Thanks for asking me to work on this. I will get started on it and keep this PR's description up to date as I form a plan and make progress.

---

This section details on the original issue you should resolve

<issue_title>feat: daily token usage analysis workflow</issue_title>
<issue_description>## Summary

Create a daily agentic workflow that mines token usage data from recent workflow runs, identifies trends and inefficiencies, and creates a summary issue with findings and optimization recommendations.

Background

PR #1539 introduced token usage tracking in the api-proxy sidecar. Each workflow run with --enable-api-proxy now produces a token-usage.jsonl artifact containing per-request records with:

{"timestamp":"...","request_id":"...","provider":"anthropic","model":"claude-sonnet-4-6","path":"/v1/messages?beta=true","status":200,"streaming":true,"input_tokens":3,"output_tokens":418,"cache_read_tokens":14044,"cache_write_tokens":26042,"duration_ms":5858,"response_bytes":2800}

This data is captured in the agent-artifacts upload (under sandbox/firewall/logs/api-proxy-logs/token-usage.jsonl) but is not yet analyzed systematically.

Proposed Workflow

Trigger

• Daily schedule (e.g., 08:00 UTC)
• workflow_dispatch for manual runs

Behavior

1. Discover runs since last analysis

   • Use gh run list to find completed agentic workflow runs since the last analysis issue
   • Cover all smoke tests, secret-digger, and other workflows that use --enable-api-proxy

2. Download and aggregate token-usage.jsonl

   • For each run, download the agent-artifacts artifact
   • Extract token-usage.jsonl from the artifact
   • Gracefully handle runs where token logs are not available (this is a new feature — older runs and runs without --enable-api-proxy won't have logs)

3. Compute per-workflow statistics

   • Total tokens (input + output + cache read + cache write)
   • Input/output ratio
   • Cache hit rate (cache_read / (cache_read + input))
   • Average request count per run
   • Average duration per request
   • Model mix (which models are used, in what proportions)

4. Identify trends

   • Compare current period vs previous period (if historical data available)
   • Flag workflows with increasing token consumption
   • Flag workflows with low cache hit rates (opportunity for prompt caching)
   • Flag workflows with high input/output ratios (may indicate bloated prompts or tool schemas)

5. Generate optimization recommendations

   • Workflows with zero cache hits → suggest enabling prompt caching
   • Workflows with >100:1 input/output ratio → suggest reducing system prompt or tool surface
   • Workflows with many small requests → suggest batching
   • Workflows with unusually high total tokens → flag for review

6. Create a summary issue

   • Title: Token Usage Report: YYYY-MM-DD
   • Body includes:
     • Overview table (workflow name, total tokens, cost estimate, cache rate)
     • Per-workflow detail sections (collapsible)
     • Historical trend chart (if prior reports exist)
     • Top optimization opportunities
     • Link to previous report issue (if any)
   • Label: token-usage-report

Graceful Degradation

Since token tracking is new:

• Skip runs that have no agent-artifacts artifact
• Skip runs where token-usage.jsonl is missing or empty
• Report which workflows are not yet instrumented
• Don't fail the workflow if no data is available — create a minimal report noting the gap

Tools Needed

• github MCP server — for listing runs, downloading artifacts, creating issues
• bash — for processing JSONL files, computing aggregates
• web-fetch — not needed (all data is in GitHub artifacts)

Permissions

• actions: read — to list workflow runs and download artifacts
• issues: write — to create summary issues

Token Usage Record Schema

Field	Type	Description
timestamp	string	ISO 8601 timestamp
request_id	string	Unique request ID
provider	string	anthropic, openai, copilot, opencode
model	string	Model name (e.g., claude-sonnet-4-6)
path	string	API endpoint path
status	number	HTTP status code
streaming	boolean	Whether response was SSE streamed
input_tokens	number	Input/prompt tokens
output_tokens	number	Output/completion tokens
cache_read_tokens	number	Prompt cache read tokens
cache_write_tokens	number	Prompt cache write tokens
duration_ms	number	Request duration in milliseconds
response_bytes	number	Response body size

Example Analysis (from smoke-claude run)

From a recent smoke-claude run (6 requests):

• Total tokens: 197,558 (775 input, 1,227 output, 152K cache read, 43K cache write)
• Cache hit rate: 99.5%
• Models: claude-haiku-4.5 (routing), claude-sonnet-4.6 (main agent)
• Estimated cost: ~$0.23
• Key finding: Anthropic prompt caching is working well (63% cost savings)

Related

• PR feat: add token usage tracking to api-proxy sidecar #1539 — Token tracking infrastructure
• PR feat: include api-proxy token logs in firewall audit artifact #1549 — Include token logs in firewall audit artifact
• PR fix: decompress gzip responses for Anthropic token extraction #1550 — Fix gzip decompression for Anthropic token extraction</issue_description>

<agent_instructions>Create a PR that addresses this issue</agent_instructions>

Comments on the Issue (you are @copilot in this section)