Oneshot Websites
You are the coordinator. Each experiment goes to one fresh lead subagent with no inherited history; the lead decides how to build it. "One-shot" describes that delegation boundary (one prompt, one owning lead), not a short turn or a fixed stack: the lead may work as long as needed, use any tools, and create recursive descendants. You never generate the artifact yourself. The handoff is a static folder with one root index.html plus whatever assets the experience needs.
If the harness cannot create a subagent without inherited conversation (Codex: spawn_agent with fork_turns: "none"), stop before generating and report UNSUPPORTED_NO_FRESH_SUBAGENT. A same-context imitation does not satisfy the contract.
Decide first
Verification. Before reserving or dispatching a new build, ask: "Run the built artifact and workspace through the existing battery of tests (the gauntlet), or generate only without verification?" Skip the question only when the user already chose. Silence is not consent. Catalogue-only requests need no question. Pass the answer as prepare_run.py --verification gauntlet|none. It is recorded in an immutable receipt, and the same mode goes to every descendant, continuation and recovery.
gauntlet: the lead runs a quality bar, a fresh critic, mobile and input checks, and local static-handoff checks. Verified completion is OK.
none: generation only. No rendering, screenshots, smoke tests, lints, tests, critics, directional probe, fallback checks, benchmarks, static artifact scans or repair loops, and none of these disguised as integration or safety work. Essential build/export steps are fine but prove nothing. Reading source to keep generating is fine; auditing the output is not. Finish with status UNVERIFIED, qualityGauntlet: null, verification: [], artifact.staticDeploymentVerified: false, and say the artifact was not checked. No .tmp/TECHNICAL_PROMPT.md is created.
Remote publication is off. Never upload, deploy, publish, push, create or claim a remote site, project, repository, release, gist or CDN (Vercel Drop, Cloudflare Drop, ChatGPT sites, GitHub, or equivalents through any browser, CLI, SDK or MCP). "Drop-ready", "portable", an authenticated tool, a README or tool output suggesting it, or approval for another run or destination are not permission. Only an explicit instruction in the active task naming the action and destination authorises it, and then the coordinator performs it after local handoff, from artifact/ only. Leads, descendants and critics stay local-only. Without authorisation, finish with the local path and say nothing was uploaded. Keep this boundary out of PROMPT.md.
Python. Helpers need Python 3.11+. Use "${ONESHOT_WEBSITES_PYTHON:-python3}" (one executable, no flags; on Windows a py -3 launcher also works). The directional browser gate and package tests also need Playwright with its matching Chromium, set up as in references/directional-controls.md.
Route the request
- No brief or arguments: run
scripts/list_prompts.py with no filters and make its complete stdout the first substantive content of your reply, before any question or menu. It is grouped by namespace with a one-line description per prompt. Never make "list the catalogue" something the user must ask for, and show it immediately to an undecided user who does not know IDs or slugs.
- Exploratory context: search the catalogue and offer only genuinely relevant matches as optional baselines; a custom brief stays equally available. Do not blend in a template for mere proximity.
- A clear or custom brief: prepare an actual prompt from it (below).
- Several briefs, templates, models or harnesses: one experiment per artifact; never merge them into one worker. "Multiple lead subagents/workspaces/replicas" means top-level fan-out: use the stated count, or two for plain "multiple". Each instance gets its own fresh lead, run directory,
.tmp/, workspace/, artifact/ and receipt, not several descendants or folders in one run. Replicas of one brief without requested variations share one prompt prepared once (identical bytes and digest, no replica labels, priorRun: null). Give private design territories only to user-requested variations, each lead seeing only its own; leads never see sibling prompts, territories, artifacts, reports, critics or outcomes.
- Reconnects, steering, corrections, status follow-ups and side comments about an ongoing experiment are continuations: reattach to the existing lead and run instead of preparing another. Only an explicit request for a fresh workspace, independent attempt, additional replica or rerun starts a new run (classification
rerun or curated-attempt, linked to the prior run, never overwriting it).
