Loading article…

Loading 4weird…

DocsVibeCodeWorker

Docs · the QA lab

Break it. Prove it. Ship it.

Evidence-driven QA: point VibeCodeWorker at a game, watch it observe → reason → act, and get runs, findings, bugs, and portable handoffs with proof. Start at /vibecodeworker.

3: steps per loop
4: severities
3: verdicts
1: portable handoff
20: steps per batch
2000: steps per export
500: bugs per export

The floor plan

Surfaces at /vibecodeworker/*

Overview, hub, run, full, phone, docs, and demo - each pairs a guide with the live surface plus a public service-status pill. If the pill reports trouble, wait before opening runs: the loop depends on the service underneath.

The assembly line

The 7-step agent loop

Signed in, every route returns success envelopes and rate-limits per user. Read the full run (trail + findings) before every next step.

Check status
Confirm service health, catalog game count, recent runs/bugs. Unhealthy? Stop here.
Pick a target
List catalog games - slug, title, genre, play URL. Every legal run target comes from this list; first-party play URLs only.
Open a run
Catalog slug + a 1-500 character goal (“verify level-2 boss spawns adds”). You get a run ID. List yours with ?game_slug= + ?status=open|completed + ?verdict=pass|fail|inconclusive + ?limit= + ?before= cursor paging; bugs filter by ?severity= + ?game_slug= + ?run_id= the same way.
Observe → reason → act
Append one iteration at a time - observation, action, or finding (text + optional ≤10 KB data object). Open runs only.
File bugs
Title + description + severity (low/medium/high/critical, default medium), optionally pinned to a run (defaults the slug).
Complete the run
Close with summary + verdict: pass / fail / inconclusive. Closed runs are read-only history.
Hand off
Export a portable markdown brief for any coding tool (defaults to your latest run). The dashboard gives recent runs + bugs in one view.

Exhibit A

What good evidence looks like

run #4821 - boss-spawn verification

verdict: pass

OBSERVE screenshot: 3 adds spawn at 0:42, HP bars visible

REASON spawn matches spec §2.1 → engage, record score events

ACT cleared wave, no console errors, 60fps held

FINDING goal met - no bug filed, evidence attached

Field notes

Writing runs + bugs that get fixed

🎯 Goals

One verifiable claim per run: “menu → new game → first checkpoint, no console errors.”

🔬 Steps

Small, timestamped, evidence-first: “screenshot shows…”, “score event fired…”.

🐞 Bugs

Expected vs. actual + repro path + severity + run link. Critical = data loss, payment error, safety.

🏁 Verdicts

Pass (evidence), fail (repro defect filed), inconclusive (blocked - say what blocked you).

Keys + gateway

Call the loop with your own tools

The run loop above uses your login session. The gateway (under /api/vcw/gateway/*) gives the same loop a key you can paste into your own agent tools - hosted on our GPUs, or routed to your own provider keys (BYOK).

Probe the gateway
No key needed: GET /api/vcw/gateway/status reports the hosted markups (15% standard, 9% enterprise) and the BYOK kinds. Start here before wiring anything.
Issue a key
Signed in: POST /api/vcw/gateway/keys with a label, scopes (vcw:read, vcw:write), and optional lifetime/daily budgets + expiry. The vcw_live_ secret is shown once and never again (max 10 active keys). List them anytime with GET - secrets never come back.
Register your providers (BYOK, optional)
POST /api/vcw/gateway/providers with kind (runpod, openai, fal, meshy, or custom + your https endpoint), a label, and your key. We store only kind + label + last-4 - your full secret is never stored, returned, or logged. DELETE removes one.
Dispatch a run
POST /api/vcw/gateway/dispatch with game_slug, compute (cpu/gpu/gpu-boosted), mode (hosted/byok), and goal - authenticated with your login, a bot key, or the gateway key (x-bot-key or Authorization: Bearer), with vcw:write scope. The MVP answers honestly: always started:false with the coin quote, the meter receipt, and the manual next step (POST /api/vcw/runs or autoplay). No worker URL is ever faked; low balance fails closed with 402.
Watch the spend
GET /api/vcw/gateway/usage shows your latest 50 metered rows. Every figure is gross with the 25% cut already inside - same as every metered surface on the site.

💰 Gateway prices (gross, cut inside)

run-open 10 · action-step 1 · bug-file 2 · handoff 5 · worker-min 6 · byok-route 2 coins. Hosted adds 15% over provider cost (9% enterprise); BYOK bills the base price. 100 coins = $1.00, always.

🔑 Key hygiene

Copy the secret once, store it like a password, set a budget + expiry, revoke keys you stop using. Gateway-key reads and writes are metered to your coins like session calls.

