feat: capture token usage from AI providers

[chubes4]

Summary

Stops dropping token usage data on the floor. Every AI provider in ai-http-client already parses prompt_tokens, completion_tokens, and total_tokens from API responses — Data Machine was just ignoring them.

The data flow

ai-http-client                     Data Machine
┌──────────────┐     ┌────────────────────────────────────────────┐
│ Provider      │     │ AIConversationLoop                         │
│ format_       │────▶│ accumulates usage across turns              │
│ response()    │     │                                            │
│ returns usage │     │     ┌──────────────┐  ┌──────────────────┐ │
│               │     │     │ AIStep       │  │ ChatOrchestrator │ │
│               │     │     │ → engine_data│  │ → metadata       │ │
│               │     │     │   .token_    │  │   .token_usage   │ │
│               │     │     │    usage     │  │                  │ │
└──────────────┘     │     └──────────────┘  └──────────────────┘ │
                      └────────────────────────────────────────────┘

Changes (3 files, 75 lines added)

File	What changed
AIConversationLoop.php	Accumulates usage from $ai_response['data']['usage'] across all turns. Adds usage to return value (including error returns).
AIStep.php	After loop completes, writes token_usage to job engine_data via read-then-merge pattern.
ChatOrchestrator.php	Stores token_usage in session metadata. Accumulates across multi-turn conversations. All three entry points covered: processChat, processContinue, processPing.

Storage format

Jobs (engine_data JSON):

{
  "task_type": "...",
  "token_usage": {
    "prompt_tokens": 2450,
    "completion_tokens": 380,
    "total_tokens": 2830
  }
}

Chat sessions (metadata JSON):

{
  "status": "completed",
  "token_usage": {
    "prompt_tokens": 15200,
    "completion_tokens": 3400,
    "total_tokens": 18600
  }
}

Testing

• All 3 files pass php -l
• PHPUnit: 906 tests, same pass/fail as main
• No schema changes — uses existing JSON fields

Future work

• CLI/REST endpoints to query usage per agent, per time period
• Budget enforcement (Enforce agent status before allowing chat #852 agent status + usage limits)
• Usage dashboard in admin UI

Closes #849.

[github-actions]

Homeboy Results — data-machine

Lint

Failure Digest

Lint Failure Digest

• No structured lint details available.
• Full lint log: https://github.com/Extra-Chill/data-machine/actions/runs/23278385005

Autofixability classification

• Overall: human_needed
• Autofix enabled: no
• Autofix attempted this run: no
• Human-needed failed commands:
  • lint
• Failed commands with available automated fixes:
  • lint
• Automated fixes are disabled for this step. Commands with available fix support in this run: lint

Machine-readable artifacts

• homeboy-lint-summary.json
• homeboy-test-failures.json
• homeboy-audit-summary.json
• homeboy-autofixability.json

⚡ Scope: changed files only

❌ lint (changed files only)

Test

⚡ Scope: changed files only

✅ test (changed files only)

Audit

⚡ Scope: changed files only

✅ audit (changed files only)

Tooling versions

• Homeboy CLI: homeboy 0.81.1+81c8f58a
• Extension: wordpress from https://github.com/Extra-Chill/homeboy-extensions
• Extension revision: unknown
• Action: Extra-Chill/homeboy-action@v1

---

Homeboy Action v1