contextcore-ai-toolkit 4.0.0

5-layer hybrid context delivery system for AI coding assistants.

CONTEXTCORE

AST-first context infrastructure for AI coding workflows.

CONTEXTCORE converts raw source trees into compact, structurally meaningful context so assistants receive less noise and more relevant signals.

Highlights

• v1 sealed: 11.38x compression, 100% eval accuracy
• v2 sealed: gate passed (10/10 subgraph accuracy, 10.8ms avg latency, 577 avg tokens)
• v3 sealed: intent routing + v3 gate + v3 integration coverage passing
• v4 sealed: RBAC + freshness gate, v4 integration, and external dogfood validation passing
• local-first by design: source code stays on your machine

Current status

Track	Status
v1	SEALED (tag: v1.0)
v2	SEALED (tag: v2.0)
v3	SEALED (intent routing + v3 gate + v3 integration)
v4	SEALED (RBAC + freshness gate + v4 integration + external dogfood evidence)
Active suite	124 passed, 1 skipped, 0 failed
Gate suite	14 passed, 0 skipped, 0 failed

Why this exists

Raw file dumps are expensive and low-signal for assistants. CONTEXTCORE enforces a layered pipeline to deliver smaller, task-relevant context with measurable quality gates.

Layer Status

• L1 Static Analysis (AST): SEALED
• L4 Dependency Graph (SQLite): SEALED
• L5 Compression Emitter (Structured Markdown): SEALED
• L3 Intent Engine (task routing): SEALED
• L2 Temporal/Freshness: Implemented staleness/freshness validation and reporting; full temporal decision graph expansion is future work.

Proven results

v4 seal (sealed)

Metric	Result	Target
RBAC correctness	0 leaks across developer/auditor/maintainer	0 unauthorized nodes
Freshness correctness	stale labeling PASS, 0 false positives, reindex clears PASS	PASS
External validation	scrapy dogfood run archived on 445 Python files	200-500 file real project

v1 gate (sealed)

Metric	Result	Target
Compression ratio	11.38x	>5x
Accuracy on eval set	10/10 (100%)	>=80%
Parse failures	0/20 files	0

v2 gate (sealed)

Metric	Result	Target
Subgraph accuracy	10/10	>=8/10
Average latency	10.8ms	<=500ms
Average tokens	577	<=600

Quick start

pip install -e ".[dev]"
python tests/run_all.py
python tests/run_all.py --gate

CLI usage

contextcore index ./sample_project
contextcore status
contextcore query "where does token counting happen"
contextcore diff

Documentation map

• PROJECT.md: roadmap and version dashboards
• CONTEXT.md: session log and benchmark tracker
• DECISIONS.md: architecture decisions (ADRs)
• tests/TEST_MANIFEST.md: test inventory and gate map
• .contextcore/context_snapshot.md: current execution snapshot
• CLAUDE.md: operational rules used during implementation

Safety and privacy

• local processing only
• no source-code telemetry
• no external API calls with source code content

License

MIT (see LICENSE)