Gpt Pro Relay

gpt-pro-relay

One prompt in, one response out. The gpt-pro wrapper composes the whole call: it generates the run-id, picks local-vs-SSH transport, submits, polls with backoff over flaky links, demuxes exit codes, and prints the response to stdout. The browser automation runs on macmini against a dedicated logged-in profile; a detached worker does the work so transport drops (or parent death) don't kill it — reconnect with gpt-pro --run-id <id>.

Quick start

The happy path, in execution order — the sections below are the authoritative detail; this is the shape:

1. Runtime — Pro runs take 5–20 min. Launch immediately; the invocation (direct, or a calling skill adding a gpt-pro lens) is the go-ahead. Never pause to confirm — mention the runtime only as passing context if useful.
2. Put the question in a file; attach local files with -f — GPT-Pro can't see anything local to you (codebase, shell, this conversation). Don't hand-cat files into the prompt: pass each one with -f <path> (repeatable; globs and @base-relative paths ok; whole dirs via --include-tree) and gpt-pro inlines them for you — validated, secret-scanned, and capped (5 MB composed) before submit, no quota burned on rejection. If ChatGPT converts that composed paste into a Pasted markdown attachment, the relay restores and verifies a short ordinary top-level execution instruction before Send. The file on stdin then holds just your question (plus any prior decisions only you know). GPT-Pro can search the public web, so for external facts include the grounding directive — see "Grounding external facts". (Hand-paste only when you need a surgical excerpt -f can't express.)
3. Run it as a Bash-tool call — mandatory on every call (a bare shell command is a trap: it blocks the turn and drops the envelope). Issue exactly:

   command:           gpt-pro < prompt.md
   run_in_background: true
   timeout:           7260000        # 121 min — see Background and timeout
   

   Response → stdout; run_id=<id> and a copy-paste recover_with=… line → stderr.
4. Wait for the completion notification — do NOT poll. A backgrounded call returns a task id immediately (expected); the harness notifies you when it exits. Don't re-run gpt-pro or re-read the task output in a loop meanwhile — the wrapper is already polling through any drops.
5. On failure, reattach (also backgrounded — recovery, not polling). If the call dies or exits 124/255, the macmini worker survives. Take the literal run-id from the task output (run_id=<id> / recover_with=… on the first stderr lines) and reattach in the same envelope: gpt-pro --run-id <id>. Then map the exit code — see "If it fails".

Invocation

Compose the prompt in a file (so $, backticks, and heredoc markers don't get mangled), then pipe it on stdin:

gpt-pro < prompt.md

That's the whole call. gpt-pro is on PATH (added in ~/.zshenv). The response is on stdout; the run-id, diagnostics, and the engine's JSONL are on stderr. Always issue it inside the Bash-tool envelope (run_in_background: true, timeout: 7260000) — see Background and timeout.

gpt-pro [-f <path>]... [--max-wait <sec>] [--dry-run] < prompt.md   # new run
gpt-pro --run-id <id> [--max-wait <sec>]                            # reattach / recover

Files are inlined through the shared filectx helper (sibling script): validated, secret-scanned (FILECTX_SECRETS=deny\|warn\|off), and capped — all before submit, so a bad file burns no quota.

Want a durable copy? Keep the streams separate: gpt-pro < prompt.md > answer.md 2> gpt-pro.log — never 2>&1 (it contaminates the answer and buries the run_id). Optional: the backgrounded task output already captures both streams, so a redirect is only for a persistent on-disk artifact.

If gpt-pro: command not found (a sandboxed or reset-env shell that didn't inherit the .zshenv PATH), run the identical command from this skill's own scripts/ directory — the gpt-pro script sits beside this SKILL.md (<this-skill-dir>/scripts/gpt-pro < prompt.md).

Prompts must be self-contained

GPT-Pro runs in a ChatGPT web tab. It cannot see anything local to you — your codebase, your shell, the files on your machine, or this conversation. Every byte of local context it needs must be inside the prompt: attach files with -f <path> (the wrapper inlines them for you, validated/secret-scanned/capped before submit) and paste any non-file context (a decision earlier in this session) into the stdin prompt. Use -f instead of hand-cat-ing — it eliminates the most error-prone step (a forgotten file, mangled $/backticks, a silently oversized prompt, an accidentally-pasted secret). It can, however, reach the public web through its own browser/search tool — so don't attach public docs it can fetch itself; tell it what to look up (see "Grounding external facts"). The rule: private-local context comes in via -f (or paste); public-external facts are searched for.

