How long does the integration take?

Change one base URL in an OpenAI-compatible client, or install the Python or TypeScript SDK; the quickstart routes a first request in a few minutes.

Will this lock us into another vendor?

No. The router is Apache-2.0 open source, your provider keys stay yours, and the API is OpenAI-compatible, so leaving is a change of base URL.

What happens when a provider fails?

The router falls back to the next model your policy allows and records the fallback in the trace, so an outage costs a retry, not an incident.

numen-pr-evidence - Prompt

Name: numen-pr-evidence
Author: THEjacob1000

Prompt file imported from THEjacob1000/dotfiles (.codex/prompts/numen-pr-evidence.md). Copyright stays with the author.

/pr-evidence

Capture and attach manual testing evidence to GitHub PR descriptions using isolated terminals and headless browsers. Use when asked to test a change manually, gather proof, capture a screenshot or recording for a PR, or add evidence to a PR. Covers real request and service-log captures, public R2 uploads, and embedding in the relevant PR section.

PR evidence

Evidence means proof that the change was manually tested: the new behaviour driven by hand against a running system and captured while it happens. A screenshot of a green test run, a coverage number, a CI badge or a lint pass is not evidence; the reviewer can see those on the PR already. If the capture does not show the changed behaviour actually happening, it is not done.

Prove the changed path, capture the proof raw, upload it, embed it. The evidence should let a reviewer verify the claim without trusting a narrated summary.

What to capture

Backend change: the real request against the public boundary plus the live logs of every changed service, as one terminal capture. If the change alters something a user sees, also capture that screen after the change; a backend PR often needs both.
Frontend change: the affected screen, before and after when the diff is visual. A recording (mp4 or gif) when the evidence is an interaction, a screenshot otherwise.
One image per claim in the PR. Cropped to what matters, large enough to read.

Terminal captures

Capture offscreen by default. Use an agent-owned PTY or a detached tmux server on a unique socket (tmux -L pr-evidence-<run>), never Jacob's existing server. Keep the request and each live service log in separate panes. Do not attach a client to his terminal, select his windows, activate applications, resize or zoom his terminal, send desktop keystrokes, or capture his monitor without an explicit request to use his display.

Save the real terminal output or recording, then render it offscreen to an image or video. For a static capture, capture-pane -p on the isolated server gives the terminal's actual text; render that verbatim with a monospace font in an agent-owned headless browser. Keep the raw transcript alongside the image. Identify it as a rendered terminal capture, not a native screenshot. Preserve literal commands, responses, wrapping and log output; no reconstructed output, invented prompts, styled reports or summary cards. Redact secrets before recording.

Exercise the public boundary when one exists, not only the downstream RPC or a unit test. Record method, URL, non-secret headers, body, status, response, and the matching method/error lines from every changed service; use a request id when the stack has one. Read the ticket first: an intentional stub error can be the correct result.

UI captures

Use an agent-owned headless browser/context for screenshots and recordings. Reuse the running dev stack and the repo's token/config tooling; don't launch a second copy of the app. Do not drive Jacob's active browser or fall back to his monitor when headless capture is unavailable; report the specific capture limitation. Inspect the saved file before uploading: a blank, truncated or wrong-state capture is worse than none.

Upload and embed

Save captures outside every repository (/tmp or the scratchpad). Worktrees stay clean of evidence assets; never add an image to the feature branch.
r2-upload /path/a.png /path/b.mp4 prints one public URL per file, in order.
Embed in the PR body's existing Screenshots / Videos or evidence section: ![settings page after the fix](URL), with the actual test scope and any simulation caveat. Before/after is two images on consecutive lines. A video URL goes on its own line as a plain link; GitHub does not inline external video. When attachment is authorized, finish by editing the description and reading it back to verify the embed, not just returning an upload link.
Editing a PR is a remote mutation and needs Jacob's approval in the current turn, same as gh pr create. The upload itself needs none.

Never upload

Everything in the bucket is world-readable with no expiry. Nothing showing a token, cookie, password, private key, customer data, or an internal hostname you would not paste into a public issue. Redact before capture, not after upload: an object is public the moment the command returns. Put a dev token in an env var and show the command referencing the variable; clear it from shell history and setup output before capturing.