Full fidelity

Read the whole trail, then compare

The run detail caps the trail for speed (steps_limit 1-200, bugs_limit 1-100), but nothing is lost: the export removes the caps for archiving, and compare answers “did the fix work” side-by-side. Detail, export, and compare are read-only, so none of them meter coins.

GET /api/vcw/runs/[id]/export

format: vcw-run-export/1

run id · game_slug · goal · status · verdict · summary · timestamps

steps every step in order, up to 2000 (id · kind · text · data · created_at)

bugs every filed bug in order, up to 500 (id · title · severity · description · created_at)

COMPARE GET /api/vcw/runs/compare?a=<uuid>&b=<uuid> - verdicts, step-kind mix, bug-severity mix, plus same_game

Batch + media

Batch steps, chain fal media

The loop usually produces a whole observe → reason → act triplet per iteration. Recording it one step at a time costs three metered round trips - batch appends 1-20 steps in one call, validated exactly like single steps.

One triplet, one call
POST /api/vcw/runs/[id]/actions/batch with { steps: [{ kind, text, data? }] }. Each item needs a legal kind, 1-5000 characters of text, and a data object that fits in 10 KB - the first bad item rejects the whole batch.
All-or-nothing metering
Every step meters 1 coin gross (action-step, 25% cut inside). If any debit fails, the whole batch rolls back - steps are never free and partial batches never linger. Low balance answers 402, closed runs answer 409.
Tag media in the trail
Write [tool: fal.generate; op=<op> prompt="..."] in an action-step text or data object and the route answers with the gross quote plus the next hop (POST /api/fal/generate with source “vcw”). Unknown ops get a hint naming the 30-op catalog instead.
Pick the right phase
Observe and reason prefer the fast text/image/audio ops; video and 3D ops are act-phase only (slow). GET /api/vcw/status lists every op grouped by phase plus whether fal is configured - probe it before tagging.

Remotes

Autoplay: real remotes, honest no's

Autoplay starts a RunPod remote that drives the play page for you - cpu for cheap, gpu for vision, gpu-boosted (pinned RTX 4090 class) for the hardest titles. It never fakes a worker URL: no stock, no credentials, or over budget comes back as started:false with the code and the manual next step.

🖥️ cpu · 0.14/min

Cheapest seat: a CPU remote drives the first-party play page. A 60-minute run lands around 9 coins.

🎮 gpu · 0.63/min

Vision seat: a GPU remote watches and plays the same on-site page. About 38 coins for a full hour.

🚀 gpu-boosted · 2.2/min

Fastest eyes, exact quote at start: pinned RTX 4090 class (5090 fallback), required for off-site Xonotic. About 132 coins per hour ceiling.

Look before you leap

Probe status + dashboard first

Two authenticated reads prove the loop is usable before you spend a coin: status for the control plane, dashboard for your own recent runs and bugs. Both are rate-limited per user and both refuse strangers with 401.

📡 GET /api/vcw/status

Service health (ok, degraded, unavailable, or unconfigured), catalog game count, your run/bug totals, every fal op grouped by observe/reason/act phase, plus the swarm tools hint for multi-agent QA.

🗂️ GET /api/vcw/dashboard

Service state plus your 10 most recent runs and 10 most recent bugs in one call - with counts for open runs, runs by verdict, and bugs by severity. The hint field points at the deeper run/bug filters.

Who may do what

Three doors, one keychain

Every gateway call resolves its caller from exactly one credential, tried in order: your login session, a bot key carrying a vcw scope, or a vcw_live_ gateway key. All failures collapse to one uniform “authentication required” - the how is never leaked.

Session first
Signed in with Supabase, you get full read + write with no key at all. Key issuance itself (POST /api/vcw/gateway/keys) needs the session plus same-origin - keys can't mint keys.
Bot keys carry scopes
A bot key with vcw:read or vcw:write rides the same loop; write implies read. Keys are looked up by prefix with constant-time comparison, and unknown prefixes still burn a dummy compare so timing leaks nothing.
Gateway keys carry budgets
Issue with a 1-80 character label, scopes (vcw:read default, vcw:write for writes), optional lifetime/daily budgets, max uses, and a future expiry date. Max 10 active keys; the full secret shows once, only its hash is stored. Present it via x-vcw-key, x-gateway-key, or Bearer.
The floor holds
Keys with a low-balance floor fail closed when the owner's coin balance can't be read or sits at/below the floor; spent budgets and expired or revoked keys refuse the same way. Every use bumps counters and refreshes the daily window.

Take it with you

Handoffs: markdown any tool can read

POST /api/vcw/handoff with an optional run_id and reason (defaults to your latest run) returns a portable markdown brief capped for pasting: trail digest, 50 steps, 50 bugs, and next actions for the coding agent. Metered at 5 coins gross - assembled only for a paid handoff.

handoff.md - goal, digest, steps, bugs, next

50 + 50 caps

# VCW run handoff game slug · goal · status + verdict · reason · run id · updated-at · summary

## Trail digest step count by kind · bug count by severity · fal.generate mentions

## Steps (50) [kind] @ timestamp + first 300 chars each

## Bugs (50) [severity] title + first 300 chars of description

NEXT replay last observe → action · fix critical → low · emit fal tags for media · complete with verdict

Loading article…

DocsVibeCodeWorker

Docs · the QA lab

Break it. Prove it. Ship it.