Err toward more context, not less — the 5 MB submission cap is generous, and a run that fails for missing context still burns 5–20 min. Put the question in a file (the Write tool sidesteps $/backtick/heredoc mangling) and attach files with -f:

gpt-pro -f src/foo.py -f src/bar.py -f 'tests/*.py' < question.md

question.md holds only the question (plus any non-file context); each -f is inlined under a fenced ## Included files section. Preview exactly what would be sent — resolved files and the composed byte total — without spending quota:

gpt-pro --dry-run -f src/foo.py --include-tree docs/ < question.md

You rarely need to assemble the prompt by hand now. If you must (a surgical excerpt -f can't express), build the whole prompt in a file with the Write tool and pipe it in with no -f.

Large composed prompts cross a ChatGPT frontend boundary: the browser turns the paste into a Pasted markdown file and leaves the ordinary message box empty. That used to work by accident, but the backend may now treat instructions inside the file as document content and merely ask what to do next. The relay handles this dynamically (no hard-coded size threshold): it detects the empty message box, adds a short top-level instruction to execute the attached prompt now, and verifies that instruction before Send. Do not work around this by dropping context or hand-splitting files. A pre-send instruction_boundary_lost_before_send means the verification failed and no quota was burned; a post-send instruction_boundary_lost means Pro still only acknowledged/planned, and its body is quarantined rather than returned.

Grounding external facts

Web research is one of gpt-pro's two first-class modes (alongside deep reasoning), not an afterthought — but it stays conditional. Local context is pasted in (above); public, external facts are searched for. When the answer turns on information not in the pasted context, GPT-Pro must search the web and ground its answer in cited sources. The directive below is conditional — a no-op on pure-reasoning tasks — so include it in prompt.md whenever the task could depend on external facts. Paste it verbatim:

## Grounding external facts

Some questions turn on facts not in the pasted context — current versions, releases, dates,
prices, API/library behavior, specs, benchmarks, CVEs, recent events, current best practices:
anything that changes over time or that you'd otherwise answer from memory. For any such fact
you MUST use web search and ground the answer in what you find — do not answer external or
time-sensitive questions from memory.

When you search:
- Corroborate each load-bearing fact across at least two independent sources (prefer
  primary/official). If sources conflict or only one exists, say so and mark it low-confidence.
- Cite the specific page URL you actually opened, inline next to the claim — never a homepage,
  a search-results page, or a URL reconstructed from memory. Didn't open it? Don't cite it.
- End with a `## Sources` list: every URL you opened, each with one line on what you took from it.

Do NOT search when the task is self-contained — pure reasoning, math, or analysis of the pasted
code/text — or to confirm stable facts you already know. If you didn't need external facts, say
so in one line ("No external sources needed — reasoning over the pasted material.") rather than
just omitting citations: that line tells the caller a deliberate skip from a missed one.

On return, verify grounding cheaply: a fact-dependent answer should carry inline URLs and a ## Sources list — spot-check one or two with the Read/WebFetch tools (does the page resolve, does it actually say that) before relying on it. A fact-heavy answer with no sources and no "no external sources needed" line is ungrounded — re-prompt or discount it.

Calibration block

For judgment tasks — analysis, recommendations, design, second opinions — have gpt-pro end with a reasons-based calibration block. Its slowness and cost make a self-graded score more tempting and no better calibrated, so ask for assumptions and failure modes, not a number. Skip it for deterministic lookups or mechanical transforms. Paste this into prompt.md (alongside the grounding directive when both apply):

## Calibration

End your response with this block exactly:
- Key assumptions: 1-3 the answer depends on ("none material" only if true)
- Most likely wrong because: the strongest failure mode, missing info, or counterargument
- Would change my conclusion: the specific fact, test, or counterexample that would flip it
- Best next check: the single source, test, or lookup that would most reduce uncertainty
- Verify before acting: specific current/high-stakes claims to check ("none" for pure reasoning)

No numeric %, probability, star rating, or High/Medium/Low label — this block is for routing and
verification, not a calibrated probability.

On return, treat it as the action plan: run the Best next check and verify the listed claims before relying on the answer. Because a re-query costs another 5–20 min, a targeted check is almost always cheaper than re-running. Discount any answer that self-asserts confidence in place of naming its assumptions.

Runtime note

Pro runs take 5–20 minutes per prompt. Invoking gpt-pro is itself the go-ahead — whether the user named it directly or a calling skill (prism, goal-loop) added a gpt-pro lens. Just launch; do not pause to confirm or wait for a "continue". Mention the ~5–20 min runtime in passing only if it's useful context, never as a gate.

