Skip to content

mcp_context_layer.md

Latest commit

 

History

History
118 lines (94 loc) · 3.78 KB

mcp_context_layer.md

File metadata and controls

118 lines (94 loc) · 3.78 KB

MCP Context Layer

CompText V7 includes a lightweight replay-aware context layer for MCP-style runtime systems. The layer preserves compact operational context for deterministic replay validation; it does not execute tools, orchestrate agents, call external APIs, or judge semantic quality.

Scope

The context layer is additive to the existing replay fixtures and contracts. It builds a compact payload from explicit trace, state, and dependency-graph fields:

{
  "task": "mcp_trace_replay_v1",
  "constraints": [
    "execute_external_action:requires_human_approval",
    "execute_external_action:requires_validation_passed"
  ],
  "required_order": [
    "capability_scope_checked",
    "tool_schema_validated",
    "read_context",
    "validate_external_action",
    "execute_external_action"
  ],
  "blockers": [
    ["capability_scope_checked", "execute_external_action"]
  ],
  "dependency_chains": [
    ["validate_external_action", "execute_external_action"]
  ],
  "recovery": [
    ["execute_external_action", "recovery_path_registered"]
  ]
}

Public API

  • build_replay_payload(trace) extracts compact deterministic commitments.
  • render_prompt_context(payload) renders a compact prompt-safe text view.
  • validate_replay_payload(payload) validates replay admissibility from explicit payload fields.
  • ContextStore(root).save_context(trace) writes a stable JSON payload.
  • ContextStore(root).load_context(task_id) restores the compact payload.
  • save_context(trace, store_dir=...) and load_context(task_id, store_dir=...) provide module-level convenience wrappers.

Deterministic checks

validate_replay_payload detects:

  • missing preserved constraints as CONSTRAINT_DRIFT
  • dependency-order drift as TOOL_ORDER_VIOLATION
  • dependency collapse as DEPENDENCY_CHAIN_BREAK
  • missing recovery paths as RECOVERY_PATH_LOSS

The validator uses exact strings, ordered lists, and explicit dependency edges. It performs no embedding lookup, fuzzy scoring, probabilistic reasoning, LLM judging, runtime blocking, or policy enforcement.

Prompt context rendering

render_prompt_context(payload) produces a token-light text form without dumping raw trace, state, or dependency-graph documents:

task: mcp_trace_replay_v1
admissible: true
constraints:
- execute_external_action:requires_validation_passed
required_order:
- validate_external_action
- execute_external_action
dependencies:
- validate_external_action -> execute_external_action
recovery:
- execute_external_action -> recovery_path_registered

Example artifact

artifacts/mcp_context_layer_example.json is a fixture-bound deterministic example generated from fixtures/mcp_trace_replay_v1/original. It contains the compact replay payload, prompt-context rendering, and validation result for the baseline MCP trace replay fixture.

CLI usage

Generate compact prompt context with validation from the baseline MCP fixture:

python scripts/mcp_context_cli.py \
  --fixture fixtures/mcp_trace_replay_v1/original \
  --render-prompt \
  --validate

Add --json to emit deterministic JSON containing the compact replay payload, optional prompt context, and optional validation result.

Local tool adapter

scripts/mcp_context_tool.py accepts one JSON request from stdin or --request-file and emits one deterministic JSON response. It is a local adapter for later MCP wrapping, not a server or runtime tool executor.

printf '%s\n' '{"tool":"build_replay_payload","params":{"fixture":"fixtures/mcp_trace_replay_v1/original","render_prompt":true,"validate":true}}' \
  | python scripts/mcp_context_tool.py

Relationship to MCP

This layer augments context integrity for MCP-compatible systems. It is not an MCP implementation and does not replace MCP transport, runtime execution, or tool semantics.