Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

Include separately approved reductions in quota reassessment

Barack Sokullu89bf5feb0cff
diff --git a/docs/compass-quota-rollout.md b/docs/compass-quota-rollout.md
index 0be60b322..628496a03 100644
--- a/docs/compass-quota-rollout.md
+++ b/docs/compass-quota-rollout.md
@@ -59,7 +59,7 @@ The report contains quota-card views, checkout opens, server-recorded claim subm
Metered consumption is recorded under the same store lock from actual completed token usage, including visual sidecars. Numeric samples use hashed device/session identifiers and retain 30 UTC days; no prompt, post body, or page content is added to these samples. Each consumption map retains at most 10,000 observations across retained dates; existing samples continue accumulating, and new keys contribute to overflow totals. Exact metered aggregates remain intact. The report discloses observation counts and overflow USD/requests in `sampling`; inspect that coverage before using the sampled distributions to change the allowance. The 14-day report shows free-session and free-device/week p50/p75/p90/p95 costs and `fraction_over_allowance` for the active amount. It retains `fraction_over_0_375` for evaluating the previously proposed reduction. Calls without a session identifier contribute to `unattributed_usd`, so report that coverage gap. Device/week distributions use completed Monday–Sunday UTC weeks independently of the rolling event/session window. With `--days=14`, verify `device_week_window` reports two completed weeks after activation, inspect its retained overflow/untagged coverage totals, and wait until both weeks have full instrumentation. A partly expired retention-boundary week is omitted; reports shorter than seven days have no completed weekly distribution. Admission-week tags keep requests that finish after reset in their original UTC week. Lifetime claims/indexes are not subject to sample retention.
-Reassess whether $0.50 covers useful completed sessions, how many device/weeks exhaust it, how often rewards lead to explicit continuation, how many users switch providers or subscribe, and how long pending reviews take. Use actual consumption distributions and recovery/conversion behavior to decide whether to keep or raise the allowance. Do not equate views or composer opens with published posts or successful tasks.
+Reassess whether $0.50 covers useful completed sessions, how many device/weeks exhaust it, how often rewards lead to explicit continuation, how many users switch providers or subscribe, and how long pending reviews take. Use actual consumption distributions and recovery/conversion behavior to decide whether to keep $0.50, raise it, or propose a separately approved reduction to $0.375 once adoption and consumption support it. Do not equate views or composer opens with published posts or successful tasks.
## Validation
diff --git a/docs/compass-quota-validation.md b/docs/compass-quota-validation.md
index 61320363d..0102c673e 100644
--- a/docs/compass-quota-validation.md
+++ b/docs/compass-quota-validation.md
@@ -43,7 +43,7 @@ The full **npm test** passed, including **2,429 legacy tests** and **60 security
The full cloud **composer check** and the real HTTP/review-CLI suite passed after the latest cloud fixes. Earlier validation blockers were resolved: shell fixtures use LF line endings, the default-model test uses an isolated root without a developer .env, the PII CLI recognizes absolute Windows lock paths, and stale provider/structured-error source assertions match the current behavior. The actual .env was untouched. Expected fault-case diagnostics remain in passing test logs.
-Further review regressions verify that the pending claim button fetches status in both browsers, concurrent rotating device headers cannot exceed the global telemetry cap, dropped events do not rewrite the ledger, and exact checkout timestamps outside 30 days cannot convert. Linux PHP durability tests verify file-then-directory sync and safe approval retry after an injected directory-sync failure. The Linux social test could not run because that PHP installation lacks PDO SQLite; the complete cloud check and social/concurrency suite passed with Windows PHP. Windows PHP cannot sync directory handles, so production power-loss durability requires a POSIX filesystem with directory fsync.
+Further review regressions verify that the pending claim button fetches status in both browsers, concurrent rotating device headers cannot exceed the global telemetry cap, dropped events do not rewrite the ledger, and exact checkout timestamps outside 30 days cannot convert. Linux PHP durability tests verify file-then-directory sync and safe approval retry after an injected directory-sync failure. The initial WSL social check lacked PDO SQLite. The approved deployment preflight subsequently ran the full cloud `composer check`, social/concurrency suite, and real HTTP/review-CLI tests at $0.50 and $0.375 in an isolated copy on the production Linux PHP platform; all passed. The Windows suite also passed. Windows PHP cannot sync directory handles, so production power-loss durability requires a POSIX filesystem with directory fsync.
Concurrent rotating session and device identifiers are bounded to 10,000 retained observations per consumption map. Overflow contributes to exact metered aggregates and disclosed report totals, while existing samples keep accumulating and expiration frees capacity. Regression tests cover concurrent capacity, exact overflow totals, unchanged weekly accounting, and retention. The extension branch includes current main and its upstream social-platform changes; the complete suite passes after integration.