Cleanup

Stop only the dev services started for this run, destroy only the agent-owned tmux server/browser, remove temporary logs containing tokens, verify the worktrees are unchanged. Never kill or reconfigure Jacob's shared tmux server. Report the tested route/result, the file paths, the URLs, whether the PR was edited, and any expected failure or environmental caveat.

Setup on a new machine

Only Jacob's credentials can write to the bucket. They live in ~/.config/r2/env, mode 0600, one copy per machine, never in dotfiles or the harness repo:

R2_ACCOUNT_ID=…
R2_BUCKET=jacob-pr-evidence
R2_PUBLIC_BASE=https://evidence.jacobdevelops.com
R2_ACCESS_KEY_ID=…
R2_SECRET_ACCESS_KEY=…

r2-upload --check verifies the file and lists the bucket without uploading. The script needs only bash and curl 7.76+ (--aws-sigv4), so it runs unchanged on linux and macOS; numen sync copies it into the dotfiles Stow tree, which exposes it at ~/.local/bin on each machine. A 403 on --check means the token was revoked or is scoped to another bucket: ask Jacob rather than trying other credentials.

Install numen-pr-evidence and it becomes one more candidate the router can pick - when it fits.

A listing is a routing target with a manifest: what it does, which domains it covers, what it costs and who publishes it. Once installed it sits beside your own models and tools, is scored like any other candidate for each request, and shows up in the trace when it wins. Ratings come from workspaces that installed it, one per account.

# /pr-evidence Capture and attach manual testing evidence to GitHub PR descriptions using isolated terminals and headless browsers. Use when asked to test a change manually, gather proof, capture a screenshot or recording for a PR, or add evidence to a PR. Covers real request and service-log captures, public R2 uploads, and embedding in the relevant PR section. # PR evidence Evidence means proof that the change was manually tested: the new behaviour driven by hand against a running system and captured while it happens. A screenshot of a green test run, a coverage number, a CI badge or a lint pass is not evidence; the reviewer can see those on the PR already. If the capture does not show the changed behaviour actually happening, it is not done. Prove the changed path, capture the proof raw, upload it, embed it. The evidence should let a reviewer verify the claim without trusting a narrated summary. ## What to capture - Backend change: the real request against the public boundary plus the live logs of every changed service, as one terminal capture. If the change alters something a user sees, also capture that screen after the change; a backend PR often needs both. - Frontend change: the affected screen, before and after when the diff is visual. A recording (mp4 or gif) when the evidence is an interaction, a screenshot otherwise. - One image per claim in the PR. Cropped to what matters, large enough to read. ## Terminal captures Capture offscreen by default. Use an agent-owned PTY or a detached tmux server on a unique socket (`tmux -L pr-evidence-<run>`), never Jacob's existing server. Keep the request and each live service log in separate panes. Do not attach a client to his terminal, select his windows, activate applications, resize or zoom his terminal, send desktop keystrokes, or capture his monitor without an explicit request to use his display. Save the real terminal output or recording, then render it offscreen to an image or video. For a static capture, `capture-pane -p` on the isolated server gives