omp Settings Web (omp-settings-web)

English | 中文

Screenshots

Desktop widget — desktop-level panel (stays beneath app windows) showing remaining quota per provider:

Usage dashboard — per-provider quota cards with force-refresh and reset-credit redemption:

Model roles — beginner notes per role, two-level picker to switch models:

All settings — 529 keys in a categorized sidebar, each with an explanation:

Model visibility — choose which models appear in the /model picker:

Features

Model roles

• Role matrix with two-level picker: provider group → model; auto-positions on the current value; @role aliases kept as a fallback entry.
• Each role shows a beginner-friendly explanation below it (which features use the role, what model to assign, and the fallback behavior when unset); a general note at the top explains the role mechanism (@aliases, *, comma candidates, custom roles).
• Ctrl+P cycle order editor (cycleOrder): move up/down, remove, add existing roles or a custom role name; roles without a model can be bound to one right there (written to modelRoles). Save / cancel / restore default — restoring removes cycleOrder so omp falls back to smol → default → slow. Empty lists, duplicates, @aliases and invalid characters are rejected.
• Failure fallback editor: ordered backup models for roles and existing model/provider chains, with add, remove, move up/down and cancel. Distinguishes unset (remove the key and use omp's normal rules), explicitly disabled ([]), and a custom chain. Existing hidden/offline entries are retained; new entries use explicit provider/model-id selectors, not @role aliases.
• Fallback entries and concrete model keys support model version tags (for example, ollama-cloud/deepseek-v4.1-flash:1017) and a thinking suffix after the tag (:1017:high); version tags are not rejected as invalid thinking levels.
• Controls for retry.modelFallback, retry.fallbackOn and retry.fallbackRevertPolicy. Exact model keys take precedence over provider wildcards, then the active role and default; model-kind roles use their own chains rather than the chat default.

All settings

• Full catalog from omp config list --json (500+ keys) with current values, types, and descriptions.
• All 529 keys carry detailed beginner-oriented Chinese names & explanations — what each value does, the default, when it is worth changing, and related settings (upstream English text is available on hover).
• Scalar settings are editable inline; arrays and nested blocks are read-only here. Use the dedicated editor on Model Roles for retry.fallbackChains.
• Sidebar category navigation: ~35 semantic categories sorted by size; search shows per-category hit counts and dims empty ones.
• Search + "only customized" filter; badges distinguish keys you have set from defaults.

Model visibility

• Choose which models appear in omp's /model picker, Ctrl+P cycling and this page's role dropdowns — per model or per provider, with search. Hidden models are not disabled: roles already assigned to them keep working. Writes enabledModels in config.yml (backed up, live-reloaded by omp); fully selected providers collapse to provider/*, the default role's model is pinned first and always shown, and patterns matching no current model (offline providers) are preserved.

Usage

• A visual dashboard mirroring omp usage: per-provider cards with progress bars for each rate-limit window (5h / 7d / weekly / monthly), used vs. remaining, live reset countdowns, account/plan info and reset-credit counts; shared windows are deduplicated. Manual refresh + per-provider or global force-refresh (omp usage invalidate, whitelist-validated) + optional 60s auto-polling, friendly empty/error states. One-click "Reset with credit" on anthropic/openai-codex cards — a confirmed click spends one saved reset credit through the same upstream protocol omp's /usage reset uses (server-side reimplementation of resets.ts; OAuth tokens are read from ~/.omp/agent/agent.db and never leave the process).

Accounts

• OAuth accounts omp can route per account, grouped by provider (identity from omp usage accounts, no tokens; remaining quota from omp usage); logins without an identity and API-key credentials are not shown. Each account gets a priority and reserve % editor that writes omp's native auth.accountPolicies in config.yml (backup + targeted block rewrite; everything outside the block stays byte-identical); leave both empty to clear. The global default retry.usageReservePct is editable on the same page. Policies that match no signed-in account are listed separately with a delete button. Priority only orders healthy, measured accounts and affects new sessions; it never overrides rate limits or reserve. Add accounts with /login in omp. A policy block containing comments or flow syntax is shown read-only.

