On this page
Every numeric limit the product enforces, the file that holds it, the external standard it traces to, and whether it meets that standard. Reviewed against the expert review 2026-11-06. A limit with no standard says so; it is a measured ratchet or an engineering choice, never presented as a standard.
The last column is machine-checked by scripts/audit/limits-registry.py (in audit:all): it fails when the constant in code differs from the Value, when a constant moves out of its pattern, or when the code file lacks the limits-and-budgets.md#anchor comment.
Proposed, pending operator#
| Item | Today | Standard says | Status |
|---|---|---|---|
| Account password minimum | 10 | 8 (NIST rev. 3, ASVS 5.0), 12 (ASVS 4.0.3), 15 (NIST rev. 4, single-factor) | Unchanged; operator decision pending |
| /chat gzip cap | 372,588 B | 365,000 B | Waived at its ratchet; split of the chat thread owed |
| /auth/signin gzip cap | 186,131 B | 180,000 B | Waived at its ratchet; split owed |
| Core Web Vitals field gate | none | p75 field data | Budgets documented, not measured in CI |
| Backup retention | 14 in repo, 7 reported | none | IT to reconcile |
Route bundle budgets#
Gate: pnpm run performance:budget (§76). A route fails only past 2% over its cap.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Default route cap, raw JS | 1,220,000 B | apps/ |
measured ratchet, no standard | Ratchet; the standard ceiling below binds first on gzip | 2026-11-06 | apps/ |
| Default route cap, gzip JS | 336,000 B | apps/ |
measured ratchet, no standard | Ratchet; 8% under the standard ceiling | 2026-11-06 | apps/ |
| Headroom over a cap before the gate fails | 2% | apps/ |
No published standard; engineering tolerance (no source found); verified 2026-11-06 | Engineering tolerance | 2026-11-06 | apps/ |
| App-wide median route cap, gzip | 264,224 B | apps/ |
measured ratchet, no standard | Creep detector; never raised to fit one route | 2026-11-06 | apps/ |
| App-wide maximum route cap, gzip | 372,588 B | apps/ |
measured ratchet, no standard | Over the standard ceiling because of /chat; see waiver | 2026-11-06 | apps/ |
| Hard ceiling, any route, gzip JS | 365,000 B | apps/ |
Alex Russell, The Performance Inequality Gap, 2024 (https://infrequently.org/2024/01/performance-inequality-gap-2024/): 365 KiB compressed JavaScript (JS-heavy site, 3 s first load, P75 device and network, not the global median); the 365,000 B cap is 2.4% under 365 KiB; verified 2026-11-06 | Enforced on measured size; /chat waived at its ratchet | 2026-11-06 | apps/ |
Hard ceiling, landing and auth routes (/, /; / is an app route), gzip JS |
180,000 B | apps/ |
No published number; engineering choice of about half the 365 KiB budget of Russell 2024 (https://infrequently.org/2024/01/performance-inequality-gap-2024/) for signed-out first visits; verified 2026-11-06 | Enforced on measured size; /auth/signin waived at its ratchet | 2026-11-06 | apps/ |
| /chat route cap, gzip JS | 260,300 B | apps/ |
measured ratchet under the 365 KiB budget (Russell 2024, https://infrequently.org/2024/01/performance-inequality-gap-2024/); verified 2026-11-06 | KEEP: thread split into a dynamic chunk (measured 257,700 B + 1 %) | 2026-11-06 | apps/ |
| / (landing) route cap, gzip JS | 179,391 B | apps/ |
measured ratchet, no standard | Under the 180,000 public ceiling by 609 B | 2026-11-06 | apps/ |
| /auth/signin route cap, gzip JS | 186,131 B | apps/ |
measured ratchet, no standard | Over the 180,000 public ceiling; waived, split owed; proposed, pending operator | 2026-11-06 | apps/ |
/dashboards/[slug] route-specific cap (first load minus / chunks), gzip JS |
175,728 B | apps/ |
measured ratchet, no standard | Shell floor cannot hide this route's own growth | 2026-11-07 | apps/ |
Core Web Vitals#
ADR 0059 D9 sets the budgets from web.dev field thresholds (75th percentile of page loads). They are documented budgets; no CI gate measures field data yet.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Largest Contentful Paint, p75 at 1440 px | 2.5 s | docs/ |
web.dev Core Web Vitals (https://web.dev/articles/vitals): LCP good <= 2.5 s at p75; page published 2020-05-04, updated 2024-10-31; verified 2026-11-06 | Matches the standard; unmeasured in CI | 2026-11-06 | docs/ |
