Limits

This page enumerates the platform’s hard limits. Each row carries a stability grade that tells you whether the number is safe to bake into automation:

Grade	Meaning
Hardcoded	Written into the code; changing it requires a release. Safe to rely on
Configurable default	Code holds a default that operations can adjust. Treat the value as indicative, do not hardcode
Ops data	Fully driven by backend configuration; this page documents structure only, not values

File uploads

Stability: hardcoded. Source: apps/web/src/lib/storage/config.ts.

Item	Limit
Image, per file	20 MB
Document, per file	512 MB
Code file, per file	50 MB
Other types, per file	512 MB
Avatar	5 MB
Workspace, per file	512 MB
Files per upload	10
Attachments per message (images + files)	10

Accepted types

Images: jpeg, png, gif, webp, svg+xml. Documents: PDF, Word (doc/docx), Excel (xls/xlsx), PowerPoint (ppt/pptx). Code and text: txt, md, json, xml, yaml/yml, js, ts, jsx, tsx, vue, svelte, py, rb, go, rs, java, kt, scala, c, cpp, h, hpp, cs, html, css, scss, less, sql, sh, bash, zsh, toml, ini, cfg, conf, and more. Binaries: archives (zip/tar/gz/rar/7z/bz2/xz/tgz), media (mp4/mp3/wav/webm/ogg/flac/aac/m4a/avi/mov/mkv), data files (db/sqlite/parquet/arrow), fonts (ttf/otf/woff/woff2), app packages (apk/ipa), Python packages (whl/egg).

The image types the sandbox can process are narrower than what upload accepts: only png, jpeg/jpg, gif, webp. SVG uploads fine but the expert cannot “look at” it — SVG is not a raster image and the vision model cannot ingest it. Convert to PNG first if the expert needs to see it.

Storage quota

Stability: configurable default. Code defaults live in packages/types/src/subscription.ts; effective values come from the backend plan catalog.

Plan	Storage quota
Free	10 GB
Plus (`pro`)	15 GB
Pro (`max`)	20 GB
Flagship	30 GB
Team	50 GB
Enterprise	50 GB

Internal plan codes and display names diverge for historical reasons: pro displays as Plus, max displays as Pro. When reading API responses, branch on the code, never the display name.

Credits and validity

Stability: grant amounts are configurable defaults; validity periods are hardcoded.

Item	Value	Stability
Signup grant	2,000 credits by default	Configurable
Signup credit validity	90 days	Hardcoded
Daily check-in	200 credits by default	Configurable
Plan-granted credit validity	One subscription cycle	Hardcoded
Inviter reward	1,000 credits by default	Configurable
Invitee reward	1,000 credits by default	Configurable
Referral reclaim on inactivity	100 days	Hardcoded

Plan-granted credits:

Plan	Granted credits
Free	0
Plus (`pro`)	3,900
Pro (`max`)	11,000
Flagship	30,000
Team	100,000 pooled (1,200,000 annual)

Referral caps by plan

Plan	Max rewarded invites
Free	10
Plus (`pro`)	15
Pro (`max`)	20
Flagship	25
Team / Enterprise	30

The counting basis is rewarded — only invites that actually paid out count. Invites blocked by risk control do not consume the allowance.

Expert configuration constraints

Stability: hardcoded. Source: request validation in services/core/src/routes/marketplace-expert/index.ts.

Item	Constraint
Expert price	Integer, 0 – 99,900 credits
Free trial turns per user	Integer, 0 – 10
Tag count	Max 10
Tag length	Max 64 characters

Price 0 makes the expert free (billingType = FREE); anything above 0 makes it unlock-based (UNLOCK). There is no third mode. Setting trial turns to 0 falls back to the global default (configurable in the backend), it does not mean “no trial.” Actually disabling trials requires the expert mode setting — see Pricing and billing.

Asset and output size

Stability: hardcoded.

Item	Limit	Notes
Sites asset directory (`web/public/assets/`)	40 MiB	Measured at three entry points: `open_preview` (warn, allow), `website_init` (422 + degraded guidance), `deploy` (reject)
`site_assets` sandbox threshold	2 MB	≤2MB writes to the sandbox returning `path`; larger or over-budget goes to object storage returning `url`
Widget HTML persistence threshold	256 KB	Above this it is persisted to object storage to survive multi-replica routing

40 MiB is the asset directory budget, not the published-bundle budget (which includes JS/CSS/HTML and is larger). Sizing your assets against the publish-side number produces the “built successfully, then rejected” waste.

Tool execution

Stability: hardcoded defaults, most overridable per call.

Tool	Item	Default
`bash`	Command timeout	120 seconds
`read`	Lines per read	2,000
`grep`	Matching lines returned	50
`memory`	Results returned	5
`knowledge`	Results returned	5
`obsidian_knowledge`	Results returned	10
`canvas_run_dag`	Concurrency	1 (sequential)
`inspect_3d`	Scene tree depth	5
`image`	Resolution	1024x1024
`video`	Resolution / duration	720p / 5 s
`music`	Duration	30 s

Bigger timeouts are not better: bash’s 120 seconds sits between “long enough for a build” and “a hang won’t burn the whole turn’s budget.” Genuinely long jobs should be backgrounded and polled, not given a multi-thousand-second timeout.

