System Design

Canon is a Python application built on FastAPI, with modular components for spec parsing, agent analysis, and ticket synchronization.

Platform Components

┌──────────────┐  ┌──────────────┐  ┌─────────────┐
│  Doc Engine   │  │ Agent Runner │  │  Knowledge  │
│               │  │              │  │    Base     │
│ Parse all md  │  │ Claude calls │  │ Vector DB   │
│ Track specs   │  │ PR analysis  │  │ BM25 index  │
│ Diff & delta  │  │ Realization  │  │ Hybrid      │
│ Verify intent │  │ Doc gen      │  │ search      │
└───────────────┘  └──────────────┘  └─────────────┘
         │               │               │
    ┌────┴───┐     ┌─────┴────┐    ┌─────┴────┐
    │ Tickets│     │  Comms   │    │Dev Tools │
    │        │     │          │    │          │
    │ Jira   │     │ Slack    │    │ MCP      │
    │ Linear │     │ Teams    │    │ VS Code  │
    │ GitHub │     │ Email    │    │ Cursor   │
    └────────┘     └──────────┘    └──────────┘

FastAPI App

File: src/canon/main.py

The main application exposes:

• POST /webhook — GitHub App webhook endpoint
• GET /healthz — Health check
• POST /auth/device/code + /auth/device/token — Device Authorization Grant for CLI login
• POST /{org}/api/tickets/* — Server-proxied ticket CRUD (used by CLI sync)
• Web routes for the SPA and fallback templates

All webhook events are verified using HMAC SHA-256 signature checking before being dispatched to handlers.

Spec Parser

Files: src/canon/parser/

The parser processes markdown spec files into structured Pydantic models:

• parse.py — Main parser (frontmatter extraction + section parsing + status comment parsing)
• models.py — Pydantic models: SpecDocument, SpecSection, AcceptanceCriterion
• writer.py — Writes back to spec files (insert ticket links, update status comments)

Parse Flow

Agent Runtime

Files: src/canon/agent/

The Claude agent runtime handles PR analysis:

• client.py — Anthropic SDK wrapper with model config and error types
• analyzer.py — Analysis orchestration, response parsing, comment formatting
• prompts.py — System prompt construction, user message builder, token estimation

Model Configuration

Default values as of the current release — see agent/client.py for the latest:

Parameter	Value
Model	Claude Sonnet 4.5
Max input tokens	128,000
Max output tokens	16,000
Temperature	0 (deterministic)

Prompt Architecture

The system prompt is assembled from three pieces:

1. Base system prompt — Agent role, rules, severity levels
2. Realization instructions — AC evaluation criteria (when enabled)
3. JSON schema — Structured output enforcement

The user message includes PR metadata, spec summaries, context docs, changed file list, and diffs (budget-limited).

GitHub Client

Files: src/canon/github/

• verify.py — HMAC SHA-256 webhook signature verification
• client.py — GitHub API client (JWT auth via GitHub App, httpx)
• spec_utils.py — Spec file detection and loading from GitHub
• handlers/ — Event handlers for push, PR, issue_comment, etc.

Ticket Sync

Files: src/canon/sync/

Bidirectional synchronization between spec sections and external ticket systems:

• engine.py — Forward/reverse sync logic
• status_map.py — Bidirectional spec-to-ticket status mapping
• models.py — Pydantic models for sync operations
• adapters/ — Per-system adapters (Jira, Linear, GitHub Issues, API Proxy)

Server-Proxied Sync

Files: src/canon/sync/adapters/api_proxy.py, src/canon/web/ticket_routes.py

When CLI users are logged in (canon login), the sync engine uses the CanonApiAdapter instead of calling GitHub directly. This adapter routes ticket CRUD through the Canon server's /api/tickets/* endpoints, which use the GitHub App installation token. The server endpoints are in web/ticket_routes.py.

Adapter Interface

class TicketAdapter:
    async def create_ticket(self, input: CreateTicketInput) -> CreateTicketResult
    async def update_ticket(self, id: str, changes: dict) -> None
    async def sync_status(self, id: str) -> str
    async def link_pr(self, ticket_id: str, pr_url: str) -> None
    async def close_ticket(self, id: str, evidence: str) -> None

Config Parser

Files: src/canon/config/

• parse.py — YAML parsing into Pydantic models with validation

Validates CANON.yaml files and provides typed access to all configuration values. Unknown keys produce warnings rather than errors for forward compatibility.

Cron Jobs

Files: src/canon/cron/

• sync_status.py — Standalone reverse sync script run as a Kubernetes CronJob

Periodically polls ticket systems for status changes and updates spec files accordingly.