Admin console

Overview, models, and runtime settings.

Overview

Plan health — is a credit still worth what we charge for it? —

loading…

System Health — last 24h, what to fix today — ↻ refresh

Loading…

Users

—

Active paid

—

Auto-renew subs

—

MRR (est.)

—

Revenue · 30d

—

Revenue · all

—

Platforms — real Mac/Windows/Linux split of signed-in users (IDE-reported, internal accounts excluded)

Loading…

Recent payments

When	User	Plan	₹	Via
Loading…

Live ops: tier-stats · provider-health · latency SLOs · queue

Provider health · last 72h — the shape, not just the current status

One cell per hour, oldest left. Grey = no traffic (not "healthy" — an idle hour and a clean hour are different facts). Window is capped at 72h because that is how long the hourly buckets live; anything longer would be invented history.

Loading…

Estimated spend — fleet-wide, outage-amplification guard

Estimates, not invoices — incremented with the failover engine's per-request estimate, so a threshold crossing is meaningful even if the absolute $ is ±30%.

Loading…

Grant plan · testing / comps

Manually set a user's plan (no payment). Use free to revoke.

User email

Plan

Days

Users

	Email	Plan	Auto-renew	Joined	Set plan
Loading…

Page size

Model usage & cost

Real input/output tokens per model, with an approximate cost (provider list prices; cached input billed ~20%). Bars show input (blue) vs output (green) volume.

Loading…

Edit success (per model)

Share of agent edit-tool calls that applied cleanly — detected server-side from edit results (the agent-quality north star). Compare the cheaper tool model vs the frontier one: if success % is comparable, keep the cheaper routing; if it drops, revert tool_call_model_cheap in Settings → Model routing.

Loading…

Skill-pack usage (the moat thesis-test)

Which domain skill packs the agent actually loads (counted server-side from the injected skill block — no Langfuse). The thesis test: if the packs are rarely invoked, the domain-depth bet isn't landing yet.

Loading…

Live in ~30 s — changes apply within the config cache TTL after saving.

Model routing

model_cheap Fast, low-cost tasks

model_agentic Agentic / tool-use workloads

model_frontier Standard quality tier

model_hard Hardest / most capable review_model The model that REVIEWS code, independent of the one that wrote it. Applies to every review call (Max pass 2, the review endpoint, deep review). Leave EMPTY to keep using the tier the review is pinned to — which today is the same model that produced the answer, so the review is a self-review. Only the code block is sent, not the conversation, so a dearer model costs far less here than in an answer slot.

max_validate_enabled Max pass 2 — after a HARD-tier answer, re-read its code with the hard-tier reviewer and append anything high-severity. Costs a second (much smaller) provider call. This is the only thing that makes Max different from Pro. on async_indexing_enabled Move repo indexing off the request path onto the background worker, so the index call returns immediately. Safe to leave ON only while a worker is running — if none is, the server detects that and indexes inline instead of queueing work nothing would ever pick up.

max_validate_pct Of the users eligible for the pass above, the share that actually gets it — a readable A/B. 100 = everyone (the plain on-switch). Set it to 50 and BOTH arms run at once, so the edit-success board can show the difference instead of you comparing two periods that differ in a dozen other ways. Assignment is stable per user.

telemetry_excluded_emails Accounts whose turns stay OUT of the learning counters (routing/edit outcomes, task-model, edit-format, skills, activation) and out of the A/B arms — but are still BILLED normally. This exists for the coding benchmark, which posts as an ordinary user on purpose: its suites are 19–32 tasks against a handful of real users, so one run would not add noise to the evidence, it would BE the evidence. Comma-separated. Leave EMPTY to record everyone.

tool_call_model Tool/function-call turns

tool_call_model_cheap Cheaper tool turns — empty falls back to tool_call_model

skill_gating_pct % of users (stable sha256(user.id) bucket) who get the NARROWED skill plan instead of the whole ~180-name catalogue. 0 = off, 100 = everyone. ⚠️ NOT a token saving — that block is an identical prefix on every request and our measured cache hit is 95.2%, so it costs ~₹0.0152/request. Gating makes the prefix vary per repo and can LOWER the hit rate. Judge it on answer quality and the Models-tab cache hit; revert if that falls.

