Source-cited context packs for AI coding agents

Stop guessing. Cite the source.

CruxAtlas installs as a skill for your coding agent, then turns a plain-English question into compact, current evidence from trusted docs, source, tests, examples, changelogs, repository examples, and API references. Every snippet is cited, freshness-aware, and scanned so third-party text stays evidence instead of becoming instructions.

Setup: one command adds the agent skill
Grounding: line-level citations over stale memory
Safety: prompt-injection-aware context packs

POST /v1/context-pack

query "Pydantic validator that runs before type coercion"

resolve pydantic@2.9 - version-sensitive select docs - source - tests - changelog verify 6 citations - injection-scanned

docs docs/migration.md

Replace @validator(pre=True) with @field_validator(mode="before"); the v1 decorator is deprecated.

source functional_validators.py

field_validator(*fields, mode="before") registers the hook before coercion runs.

tests tests/test_validators.py

Asserts a before validator receives the raw input, not the coerced value.

answerability: sufficient citations verified version-sensitive prompt-injection scan: clear

Basic setup

Set up CruxAtlas in one command.

Requires Node.js 20 or newer. Setup auto detects Codex, Claude Code, Cursor, Windsurf, Gemini CLI, GitHub Copilot CLI, Cline, Roo Code, Kilo Code, Continue, Zed, OpenCode, Amp, Antigravity, Universal skills, and Generic MCP.

npx --yes --package cruxatlas@0.0.2 -- cruxatlas setup

Public catalog packs work from the hosted API.

Self-serve setup never writes secrets into committable config.

Restart or reload each selected agent after setup prints its product-specific instruction.

Enterprise private deployments use the same client contract and are arranged separately.

The problem

A coding agent can sound certain while its library memory is stale or its retrieved context is hostile.

Stale model memory

Fast-moving libraries outrun training data. Your agent suggests a deprecated call, a renamed option, or a removed API, then you find out only after the edit breaks.

Evidence scattered everywhere

Public APIs want docs. Setup wants examples. Edge cases live in tests. Migrations live in changelogs. Repeated searching flattens those roles into noise and burns the same context tokens your agent needs for the actual fix.

Untrusted text has a blast radius

Agents read third-party docs with file, terminal, and network tools nearby. Retrieved text should arrive framed as evidence, not as commands to obey.

How it works

The answer is not another confident paragraph. It is a packet of evidence.

Four deterministic stages turn a vague question into context an agent can use and a human can verify.

01
Resolve
Map the question to the right library, ecosystem, and version using names, aliases, packages, and version clues from the project.
02
Retrieve
Pull from curated source roles with role-aware ranking and deliberate diversity.
03
Qualify
Attach citations and freshness signals, flag conflicts, and check whether the pack can answer the question.
04
Guard
Scan third-party text for prompt-injection patterns and return it with provenance.

One API, many agents

Ask for context. Get back receipts your agent can quote.

The hosted API returns compact public-catalog packs with citations, freshness, conflicts, answerability, and prompt-injection scan signals in the same response.

# one call, source-backed answer
curl https://api.cruxatlas.com/v1/context-pack \
  -H "Authorization: Bearer ca-demo_not-a-real-key" \
  -H "content-type: application/json" \
  -d '{ "query": "FastAPI lifespan cleanup", "maxTokens": 1600, "requestId": "req_home_api" }'

# returns answerability, citations, warnings, and billable status

The difference

Same question. Very different context.

Model memory or raw search

Confident guidance from outdated training data.
Generic snippets with no line-level source to inspect.
Version changes discovered only after the edit breaks.
Unvetted third-party text lands next to powerful agent tools.

With CruxAtlas

One command gives supported agents the CruxAtlas skill.
Curated evidence chosen for the job at hand.
Line-level citations you can open and check yourself.
Freshness, conflict, answerability, and injection signals ride along.

EVIDENCE NOT INSTRUCTIONS

First class security and prompt-injection mitigation

CruxAtlas assumes retrieved docs, examples, comments, and tests can be stale, wrong, or adversarial. Packs carry provenance, citation checks, freshness/conflict warnings, and prompt-injection scan signals before your agent uses them.

Tenant isolation

Account, tenant, session, usage, billing, support, and audit routes are scoped by tenant membership and role checks.

API-key storage

API keys are hashed server side, displayed once at creation, and never stored in browser local storage by the account shell.

Query-text retention

Browser telemetry is disabled in the web shell. Raw query text is not allowed in browser analytics or route registry telemetry fields.

Source controls

Catalog membership authorizes public serving. Explicit disable, takedown, access denial, quarantine, citation-integrity, revision-membership, and cache controls can block evidence; license and source-rights metadata is preserved with the source but does not decide serving by itself.

Vulnerability disclosure

Security reports route through the contact form or security sender and are handled without attachments until scanning policy is approved.

Security assurance

Security controls are documented and tested. CruxAtlas makes certification claims only when the corresponding assessment has been completed.

Research-backed agent safety

CruxAtlas catalog curation and context-pack design are shaped by prompt-injection research. Supply-chain incidents such as Shai-Hulud are one reason source text is treated as hostile evidence: cited, scanned, bounded, and never agent instructions.

This is risk reduction, not a replacement for dependency scanning, sandboxing, credential hygiene, code review, incident response, or a public status process.