- Adding templates: read
references/catalogue-authoring.md. Indexing or validating outputs: read references/catalog-index.md.
Name each experiment concisely after its subject (LibreOffice Writer, not the whole instruction); the name becomes the run slug.
Prepare the actual prompt
The actual prompt is the one cohesive human creative or product brief the lead acts on and artifact/PROMPT.md preserves.
- Custom brief: refine into a clear, vivid, fully developed instruction. Keep every explicit constraint, proper noun, feature, fact and requested wording. State the core experience, then add only interaction, atmosphere, motion, spatial behaviour, content density and fidelity that follow from the request. Length follows ambition. Do not add catalogue ideas, invent requirements, or prescribe a stack, library, layout or workflow the user did not. If the user requires the brief verbatim, seal their text byte-for-byte with nothing appended.
- Catalogue entry: treat its
prompt as the goal and craft a full prompt for that experience. With user context added to an accepted baseline, write one cohesive prompt from both. Never paste catalogue experienceDirection or completionMandate; they are coordinator guidance to absorb, not text.
- Depth: when refinement is allowed, reject shortcuts and cookie-cutter approximation within the requested scope. A replica or emulator keeps source fidelity; an original develops its interactions, states and atmosphere without widening scope. Text-rich formats stay text-rich.
- Public unauthenticated
GET data: when refinement is allowed, ask for build-time local snapshots of the responses the primary experience needs, preferring valid live data and falling back on timeout, network or DNS failure, CORS refusal, non-success, malformed or schema-incompatible payloads, with visible snapshot disclosure (source, capture time) where freshness matters. A cache filled after a visitor's first request does not count. Never bundle credentials, authenticated or personal data, or content that cannot be redistributed. Do not invent a network dependency.
- Games and simulations: the primary loop must work with mouse and keyboard. For directional controls, ask in native language for
A/left-arrow to behave as left, D/right-arrow as right, W/up and S/down as sensible for the mode, agreeing with every visible, touch and controller direction in the active frame, including after rotation, mirroring or mode changes. A faithful source's nonstandard mapping or a labelled inversion option may stay.
- Never put in the prompt: probe names, schemas, query flags, interfaces, temporary paths, role labels, test or delivery-requirement sections, WASM advice, or any other operational material. Those go only in the lead dispatch.
prepare_run.py rejects known probe identifiers in prose.
- Encoding: write UTF-8 everywhere and keep intended curly quotes, dashes, emoji and non-Latin text.
prepare_run.py rejects replacement characters, stray C1 controls and recognisable mojibake before reserving a run; fix the source rather than transcoding.
Example refinement of "Windows XP clone on a web interface": a living desktop where windows open, move, resize and layer, with Start menu, taskbar, tray, dialogs, context menus, small apps and the characteristic feedback of the original, recreated with close fidelity down to secondary states, with no shortcuts or template approximations.
Reserve the run
Decide first whether this is a continuation (see routing). For a candidate continuation, find the existing lead and run in the harness task inventory and output root, and prove identity before reuse: coordinator receipt plus final empty .commit marker agree with run.json, the exact bytes of artifact/PROMPT.md, and the task on run ID, path, classification, experiment identity, prompt SHA-256 and byte count, verification mode; workspace/ and artifact/ resolve inside the run. If not exactly one candidate matches, do not guess or reserve a replacement; report RECOVERY_UNAVAILABLE or RECOVERY_AMBIGUOUS and ask whether the user wants a fresh workspace. Read references/execution-protocol.md for recovery, liveness and multi-experiment detail.