the terminal's actual text; render that verbatim with a monospace font in an agent-owned headless browser. Keep the raw transcript alongside the image. Identify it as a rendered terminal capture, not a native screenshot. Preserve literal commands, responses, wrapping and log output; no reconstructed output, invented prompts, styled reports or summary cards. Redact secrets before recording. Exercise the public boundary when one exists, not only the downstream RPC or a unit test. Record method, URL, non-secret headers, body, status, response, and the matching method/error lines from every changed service; use a request id when the stack has one. Read the ticket first: an intentional stub error can be the correct result. ## UI captures Use an agent-owned headless browser/context for screenshots and recordings. Reuse the running dev stack and the repo's token/config tooling; don't launch a second copy of the app. Do not drive Jacob's active browser or fall back to his monitor when headless capture is unavailable; report the specific capture limitation. Inspect the saved file before uploading: a blank, truncated or wrong-state capture is worse than none. ## Upload and embed 1. Save captures outside every repository (`/tmp` or the scratchpad). Worktrees stay clean of evidence assets; never add an image to the feature branch. 2. `r2-upload /path/a.png /path/b.mp4` prints one public URL per file, in order. 3. Embed in the PR body's existing Screenshots / Videos or evidence section: `![settings page after the fix](URL)`, with the actual test scope and any simulation caveat. Before/after is two images on consecutive lines. A video URL goes on its own line as a plain link; GitHub does not inline external video. When attachment is authorized, finish by editing the description and reading it back to verify the embed, not just returning an upload link. 4. Editing a PR is a remote mutation and needs Jacob's approval in the current turn, same as `gh pr create`. The upload itself needs none. ## Never upload Everything in the bucket is world-readable with no expiry. Nothing showing a token, cookie, password, private key, customer data, or an internal hostname you would not paste into a public issue. Redact before capture, not after upload: an object is public the moment the command returns. Put a dev token in an env var and show the command referencing the variable; clear it from shell history and setup output before capturing. ## Cleanup Stop only the dev services started for this run, destroy only the agent-owned tmux server/browser, remove temporary logs containing tokens, verify the worktrees are unchanged. Never kill or reconfigure Jacob's shared tmux server. Report the tested route/result, the file paths, the URLs, whether the PR was edited, and any expected failure or environmental caveat. ## Setup on a new machine Only Jacob's credentials can write to the bucket. They live in `~/.config/r2/env`, mode 0600, one copy per machine, never in dotfiles or the harness repo: ``` R2_ACCOUNT_ID=… R2_BUCKET=jacob-pr-evidence R2_PUBLIC_BASE=https://evidence.jacobdevelops.com R2_ACCESS_KEY_ID=… R2_SECRET_ACCESS_KEY=… ``` `r2-upload --check` verifies the file and lists the bucket without uploading. The script needs only bash and curl 7.76+ (`--aws-sigv4`), so it runs unchanged on linux and macOS; `numen sync` copies it into the dotfiles Stow tree, which exposes it at `~/.local/bin` on each machine. A 403 on `--check` means the token was revoked or is scoped to another bucket: ask Jacob rather than trying other credentials.

