Entitystore

entitystore

The EntityStore engine — the machinery behind a provenance-tracked company brain. Carved out of context-engineering so the engine is reusable and schema-agnostic: it never embeds an entity schema, it reads one by path.

Why it exists (the split)

| Layer | Owns | Lives in | |---|---|---| | entitystore (this skill) | engine: raw/events/wiki tiers, Source ABC, consolidator, auditor, packer, MCP | here (sibling of context-engineering) | | schema + canonical entities | entity.schema.json + the git source-of-truth entities | company-brain | | domain Sources / connectors | company-knowledge→events, HubSpot / Notion / Granola adapters | syroco-product-ops |

Because the schema lives with the data (not the engine), adding a new entity kind or property is a schema + data change with ZERO engine change. The engine validates and packs whatever schema company-brain hands it.

Storage: JSON-native, not CE's three-tier raw/events/wiki

Superseded by the JSON-native pivot (2026-05-28, see SURFACE.md "Status"): entitystore has no raw/ or events/ tier and no events-log consolidator. corpora//entities/**/*.json (git-committed, enricher-written) is the one source of truth every endpoint reads directly.

corpora//wiki/*.md DOES exist again as of M11 — but it's this repo's own scripts/wiki_init.py deriving pages straight from the JSON entities (a regenerable projection), not CE's events-log consolidator of the same name. See SURFACE.md "Wiki pages (M11)" for the contract; don't confuse the two wiki_init.pys if you're cross-referencing context-engineering/scripts/wiki/.

• Auditor (cb_engine.py's wiki_audit) surfaces contradictions / dead

links / freshness expiries / orphans / schema-invalid, plus (M11) merge / split / stale-supersession / last_verified_at freshness lints — render_proposals writes them to audit/proposals.md.

• Packer (wiki_pack in cb_engine.py) assembles a depth-banded,

budget-bounded context bundle.

Schema-injection contract

Every validation / emission entrypoint takes --schema . Reference impl: scripts/validate_corpus.py --schema --corpus . The engine holds no opinion on kinds, claims, or overlays — those are the schema's job.

Status — JSON-native MCP SHIPPED + audit-pass landed (2026-05-28)

The engine pivoted to JSON-native after empirics overturned the "build on gbrain as library" call. Same day, a /code-review pass reversed three "wrong defers" (semantic, depth-banding, git commit-through) and a /simplify pass deleted ~3000 LOC of CE markdown carcass. See SURFACE.md for the locked v1 contract and the memory files (project_syroco_company_brain, gbrain-vs-anabasis) for the premortem and revised verdict.

• Shipped (v1, post-audit): scripts/cb_engine.py (six JSON-native

endpoints) + scripts/cb_mcp.py (local stdio FastMCP server) + scripts/cb_embed.py (semantic resolver — Mistral default, OpenAI fallback) + scripts/cb_mcp_smoke.py (live wire-protocol smoke test) + scripts/validate_corpus.py (schema-injection seam). Engine self-test 15/15 PASS (incl. negative test for contradiction detector + path-traversal guard, all running in a tempdir copy of the corpus). MCP smoke test 17/17 PASS over real JSON-RPC against the live 248-entity syroco-commercial corpus. Charter-aware auditor logic ported from company-brain/scratch/promote_gate.py. Schema-injection seam preserved (CB_SCHEMA_PATH env). Registered in ~/.claude.json as companybrain MCP.

• Six endpoints: wiki_ask (substring | semantic | hybrid + neighborhood) /

wiki_pack (depth-banded budget-bounded bundle: Full/Detail/Summary/ Headlines/Mention) / wiki_audit (charter-aware contradictions + deadlinks + freshness + orphans + schemainvalid) / wiki_add (validate + write + git commit-through, path-traversal-safe) / stats (counts + freshness + embedding-provider status) / resolve (slug/alias/name).

• Wrong-defer reversals (live-proven): semantic via Mistral, depth-banded

wiki_pack (43 items packed into 3997/4000 token budget), git commit-through (real commit sha returned via subprocess git invocation).

• Honest deferrals to v1.1: structured-query (filter on claim values),

cross-corpus, Vercel/HTTP + OAuth, embedding-cache management UI.

Wiki/document-entity layer (M11)

scripts/wiki_init.py seeds corpora//wiki/.md — one page per entity, derived straight from the JSON corpus (not an events log; that was CE's dropped model). A page is a regenerable projection, never a second source of truth: --rebuild reproduces it byte-for-byte from the entities, nothing here writes an entity or a claim (THE WRITER RULE holds). Cap-aware, pointer-based (links to ids/refs, never copies claim evidence quotes), carries last_verified_at verbatim + decision-continuity fields (supersedes/superseded_by/valid_until), and never stores a freshness_score/confidence/trust/tier field.

scripts/freshness_policy.py computes freshness on read only from last_verified_at with per-kind half-lives; a missing timestamp is score=None, status="pre-rule, never verified" — never 0.0, never an error (most of the live corpus is pre-rule as of M11 — person 0/331, vessel 0/131 last_verified_at coverage).

cb_engine.py's wiki_audit gained four more report-only lints — merge candidates, split candidates, stale supersessions, and the last_verified_at-first freshness lint — and cb_engine.py wiki-audit --proposals renders the full result to /audit/proposals.md. None of this is an MCP tool addition; it's a CLI/library layer SURFACE.md's six MCP endpoints don't depend on. See SURFACE.md "Wiki pages (M11)" for the full contract.

Not in scope

Code-context packing, code-AST indexing, code-knowledge-graph viz, CSB benchmarks — those stay in context-engineering, which depends on this engine.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: VictorGjn
• Source: VictorGjn/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.