Evidence-driven QA: point VibeCodeWorker at a game, watch it observe → reason → act, and get runs, findings, bugs, and portable handoffs with proof. Start at /vibecodeworker.

3: steps per loop
4: severities
3: verdicts
1: portable handoff
20: steps per batch
2000: steps per export
500: bugs per export

The floor plan

Surfaces at /vibecodeworker/*

The assembly line

The 7-step agent loop

Signed in, every route returns success envelopes and rate-limits per user. Read the full run (trail + findings) before every next step.

Check status
Confirm service health, catalog game count, recent runs/bugs. Unhealthy? Stop here.
Pick a target
List catalog games - slug, title, genre, play URL. Every legal run target comes from this list; first-party play URLs only.
Open a run
Catalog slug + a 1-500 character goal (“verify level-2 boss spawns adds”). You get a run ID. List yours with ?game_slug= + ?status=open|completed + ?verdict=pass|fail|inconclusive + ?limit= + ?before= cursor paging; bugs filter by ?severity= + ?game_slug= + ?run_id= the same way.
Observe → reason → act
Append one iteration at a time - observation, action, or finding (text + optional ≤10 KB data object). Open runs only.
File bugs
Title + description + severity (low/medium/high/critical, default medium), optionally pinned to a run (defaults the slug).
Complete the run
Close with summary + verdict: pass / fail / inconclusive. Closed runs are read-only history.
Hand off
Export a portable markdown brief for any coding tool (defaults to your latest run). The dashboard gives recent runs + bugs in one view.

Exhibit A

What good evidence looks like

run #4821 - boss-spawn verification

verdict: pass

OBSERVE screenshot: 3 adds spawn at 0:42, HP bars visible

REASON spawn matches spec §2.1 → engage, record score events

ACT cleared wave, no console errors, 60fps held

FINDING goal met - no bug filed, evidence attached

Field notes

Writing runs + bugs that get fixed

🎯 Goals

One verifiable claim per run: “menu → new game → first checkpoint, no console errors.”

🔬 Steps

Small, timestamped, evidence-first: “screenshot shows…”, “score event fired…”.

🐞 Bugs

Expected vs. actual + repro path + severity + run link. Critical = data loss, payment error, safety.

🏁 Verdicts

Pass (evidence), fail (repro defect filed), inconclusive (blocked - say what blocked you).

Keys + gateway

Call the loop with your own tools

Probe the gateway
No key needed: GET /api/vcw/gateway/status reports the hosted markups (15% standard, 9% enterprise) and the BYOK kinds. Start here before wiring anything.
Issue a key
Signed in: POST /api/vcw/gateway/keys with a label, scopes (vcw:read, vcw:write), and optional lifetime/daily budgets + expiry. The vcw_live_ secret is shown once and never again (max 10 active keys). List them anytime with GET - secrets never come back.
Register your providers (BYOK, optional)
POST /api/vcw/gateway/providers with kind (runpod, openai, fal, meshy, or custom + your https endpoint), a label, and your key. We store only kind + label + last-4 - your full secret is never stored, returned, or logged. DELETE removes one.
Dispatch a run
POST /api/vcw/gateway/dispatch with game_slug, compute (cpu/gpu/gpu-boosted), mode (hosted/byok), and goal - authenticated with your login, a bot key, or the gateway key (x-bot-key or Authorization: Bearer), with vcw:write scope. The MVP answers honestly: always started:false with the coin quote, the meter receipt, and the manual next step (POST /api/vcw/runs or autoplay). No worker URL is ever faked; low balance fails closed with 402.
Watch the spend
GET /api/vcw/gateway/usage shows your latest 50 metered rows. Every figure is gross with the 25% cut already inside - same as every metered surface on the site.

💰 Gateway prices (gross, cut inside)

🔑 Key hygiene

Copy the secret once, store it like a password, set a budget + expiry, revoke keys you stop using. Gateway-key reads and writes are metered to your coins like session calls.

