Trikon

The verification layer for autonomous AI coding agents.

Autonomous AI agents (Claude Code, Codex, Cursor, Unideploy autopilot, custom nightly bots) are increasingly committing, merging, and deploying code without a human in the loop. Trikon sits between the agent and production. Before a change is merged or deployed, Trikon:

1. Analyzes the change's blast radius using an AST-based dependency graph.
2. Executes the targeted subset of tests, static checks, and policy rules that the change actually affects.
3. Produces a machine-readable Verdict — ALLOW, BLOCK, or REQUIRE_HUMAN — with structured evidence.
4. Records the verdict in a hash-chained audit log for compliance.

Install

pip install trikon

Trikon requires Python 3.11+ and (optionally) Docker Desktop for the sandboxed verification backend. Verify your environment with trikon doctor.

Why this exists

The bottleneck in software engineering has moved from writing code to verifying it. Human review does not scale with agent output. Human review is also the wrong tool for unattended agents that run at 3 a.m.

The market is full of tools that ask an LLM whether a diff looks correct. Trikon executes the code and produces evidence.

Positioning

Relationship to Unideploy

Trikon is a separate product, not a Unideploy feature. It reuses Unideploy's engine as internal infrastructure:

• Unideploy's warden sandbox provides isolated test execution.
• Unideploy's AWS backend pattern (Lambda + DynamoDB + API Gateway + Cognito) is copied for the Trikon control plane.
• Unideploy's BYOM licensing flow is mirrored for Trikon subscriptions.
• Unideploy's autopilot is the first customer of Trikon (verifies its own actions before touching prod).

See ARCHITECTURE.md for the full design.

Status

Pre-alpha. Phase 1 (change intelligence) is implemented end-to-end; the verification runner (Phase 2), policy engine (Phase 3), and hosted control plane come later. See EXECUTION_PLAN.md for the phase-by-phase map.

Phase 1: what ships today

• Change-intelligence pipeline — parse_diff → AST indexer → symbol resolver → dep graph → compute_impact, producing a structured ImpactSet (impacted files, symbols, modules, public APIs, tests, blast-radius score). Frozen contract in .kiro/specs/change-intelligence/design.md.

• CLI: trikon debug impact — runs the whole pipeline against a git repo and prints the ImpactSet as JSON. The first end-to-end runnable checkpoint.

• Never-fail-open guarantee — every exception raised inside trikon/change_intel/** is a subclass of ChangeIntelError. The SDK boundary catches the base class and returns a Verdict with decision="require_human" and blast_radius_score="HIGH". allow is never emitted on an error path. See docs/change_intel.md.

• Verification (Phase 2) — the same pipeline now executes the ImpactSet. Targeted pytest runs, static analysis (ruff + mypy), and custom .trikon/checks/*.py plugins all run inside a pinned, network-isolated Docker sandbox (trikon/sandbox:0.1.0). Every raise site under trikon/verify/** is a VerificationRunnerError subclass, so sandbox failures still land on require_human — the never-fail-open guarantee extends to Phase 2. Full walkthrough in docs/verification.md. One-line demo:

  trikon debug verify --repo examples/sample_repo --base HEAD~1 --head HEAD
  

Try it locally:

pip install -e ".[dev]"
trikon debug impact --repo examples/sample_repo --base HEAD --head HEAD

For the Phase 2 verification runner, the warm path (CoverageMap-hit, sub-15 s verdicts) requires a one-time coverage-map build per repo before the first trikon debug verify invocation:

# 1. Build the coverage map (one-time setup per repo)
trikon coverage build --repo .

# 2. Run verification on a change
trikon debug verify --repo . --base HEAD~1 --head HEAD

See docs/verification.md for the full verification pipeline (sandbox, test selection, static checks, plugins).

Phase 2 (targeted test execution + sandbox) and Phase 3 (policy engine + trikon verify) are next. Until they land, sdk.verify() returns require_human because there is no verification evidence to grade yet.

MCP integration

Trikon exposes a trikon_verify MCP tool that any MCP-capable AI agent (Claude Code, Cursor, Kiro) can call:

trikon mcp serve --transport stdio

See docs/mcp-integration.md for per-editor configs.

Documentation

Immediate next steps

1. Engine language: Python 3.11+ (locked, see ARCHITECTURE.md §7).
2. MVP scope: Python + pytest + Docker sandbox + CLI + MCP tool. No hosted backend in v0.1. See EXECUTION_PLAN.md.
3. First design partner: needed. One Python codebase with an active AI agent generating unattended PRs. This is the only truly blocking decision.

License

TBD — likely Apache-2.0 for the core engine, commercial for the hosted backend.