Skip to content

architecture.md

Latest commit

 

History

History
90 lines (70 loc) · 3.27 KB

architecture.md

File metadata and controls

90 lines (70 loc) · 3.27 KB

Architecture

Data flow

~/.copilot/session-state/<uuid>/events.jsonl
~/.copilot/session-state/<uuid>/workspace.yaml
           │
           ▼
       harvest.py
  - Scans all session directories for target date
  - Extracts user instructions from user.message events (filters approvals + injected context)
  - Captures tool summaries from assistant.message.toolRequests[].intentionSummary
  - Reads token breakdown, premium requests, AI time, code changes from session.shutdown
  - Reads project summary, cwd, repo, branch from workspace.yaml + session.start
  - Returns: list of session dicts
           │
           ▼
       analyze.py
  - Builds a structured transcript from session data (includes code impact, workspace summary)
  - Calls GitHub Models API (gpt-4o-mini) using gh CLI token
  - Returns: goals[] with tasks[], skills, hours, docs_referenced
  - Caches result to <install-dir>/cache/YYYY-MM-DD.json
           │
           ▼
       report.py
  - Generates Outlook-compatible HTML
  - Layout: header → narrative → KPI cards → goals table → activity bar → token bar → task accordion
           │
           ▼
   email_send.py (optional)
  - Writes HTML to temp file
  - PowerShell Outlook COM automation sends it

Session file format

Copilot writes one directory per session at ~/.copilot/session-state/<uuid>/.

events.jsonl

Each line is a JSON object. Relevant event types:

Type Content
session.start data.context: cwd, gitRoot, repository, branch, headCommit
user.message data.content: raw user instruction (may include injected <current_datetime> tags)
assistant.message data.toolRequests[]: name, intentionSummary (human-readable tool call summary)
tool.execution_start data.toolName, data.arguments
tool.execution_complete data.model, data.success, data.result
session.shutdown data.totalPremiumRequests, data.totalApiDurationMs, data.codeChanges, data.modelMetrics

workspace.yaml

Simple key:value file with fields: id, cwd, git_root, repository, host_type, branch, summary, created_at, updated_at.

The summary field is a Copilot-generated session title (e.g. "Create Exec Deck From PBIP").

Token cost model

Token data is in session.shutdown.modelMetrics.<model>.usage. Uses Anthropic's published per-token rates (sessions use Claude models via Copilot):

Token type Field Rate
Input inputTokens $3.00 / 1M
Output outputTokens $15.00 / 1M
Cache read cacheReadTokens $0.30 / 1M
Cache creation cacheWriteTokens $3.75 / 1M

Update these in report.py → _cost() if rates change.

Leverage metric

human_value    = total_human_hours × HOURLY_RATE   ($72/hr blended rate)
seat_cost/mo   = $39/mo enterprise plan
leverage       = human_value / seat_cost_per_month

Example: 29h × $72 = $2,088 human value ÷ $39/mo seat = 54×

This measures return on Copilot seat investment per day used.

GitHub Models API

  • Endpoint: https://models.inference.ai.azure.com/chat/completions
  • Auth: Authorization: Bearer <github_token> (from gh auth token)
  • Model: gpt-4o-mini (OpenAI-compatible request body)
  • No extra credentials — uses the same GitHub token as gh CLI