| Interaction to Next Paint | 200 ms | docs/ |
web.dev Core Web Vitals (https://web.dev/articles/vitals): INP good <= 200 ms at p75; page updated 2024-10-31; verified 2026-11-06 | Matches the standard; unmeasured in CI | 2026-11-06 | docs/ |
| Cumulative Layout Shift | 0.1 | docs/ |
web.dev Core Web Vitals (https://web.dev/articles/vitals): CLS good <= 0.1 at p75; page published 2020-05-04; verified 2026-11-06 | Matches the standard; unmeasured in CI | 2026-11-06 | docs/ |
| Time to First Byte | 0.8 s | docs/ |
web.dev TTFB (https://web.dev/articles/ttfb): good <= 0.8 s, a rough guide and not a Core Web Vital; published 2021-10-26, updated 2025-11-18; verified 2026-11-06 | Matches the standard; unmeasured in CI | 2026-11-06 | docs/ |
Request body caps#
§138. Every cap is enforced while the body is read.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Inline tRPC upload, one file decoded | 8 MiB | apps/ |
OWASP API Security Top 10 2023, API4 Unrestricted Resource Consumption (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "define and enforce a maximum size of data on all incoming parameters and payloads"; no number; verified 2026-11-06 | Engineering limit; a number is defined, as the standard asks | 2026-11-06 | apps/ |
| Proxy body clone cap | 24 MiB | apps/ |
OWASP API Security Top 10 2023, API4 Unrestricted Resource Consumption (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "define and enforce a maximum size of data on all incoming parameters and payloads"; no number; verified 2026-11-06 | Must exceed 2 base64 files plus the envelope; test-guarded | 2026-11-06 | apps/ |
| Proxy cap ceiling held by the guard test | 32 MiB | apps/ |
OWASP API Security Top 10 2023, API4 Unrestricted Resource Consumption (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "define and enforce a maximum size of data on all incoming parameters and payloads"; no number; verified 2026-11-06 | Engineering limit | 2026-11-06 | apps/ |
| Webhook and internal JSON body | 256 KiB | apps/ |
OWASP API Security Top 10 2023, API4 Unrestricted Resource Consumption (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "define and enforce a maximum size of data on all incoming parameters and payloads"; no number; verified 2026-11-06 | Engineering limit; real vendor envelopes are a few KiB | 2026-11-06 | apps/ |
| Org-data archive upload, streamed | 2 GiB | apps/ |
OWASP API Security Top 10 2023, API4 Unrestricted Resource Consumption (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "define and enforce a maximum size of data on all incoming parameters and payloads"; no number; verified 2026-11-06 | Engineering limit; stream is metered, admin-only, rate-limited 10/60 s | 2026-11-06 | apps/ |
Rate limits#
Fixed-window counters in Redis through rateLimit. Counts are requests per window in seconds.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Sign-in per IP and email | 10 per 300 s | apps/ |
NIST SP 800-63B rev. 3 s5.2.2 (https://pages.nist.gov/800-63-3/sp800-63b.html) and rev. 4 s3.2.2 (https://pages.nist.gov/800-63-4/sp800-63b/authenticators/): at most 100 consecutive failed attempts per account; ASVS 4.0.3 V2.2.1 (no more than 100 failed attempts per hour on a single account); verified 2026-11-06 | Within the standard | 2026-11-06 | apps/ |
| Sign-in per account, any IP | 20 per 900 s | apps/ |
NIST SP 800-63B rev. 3 s5.2.2 (https://pages.nist.gov/800-63-3/sp800-63b.html) and rev. 4 s3.2.2 (https://pages.nist.gov/800-63-4/sp800-63b/authenticators/): at most 100 consecutive failed attempts per account; ASVS 4.0.3 V2.2.1 (no more than 100 failed attempts per hour on a single account); verified 2026-11-06 | Within the standard; throttle, not a lockout at 100 consecutive failures | 2026-11-06 | apps/ |
| Sign-in per IP, any account | 30 per 300 s | apps/ |
No numeric standard for per-IP limits; NIST SP 800-63B s5.2.2 and the OWASP Authentication Cheat Sheet (https://cheatsheetseries.owasp.org/cheatsheets/Authentication_Cheat_Sheet.html) key the failure counter to the account, not the IP; verified 2026-11-06 | Within the standard | 2026-11-06 | apps/ |
| Magic-link request per IP | 5 per 60 s | apps/ |
