webgpt

Imported from grahama1970/agent-skills (skills/webgpt/SKILL.md). Install upstream with npx skills add grahama1970/agent-skills --skill webgpt. Copyright stays with the author.

Commands

All commands default to --background (no KDE switch, no window focus).

Hard tab-id preservation rule

When the human gives an explicit ChatGPT or Kimi tab id, that tab id is authoritative. The project agent does not select a provider. WebGPT resolves the live URL for that exact tab and dispatches by hostname:

Unsupported hosts fail with BLOCKED_WEBGPT_PROVIDER_UNSUPPORTED. Do not create, select, discover, or reuse any other tab.

Required behavior before submitting:

1. Resolve the live URL from the exact tab id without activation.
2. Infer the provider from that URL; --expect-url is an optional assertion.
3. Require a ready Browser Oracle binding for the inferred backend and exact tab.
4. Submit only with that exact tab id and resolved URL.
5. Do not pass --create-tab.
6. Do not rely on project binding alone or provider-specific state files.
7. If the wrapper cannot expose explicit tab-id submission, use the composed Surf transport directly:

skills/surf/run.sh webgpt.preflight \
  --tab-id <HUMAN_TAB_ID> \
  --expect-url "<EXPECTED_CHATGPT_CONVERSATION_URL>" \
  --no-activate \
  --json

skills/surf/run.sh webgpt.submit \
  --input bundle.md \
  --output response.md \
  --raw-output response.raw.md \
  --meta-output response.meta.json \
  --receipt-output response.receipt.json \
  --tab-id <HUMAN_TAB_ID> \
  --expect-url "<EXPECTED_CHATGPT_CONVERSATION_URL>" \
  --no-activate \
  --no-remember

# Kimi uses the same exact-tab rule through its provider transport.
skills/surf/run.sh kimi.submit \
  --input bundle.md \
  --output response.md \
  --meta-output response.meta.json \
  --tab-id <HUMAN_TAB_ID> \
  --no-activate

The proof metadata must show:

requested_tab_id == controlled_tab_id == <HUMAN_TAB_ID>
controlled_tab_id_mismatch == false
tab_was_created == false

If any of those checks fail, stop and report the routing failure. Do not resubmit through a path that can create a new tab.

Mandatory Browser Oracle and source-provenance preflight

For an explicit --tab-id, submit performs a non-mutating Surf inventory to resolve the live URL, then verifies the matching webgpt or webkimi Browser Oracle binding before repository validation or prompt injection. The binding's tab id and conversation URL must exactly match the live tab. WebGPT does not silently create, replace, or bypass a missing binding.

Every repository-backed mode (assess, plan, code, and all) also requires:

1. --repo-root points inside a Git repository with a network upstream.
2. At least one repeatable --source-path is supplied.
3. Every source path is repository-relative, tracked in HEAD, and clean.
4. The current HEAD is reachable from the freshly fetched upstream branch.
5. The upstream URL is HTTP(S) or SSH and can be represented as a credential-free clone URL for WebGPT.

Unrelated dirty paths are allowed. Declared source paths are not. On success, the CLI writes <bundle>.source-provenance.json and injects the repository URL, exact commit SHA, relative paths, and detached-checkout commands into the submitted prompt. WebGPT can then inspect the exact pushed revision rather than depending on workstation-only absolute paths.

python scripts/webgpt_cli.py submit review.md -p <project> \
  --repo-root /path/to/project \
  --source-path src/service.py \
  --source-path tests/test_service.py

--output-contract none is the explicit browser-only exemption from source provenance. It still requires Browser Oracle readiness and exact-tab identity. It does not imply any file-download path. If the submitted prompt asks for inline JSON, Markdown, or prose, diagnose failures as submit, capture, parser, or content-contract failures unless a separate download, --auto-download, or --require-attachment path was actually used.

Routing, content, and download proof boundaries

Do not collapse WebGPT proof layers. Exact-tab routing, sentinel capture, content parsing, deliverable validation, and Chrome file saving are separate claims.

