<br/>.openlore/analysis/call-graph.db)]
    Analyze --> Digest[CODEBASE.md<br/>~600-token structural digest]

    subgraph shared["Projected onto shared node + edge primitives"]
      direction LR
      CodeNodes[functions + call edges]
      Iac[IaC resources + references]
      Dec[decisions + affects edges]
    end
    Analyze --> CodeNodes
    Analyze --> Iac
    Analyze -. active decision store .-> Dec
    CodeNodes --> DB
    Iac --> DB
    Dec --> DB

    DB --> MCP[50 MCP tools<br/>orient · BFS · search · analyze_impact]
    MCP --> Agent((Coding Agent))

    Code -. optional, API key .-> Gen[openlore generate]
    Gen --> Specs[openspec/specs/*.md<br/>RFC 2119 living specs]
    Code --> Drift[openlore drift<br/>spec/code drift, ms, no API]
    Agent -. record_decision .-> Gate[decisions pre-commit gate]
    Gate --> Specs

The graph and the OpenSpec spec layer are co-equal: the graph makes orientation fast, the specs make it semantically grounded. Drift detection and decision gates connect both. Crucially, application code, Infrastructure-as-Code, and architectural decisions all project onto one shared set of node/edge primitives — so a single traversal answers questions that span all three. See docs/ARCHITECTURE.md for the full pipeline diagram.

Decisions on the graph (Spec 16) — a decision becomes a node joined to the files it governs by affects edges, so impact analysis returns governance as a neighbor:

flowchart LR
    D["decision::c6d1ad07<br/>North-star substrate"]:::dec
    D -- affects --> F1[src/cli/export/scip.ts]
    D -- affects --> F2[src/core/analyzer/iac/project.ts]
    F1 -- calls --> G[exportScip]
    classDef dec fill:#6f42c1,stroke:#4b2e83,color:#fff

Design Decisions

OpenLore dogfoods its own decision system. These ADRs were recorded with record_decision, gated at commit, and synced into openspec/specs/ — and (per Spec 16) are now projected onto the graph itself. They are the load-bearing constraints behind the architecture above:

Decision	Rationale	Where
North star is a deterministic structural context substrate	Local-first plumbing (like tree-sitter/SCIP/LSP) that agents build on; every feature must make the coding-agent case more useful and stay grounded in static analysis, not LLM guessing	ADR-0001
IaC resources project onto the existing graph primitives	One projector maps infrastructure onto FunctionNode/CallEdge so every MCP tool works on IaC with zero new tooling	analyzer spec · src/core/analyzer/iac/project.ts
Decisions project onto the graph the same way	A parser→projector split turns the decision store into decision:: nodes + affects edges — governance becomes a deterministic graph join	analyzer spec · src/core/decisions/project.ts (Spec 16)
EdgeStore uses SCHEMA_VERSION rebuild-on-bump, not migrations	The graph is fully derivable from source, so a schema change drops and rebuilds — no migration code, no drift	analyzer spec · src/core/services/edge-store.ts
BM25 keyword retrieval is the zero-network floor	orient/search_code work with no API key or embedding server; dense embeddings are an optional upgrade, never a requirement	analyzer spec · Spec 06
SCIP is a one-way export, not a round-trip format	The SQLite graph stays canonical; SCIP exports only the subset it can model, avoiding a lossy bidirectional contract	cli spec · src/cli/export/scip.ts
MCP exposes a curated navigation preset, not all 50 tools	A lean graph-traversal surface is what wins the Spec 14 agent benchmark; the full set stays available opt-in	cli spec · Spec 14
The tools/list prefix is trimmed losslessly + bounded by a guard, not byte-shaved	Spec 28 measured it: MCP has no server-side schema deferral and the lossless byte-lever is ~2%; the real levers are the client (deferred schemas) and tool count, so we trim safely, guard against bloat, and report the limit	cli spec · Spec 28
Lean orientation skips enrichment compute, not just its payload	orient --lean returns the navigation core for shallow lookups and skips the work behind the dropped blocks (extra embedding search, manifest/git reads); the rich default is unchanged	cli spec · Spec 27
Decision consolidation is serialized with a cross-process file lock	Concurrent record_decision calls were losing drafts; a lock makes consolidation safe and every commit instant	cli spec · Spec 15

This table is not aspirational documentation — it is the live decision log the pre-commit gate enforces and Spec 16 makes queryable. See docs/governance-dogfooding.md.

Interop

OpenLore exports SCIP (Source Code Intelligence Protocol). Plug it into Sourcegraph code nav, GitHub stack graphs, Glean importers, or any SCIP-aware tool:

openlore analyze            # build the graph (if you haven't already)
openlore export scip        # writes ./index.scip

The SQLite graph stays canonical; SCIP is a one-way export of the subset SCIP can model (functions → symbols, call edges → occurrences). See docs/scip-export.md for what is and isn't exported and how to consume it.

Federation (cross-repo)

The hardest agent-orientation questions cross repo boundaries: who calls BillingService.refund, where is event X consumed, how does data flow from service A to service B. OpenLore's answer is "SBOM-of-cognition" — every repo publishes a small, public, deterministic manifest describing what it exposes:

openlore manifest emit        # writes ./.well-known/openlore.json
openlore manifest validate .well-known/openlore.json

The manifest captures the public API surface, HTTP routes, stats, dependencies, and spec state in a versioned schema. A future OpenLore federation index will read these manifests across many repos to answer cross-repo orient() questions, staying a thin merger rather than a giant analyzer. See docs/federation.md.