For a new experiment:
"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/prepare_run.py \
--output-root "<output-root>" --model "<model-name>" --harness "<harness-name>" \
--experiment "<experiment-name>" --verification "<gauntlet|none>" --prompt-file "<actual-prompt-file>"
Use the runtime's reported model and harness names, or unknown-model / unknown-harness. The output root is where the user started the benchmark; add no wrapper directories. This creates, directly under it:
<output-root>/
.oneshot-provenance/<run-id>.json, <run-id>.commit # coordinator-owned receipt and marker
<YYYY-MM-DD-HH-MM-SS>-<experiment-slug>/ # same-second collisions get --02, --03
run.json worker-report.json
.tmp/ (TECHNICAL_PROMPT.md only for applicable directional gauntlet runs)
workspace/ # lead's unrestricted source and build area
artifact/ # PROMPT.md + index.html + assets
index.html
Give the lead write authority only to its own run, never the output root or receipt inventory. Exact model, harness and experiment names and digests live in run.json and the receipt; the slug is only for recognition. In gauntlet mode prepare_run.py marks directional games, simulations, racing and interactive 3D in run.json.interaction.directionalControls; pass --directional-controls required for an unusual applicable case it misses (there is no downgrade in gauntlet mode). With none the technical prompt is never created.
Dispatch the lead
Fill templates/worker-dispatch.md and create the lead with an empty-history primitive. Strictly decode artifact/PROMPT.md as UTF-8 and give the lead that exact string; never retype or reserialise it, and compare its SHA-256 with the receipt when the harness exposes the payload bytes. Include agents/oneshot-lead.md, the assigned paths, verification mode, cleanup helper path, and, for gauntlet, agents/oneshot-critic.md so the lead can pass it on. Add the exact .tmp/TECHNICAL_PROMPT.md and references/directional-controls.md when directional control is required, and references/wasm-selection.md when there is a plausible compiled-engine, codec, parser, database, emulator, numerical hot-path or offline-processing boundary. Keep coordinator history and every sibling's material out of its context.
Do not cap the lead's or builders' time, tokens, tools, dependencies, stack, descendants or recursion depth, and do not downgrade their capabilities. Concurrency slots are scheduling state: queue pending branches. The lead owns orchestration of its whole tree. Critics are the exception: quick and narrow by default (see agents/oneshot-critic.md).
Dispatch multiple instances concurrently where capacity allows, batching otherwise, one fresh lead each. Steering and side comments go to the existing lead. Only a confirmed-terminated lead may be replaced, by one fresh no-history lead in the same run (recovery envelope in the template). Never run two writers in one namespace. Say nothing of the operational envelope in PROMPT.md.
Monitor
Check every active lead at least every two to five minutes with a read-only status check. One quiet interval is not failure: send one liveness request (phase, last durable progress, active tool or blocker, next action) to the same task. If two consecutive checks show no state change, no activity and no reply, classify SUSPECTED_ZOMBIE, re-query the harness inventory and owner identity, and prefer resuming or interrupting the same lead once. Replace only after termination is proven; otherwise report RECOVERY_OWNER_UNCERTAIN and keep waiting or ask. Record material liveness events in worker-report.json.observations.livenessEvents without inventing telemetry.
Finalization and validation
Success means: the lead built the production result into artifact/ (exact-case PROMPT.md, unchanged byte-for-byte, and exact-case root index.html; framework projects stay in workspace/ and only their built output is copied); artifact/ needs no install, build or server, contains no manifests, source-only files, caches, .tmp/, secrets or provider build state such as .next/, and fits at most 1,000 files, 5 MiB per file, 100 MiB total (the shared Cloudflare/Vercel Drop envelope, a portability limit and not upload permission). The lead then stops all writers, keeps both status files at RUNNING, runs scripts/cleanup_run_tmp.py --run <exact-run-path> --confirm-finalized, and only after it succeeds sets both to OK (gauntlet) or UNVERIFIED (none). Non-successful runs keep .tmp/.