tool_compress_enabled Rewrites tool output before it is re-sent to the model (~22% of tokens). Audited 2026-08-08: ten silent signal losses found and fixed. Turn this OFF here the moment a tool result looks wrong — it must never need a deploy. on

tool_compress_shadow Measure only — compute the saving and report it, send the model the ORIGINAL text. The safe way to watch a filter change before trusting it. on

cost_aware_failover_enabled Let the router fall back to a cheaper healthy model when the chosen one is failing, instead of retrying the same one. on

vision_model Routed live, like every other slot. A model put here counts as sighted even if it is not in the model registry, so a new multimodal model can be onboarded before it lands there.

vision_tool_model Where a screenshot goes when the turn ALSO carries tools — every image pasted in Agent mode. Honoured only if the model registry says the id can both see and call tools; otherwise the sighted vision_model keeps the turn. Empty = keep vision_model. Until 2026-09-08 these turns were silently sent to the text tool model and the image was dropped.

Plan daily limits — requests per day per plan

free

trial

pro

proplus

ultra

team

Monthly credits per plan — the depth-weighted credit pool per plan (react to a margin surprise without a redeploy)

free

trial

pro

proplus

ultra

Credit cost per tier — credits charged per request at each routing tier (all four required)

cheap

agentic

frontier

hard

Defaults & rate limits

default_daily_limit Fallback if plan not matched

auth_rate_limit Max auth attempts per window

fx_inr_per_usd USD→INR used to price our COGS against rupee plan prices. A 10% move can flip a plan from profitable to loss-making at cap — see plan_health on /admin/quality. Set it from an invoice, not from memory.

credits_daily_ratio Share of the WEEKLY allowance a single day may use — the anti-binge rail under the weekly gate. At 0.6 a Pro+ day is 1,106 credits, while one p90 agent session (223 turns on the frontier tier) costs 892 — 81% of the day, which is how a Pro+ customer was walled mid-task with 40% of their week still unspent. Raising it is margin-neutral: day, week AND month are all enforced atomically, so the monthly cap still bounds total spend and this only changes pacing inside a week. Do not approach 1.0 — one runaway day would consume the entire weekly allowance.

daily_cap_enabled OFF by default since 2026-08-23, by the owner's decision. The pacing model we match uses a five-hour session window plus a weekly limit and has NO daily cap; this rail was the only place the two differed, and it is the gate a user actually hit. Turning it off cannot cost margin: week AND month are still enforced atomically, so the maximum a plan can spend is unchanged. Switch it back on if a doom-loop turns up — the rail is disabled, not deleted, and credits_daily_ratio above still sizes it. on

auth_rate_window_sec Window size in seconds

login_attempt_limit Failed sign-ins allowed per ACCOUNT before lockout (min 1 — 0 would lock everyone out)

login_attempt_window_sec Lockout window in seconds (a successful sign-in clears the count)

chat_max_concurrent Max in-flight chats per user

support_notify_email New-ticket notifications. Empty = nobody is told; tickets still appear in /tickets.

Spend guard — fleet-wide outage-amplification guard, not per-user billing

Estimated USD. These exist so an outage that re-routes the fleet onto an expensive fallback can be capped without a deploy — that is why they are editable here. The ceiling drops degraded fallbacks for free/trial only; paying tiers are never spend-blocked and no request is ever refused.

spend_hourly_alert_usd ~6x a normal hour → ops email

spend_daily_alert_usd ~10x a normal day → ops email

spend_daily_ceiling_usd Above this, degraded fallbacks are dropped

Downloads — fix the /download page without a redeploy

download_version label shown on the download page (e.g. 1.4.0)

download_urls JSON map of artifact key → https URL. Keys: mac-arm64, mac-x64, mac-universal, win-x64-system, win-x64-user, win-arm64-system, win-arm64-user, linux-tarball-x64/arm64, linux-deb-x64/arm64, linux-rpm-x64/arm64, linux-appimage-x64/arm64

download_checksums JSON map of the SAME artifact keys → sha256 (64 hex). Rendered at /download/checksums — the only way a visitor can verify an installer, since ours are not Authenticode-signed. Update this whenever you replace a file: a stale hash tells someone their good download is tampered.

