Cost Management

Autowright treats AI spend as a first-class output. Every fix attempt records what it actually cost — not an estimate — so you always know what self-healing is worth to you.

The cost model

Each fix attempt captures, from the agent run itself:

Metric	What it is
Cost (USD)	Real dollars for the attempt
Tokens	Input, output, and cache tokens separately
Turns	Agent conversation turns used
Duration	Wall-clock time for the attempt

A trivial selector fix lands around $0.09. Harder fixes cost more turns and more tokens — and you’ll see exactly how many.

Where costs surface

Slack — every fix-proposed and needs-attention message carries the attempt’s cost, turns, and duration.
Dashboard overview — running totals: Total cost across all fixes and Avg/fix with total turns.
Fix detail page — the full breakdown per fix: USD, turns, duration, and input/output/cache token counts.

The levers

Deduplication is the big one

The largest cost risk isn’t an expensive fix — it’s paying for the same fix repeatedly. A cron script hitting a broken selector all night reports dozens of failures; dedup by error hash collapses them into one fix job. The window is AUTOWRIGHT_DEDUPE_TTL_MS (default 24 hours) — lengthen it for slow-moving fleets, shorten it if sites you scrape change multiple times a day.

Concurrency caps burst spend

MAX_CONCURRENT_FIXES (default 2) bounds how many agent runs can bill at once. A site redesign that breaks ten scripts simultaneously queues ten jobs — the cap decides whether that’s ten parallel agent bills or a steady trickle.

Turns are bounded

Every agent run is capped at 50 turns. An attempt that can’t converge stops there and reports CANNOT_FIX with what it learned — it can’t spiral.

The confidence gate prevents churn

Low-confidence attempts don’t open PRs, so you never pay review time (or CI minutes) on speculative fixes. A CANNOT_FIX verdict still costs its attempt — but it delivers a diagnosis, which is usually the cheap part of the incident anyway.