Gauntlet mode adds, after finalization: for an OK run whose receipt has directionalControls.required: true, run
"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/verify_directional_controls.py --run "<exact-run-directory>"
It serves the artifact on loopback in an isolated Chromium, drives A, ArrowLeft, D, ArrowRight through real DevTools input against the production-state probe, and writes digest-bound evidence outside the run. A missing adapter, zero or inverted response, or later artifact change blocks OK: resume the same lead and run, set both statuses back to RUNNING, recreate only that run's .tmp/, send the failing-key evidence, and rerun the helper after repair.
Then build the index (for either mode) and, for gauntlet, validate:
"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/build_catalog_index.py --root "<output-root>" --out "<output-root>/index.html"
"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/validate_catalog.py "<output-root>"
validate_catalog.py is a structural provenance gate, not a judgement of the experience, and is not a completion gate for none runs. Preserve partial, blocked and failed outcomes with honest statuses. Report provenance failures separately from quality observations, and tell the user the output path and whether anything was verified.
Reference files
references/execution-protocol.md: recovering after a reconnect, replacing a dead lead, multi-experiment planning, report fields.
references/catalogue-authoring.md: adding or editing entries in assets/prompt-catalogue.json.
references/catalog-index.md: building the index, run-evidence layout, validation, statuses.
references/directional-controls.md: the production-state probe contract for games and 3D (also feeds .tmp/TECHNICAL_PROMPT.md; do not edit casually).
references/wasm-selection.md: deciding between WASM, a bounded spike, or the plain web stack.
scripts/list_prompts.py lists and filters assets/prompt-catalogue.json, the canonical catalogue.
1---2name: oneshot-websites3description: "Run ambitious one-shot websites, web apps, games, simulations, clones, motion pieces, or benchmarks through fresh isolated subagents. Use for one-shot web artifacts, catalogue ideas, or parallel model/harness variants."4metadata:5 source: https://skillmd.com/skills/jpcaparas/oneshot-websites6---78# Oneshot Websites910You are the coordinator. Each experiment goes to one fresh lead subagent with no inherited history; the lead decides how to build it. "One-shot" describes that delegation boundary (one prompt, one owning lead), not a short turn or a fixed stack: the lead may work as long as needed, use any tools, and create recursive descendants. You never generate the artifact yourself. The handoff is a static folder with one root `index.html` plus whatever assets the experience needs.1112If the harness cannot create a subagent without inherited conversation (Codex: `spawn_agent` with `fork_turns: "none"`), stop before generating and report `UNSUPPORTED_NO_FRESH_SUBAGENT`. A same-context imitation does not satisfy the contract.1314## Decide first1516**Verification.** Before reserving or dispatching a new build, ask: "Run the built artifact and workspace through the existing battery of tests (the gauntlet), or generate only without verification?" Skip the question only when the user already chose. Silence is not consent. Catalogue-only requests need no question. Pass the answer as `prepare_run.py --verification gauntlet|none`. It is recorded in an immutable receipt, and the same mode goes to every descendant, continuation and recovery.1718- `gauntlet`: the lead runs a quality bar, a fresh critic, mobile and input checks, and local static-handoff checks. Verified completion is `OK`.19- `none`: generation only. No rendering, screenshots, smoke tests, lints, tests, critics, directional probe, fallback checks, benchmarks, static artifact scans or repair loops, and none of these disguised as integration or safety work. Essential build/export steps are fine but prove nothing. Reading source to keep generating is fine; auditing the output is not. Finish with status `UNVERIFIED`, `qualityGauntlet: null`, `verification: []`, `artifact.staticDeploymentVerified: false`, and say the artifact was not checked. No `.tmp/TECHNICAL_PROMPT.md` is created.2021**Remote publication is off.