Background and timeout

On a remote machine gpt-pro runs its own poll loop for up to --max-wait (default 120 min); on macmini it blocks on the engine directly (the engine's cap and the Bash-tool timeout bound it there). The default is deliberately generous — a queued or long Pro run killed by a tight timeout wastes the tokens it already spent, so favor completion. Either way, always wrap the Bash invocation in:

• run_in_background: true
• timeout: 7260000 — Bash-tool timeout in milliseconds = 121 min, 60 s of slack over the default --max-wait 7200 (= 120 min) so the tool doesn't preempt the wrapper's final diagnostic. If you change --max-wait, set this to at least (--max-wait + 60) × 1000.

Three clocks, inner → outer: the engine's 60-min per-run cap → the wrapper's --max-wait poll deadline (default 120 min, SSH path) → the Bash-tool timeout (121 min). A status: "timeout" (exit 3) is the engine cap firing, not --max-wait.

Wait for the completion notification. Do NOT poll from the agent side — the wrapper is already polling (the affirmative wait primitive is in Quick start step 4).

Recovery

gpt-pro prints run_id=<id> and a ready-to-run recover_with=gpt-pro --run-id <id> line to stderr at the start — both land in the background task's output file. Capture the literal id from there; a $RUN_ID shell variable does not survive into a later Bash-tool call. If the caller dies, or the wrapper exits 124 (timed out, still pending) or 255 (SSH transport unknown), the worker on macmini keeps running. Reattach with the same id, inside the same envelope (run_in_background: true, timeout: 7260000) — this is recovery, not the forbidden polling:

gpt-pro --run-id ask-20260611T002710Z-…    # the literal id, not "$RUN_ID"

This skips submit and waits for the existing run to complete (blocking fetch on macmini, poll loop over SSH). If the run already finished but you lost the output, read it off macmini directly — but read result.json FIRST: only status: "ok" makes response.md usable. Quote the remote command so ~ expands on macmini, not locally:

ssh macmini "jq -r '.status, .reason, .model_audit' ~/.gpt-pro/runs/<run_id>/result.json"   # gate — must print ok
ssh macmini "cat ~/.gpt-pro/runs/<run_id>/response.md"                                      # ONLY when status == ok

On status: "error" any extracted body is QUARANTINED — diagnostic only, never an answer. A rejected turn can be complete, fluent, and on-topic, so "the text looks fine" is not a reason to use it. Model-audit failures publish response.rejected.md; an attachment acknowledgement instead of completed work publishes response.incomplete.md; a failure before adjudication may leave response.pending.md. None is response.md, and all are void. Recovery is a fresh run (new quota → the user decides) or an honest failure report — never a cat.

Stopping a run

To interrupt a run you launched — you changed your mind, or a better answer arrived elsewhere — use --stop with the literal run-id (same envelope isn't needed; stop is quick and bounded):

gpt-pro --stop ask-20260611T002710Z-…    # the literal id, not "$RUN_ID"

It's graceful and worker-driven: the command writes a stop signal into the run_dir and the owning worker consumes it at its next phase gate. If the prompt hasn't been sent yet, the run is dequeued (no Pro quota spent). If it was already sent, the worker clicks ChatGPT's Stop button on the live turn to halt generation. Either way the run finalizes status: stopped and no response is returned (the partial is discarded, not published). All output is JSONL on stderr.

• stopped / already_finished / pending → exit 0. pending means the worker is alive and will consume the stop; poll fetch to confirm.
• no_live_worker → exit 2. No live worker consumed the signal (the worker process itself died, rare — workers survive SSH drops). Server-side generation may still be running; use the manual teardown path if you must halt it: ssh macmini gpt-pro-relay close-chrome --force is a blunt last resort (kills all tabs).
• not_found → exit 4. Unknown run-id.

Stop is not resubmit-safe cover for a mistake: a dequeued run spent no quota, but a run stopped after send already burned its reasoning up to the interrupt. Don't stop-then-resubmit reflexively.

Concurrency

Up to GPT_PRO_MAX_PARALLEL (default 6, clamped to a ceiling of 10) gpt-pro calls run in parallel — each worker gets its own tab in a single shared Chrome process. Beyond the cap, additional workers queue on a file-lock semaphore in ~/.gpt-pro/slots/ and wait for a slot to free up (the worker logs slot_queued, then slot_acquired when it gets in). A queued run can wait 15+ min before it even reaches sent (961 s observed), so total wall-clock = queue wait + the 5–20 min run — don't set --max-wait (or the Bash-tool timeout) shorter than that.