Desktop widget (widget.py)

• A borderless dark mini window pinned to the desktop layer (never covers other apps) showing every provider's rate-limit windows from omp usage --json — compact bars showing remaining quota (colored by how much is left), remaining %, reset countdown, and a badge for saved reset credits. Runs standalone (no dependency on the web server), refreshes every 60 s, drag anywhere to move (position is remembered; the window is always kept fully inside the nearest monitor's work area, including after resolution/monitor/taskbar changes), right-click for refresh / force-refresh / open web panel / quit. Single-instance guarded; Win11 rounded corners; per-monitor DPI aware; omp subprocesses run without console windows.
• Custom provider order: right-click a provider header for Move up / Move down / Reset order. The order is saved per machine in widget-state.json as "order": ["openai-codex", "google-antigravity", ...] (can also be edited by hand while the widget is closed); listed providers come first, unlisted ones keep omp's order after them.

Requirements

• Python 3.8+ (standard library only, zero dependencies)
• omp installed and on PATH (used for omp models --json)
• Windows for the desktop widget (widget.py): it pins itself to the desktop layer, keeps Win11 rounded corners and is per-monitor DPI aware through Win32 calls, none of which exist elsewhere. The web panel (server.py) is plain Python and runs on any OS.

Usage

python server.py            # starts http://127.0.0.1:8788 and opens your browser
python server.py --no-browser

Desktop widget (Windows):

pythonw widget.py           # no console window; right-click the widget for the menu

Open http://127.0.0.1:8788, click Switch on a role, pick a provider group, then a model, Save. The change applies immediately — running omp sessions pick it up, no restart needed.

How it stays safe

omp config set rewrites the whole file and strips your comments. This tool never calls it. Instead:

1. Targeted line edit — reads ~/.omp/agent/config.yml and regex-replaces only the target key's line; keys not yet present are inserted at the correct nesting level.
2. Byte-precise I/O — read_bytes/write_bytes with original line-ending detection (LF vs CRLF), so everything except the edited/inserted lines stays byte-identical.
3. Automatic backup — every write first copies the config to config.yml.bak-settings-web-<timestamp>.
4. Type validation — scalar values follow the type reported by omp config list. The dedicated fallback API validates keys and selectors, supports block maps/lists and single-line maps, and refuses unsupported complex YAML rather than overwriting it.

Server binds to 127.0.0.1 only. Value/role inputs are validated against a whitelist regex before touching the file.

⚠️ Pitfall worth knowing (and the reason for #2): on Windows, Python's Path.write_text() silently translates LF to CRLF, which would rewrite your entire config's line endings.

API

Endpoint	Method	Purpose
/api/state	GET	Current roles (parsed from config.yml) + full model catalog (omp models --json)
/api/set	POST	{"role": "...", "value": "..."} — backup + targeted write
/api/settings	GET	Full settings catalog (omp config list --json) + zh annotations + __groupmap
/api/set-key	POST	{"key": "...", "value": "..."} — type-coerced scalar write, backup + targeted line edit
/api/fallback	GET	Fallback chains and modelFallback / fallbackOn / fallbackRevertPolicy
/api/fallback	POST	{"key":"default","chain":["provider/model"]}; [] disables fallback, null removes the key. Updates only that chain, with a backup.
/api/cycle-order	GET	Ctrl+P cycle order {order, custom, default}; omp's default order when unset
/api/cycle-order	POST	{"order":["smol","default","e"]} writes it; {"order": null} removes the key to restore the default; backed up
/api/usage	GET	omp usage --json passthrough — per-provider quota/usage reports
/api/usage/invalidate	POST	`{"provider": "all"
/api/usage/redeem	POST	{"provider": "anthropic"|"openai-codex"} — spends one saved reset credit via the same upstream calls as omp's /usage reset (two-step request_id confirm for Anthropic, single idempotent consume for Codex); structured {ok, code, error} responses, tokens never exposed
/api/accounts	GET	Accounts per provider (identity, quota, matching policy), unmatched policies, default reserve
/api/accounts/policy	POST	{"provider", "identityKey", "priority", "reservePct"} — upsert/clear the policy of a signed-in OAuth account; validated, backed up
/api/accounts/policy-delete	POST	{"index", "provider", "account"} — delete one policy; rejected if the entry no longer matches

License

MIT — not affiliated with the omp project.