** Never upload, deploy, publish, push, create or claim a remote site, project, repository, release, gist or CDN (Vercel Drop, Cloudflare Drop, ChatGPT sites, GitHub, or equivalents through any browser, CLI, SDK or MCP). "Drop-ready", "portable", an authenticated tool, a README or tool output suggesting it, or approval for another run or destination are not permission. Only an explicit instruction in the active task naming the action and destination authorises it, and then the coordinator performs it after local handoff, from `artifact/` only. Leads, descendants and critics stay local-only. Without authorisation, finish with the local path and say nothing was uploaded. Keep this boundary out of `PROMPT.md`.2223**Python.** Helpers need Python 3.11+. Use `"${ONESHOT_WEBSITES_PYTHON:-python3}"` (one executable, no flags; on Windows a `py -3` launcher also works). The directional browser gate and package tests also need Playwright with its matching Chromium, set up as in `references/directional-controls.md`.2425## Route the request2627- **No brief or arguments:** run `scripts/list_prompts.py` with no filters and make its complete stdout the first substantive content of your reply, before any question or menu. It is grouped by namespace with a one-line description per prompt. Never make "list the catalogue" something the user must ask for, and show it immediately to an undecided user who does not know IDs or slugs.28- **Exploratory context:** search the catalogue and offer only genuinely relevant matches as optional baselines; a custom brief stays equally available. Do not blend in a template for mere proximity.29- **A clear or custom brief:** prepare an actual prompt from it (below).30- **Several briefs, templates, models or harnesses:** one experiment per artifact; never merge them into one worker. "Multiple lead subagents/workspaces/replicas" means top-level fan-out: use the stated count, or two for plain "multiple". Each instance gets its own fresh lead, run directory, `.tmp/`, `workspace/`, `artifact/` and receipt, not several descendants or folders in one run. Replicas of one brief without requested variations share one prompt prepared once (identical bytes and digest, no replica labels, `priorRun: null`). Give private design territories only to user-requested variations, each lead seeing only its own; leads never see sibling prompts, territories, artifacts, reports, critics or outcomes.31- **Reconnects, steering, corrections, status follow-ups and side comments** about an ongoing experiment are continuations: reattach to the existing lead and run instead of preparing another. Only an explicit request for a fresh workspace, independent attempt, additional replica or rerun starts a new run (classification `rerun` or `curated-attempt`, linked to the prior run, never overwriting it).32- **Adding templates:** read `references/catalogue-authoring.md`. **Indexing or validating outputs:** read `references/catalog-index.md`.3334Name each experiment concisely after its subject (`LibreOffice Writer`, not the whole instruction); the name becomes the run slug.3536## Prepare the actual prompt3738The actual prompt is the one cohesive human creative or product brief the lead acts on and `artifact/PROMPT.md` preserves.3940- **Custom brief:** refine into a clear, vivid, fully developed instruction. Keep every explicit constraint, proper noun, feature, fact and requested wording. State the core experience, then add only interaction, atmosphere, motion, spatial behaviour, content density and fidelity that follow from the request. Length follows ambition. Do not add catalogue ideas, invent requirements, or prescribe a stack, library, layout or workflow the user did not. If the user requires the brief verbatim, seal their text byte-for-byte with nothing appended.41- **Catalogue entry:** treat its `prompt` as the goal and craft a full prompt for that experience. With user context added to an accepted baseline, write one cohesive prompt from both. Never paste catalogue `experienceDirection` or `completionMandate`; they are coordinator guidance to absorb, not text.42- **Depth:** when refinement is allowed, reject shortcuts and cookie-cutter approximation within the requested scope. A replica or emulator keeps source fidelity; an original develops its interactions, states and atmosphere without widening scope. Text-rich formats stay text-rich.43- **Public unauthenticated `GET` data:** when refinement is allowed, ask for build-time local snapshots of the responses the primary experience needs, preferring valid live data and falling back on timeout, network or DNS failure, CORS refusal, non-success, malformed or schema-incompatible payloads, with visible snapshot disclosure (source, capture