Auto-renew (Razorpay Subscriptions)

One-time setup: creates a monthly Razorpay plan per paid tier so auto-renew checkout works. Idempotent — safe to click again.

Live promo codes — add, edit, or remove instantly (no redeploy). percent = % off checkout; free_days = grants a plan free for N days with no payment. Removing a code stops it immediately; redemption counts are preserved.

Active coupons

Code	Type	Value	Plan	Expires	Used / Max
Loading…

Add / update a coupon

Code UPPERCASE, alphanumeric

Type

Value percent 1–100, or number of days

Plan required for free_days; blank = any (percent)

Expires YYYY-MM-DD, blank = never

Max uses 0 = unlimited

Measured model quality — the benchmark engine scores each tier/model on the coding-eval and ranks by pass@1. This is the source of truth for routing — measured, not vendor claims. Runs are stored versioned (never overwritten).

Coverage — do the models actually serving traffic have a score in this run?

Loading…

Routing recommendations

Advisory only. Nothing switches automatically — the engine ranks by pass@1, breaks ties on $/solve, and you flip the slot yourself in Settings → Model routing. Benchmark scores are compared against live production signals (TTFT p95, edit-success) because the offline eval is blind to latency.

Loading…

Leaderboard

#	Target	pass@1	Passed	Wall s
Loading…

Capability matrix — pass@1 per difficulty (measured)

Target	Overall	Easy	Medium	Hard
—

Run history — newest first, winner per run (regression tripwire)

When	Label	Winner	pass@1	Targets
—

Context Engine — is semantic_search actually being used? The IDE tool has no logging and falls back to grep silently, so a broken engine and an unused one look identical from the outside. This is the only thing that tells them apart.

Loading…

Live Razorpay mandates — we store ONE subscription id per user and overwrite it with whichever mandate charged last, so a second mandate made the first invisible: on no screen, cancellable through nothing, still charging. New ones are refused now; the ones already out there can only be found by asking Razorpay. Does not run on its own — it calls the Razorpay API.

Did anything break? — user-facing failures, so a silent breakage shows up here instead of waiting for a customer to report it. ide = reported by the app (the class the server can't see); server = what the backend knows (empty_response_healed = a user almost saw an error; _unhealed = they did).

Last 7 days — a spike means investigate now

TOTAL ERRORS

—

FROM APP (IDE)

—

FROM BACKEND

—

By kind — what is actually failing

Kind	Count
Loading…

By model — which model produced the failure. A heal is a SECOND full provider call we pay for, so a model high on this list is costing twice per affected turn.

Model	Kind	Count
—

By release — which build regressed

Version	Errors
—

Ops error log — provider degraded/failed, reliability, and billing alerts land here instead of email. Each issue is deduped to once per hour, newest first. Email is off by default — flip ops_email_alerts in Settings to also email during a critical window.

Recent alerts

Loading…

What actually serves a request. These are the live values the router reads, not the defaults baked into the deploy. An A/B rule sends a slice of one tier to a challenger model — it was the only routing control with no page at all, so stopping one meant a database or a script.

A/B challenger

Loading…

Sets model_ab to {}. Takes effect on the next request.

Tier → model (live)

Edit these on the Settings tab. They are read on every request, so a change is live without a deploy.

Budgets

Voice is billed per audio second and vision per image, so each has its own daily ceiling per plan.

Did users LIKE it? — every other board measures machinery (edit-success = the edit applied; the eval = we pass tests). This is the only one that says the human was happy. A 👎 is also a real failure case worth turning into an eval task.

Last 30 days

SATISFACTION

—

👍 UP

—

👎 DOWN

—

BY TIER

—

What users said went wrong — their own words; the most actionable list here (and real eval tasks)

Loading…

By release — did a build make people unhappy

Version	Ratings
—