Fix agents

The agents#

Tiers and prices#

Every agent runs on one of three tiers. The tier sets the model alias, and each agent has a price range per tier.

• The quote’s maximum is the top of the range. Its estimate depends on how much there is to read.
• You’re charged for the work the run did, the tokens it read and wrote, from the bottom of the range up to the maximum. Never more than the maximum.
• A run that fails is charged nothing. It still gets a receipt.

What to send#

A fix is only as good as what the agent can read. Say in a few sentences what’s going wrong, paste the error as it was printed, and attach the files it points at. In the console, files are attached from your computer; over the API the description goes in description and files in files, as a path and its text.

TypeError: Cannot read properties of undefined (reading 'map')
    at OrderList (src/components/order-list.tsx:14:22)
    at renderWithHooks (react-dom.development.js:16305:18)

> next build
./app/(shop)/cart/page.tsx:3:10
Type error: Module '"next/navigation"' has no exported member 'useRouterState'.
Error: Command "npm run build" exited with 1

File types

Any text file: source code (.ts, .tsx, .js, .jsx, .mjs, .py, .go, .rs, .java, .php, .rb, .vue, .svelte…), config (.json, .yaml, .toml) and Dockerfiles.

Leave out

Leave out .env and key files (they’re refused), node_modules, build output like dist/ or .next/, and images or other binary files.

Limits in the console

Up to 10 files, 200 KB each. Lockfiles may be up to 1.5 MB, since only the versions they pin are read.

Limits per request

Quick Fix reads up to 60,000 characters and Build Doctor up to 160,000, log and files together, after lockfiles are reduced. The whole request may be up to 2 MB.

How a fix runs#

1. Ask for a quote

   Send the error log and the files it points at, and pick an agent and a tier. You get an estimate and a hard maximum. Quotes are free and last five minutes.

2. Approve it

   Approving holds the maximum from your balance. If your available balance (your balance minus what’s held for other runs and requests) is below the maximum, the run won’t start and the quote stays usable.

3. The agent works

   Each agent follows its own procedure. Quick Fix finds the stack frames in your own code and reads those files. Build Doctor reads package.json and the lockfile, lists the versions actually installed of the packages in the error, and checks them for known breaking changes. The lockfile’s text isn’t sent to the model, only those versions.

4. The patch is checked

   Every hunk of the patch must match the files you sent. If one doesn’t, the agent is asked once to correct it. A patch that still doesn’t match comes back marked patch_applies: false with low confidence. The patch and the verify steps are also scanned for things a fix never needs (see below).

5. You get the result and a receipt

   You’re charged the run’s actual cost and the rest of the hold is released. The receipt shows the quote, the hold, the charge and your balance before and after.

API#

Fix calls need a key with the fix scope. If the key is limited to certain models, it must allow the tier’s alias (for example inferiq/balanced for Balanced), or the quote fails with model_not_allowed. See Authentication.

123456789101112131415161718192021curl $INFERIQ_BASE_URL/fix/agents \
  -H "Authorization: Bearer $INFERIQ_API_KEY"

{
  "object": "list",
  "data": [
    {
      "id": "quick-fix",
      "name": "Quick Fix",
      "steps": ["Reading the error and stack trace", "Opening the files it points at", "Finding the cause and writing a patch", "Checking the patch"],
      "max_input_chars": 60000,
      "max_files": 10,
      "tiers": [
        { "id": "economy", "model": "inferiq/fast", "available": true, "price": { "min_micro": 20000, "max_micro": 35000 } },
        { "id": "balanced", "model": "inferiq/balanced", "available": true, "price": { "min_micro": 35000, "max_micro": 100000 } },
        { "id": "strongest", "model": "inferiq/best", "available": true, "price": { "min_micro": 100000, "max_micro": 280000 } }
      ]
    },
    { "id": "build-doctor", … }
  ]
}

12345678910111213141516171819202122232425262728293031323334353637curl $INFERIQ_BASE_URL/fix/quote \
  -H "Authorization: Bearer $INFERIQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": "build-doctor",
    "tier": "balanced",
    "description": "npm run build fails on Vercel since I upgraded Next.js to 16. It still builds locally.",
    "log": "Type error: Module \"next/navigation\" has no exported member useRouterState …",
    "files": [
      {"path": "package.json", "content": "…"},
      {"path": "package-lock.json", "content": "…"},
      {"path": "app/(shop)/cart/page.tsx", "content": "…"}
    ]
  }'

HTTP/1.1 201 Created
{
  "object": "fix.run",
  "id": "fix_Q2vR8sLm4TxN0bWk",
  "agent": "build-doctor",
  "tier": "balanced",
  "status": "quoted",
  "title": "Type error: Module \"next/navigation\" has no exported member useRouterState …",
  "source": "api",
  "model": "inferiq/balanced",
  "steps": ["Reading the build log", "Checking package.json and the installed versions", "Finding the cause and writing a patch", "Checking the patch"],
  "quote": { "estimate_micro": 196000, "max_micro": 450000, "expires_at": "2026-10-17T14:37:00.000Z" },
  "cost_micro": 0,
  "receipt_id": null,
  "result": null,
  "error": null,
  "created_at": "2026-10-17T14:32:00.000Z",
  "started_at": null,
  "finished_at": null,
  "duration_ms": null,
  "input": { "chars": 18420, "files": 3, "secrets_removed": 0, "left_out": [] }
}

agent

Required. quick-fix or build-doctor.

tier

Optional. economy, balanced (default) or strongest.

description

Optional, up to 4,000 characters. In your own words: what you were doing, what you expected, what happened and what you tried. The agent reads it before the log. Secrets are removed from it too.

log

Required. The error output, as plain text.

files

Optional. Up to 10 files with content, as { path, content }. A path without content tells the agent the error mentions it. Send only what the error points at; more input raises the estimate.

input.secrets_removed

How many secrets were replaced with [redacted] before anything was stored.

input.left_out

Files dropped because they only hold secrets, such as .env or key files.