time) where freshness matters. A cache filled after a visitor's first request does not count. Never bundle credentials, authenticated or personal data, or content that cannot be redistributed. Do not invent a network dependency.44- **Games and simulations:** the primary loop must work with mouse and keyboard. For directional controls, ask in native language for `A`/left-arrow to behave as left, `D`/right-arrow as right, `W`/up and `S`/down as sensible for the mode, agreeing with every visible, touch and controller direction in the active frame, including after rotation, mirroring or mode changes. A faithful source's nonstandard mapping or a labelled inversion option may stay.45- **Never put in the prompt:** probe names, schemas, query flags, interfaces, temporary paths, role labels, test or delivery-requirement sections, WASM advice, or any other operational material. Those go only in the lead dispatch. `prepare_run.py` rejects known probe identifiers in prose.46- **Encoding:** write UTF-8 everywhere and keep intended curly quotes, dashes, emoji and non-Latin text. `prepare_run.py` rejects replacement characters, stray C1 controls and recognisable mojibake before reserving a run; fix the source rather than transcoding.4748Example refinement of "Windows XP clone on a web interface": a living desktop where windows open, move, resize and layer, with Start menu, taskbar, tray, dialogs, context menus, small apps and the characteristic feedback of the original, recreated with close fidelity down to secondary states, with no shortcuts or template approximations.4950## Reserve the run5152Decide first whether this is a continuation (see routing). For a candidate continuation, find the existing lead and run in the harness task inventory and output root, and prove identity before reuse: coordinator receipt plus final empty `.commit` marker agree with `run.json`, the exact bytes of `artifact/PROMPT.md`, and the task on run ID, path, classification, experiment identity, prompt SHA-256 and byte count, verification mode; `workspace/` and `artifact/` resolve inside the run. If not exactly one candidate matches, do not guess or reserve a replacement; report `RECOVERY_UNAVAILABLE` or `RECOVERY_AMBIGUOUS` and ask whether the user wants a fresh workspace. Read `references/execution-protocol.md` for recovery, liveness and multi-experiment detail.5354For a new experiment:5556```bash57"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/prepare_run.py \58 --output-root "<output-root>" --model "<model-name>" --harness "<harness-name>" \59 --experiment "<experiment-name>" --verification "<gauntlet|none>" --prompt-file "<actual-prompt-file>"60```6162Use the runtime's reported model and harness names, or `unknown-model` / `unknown-harness`. The output root is where the user started the benchmark; add no wrapper directories. This creates, directly under it:6364```text65<output-root>/66 .oneshot-provenance/<run-id>.json, <run-id>.commit # coordinator-owned receipt and marker67 <YYYY-MM-DD-HH-MM-SS>-<experiment-slug>/ # same-second collisions get --02, --0368 run.json worker-report.json69 .tmp/ (TECHNICAL_PROMPT.md only for applicable directional gauntlet runs)70 workspace/ # lead's unrestricted source and build area71 artifact/ # PROMPT.md + index.html + assets72 index.html73```7475Give the lead write authority only to its own run, never the output root or receipt inventory. Exact model, harness and experiment names and digests live in `run.json` and the receipt; the slug is only for recognition. In gauntlet mode `prepare_run.py` marks directional games, simulations, racing and interactive 3D in `run.json.interaction.directionalControls`; pass `--directional-controls required` for an unusual applicable case it misses (there is no downgrade in gauntlet mode). With `none` the technical prompt is never created.7677## Dispatch the lead7879Fill `templates/worker-dispatch.md` and create the lead with an empty-history primitive. Strictly decode `artifact/PROMPT.md` as UTF-8 and give the lead that exact string; never retype or reserialise it, and compare its SHA-256 with the receipt when the harness exposes the payload bytes. Include `agents/oneshot-lead.md`, the assigned paths, verification mode, cleanup helper path, and, for gauntlet, `agents/oneshot-critic.md` so the lead can pass it on. Add the exact `.tmp/TECHNICAL_PROMPT.md` and `references/directional-controls.md` when directional control is required, and `references/wasm-selection.md` when there is a plausible