No numeric standard; NIST SP 800-63B s5.2.2 (https://pages.nist.gov/800-63-3/sp800-63b.html) throttles failed authentications per account only; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| Magic-link request per email | 3 per 300 s | apps/ |
No numeric standard; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| MCP tools listing per IP | 60 per 60 s | apps/ |
OWASP API Security Top 10 2023, API4 (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "implement a limit on how often a client can interact with the API within a defined timeframe"; no number; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| MCP tool execution per IP | 30 per 60 s | apps/ |
OWASP API Security Top 10 2023, API4 (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "implement a limit on how often a client can interact with the API within a defined timeframe"; no number; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| Chat turn per org and user | 30 per 60 s | apps/ |
OWASP API Security Top 10 2023, API4 (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "implement a limit on how often a client can interact with the API within a defined timeframe"; no number; model cost bound is an engineering choice; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| Connector webhook per type and IP | 600 per 60 s | apps/ |
No numeric standard; vendor burst rates; verified 2026-11-06 | Measured against vendor bursts | 2026-11-06 | apps/ |
| Share-link wrong password per link and viewer | 10 per 300 s | apps/ |
NIST SP 800-63B rev. 3 s5.2.2 (https://pages.nist.gov/800-63-3/sp800-63b.html) and rev. 4 s3.2.2 (https://pages.nist.gov/800-63-4/sp800-63b/authenticators/): at most 100 consecutive failed attempts per account; ASVS 4.0.3 V2.2.1 (no more than 100 failed attempts per hour on a single account); verified 2026-11-06 | Within the standard | 2026-11-06 | apps/ |
| Retry-After on a procedure-level 429 with no exact wait | 60 s | apps/ |
RFC 6585 s4 (https://www.rfc-editor.org/rfc/rfc6585#section-4): a 429 MAY include Retry-After; RFC 9110 s10.2.3 (https://www.rfc-editor.org/rfc/rfc9110#section-10.2.3) defines the header and names 503 and 3xx, not 429; verified 2026-11-06 | Every 429 carries Retry-After; exact limiter value where the route has it | 2026-11-06 | apps/ |
| Redis call deadline in the limiter | 500 ms | apps/ |
No published standard; measured; verified 2026-11-06 | Fails open past the deadline | 2026-11-06 | apps/ |
AI token budgets#
Per-org monthly allowances, metered on ai_usage_log (input + output tokens) per ICT calendar month. NULL means unlimited and not metered; 0 refuses everything. No allowance has a default: the operator sets one per org in the Admin CP org drawer. A spent chat allowance answers 429 with Retry-After at the next ICT month start.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
Chat assistant tokens per org per month (organizations.), settable range |
No default (NULL = unlimited); 0 to 2,000,000,000 | apps/ |
OWASP API Security Top 10 2023, API4 (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): "Configure spending limits for all service providers/API integrations"; no number, so no default cap (ADR 0060 S5); the upper bound stays under the PostgreSQL integer maximum 2,147,483,647 (https://www.postgresql.org/docs/current/datatype-numeric.html); verified 2026-11-07 | Mechanism meets the standard; the number is the operator's per org; no-default guarded by org- |
2026-11-07 | apps/ |
Sessions#
Operator decision 2026-11-06: KEEP 14 days with a rolling 24 hours for all roles, no admin-specific cut. ASVS L2 would be 12 hours or 30 minutes idle; the gap is accepted.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Session lifetime | 14 days | apps/ |
NIST SP 800-63B rev. 3 s4.1.3 (https://pages.nist.gov/800-63-3/sp800-63b.html): AAL1 reauthentication SHOULD repeat at least once per 30 days; rev. 4 s2.1.3 (https://pages.nist.gov/800-63-4/sp800-63b/aal/): overall timeout SHOULD be no more than 30 days at AAL1; ASVS 4.0.3 V3.3.2 (https://github.com/OWASP/ASVS/blob/v4.0.3/4.0/en/0x12-V3-Session-management.md): L1 30 days, L2 12 hours or 30 min idle (2FA optional), L3 12 hours or 15 min idle with 2FA; ASVS 5.0 V7.3.1/V7.3.2 (L2) set no number; verified 2026-11-06 | KEEP (operator decision 2026-11-06): within L1 and AAL1, outside ASVS L2 | 2026-11-06 | apps/ |