Documentation

Topic	Doc
MCP tools reference (50 tools + parameters)	docs/mcp-tools.md
Agent setup (Claude Code, Cline, OpenCode, Vibe…)	docs/agent-setup.md
openlore install — auto-configure agent surfaces	docs/install.md
LLM providers + embedding config	docs/providers.md
Drift detection in depth	docs/drift-detection.md
Spec-driven tests + spec digest	docs/spec-tests.md
CI/CD integration	docs/ci-cd.md
Preflight CI staleness gate	docs/preflight.md
SCIP export (Sourcegraph/Glean interop)	docs/scip-export.md
Cross-domain impact (code ↔ infrastructure)	docs/cross-domain-impact.md
Local provenance (git/PR, no OAuth)	docs/provenance.md
Test impact selection (which tests to run)	docs/test-impact-selection.md
Reachability & dead-code analysis	docs/reachability-dead-code.md
Structural change analysis (graph diff)	docs/structural-diff.md
Change-coupling & volatility (git-mined)	docs/change-coupling.md
Architecture invariant guardrails (pre-edit)	docs/architecture-invariants.md
Federation manifest (cross-repo)	docs/federation.md
CLI command reference	docs/cli-reference.md
Interactive graph viewer	docs/viewer.md
Analysis output files	docs/output.md
Configuration reference	docs/configuration.md
Programmatic API	docs/api.md
Pipeline architecture	docs/pipeline.md
Internal design	docs/ARCHITECTURE.md
Algorithms	docs/ALGORITHMS.md
Agentic workflows (BMAD, Vibe, GSD, spec-kit)	docs/agentic-workflows.md
Troubleshooting	docs/TROUBLESHOOTING.md
Philosophy	docs/PHILOSOPHY.md
Telemetry & cognitive metrics	docs/telemetry.md

Known Limitations

• Incremental call graph updates are depth-1 only: the MCP file watcher (--watch-auto, on by default) re-indexes signatures and edges on save for the changed file and its direct callers. Transitive callers (A→B→C, C changes, A stays stale) are only refreshed by the next analyze --force. For hub files with 100+ callerFiles, re-parse may take several seconds. The watcher prunes build/dependency directories (target/, node_modules/, dist/, .venv/, vendor/, …) so it stays light even on large repos; turn it off entirely with openlore mcp --no-watch-auto.
• Static analysis only: dynamic dispatch, runtime metaprogramming, and eval-based patterns are not captured in the call graph.
• LLM spec quality varies: generated specs reflect the model's understanding. Review sections covering complex business logic before treating them as authoritative.
• Embedding is optional: plain openlore analyze (no --embed, no EMBED_*) builds a keyword (BM25) search index out of the box, so orient, search_code, suggest_insertion_points, and search_specs work immediately. Configure an embedding endpoint (EMBED_BASE_URL/EMBED_MODEL or an embedding block in .openlore/config.json) to upgrade to hybrid dense+BM25 search, which is more accurate for semantic queries.
• Large monorepos: openlore analyze on large codebases may take several minutes. Graph storage itself has no practical limit — the pipeline (AST parsing, symbol extraction) is the bottleneck.
• node:sqlite experimental warning on Node 22: Node.js 22 prints ExperimentalWarning: SQLite is an experimental feature to stderr. The warning is gone on Node 24+. Suppress on Node 22 with NODE_NO_WARNINGS=1 openlore analyze.

Requirements

• Node.js 22.5+
• API key for generate, verify, and drift --use-llm:

  export ANTHROPIC_API_KEY=sk-ant-...    # default provider
  export OPENAI_API_KEY=sk-...           # OpenAI
  export GEMINI_API_KEY=...              # Google Gemini
  

  Or use a CLI-based provider (claude-code, gemini-cli, mistral-vibe, cursor-agent) — no API key, just the CLI on your PATH.
• analyze, drift, mcp, and init require no API key

Languages supported: TypeScript · JavaScript · Python · Go · Rust · Ruby · Java · C++ · Swift · C# · Kotlin · PHP · C · Scala · Dart · Lua · Elixir · Bash — call graphs ride the same node/edge primitives for every language. See docs/languages.md for per-language extraction limits and the .h C/C++ rule.

Infrastructure-as-Code: Terraform/HCL · Kubernetes · Helm · CloudFormation · Ansible · Pulumi · AWS CDK · CDKTF — IaC resources and their references are projected onto the same graph as application code, so orient, search_code, get_subgraph, and analyze_impact answer "what is the blast radius of changing this security group / ConfigMap / IAM role?" with zero new tooling. See docs/iac.md.

Cross-domain impact (Spec 17): for embedded IaC (Pulumi/CDK/CDKTF), the code that provisions a resource is linked to it by a references edge, so analyze_impact traverses the code↔infra boundary end-to-end — "what infrastructure does this handler reach?" and the reverse, "what code breaks if I change this resource?". Infra neighbors are surfaced as a typed, ecosystem-tagged crossDomain block, distinct from the code blast radius. A code-only navigator structurally cannot answer this. Reproducible example: docs/cross-domain-impact.md.

Development

npm install
npm run build
npm test          # 2900+ unit tests
npm run typecheck

Links

• OpenSpec — spec-driven development framework
• AGENTS.md — system prompt for direct LLM prompting
• Examples — BMAD, Vibe, GSD, drift-demo, spec-kit integrations