Full fidelity

Read the whole trail, then compare

GET /api/vcw/runs/[id]/export

format: vcw-run-export/1

run id · game_slug · goal · status · verdict · summary · timestamps

steps every step in order, up to 2000 (id · kind · text · data · created_at)

bugs every filed bug in order, up to 500 (id · title · severity · description · created_at)

COMPARE GET /api/vcw/runs/compare?a=<uuid>&b=<uuid> - verdicts, step-kind mix, bug-severity mix, plus same_game

Batch + media

Batch steps, chain fal media

One triplet, one call
POST /api/vcw/runs/[id]/actions/batch with { steps: [{ kind, text, data? }] }. Each item needs a legal kind, 1-5000 characters of text, and a data object that fits in 10 KB - the first bad item rejects the whole batch.
All-or-nothing metering
Every step meters 1 coin gross (action-step, 25% cut inside). If any debit fails, the whole batch rolls back - steps are never free and partial batches never linger. Low balance answers 402, closed runs answer 409.
Tag media in the trail
Write [tool: fal.generate; op=<op> prompt="..."] in an action-step text or data object and the route answers with the gross quote plus the next hop (POST /api/fal/generate with source “vcw”). Unknown ops get a hint naming the 30-op catalog instead.
Pick the right phase
Observe and reason prefer the fast text/image/audio ops; video and 3D ops are act-phase only (slow). GET /api/vcw/status lists every op grouped by phase plus whether fal is configured - probe it before tagging.

Remotes

Autoplay: real remotes, honest no's

🖥️ cpu · 0.14/min

Cheapest seat: a CPU remote drives the first-party play page. A 60-minute run lands around 9 coins.

🎮 gpu · 0.63/min

Vision seat: a GPU remote watches and plays the same on-site page. About 38 coins for a full hour.

🚀 gpu-boosted · 2.2/min

Fastest eyes, exact quote at start: pinned RTX 4090 class (5090 fallback), required for off-site Xonotic. About 132 coins per hour ceiling.

Look before you leap

Probe status + dashboard first

📡 GET /api/vcw/status

Service health (ok, degraded, unavailable, or unconfigured), catalog game count, your run/bug totals, every fal op grouped by observe/reason/act phase, plus the swarm tools hint for multi-agent QA.

🗂️ GET /api/vcw/dashboard

Who may do what

Three doors, one keychain

Session first
Signed in with Supabase, you get full read + write with no key at all. Key issuance itself (POST /api/vcw/gateway/keys) needs the session plus same-origin - keys can't mint keys.
Bot keys carry scopes
A bot key with vcw:read or vcw:write rides the same loop; write implies read. Keys are looked up by prefix with constant-time comparison, and unknown prefixes still burn a dummy compare so timing leaks nothing.
Gateway keys carry budgets
Issue with a 1-80 character label, scopes (vcw:read default, vcw:write for writes), optional lifetime/daily budgets, max uses, and a future expiry date. Max 10 active keys; the full secret shows once, only its hash is stored. Present it via x-vcw-key, x-gateway-key, or Bearer.
The floor holds
Keys with a low-balance floor fail closed when the owner's coin balance can't be read or sits at/below the floor; spent budgets and expired or revoked keys refuse the same way. Every use bumps counters and refreshes the daily window.

Take it with you

Handoffs: markdown any tool can read

handoff.md - goal, digest, steps, bugs, next

50 + 50 caps

# VCW run handoff game slug · goal · status + verdict · reason · run id · updated-at · summary

## Trail digest step count by kind · bug count by severity · fal.generate mentions

## Steps (50) [kind] @ timestamp + first 300 chars each

## Bugs (50) [severity] title + first 300 chars of description

NEXT replay last observe → action · fix critical → low · emit fal tags for media · complete with verdict

Break it. Prove it. Ship it.

Surfaces at /vibecodeworker/*

The 7-step agent loop

Check status

Pick a target

Open a run

Observe → reason → act

File bugs

Complete the run

Hand off

What good evidence looks like

Writing runs + bugs that get fixed

Call the loop with your own tools

Probe the gateway

Issue a key

Register your providers (BYOK, optional)

Dispatch a run

Watch the spend

Read the whole trail, then compare

Batch steps, chain fal media

One triplet, one call

All-or-nothing metering

Tag media in the trail

Pick the right phase

Autoplay: real remotes, honest no's

Probe status + dashboard first

Three doors, one keychain

Session first

Bot keys carry scopes

Gateway keys carry budgets

The floor holds

Handoffs: markdown any tool can read

Break it. Prove it. Ship it.

Surfaces at /vibecodeworker/*

The 7-step agent loop

Check status

Pick a target

Open a run

Observe → reason → act

File bugs

Complete the run

Hand off

What good evidence looks like

Writing runs + bugs that get fixed

Call the loop with your own tools

Probe the gateway

Issue a key

Register your providers (BYOK, optional)

Dispatch a run

Watch the spend

Read the whole trail, then compare

Batch steps, chain fal media

One triplet, one call

All-or-nothing metering

Tag media in the trail

Pick the right phase

Autoplay: real remotes, honest no's

Probe status + dashboard first

Three doors, one keychain

Session first

Bot keys carry scopes

Gateway keys carry budgets

The floor holds

Handoffs: markdown any tool can read