| Session rolling refresh | 24 hours | apps/ |
Same sources; no numeric rule for refresh; verified 2026-11-06 | KEEP (operator decision 2026-11-06) | 2026-11-06 | apps/ |
Passwords#
Account password minimum is unchanged (proposed, pending operator). Share-link minimum rose from 4 to 8 for new and changed passwords on 2026-11-06.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Account password minimum | 10 characters | apps/ |
NIST SP 800-63B rev. 3 s5.1.1.2 (https://pages.nist.gov/800-63-3/sp800-63b.html): minimum 8; rev. 4 s3.1.1.2 (https://pages.nist.gov/800-63-4/sp800-63b/authenticators/): 15 for a single-factor password, 8 when only used with MFA; ASVS 4.0.3 V2.1.1: 12; ASVS 5.0 V6.2.1 (https://github.com/OWASP/ASVS/blob/v5.0.0/5.0/en/0x15-V6-Authentication.md): 8; verified 2026-11-06 | Meets NIST rev. 3 and ASVS 5.0; 2 under ASVS 4.0.3; 5 under NIST rev. 4 if the password is the only factor; proposed, pending operator | 2026-11-06 | apps/ |
| bcrypt cost factor | 10 | apps/ |
OWASP Password Storage Cheat Sheet (https://cheatsheetseries.owasp.org/cheatsheets/Password_Storage_Cheat_Sheet.html): bcrypt work factor "a minimum of 10"; verified 2026-11-06 | At the floor | 2026-11-06 | apps/ |
| Share-link password minimum, new and changed | 8 characters | apps/ |
NIST SP 800-63B rev. 3 s5.1.1.2 (https://pages.nist.gov/800-63-3/sp800-63b.html): minimum 8; ASVS 5.0 V6.2.1: 8; verified 2026-11-06 | Meets the standard; existing shares keep opening at any length | 2026-11-06 | apps/ |
Database#
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| RLS transaction statement and idle timeout (RLS_TX_TIMEOUT_MS) | 60 s | packages/ |
No standard; PostgreSQL statement_timeout defaults to 0, disabled (https://www.postgresql.org/docs/current/runtime-config-client.html); verified 2026-11-06 | Measured against the heaviest rollup | 2026-11-06 | packages/ |
| AI failover cooldown for a candidate that failed with authentication, quota or model-not-found | 60 s | packages/ |
No numeric standard (proposal). Pattern: Azure Circuit Breaker (https://learn.microsoft.com/en-us/azure/architecture/patterns/circuit-breaker) and AWS Builders Library timeouts/retries (https://aws.amazon.com/builders-library/timeouts-retries-and-backoff-with-jitter/) give no number; nearest default is Envoy outlier detection base_ejection_time 30 s (https://www.envoyproxy.io/docs/envoy/latest/api-v3/config/cluster/v3/outlier_detection.proto); verified 2026-11-07 | KEEP as proposal: 2x the Envoy base, process-local, last candidate always tried | 2026-11-07 | packages/ |
| AI per-attempt first-token timeout when the caller sets none, a lone candidate included | 30 s | packages/ |
No published standard; engineering default; AWS Builders Library timeouts/retries (https://aws.amazon.com/builders-library/timeouts-retries-and-backoff-with-jitter/) gives no number; verified 2026-11-07 | Engineering default; callers pass 30 s (chat, MSO summary) or 60 s (masking, worker jobs) | 2026-11-07 | packages/ |
| App pool size (DB_POOL_MAX) | 10 | packages/ |
No standard; sized against Postgres max_connections; verified 2026-11-06 | Measured | 2026-11-06 | packages/ |
| Auth pool size (AUTH_DB_POOL_MAX) | 5 | packages/ |
No standard; verified 2026-11-06 | Measured | 2026-11-06 | packages/ |
List caps and page size#
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Rows an unbounded admin list may return (§1) | 500 | apps/ |
OWASP API4:2023 (https://api-security.owasp.org/editions/2023/en/0xa4-unrestricted-resource-consumption): validate the parameter that controls records returned, no number; Google AIP-158 Pagination (https://google.aip.dev/158): the API enforces its own maximum page size and coerces larger requests, no mandated number; verified 2026-11-06 | A bound exists; the number is an engineering choice | 2026-11-06 | apps/ |
| Commerce report window, inclusive ICT calendar days | 400 | apps/ |
No published numeric standard; engineering ceiling aligned with the existing 400-day creator/audience series limits. Preserves annual, leap-year and last-12-month presets while bounding daily zero padding to 400 rows per platform; larger or reversed windows are rejected before queries | Engineering ceiling | 2026-11-13 | apps/ |
