Context OS for AI coding agents: task-scoped context, AST code streams, verified mutation, and rollback. Auto-governs context via MCP, reducing context window waste by 90%+
ContextOS defines 5 tools with reasonable naming (explore, change, verify, ship, ops) but suffers from critical gaps in schema completeness and parameter documentation. Tool descriptions are present but generic (50-150 chars, below the 194-char baseline). Input schemas visible in code show type definitions for most parameters, but lack comprehensive descriptions for individual parameters, critical for LLM parameter selection. The 'ops' tool is a catch-all passthrough that violates single-responsibility principle. No output schemas documented. Error handling guidance absent. STDIO transport caps protocol readiness at 50, further limiting production viability.
Modify or create code. The OS locates the anchor, applies a surgical AST edit, re-anchors symbols and records the touched files.
Understand, locate or resume. The OS inspects git state, picks candidate modules, outlines symbols and injects the applicable rules, all inside one context budget.
Manual passthrough to the internal capabilities (legacy facades, session, profile, system). Use only when the intent-level tools are not enough.
Close the loop: collect touched files and receipts, optionally verify first, export the architecture graph and archive the session.
Prove the change runs. With no commands the OS uses .contextos/profile.json or package.json scripts. Output stays out of context; only receipts and failures come back.
ops tool violates single-responsibility principle, it is a generic passthrough to internal capabilities rather than a focused, composable tool. LLMs cannot determine when to use it vs. intent-level tools.
Parameter descriptions are missing or minimal. 'intent', 'paths', 'depth', 'projectRoot' appear in schemas but lack LLM-facing descriptions explaining what values are valid, when to use them, or dependencies between parameters.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields (e.g., touched file paths from 'change', process IDs from 'verify') without knowing response structure.
Tool descriptions lack recovery guidance and error classification. No indication of what happens on failure, whether operations are retryable, or what the LLM should do next.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
Destructive operations (change, verify, ship) lack dry-run or confirmation patterns. 'change' has dryRun parameter, but 'ship' and 'verify' do not, agents could accidentally close sessions or execute unintended commands.