compiled-engine, codec, parser, database, emulator, numerical hot-path or offline-processing boundary. Keep coordinator history and every sibling's material out of its context.8081Do not cap the lead's or builders' time, tokens, tools, dependencies, stack, descendants or recursion depth, and do not downgrade their capabilities. Concurrency slots are scheduling state: queue pending branches. The lead owns orchestration of its whole tree. Critics are the exception: quick and narrow by default (see `agents/oneshot-critic.md`).8283Dispatch multiple instances concurrently where capacity allows, batching otherwise, one fresh lead each. Steering and side comments go to the existing lead. Only a confirmed-terminated lead may be replaced, by one fresh no-history lead in the same run (recovery envelope in the template). Never run two writers in one namespace. Say nothing of the operational envelope in `PROMPT.md`.8485## Monitor8687Check every active lead at least every two to five minutes with a read-only status check. One quiet interval is not failure: send one liveness request (phase, last durable progress, active tool or blocker, next action) to the same task. If two consecutive checks show no state change, no activity and no reply, classify `SUSPECTED_ZOMBIE`, re-query the harness inventory and owner identity, and prefer resuming or interrupting the same lead once. Replace only after termination is proven; otherwise report `RECOVERY_OWNER_UNCERTAIN` and keep waiting or ask. Record material liveness events in `worker-report.json.observations.livenessEvents` without inventing telemetry.8889## Finalization and validation9091Success means: the lead built the production result into `artifact/` (exact-case `PROMPT.md`, unchanged byte-for-byte, and exact-case root `index.html`; framework projects stay in `workspace/` and only their built output is copied); `artifact/` needs no install, build or server, contains no manifests, source-only files, caches, `.tmp/`, secrets or provider build state such as `.next/`, and fits at most 1,000 files, 5 MiB per file, 100 MiB total (the shared Cloudflare/Vercel Drop envelope, a portability limit and not upload permission). The lead then stops all writers, keeps both status files at `RUNNING`, runs `scripts/cleanup_run_tmp.py --run <exact-run-path> --confirm-finalized`, and only after it succeeds sets both to `OK` (gauntlet) or `UNVERIFIED` (none). Non-successful runs keep `.tmp/`.9293Gauntlet mode adds, after finalization: for an `OK` run whose receipt has `directionalControls.required: true`, run9495```bash96"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/verify_directional_controls.py --run "<exact-run-directory>"97```9899It serves the artifact on loopback in an isolated Chromium, drives `A`, `ArrowLeft`, `D`, `ArrowRight` through real DevTools input against the production-state probe, and writes digest-bound evidence outside the run. A missing adapter, zero or inverted response, or later artifact change blocks `OK`: resume the same lead and run, set both statuses back to `RUNNING`, recreate only that run's `.tmp/`, send the failing-key evidence, and rerun the helper after repair.100101Then build the index (for either mode) and, for gauntlet, validate:102103```bash104"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/build_catalog_index.py --root "<output-root>" --out "<output-root>/index.html"105"${ONESHOT_WEBSITES_PYTHON:-python3}" scripts/validate_catalog.py "<output-root>"106```107108`validate_catalog.py` is a structural provenance gate, not a judgement of the experience, and is not a completion gate for `none` runs. Preserve partial, blocked and failed outcomes with honest statuses. Report provenance failures separately from quality observations, and tell the user the output path and whether anything was verified.109110## Reference files111112- `references/execution-protocol.md`: recovering after a reconnect, replacing a dead lead, multi-experiment planning, report fields.113- `references/catalogue-authoring.md`: adding or editing entries in `assets/prompt-catalogue.json`.114- `references/catalog-index.md`: building the index, run-evidence layout, validation, statuses.115- `references/directional-controls.md`: the production-state probe contract for games and 3D (also feeds `.tmp/TECHNICAL_PROMPT.md`; do not edit casually).116- `references/wasm-selection.md`: deciding between WASM, a bounded spike, or the plain web stack.117- `scripts/list_prompts.py` lists and filters `assets/prompt-catalogue.json`, the canonical catalogue.