| Default table page size | 20 | packages/ |
Google AIP-158 Pagination (https://google.aip.dev/158): page size optional with a documented finite default, no mandated number; operator standing rule; verified 2026-11-06 | Operator rule | 2026-11-06 | packages/ |
| Rows of one trend-feed slice served in one read | 5000 | packages/ |
No published standard; engineering ceiling mirrored from the contract constant MAX_ (packages/, ADR 0061 D4: a slice is hundreds of rows, 446 trends held on 2026-11-12); a larger slice throws instead of truncating (§57) |
Engineering ceiling | 2026-11-12 | packages/ |
| Rows of one trend-feed history window served in one read | 30000 | packages/ |
No published standard; engineering ceiling derived from the writer: the connector stores the whole usable general list per cycle (TREND_ = 1000, packages/) and keeps 30 cycles of 24h (retention in api/), so the largest window is 30 x 1000 = 30,000 rows; a larger window throws instead of truncating (§57) |
Engineering ceiling | 2026-11-12 | packages/ |
| Vendor ids per focused trend-feed history read | 200 | packages/ |
No published standard; the board shows 50 trends (FEED_), the cap leaves headroom; 200 ids x 30 cycles stays under 6,000 rows |
Engineering ceiling | 2026-11-12 | packages/ |
Cache lifetimes#
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| OIDC discovery document | 24 hours | apps/ |
No numeric standard; OpenID Connect Discovery 1.0 (https://openid.net/specs/openid-connect-discovery-1_0.html) does not mention caching; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| OIDC signing keys (JWKS) | 1 hour | apps/ |
No numeric standard; RFC 9111 HTTP Caching, June 2022 (https://www.rfc-editor.org/rfc/rfc9111); verified 2026-11-06 | Engineering default; a rotated key is picked up within the hour | 2026-11-06 | apps/ |
| Dashboard binding cache | 60 s | apps/ |
measured ratchet, no standard; verified 2026-11-06 | Matches the deck refetch rhythm | 2026-11-06 | apps/ |
Worker retries and backoff#
Exponential delay doubles per attempt; jitter draws each delay from [delay x (1 - j), delay]. Five attempts at 30 s span 450 s (7.5 min) before jitter.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Backoff jitter share | 0.5 | packages/ |
Google SRE book ch. 22, Addressing Cascading Failures (https://sre.google/sre-book/addressing-cascading-failures/): "Always use randomized exponential backoff when scheduling retries"; Marc Brooker, Exponential Backoff And Jitter, AWS Architecture Blog, 2015-03-04 (https://aws.amazon.com/blogs/architecture/exponential-backoff-and-jitter/), whose Full Jitter draws over [0, delay], wider than this 0.5 share; no published jitter share; verified 2026-11-06 | Applied to every exponential backoff in worker and outbox defaults | 2026-11-06 | packages/ |
| Sync job attempts and base delay | 5 attempts, 30 s | apps/ |
Same sources; attempt counts have no standard; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| AI job attempts and base delay | 3 attempts, 10 s | apps/ |
Same sources; no standard for attempts or base delay; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
| Scheduled tick attempts and base delay | 3 attempts, 30 s | apps/ |
Same sources; no standard for attempts or base delay; verified 2026-11-06 | Engineering default | 2026-11-06 | apps/ |
Eventbus pending recovery#
Consumer-group delivery is at least once. Set the idle threshold above normal handler duration; handlers must tolerate duplicate delivery after a slow consumer or restart.
| Limit | Value | Code | Standard | Verdict | Date | Check |
|---|---|---|---|---|---|---|
| Pending event idle time before reclamation | 60 s | packages/ |
No published numeric standard; engineering recovery window. Redis XAUTOCLAIM reassigns entries idle for the caller's threshold (https://redis.io/docs/latest/commands/xautoclaim/). | Bounded recovery; callers may set a positive safe-integer millisecond threshold | 2026-11-13 | packages/ |
Decision engine#
Backups#
Owned by IT. The repository default is 14; the expert review reported 7 on a host. No 7 exists in the repository, so the host value must be read and reconciled by IT.