1. Routing proof only. Metadata showing requested_tab_id == controlled_tab_id, controlled_tab_id_mismatch == false, and tab_was_created == false proves the intended browser tab was controlled. It does not prove that ChatGPT produced the requested JSON, diff, zip, image, or other artifact.
2. Transport proof only. A sentinel-bearing raw response proves WebGPT captured an assistant response from the controlled tab. It does not prove the response satisfies the selected output contract or the caller's schema.
3. Content proof. The caller must inspect the response file, raw file, meta JSON, and any validator output. Sentinel-only output, empty clean output, helper code, prose where JSON was requested, or text after a terminal marker is a content/parser failure even when routing proof is clean.
4. Download proof. A downloadable-artifact claim is proven only by a local downloaded file plus a format-specific sanity check such as checksum, unzip -l, JSON parse, image dimensions, or the requested verifier. Text saying that a file was created is not download proof.
5. Inline-output proof. For --output-contract none prompts that request inline JSON or prose, there is no Chrome download step. Do not blame ~/Downloads, Chrome save settings, or webgpt.download unless a download command was actually executed and expected to return a local file.
6. Conversation-limit rollover. If ChatGPT displays You've reached the maximum length for this conversation, but you can keep talking by starting a new chat., this is a conversation state, not a download or sentinel-parser defect. The Surf transport polls for this text, clicks the controlled tab's Start new chat button, and resubmits the same prepared prompt once. If same-tab rollover cannot be clicked, Surf opens a fresh ChatGPT tab and resubmits once. Check metadata fields conversation_max_length_detected and conversation_max_length_rollover before deciding the next action.
7. Too-many-requests cooldown. If ChatGPT displays a Too many requests modal saying requests are being made too quickly and access is temporarily limited, this is provider throttling, not a routing, download, sentinel, or parser defect. The Surf transport clicks Got it when possible, waits SURF_WEBGPT_RATE_LIMIT_WAIT_SECONDS (default 300), and retries the same prepared prompt once on the same controlled tab. Check chatgpt_too_many_requests_detected and chatgpt_rate_limit metadata. If proof_status is rate_limited, do not launch parallel WebGPT attempts; let the outer scheduler back off and requeue later.

When a project agent reports success or failure, it must name which layer was proved and which layer failed. Example: "routing and sentinel capture passed; JSON extraction failed because the clean response was one byte" is acceptable. "The skills are working" is incomplete unless the requested content/artifact contract was also verified.

Execution-gate and deliverable contract

WebGPT is an assessor, architect, code creator, or bounded reviewer — but only when the request names one current gate. Its output is not evidence that the gate passed, and an architecture response is not progress on an unresolved deployment, API, persistence, or UI defect.

WebGPT normally runs as a single bounded skill node. When the human explicitly authorizes architecture work, all composes three bounded submissions in order: assess, then plan, then code. Each stage keeps its own deliverable and stops fail-closed before the next stage when its contract is missing. Longer iteration remains a Tau DAG responsibility.

Select the per-submission deliverable with --output-contract:

A missing deliverable blocks with BLOCKED_WEBGPT_<MODE>_DELIVERABLE_MISSING (exit 4). Routing-proof failures exit 3. Single-stage assess/code use remains ungated. Direct plan and all invocations require explicit human authorization through --architecture-authorized.

For a code request, the bundle must state:

current_gate
one blocking defect
allowed files or module boundary
required live proof
stop condition
forbidden adjacent scope

Run submit with its default --output-contract code. That contract requires either a unified diff in the response or a non-empty finished-file zip. A roadmap, staged architecture, status analysis, or prose-only implementation plan fails as BLOCKED_WEBGPT_CODE_DELIVERABLE_MISSING.

Research contract

Research is a two-sided requirement:

1. Project agent (before calling WebGPT). The project agent runs /brave-search first, distills the findings, and embeds them in the bundle under a ## Research context section. WebGPT does not call /brave-search; that is the caller's pre-step.
2. WebGPT (during the answer). Every submission is prepended with a research directive instructing ChatGPT to use its own web search for current, authoritative sources and to cite the source URLs it relied on. The embedded ## Research context is a starting point, not a limit.

The directive and the per-mode output contract are injected automatically at submit time (text bundles are augmented in place as <bundle>.submitted-<mode>.md; zip bundles are attached unmodified).

Every non-none submission is also wrapped in a GOAL LOCK — a preamble at the very top and a reminder at the very bottom (last instruction wins) — that forbids WebGPT from drifting into easier, adjacent, or off-goal side quests and tells it to work only on the one stated gate. If it cannot make real progress on that gate, it must return the contract's block/ruling rather than solve an easier, unrelated problem.

After applying code, reconcile it against repository and live evidence. Review the current gate only and return exactly one ruling:

PASS_CURRENT_GATE
BLOCKED_CURRENT_GATE: <one concrete blocker>
REJECTED_SCOPE_EXPANSION

Do not request or create another architecture while a current gate has an unexecuted live proof, unless the human explicitly asks for a diagram of that gate. Do not credit fixture results, committed source, or WebGPT output as live deployment proof.

