veridia

Model-agnostic quality through mechanics, not through the model.

veridia is a CLI tool that classifies a task, assesses how verifiable it is, routes it to the cheapest sufficient process, runs verification gates, and measures its own misses — so it stops guessing blindly next time.

npm install -g veridia
veridia "add user authentication"

Why veridia?

Quality comes from the process, not from which model you happen to plug in. veridia implements a deterministic triage pipeline:

TASK → CLASSIFY → ASSESS → ROUTE → ASK? → VERIFY → MEASURE
       type       level     plan    clarify  check    learn

Every mechanism is local, deterministic, and testable. AI model integration is optional — used only for execution, never for analysis.

Quick Start

Install

# Global install
npm install -g veridia

# Local install
npm install veridia
npx veridia --help

Requires Node ≥ 22.12.

Run a task

# Analyse a task description
veridia "fix login timeout"

# Analyse a specific project
veridia "add dark mode" --target /path/to/project

# Full triage loop with human-readable output
veridia run "add user authentication"

Output

{
  "type": "feature",
  "confidence": 0.29,
  "level": 3,
  "plan": { "depth": "full-tdd", "tier": "cheapest" },
  "verdict": "FAIL"
}

| Field | Meaning | |-------|---------| | type | bugfix / feature / refactor / doc / explore / open | | level | 0–3 verifiability (3 = tests + TS + CI) | | plan | full-tdd / tdd-where-possible / minimal / just-do-it | | verdict | PASS / FAIL / HUMAN (needs manual review) |

Commands

| Command | Description | |---------|-------------| | veridia <task> | Full triage pipeline (JSON output) | | veridia run <task> | Full triage with human-readable output | | veridia classify <task> | Classify task type | | veridia assess --target <path> | Assess verifiability level 0–3 | | veridia route --type <t> --level <l> | Route to orchestration plan | | veridia ask --type <t> --level <l> | Get clarifying questions | | veridia plan --type <t> --level <l> | Generate execution plan | | veridia verify --target <p> --type <t> --level <l> | Run verification gates | | veridia fix --target <p> [--dry-run] [--force] | Auto-fix console.log/TODO | | veridia measure --history | History summary | | veridia learn | Analyse history, recommendations | | veridia develop --change <name> | Dogfooding: run triage against a change | | veridia init | Wire veridia into AI agents | | veridia agents --list | List supported agents |

Streaming output

# See gate output in real-time
veridia verify --verbose --target . --type refactor --level 3
veridia run --verbose "fix login bug"

Debug logging

VERIDIA_DEBUG=1 veridia run "add auth"

Architecture

veridia is built around six deterministic mechanisms:

1. Classify — pattern-based task type detection (bugfix/feature/refactor/doc/explore/open)
2. Assess — probe target repo for tests, type-check, lint, CI → verifiability level 0–3
3. Route — map (type, level) → orchestration depth + model tier
4. Ask — generate clarifying questions when level is 0/1
5. Verify — run gates (vitest, tsc, lint), weigh results, derive verdict
6. Measure — record outcomes, compute precision, detect drift

All mechanisms are deterministic and unit-testable. AI models are optional and only used for execution.

Security

When no AI model is configured, run/execute may delegate verification gates to the local shell. Control this with:

VERIDIA_SHELL_DELEGATION=deny veridia run "..."
VERIDIA_SHELL_DELEGATION=ask veridia run "..."

Docs

• Usage — setup, commands, session mode
• Mechanics — the six mechanisms
• Verifiability — the level 0–3 model
• Roadmap — project status and plans

License

MIT