A backgrounded call with no exit code is not a failed call. Empty output + no completion notification = still running. Never diagnose it as lost, and never fresh-submit over it — that double-submits a live run and double-burns quota. Confirm liveness from the stage trace before concluding anything:

ssh macmini "tail -5 ~/.gpt-pro/runs/<run_id>/worker.stderr"

slot_queued with no slot_acquired = queued, waiting its turn. sent with no finished = generating. Either way: wait.

Chrome stays alive between runs (no per-call launch cost after the first) — an invoking agent never needs to tear it down. ssh macmini gpt-pro-relay close-chrome is operator maintenance only (it refuses by default if any worker is in flight; --force kills anyway) — don't run it as part of normal use or recovery. Raising GPT_PRO_MAX_PARALLEL above the default is a knob, not a free upgrade: parallel bursts on one ChatGPT Pro session are an account-side anti-abuse signal. If network.json starts showing 429s, captcha redirects, or unexplained needs_reauth after parallel use, drop it back to 1.

If it fails

The wrapper's exit code is the agent's decision key — every code maps to one action. Empty prompts, oversized prompts (>5 MB), and malformed run-ids are caught before any submission (exit 2 — no quota burned).

On a non-zero exit the engine's terminal stderr JSON carries a reason. Exit 1 is the catch-all — every engine error returns it (err() hardcodes exit_code: 1; only ok → 0 and the two timeouts → 3 differ), so for a failed run the reason, not the exit code, is the decision key.

The chip proves EFFORT; the served slug proves MODEL. Since the 2026-07 GPT-5.6 redesign the composer chip carries the reasoning-effort tier only — the model has no chip signal at all (no aria-label, no dataset key). A chip that read "Pro" is therefore not evidence the model was right; only the post-send served-slug audit is. The axes are gated separately, which is why a run can pass every pre-send check and still be rejected after completing.

The sent? column is the resubmit-safety key: pre-send = the failure happened before the prompt went in, no quota burned, a fresh run is safe. post-send / ambiguous = quota may be spent and a live run may exist — never auto-resubmit (it double-submits); surface to the user.

model_audit also appears in a successful result — these are the fail-open verdicts, not errors: verified (slug present and allowlisted — the normal case), model_ok_slug_missing (slug absent but the menu confirms Sol; model confirmed, effort unverified), unverified_missing_slug (slug absent and menu unreadable — a double selector break degrades rather than bricking the tool). The two FATAL verdicts (slug_mismatch, menu_mismatch) never reach a status: ok; they surface as the two mismatch reasons above.

Reasons you may see in worker.stderr's stage trace but never as a caller-visible failure — they are internal log lines or other subcommands, don't treat them as your decision key: chip_menu_open_failed, chip_menuitem_missing, closed_during_nav, browser_pid_not_found, shell_missing (an internal tab-recovery verdict), missing_prompt (the detached _run worker), run_already_claimed (a duplicate _run losing the run's claim to the worker that owns it — it exits touching nothing, and the owner's result.json is what your poll returns), workers_in_flight (close-chrome).

Run artifacts

run_dir lives on macmini at ~/.gpt-pro/runs/<run_id>/:

• prompt.md, meta.json, result.json
• the answer body — under exactly one name, and the name is the verdict. response.md only when result.json says status: "ok"; otherwise response.rejected.md (a model-audit reject), response.incomplete.md (the attached task was acknowledged rather than executed), response.partial.md (timed out — never passed the completion gate), or response.pending.md (the run died before any verdict, so the body was never adjudicated at all). A failure before extraction publishes none of them. Only response.md is ever an answer; the other four are diagnostics that may look plausible — that is precisely why they are not named response.md.
• pre-send.png, streaming-NNN.png, final.png, error-*.png
• final.html, network.json
• worker.stdout — detached worker's stdout (usually empty)
• worker.stderr — structured JSONL stage trace: one line per stage (start, slot_queued/slot_acquired, chrome_cdp_ready/chrome_connected/chrome_activated, logged_in, model_verified, prompt_typed, sent, extracted, finished, plus error / orphan_kill_* / *_skipped). When something fails mid-run, this is the fastest path to the failure point — the last stage before the error line tells you where it died.

Reach for them via ssh macmini cat <run_dir>/<file> or ssh macmini ls <run_dir> when diagnosing.

When gpt-pro-relay fits

Multi-turn

gpt-pro is one-shot per invocation — every call is a fresh ChatGPT conversation. To continue a thread, paste the prior response into the next prompt yourself. The dedicated profile retains login but does not persist conversation context across calls.