Quick Fix reads up to 60,000 characters and Build Doctor up to 160,000, log and files together. Over that, or over 2 MB per request, the quote fails with context_too_large.

123456789101112131415161718192021222324252627282930curl $INFERIQ_BASE_URL/fix/runs \
  -H "Authorization: Bearer $INFERIQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"quote_id": "fix_Q2vR8sLm4TxN0bWk"}'

{
  "object": "fix.run",
  "id": "fix_Q2vR8sLm4TxN0bWk",
  "status": "completed",
  "model": "qwen/qwen3.8-27b",
  "quote": { "estimate_micro": 196000, "max_micro": 450000, "expires_at": "2026-10-17T14:37:00.000Z" },
  "cost_micro": 188500,
  "receipt_id": "rcpt_hT3kWq9ZbN2xLp5R",
  "result": {
    "cause": "`useRouterState` was never part of `next/navigation`; it comes from TanStack Router …",
    "patch": "--- a/app/(shop)/cart/page.tsx\n+++ b/app/(shop)/cart/page.tsx\n@@ -1,8 +1,8 @@\n …",
    "verify": ["Run `npm run build`: type checking passes."],
    "confidence": "high",
    "patch_applies": true,
    "patch_problems": [],
    "warnings": []
  },
  "error": null,
  "usage": { "input_tokens": 6820, "output_tokens": 912 },
  "balance": { "before_micro": 12500000, "after_micro": 12311500 },
  "started_at": "2026-10-17T14:32:40.000Z",
  "finished_at": "2026-10-17T14:33:01.000Z",
  "duration_ms": 21000,
  …
}

Without stream the call answers once the run is done, usually within a minute. With stream: true it sends server-sent events: one step per step in the agent’s steps list, then done with the run.

1234567891011121314curl -N $INFERIQ_BASE_URL/fix/runs \
  -H "Authorization: Bearer $INFERIQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"quote_id": "fix_Q2vR8sLm4TxN0bWk", "stream": true}'

data: {"type":"step","index":0}

data: {"type":"step","index":1}

data: {"type":"step","index":2}

data: {"type":"step","index":3}

data: {"type":"done","run":{"object":"fix.run","id":"fix_Q2vR8sLm4TxN0bWk","status":"completed",…}}

If the run can’t start, the stream sends one error event instead:

data: {"type":"error","error":{"message":"Available balance $0.120000 doesn't cover this fix's maximum of $0.450000. Add funds, or pick a cheaper tier.","type":"insufficient_quota","code":"insufficient_balance","retryable":false,"request_id":null}}

GET /v1/fix/runs/{id} returns the run with its status (quoted, running, completed, failed or expired) and, for 24 hours, its result. Use it if your connection drops during a run: the run finishes either way, and its receipt is written.

A quote runs once. Running it again fails with invalid_request, and running it after five minutes fails with quote_expired. From an editor, the same flow runs through the MCP server: inferiq_quote, then inferiq_fix once you approve the price (Connect your tool).

The result#

result.cause

Why it fails, in a few sentences, naming files and lines.

result.patch

A unified diff. Nothing is applied for you. Empty when no code change is needed.

result.verify

One to four steps or commands that confirm the fix.

result.confidence

high, medium or low. Always low when the patch doesn’t match your files.

result.patch_applies

true when every hunk matched the files you sent, false when one didn’t, null when there was nothing to check (no patch, or it edits files you didn’t send).

result.warnings

What the safety scan found, each with a level (danger or caution) and a message. Empty when nothing was flagged. Confidence is always low when there’s a danger warning.

cost_micro

What you were charged, in micro-dollars (1,000,000 = $1).

receipt_id

The fix.run receipt. Its fix object holds the agent, tier, title and quote.

error

On a failed run: the code and message. A failed run costs nothing.

What a fix agent never does#

• Run commands or change files on your machine. It returns a patch; you or your own agent apply it.
• Keep your code. It’s kept only until the run ends or the quote expires. Results are deleted after 24 hours; receipts keep only metadata.
• Charge more than the maximum you approved.
• Treat instructions inside your logs or files as orders. Error output is data, not instructions.

Code, logs and files can contain text written to steer an AI (prompt injection), for example in a dependency’s output or a file someone else wrote. Inferiq guards against it in three ways:

• Your input is fenced with tags that change on every run, so text inside it can’t pose as instructions, and the agent is told to use it only as evidence. Invisible characters (zero-width, Unicode tags, bidi overrides) are removed first, and you’re told when there were any.
• Every patch and verify step is scanned before you get it. Danger: downloading and running scripts, sending environment variables, secrets or cookies over the network, reverse shells, install scripts in package.json, deleting files outside the project, editing CI, git or shell config you didn’t send, writing outside the project, or hiding code in encoded strings. Caution: dependencies from outside the npm registry, TLS checks turned off, edits to files you didn’t send, long encoded blobs.
• If the scan finds danger, the agent is asked once to fix the error without it. If it still can’t, you get the patch with warnings, low confidence and a red notice on the Fix page: read it before applying anything.
• The explanation is checked too: links to sites that aren’t usual developer sites are flagged, an answer that repeats the agent’s hidden instructions is removed, and invisible bidi characters added by a patch (“Trojan Source”) count as danger.

Before anything is stored or sent to a model, Inferiq replaces common secret patterns (API keys, tokens, passwords, private keys, passwords in connection strings) with [redacted], and leaves out .env and key files entirely. The quote tells you how many were removed. Still, don’t paste secrets.