# One command: submit + wait + download (default --output-contract code)
python scripts/webgpt_cli.py submit bundle.md \
  --repo-root /path/to/project --source-path src/relative/path.py

# Ask WebGPT to diagnose where the project agent is stuck (no code)
python scripts/webgpt_cli.py submit bundle.md --output-contract assess

# Ask for a bounded architectural task plan
python scripts/webgpt_cli.py submit bundle.md --output-contract plan --architecture-authorized

# Human-authorized diagnose -> plan -> code composition on one exact tab
python scripts/webgpt_cli.py submit bundle.md --output-contract all \
  --architecture-authorized --tab-id 837358116 \
  --expect-url "https://chatgpt.com/c/..."

# Re-submit latest bundle (auto-finds creation-bundle*.md)
python scripts/webgpt_cli.py submit

# Activate tab (KDE switch, close duplicates, release CDP, clear drafts)
python scripts/webgpt_cli.py activate                  # background, no window steal
python scripts/webgpt_cli.py activate --no-background   # foreground (KDE switch)

# Download solution zip (finds button, clicks it, waits for file)
python scripts/webgpt_cli.py download

# Listen for WebGPT response
python scripts/webgpt_cli.py listen --timeout 300

# Project binding
python scripts/webgpt_cli.py config --tab-id 837356566 --url "https://chatgpt.com/..." --kde-desktop 2

Project-agent usage (unambiguous)

A project agent calls webgpt as a single bounded skill node: one submission, one deliverable contract. Do not loop webgpt yourself — iteration against an immutable goal (bounded rounds, drift/fail-closed, receipts, human goal changes) is a Tau DAG responsibility via a webgpt skill node.

Step 1 — declare pushed source. Pass the project repository and every relevant path. Paths must be repository-relative, committed, and pushed:

git push
python scripts/webgpt_cli.py submit bundle.md -p <project> \
  --repo-root "$PWD" --source-path src/service.py --source-path tests/test_service.py

Step 2 — pass the exact tab ID (never choose a provider). Browser Oracle must already contain a live binding for that tab under the inferred backend. --expect-url is optional and only asserts that the tab did not navigate:

# Provider-neutral: works for a bound ChatGPT or Kimi tab.
python scripts/webgpt_cli.py submit bundle.md --tab-id <HUMAN_TAB_ID>

# Optional URL race assertion.
python scripts/webgpt_cli.py submit bundle.md --tab-id <HUMAN_TAB_ID> \
  --expect-url "https://www.kimi.com/chat/<id>"

Step 3 — choose exactly one mode. Modes, the flag they need, the required deliverable, and where the response is written (relative to the bundle):

Step 4 — branch on the exit code / stderr marker. These are stable and machine-checkable:

Step 5 — verify routing proof. Always confirm the run hit the intended tab by reading <bundle>[-<mode>]-response.meta.json:

requested_tab_id == controlled_tab_id == <HUMAN_TAB_ID>
controlled_tab_id_mismatch == false
tab_was_created == false

Research (both sides). Before calling webgpt, the project agent runs /brave-search and embeds the distilled findings under a ## Research context heading in the bundle. webgpt separately injects a directive telling ChatGPT to run its own web search and cite source URLs. webgpt never calls /brave-search.

What webgpt does NOT do. No goal memory, no multi-round loop, no retry policy, no stall/mutation detection, no receipts ledger. Route all of that through a Tau tau.dag_contract.v1 with a webgpt skill node.

Failure reporting

When submit, download, or listen fail, the CLI automatically files a GitHub issue on anomalyco/agent-skills with label bug + webgpt. The issue includes:

• Command that failed and the error message
• Full stderr output
• Current tab list (all browser tabs)
• Project binding contents
• Surf CLI version
• Environment variables (DISPLAY, KDE session, etc.)

Failure filing is deduplicated: before creating a new issue, the CLI checks a local cooldown marker and searches for an already-open issue with the same title. Set WEBGPT_AUTO_FILE_ISSUES=0 to disable auto-filing for a run, or WEBGPT_ISSUE_COOLDOWN_SECONDS to tune the duplicate window. Failures still print stderr and local artifacts; no silent failures.

What's hidden

Browser Oracle binding

skills/browser-oracle/run.sh bind <name> --backend webgpt \
  --tab-id <id> --url "https://chatgpt.com/c/<id>" --manual

skills/browser-oracle/run.sh bind <name> --backend webkimi \
  --tab-id <id> --url "https://www.kimi.com/chat/<id>" --manual

Stored in ~/.pi/webgpt-projects/ or ~/.pi/webkimi-projects/.