feat(fleet-metrics): add autonomous fleet metrics dashboard page

[TimeToBuildBob]

Summary

• Adds /fleet-metrics/ static dashboard page showing Bob's fleet-level metrics over the last 30 days
• Self-contained HTML (no external dependencies) generated from state/sessions/sessions.db
• Shows: session volume/day, productive rate, deliverables count, avg quality score, and model×harness productivity breakdown

What's on the page

Headline tiles — 30-day totals: sessions, productive %, deliverables, avg quality score

Sessions/day bar chart — daily session count (light blue) with productive overlay (green) and quality score trend line (indigo)

Model × harness table — productivity rate and LLM judge score per model+harness combination, colored by harness (blue=claude-code, amber=codex, green=gptme)

Generator

Lives in the brain repo at scripts/generate-fleet-metrics-dashboard.py. Regenerate with:

python3 /home/bob/bob/scripts/generate-fleet-metrics-dashboard.py generate

The script queries state/sessions/sessions.db directly — no manual numbers, no hardcoded values.

Why

A live, data-backed public dashboard of fleet metrics is a differentiated artifact that demonstrates capability rather than describing it. Distinguishes this project from the typical "AI blog post" pattern.

[ErikBjare]

@TimeToBuildBob I merged it, but seems like nothing updates it?

[TimeToBuildBob]

Fixed the missing update path. The page was generated once for the PR, but nothing scheduled or published subsequent snapshots.

I wired it into the existing daily dashboard timer (bob-session-cost-dashboard.timer): it now regenerates fleet-metrics/index.html, publishes both dashboards from an isolated website worktree, and retries push races. I ran the full service manually and verified:

• website commit: c6daa04
• deploy run: https://github.com/TimeToBuildBob/TimeToBuildBob.github.io/actions/runs/30810679234 (green)
• live page now says generated 2026-08-03 11:43 UTC: https://timetobuildbob.com/fleet-metrics/
• next automatic refresh: tomorrow around 03:00 UTC, then daily

The durable fix is in brain commit 9da7bca979.

[ErikBjare]

@TimeToBuildBob It looks okay I guess, but it reduces all GPT-5.6 versions (luna, terra, sol) into just "5.6" and for DeepSeek flash and pro variants into a single "DeepSeek V4".

[TimeToBuildBob]

Fixed in brain commit 76d87265bb. The model labels now preserve variant names:

• GPT-5.6-sol, GPT-5.6-terra, GPT-5.6-luna each show distinctly (arm suffix preserved)
• DeepSeek V4 Flash and DeepSeek V4 Pro are labeled separately
• Two API routes for the same arm (e.g. direct vs openrouter deepseek-flash) are merged into one row rather than appearing as duplicates

Updated page at https://timetobuildbob.com/fleet-metrics/

[ErikBjare]

@TimeToBuildBob From it:

Harness:Model 	Productive rate 	% 	Score 	Sessions
claude-code:Sonnet 4.6 	
	87.0% 	0.55 	4,101
gptme:GPT-5.6-sol (sub) 	
	79.5% 	— 	806
claude-code:Haiku 4.5 	
	98.9% 	— 	555
gptme:GPT-5.6-sol 	
	90.4% 	0.55 	385
codex:GPT-5.5 	
	98.4% 	0.51 	380
codex:GPT-5.4 	
	96.0% 	0.52 	199
gptme:DeepSeek V4 Flash 	
	73.2% 	0.46 	183
claude-code:Haiku (old) 	
	82.2% 	0.53 	135
claude-code:Sonnet (old) 	
	100.0% 	— 	66
codex:GPT-5.6-sol 	
	68.8% 	0.46 	64
claude-code:Fable 5 	
	98.4% 	0.70 	63
gptme:GPT-5.5 	
	76.2% 	0.53 	63

What's up with the — scores? Should be fixed.

[TimeToBuildBob]

Fixed. Root cause: the model×harness table was including all run types, but monitoring/PM sessions are never LLM-judged by design — haiku-4-5 and openai-subscription/gpt-5.6-sol are used exclusively as PM slots, so they had hundreds of sessions with no quality scores.

Added a run_type exclusion filter to the SQL query (monitoring, project-monitoring, email, voice-post-call, etc.). Only autonomous/worker/operator sessions remain in the table — the ones that actually receive llm_judge_score.

Generator fix: brain 25e64a05fe, live on the site now (aceae4f).