Withdrawal

Stability: threshold and freeze periods are configurable defaults; the rate limit is hardcoded.

Withdrawal is currently hard-disabled in code (WITHDRAWAL_CREATE_ENABLED = false); every request is rejected with WITHDRAWAL_NOT_OPEN. The table below describes the rules that apply once it opens. See Withdrawal.

Item	Default	Stability
Minimum withdrawal	10,000 diamonds	Configurable
Exchange rate	100 diamonds = 1 CNY	Configurable
Buyout earnings freeze	30 days	Configurable
Usage-share earnings freeze	30 days	Configurable
Withdrawal frequency	3 per 24 hours	Hardcoded

The verbatim error on hitting the rate limit:

WITHDRAWAL_RATE_LIMIT: Maximum 3 withdrawal requests per 24 hours

Stability: configurable defaults. Shares are stored as ppm (parts per million).

Level	Threshold (cumulative diamonds)	Buyout share	Usage share
`rising`	0	75%	10%
`growing`	300,000	85%	15%
`professional`	1,000,000	90%	20%
`master`	5,000,000	95%	20%

The usage share stops rising after professional — 20% is the current ceiling. Beyond that, revenue grows through volume, not level.

Delegation and collaboration

Stability: hardcoded.

Item	Limit
Delegation whitelist size	Capped (`DELEGATION_WHITELIST_MAX`); entries beyond it never reach the roster
Subagent tool allowlist	Three-pass resolution: L1 reserved names → glob expansion → validation. Specs that fail resolution are silently dropped

“Silently dropped” is the classic trap when a subagent seems to do nothing: a misspelled tool name raises no error, the subagent simply never appears. See Subagents and delegation.

Self-evolution

Stability: configurable defaults via environment variables.

Item	Default
Trigger threshold (conversation turns)	3
Cooldown	24 hours
Failure backoff	6 hours
Max pending suggestions	5
Candidates per batch	10
Per-run timeout	900 seconds
Memory consolidation interval	Every 5
Skill review interval	Every 10

The context_footprint scoring curve (hardcoded): a skill document at ≤3,000 characters scores full marks, ≥24,000 scores zero, linear in between.

Content and plaza

Stability: hardcoded.

Item	Limit
Note view dedup window	24 hours (per device/user)
View count flush interval	15 seconds (Redis → DB)
Sites design directions	Exactly 3 (`direction_selection` mode)
Sites design brief length	1 – 4,000 characters
Site slug	3 – 63 chars, lowercase letters / digits / hyphens; must start and end alphanumeric
Site slug reserved words	`api` / `app` / `www` / `admin` / `profy` / `test`
Site slug ownership	One owner per slug, globally; publishing under another user’s slug returns 409

What this page deliberately omits

The following are ops data; only structure is documented:

Item	Why no number
Per-model token rates	Change with supplier pricing and model lifecycle; any written value expires
Top-up bonus tiers	Promotional, adjusted per campaign
Actual plan prices	Backend catalog-driven, and split across China and global catalogs
Per-call cost of creative tools	Varies by model and resolution

How they are computed is in Billing formulas — the formulas are stable, the rates fed into them are not.

Billing formulas

The five billing unit algorithms

Tool Catalog

Every parameter default

Withdrawal

Full withdrawal rules and state machine

Revenue system

How shares are calculated

Verified 2026-08-11. Sources: apps/web/src/lib/storage/config.ts, packages/types/src/subscription.ts, packages/types/src/referral-risk.ts, services/core/src/db/service/platform-config.ts, services/core/src/db/service/withdrawal.ts, services/core/src/db/service/credit-consume.ts, services/core/src/routes/marketplace-expert/index.ts, services/core/src/config/env.ts, services/agent-runtime/src/base/tools/sandbox.py.

Get Started

Core Concepts

Guides

Platform & Account

Deep Dive

Limits

Limits

File uploads

Accepted types

Storage quota

Credits and validity

Referral caps by plan

Expert configuration constraints

Asset and output size

Tool execution

Withdrawal

Delegation and collaboration

Self-evolution

Content and plaza

What this page deliberately omits

Billing formulas

Tool Catalog

Withdrawal

Revenue system

​Limits

​File uploads

​Accepted types

​Storage quota

​Credits and validity

​Referral caps by plan

​Expert configuration constraints

​Asset and output size

​Tool execution

​Withdrawal

​Creator levels and revenue share

​Delegation and collaboration

​Self-evolution

​Content and plaza

​What this page deliberately omits

​Related pages

Billing formulas

Tool Catalog

Withdrawal

Revenue system

Limits

File uploads

Accepted types

Storage quota

Credits and validity

Referral caps by plan

Expert configuration constraints

Asset and output size

Tool execution

Withdrawal

Creator levels and revenue share

Delegation and collaboration

Self-evolution

Content and plaza

What this page deliberately omits

Related pages