{ "prompt": "# /pr-evidence\n\nCapture and attach manual testing evidence to GitHub PR descriptions using isolated terminals and headless browsers. Use when asked to test a change manually, gather proof, capture a screenshot or recording for a PR, or add evidence to a PR. Covers real request and service-log captures, public R2 uploads, and embedding in the relevant PR section.\n\n# PR evidence\n\nEvidence means proof that the change was manually tested: the new behaviour driven by hand against a running system and captured while it happens. A screenshot of a green test run, a coverage number, a CI badge or a lint pass is not evidence; the reviewer can see those on the PR already. If the capture does not show the changed behaviour actually happening, it is not done.\n\nProve the changed path, capture the proof raw, upload it, embed it. The evidence should let a reviewer verify the claim without trusting a narrated summary.\n\n## What to capture\n- Backend change: the real request against the public boundary plus the live logs of every changed service, as one terminal capture. If the change alters something a user sees, also capture that screen after the change; a backend PR often needs both.\n- Frontend change: the affected screen, before and after when the diff is visual. A recording (mp4 or gif) when the evidence is an interaction, a screenshot otherwise.\n- One image per claim in the PR. Cropped to what matters, large enough to read.\n\n## Terminal captures\nCapture offscreen by default. Use an agent-owned PTY or a detached tmux server on a unique socket (`tmux -L pr-evidence-<run>`), never Jacob's existing server. Keep the request and each live service log in separate panes. Do not attach a client to his terminal, select his windows, activate applications, resize or zoom his terminal, send desktop keystrokes, or capture his monitor without an explicit request to use his display.\n\nSave the real terminal output or recording, then render it offscreen to an image or video. For a static capture, `capture-pane -p` on the isolated server gives the terminal's actual text; render that verbatim with a monospace font in an agent-owned headless browser. Keep the raw transcript alongside the image. Identify it as a rendered terminal capture, not a native screenshot. Preserve literal commands, responses, wrapping and log output; no reconstructed output, invented prompts, styled reports or summary cards. Redact secrets before recording.\n\nExercise the public boundary when one exists, not only the downstream RPC or a unit test. Record method, URL, non-secret headers, body, status, response, and the matching method/error lines from every changed service; use a request id when the stack has one. Read the ticket first: an intentional stub error can be the correct result.\n\n## UI captures\nUse an agent-owned headless browser/context for screenshots and recordings. Reuse the running dev stack and the repo's token/config tooling; don't launch a second copy of the app. Do not drive Jacob's active browser or fall back to his monitor when headless capture is unavailable; report the specific capture limitation. Inspect the saved file before uploading: a blank, truncated or wrong-state capture is worse than none.\n\n## Upload and embed\n1. Save captures outside every repository (`/tmp` or the scratchpad). Worktrees stay clean of evidence assets; never add an image to the feature branch.\n2. `r2-upload /path/a.png /path/b.mp4` prints one public URL per file, in order.\n3. Embed in the PR body's existing Screenshots / Videos or evidence section: `![settings page after the fix](URL)`, with the actual test scope and any simulation caveat. Before/after is two images on consecutive lines. A video URL goes on its own line as a plain link; GitHub does not inline external video. When attachment is authorized, finish by editing the description and reading it back to verify the embed, not just returning an upload link.\n4. Editing a PR is a remote mutation and needs Jacob's approval in the current turn, same as `gh pr create`. The upload itself needs none.\n\n## Never upload\nEverything in the bucket is world-readable with no expiry. Nothing showing a token, cookie, password, private key, customer data, or an internal hostname you would not paste into a public issue. Redact before capture, not after upload: an object is public the moment the command returns. Put a dev token in an env var and show the command referencing the variable; clear it from shell history and setup output before capturing.\n\n## Cleanup\nStop only the dev services started for this run, destroy only the agent-owned tmux server/browser, remove temporary logs containing tokens, verify the worktrees are unchanged. Never kill or reconfigure Jacob's shared tmux server. Report the tested route/result, the file paths, the URLs, whether the PR was edited, and any expected failure or environmental caveat.\n\n## Setup on a new machine\nOnly Jacob's credentials can write to the bucket. They live in `~/.config/r2/env`, mode 0600, one copy per machine, never in dotfiles or the harness repo:\n\n```\nR2_ACCOUNT_ID=…\nR2_BUCKET=jacob-pr-evidence\nR2_PUBLIC_BASE=https://evidence.jacobdevelops.com\nR2_ACCESS_KEY_ID=…\nR2_SECRET_ACCESS_KEY=…\n```\n\n`r2-upload --check` verifies the file and lists the bucket without uploading. The script needs only bash and curl 7.76+ (`--aws-sigv4`), so it runs unchanged on linux and macOS; `numen sync` copies it into the dotfiles Stow tree, which exposes it at `~/.local/bin` on each machine. A 403 on `--check` means the token was revoked or is scoped to another bucket: ask Jacob rather than trying other credentials.", "variables": [], "notes": "Capture and attach manual testing evidence to GitHub PR descriptions using isolated terminals and headless browsers. Use when asked to test a change manually, gather proof, capture a screenshot or recording for a PR, or add evidence to a PR. Covers real request and service-log captures, public R2 uploads, and embedding in the relevant PR section.", "metadata": { "source": { "provider": "github-codex-prompts", "repository": "https://github.com/THEjacob1000/dotfiles", "path": ".codex/prompts/numen-pr-evidence.md", "ref": "8cb1feccf64642f418d3c1670e2f7119c201bdee", "url": "https://github.com/THEjacob1000/dotfiles/blob/8cb1feccf64642f418d3c1670e2f7119c201bdee/.codex/prompts/numen-pr-evidence.md", "key": "THEjacob1000/dotfiles/.codex/prompts/numen-pr-evidence.md" } } }

numen-pr-evidence

About

/pr-evidence

PR evidence

What to capture

Terminal captures

UI captures

Upload and embed

Never upload

Cleanup

Setup on a new machine

Install numen-pr-evidence and it becomes one more candidate the router can pick - when it fits.

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Use it

Manifest

Reviews