[thebrandonlucas]

Fixes DEV-2955

Problem

When a person or an agent hit a CLI bug, there was no short path to a useful GitHub issue. Reports had to be written by hand, with no build or platform, and could easily carry IDs or keys copied from a failing command.

Behavior

voltage feedback [--title TEXT] [--no-browser] [-- COMMAND...] prints a link to a prefilled new issue, and a link to search existing issues first.

Where it runs	Browser	Output	Report marked
Terminal (stdout and stdin interactive, human output)	Opens, unless --no-browser	› Search existing issues first: URL and › File the issue: URL (or Opened in your browser: URL)	Person
Off a terminal, --no-input, --json, or --output-file	Never	The report: title, labels, reporter, body, url, search_url, opened	Agent ([x] An AI agent wrote this report)

• The command after -- is reduced to voltage, its subcommand words, and flag names, so -- wallets get WALLET --profile prod --json becomes voltage wallets get --profile --json. No positional or flag value is kept.
• The body has What happened, Steps to reproduce (with placeholders), Environment (voltage --version output, OS and architecture, command), and Checks: a search-first box and the agent box. A hidden comment asks agents to search first, keep the agent box checked, and give steps a person can run.
• It needs no configuration or credentials and makes no request.

After a panic, the default panic message is followed by This is a bug in voltage. Report it: URL, with the source file:line and a scrubbed panic message. A dependency's absolute source path keeps only its file name, so the builder's home directory is not included.

Scrubbing and caps. Report text passes through scrub_line (control and bidi characters), and then any word that contains @ or ://, is 32 or more characters, or is 16 or more characters with a digit becomes [REDACTED] (the existing redaction marker). That covers UUIDs, API keys, tokens, invoices, addresses, emails, and URLs. The title is capped at 120 characters, the command at 300, and the panic message at 500. If the encoded URL is still more than 6,000 bytes, the body is trimmed until it fits.

Template. .github/ISSUE_TEMPLATE/bug_report.md gives manually filed reports the same sections and the bug label. Blank issues stay enabled, because the prefilled link uses title/body/labels parameters.

Design notes

• feedback is a new module (prefilled reports, scrubbing, and the panic hook). cli::outline walks the clap command tree without parsing, so it also works on a line that does not parse. startup::run installs the hook, and execute sends feedback to file_feedback before the configuration is read.
• Agent marking uses the same test as the rest of the CLI: no interactive terminal, or machine output, almost always means an agent or script. The checkbox is the reliable marker. GitHub applies labels= from a URL only for users with triage access, so the label is best effort.
• No new dependency (url and webbrowser are already used). output::REDACTED becomes pub(crate) so the scrubber reuses the marker.

Try it

From the repository root, on this branch, with .env in place:

alias voltage="$PWD/scripts/voltage-staging"
voltage --build
voltage login --account me --credential-store file

voltage feedback --no-browser --title Balance-looks-wrong -- wallets get 00000000-0000-4000-8000-000000000000 --account me --json

Expect two › lines: a search link, and a File the issue: link. Open the second link. The form shows the title, the body with your version and platform, Command: voltage wallets get --account --json (no UUID, no me), and an unchecked agent box.

voltage feedback --title Balance-looks-wrong -- wallets get 00000000-0000-4000-8000-000000000000 --json

Expect the browser to open the same form.

voltage feedback --json -- payments send lnbc1example --execute | cat
voltage feedback --no-input

The first prints JSON with "reporter": "agent", "opened": false, and a body with [x] An AI agent wrote this report and Command: voltage payments send --execute. The second, on a terminal, shows the report fields as a card, with reporter agent and no › links. No browser opens for either command.

The panic link can't be triggered on a normal build. The unit tests cover it, and I checked it by hand with a temporary panic! that echoed a wallet UUID. The link came out with src/startup.rs:LINE, the UUID as [REDACTED], and exit code 101.

Checks

• User-visible change: docs/guides/troubleshooting.md (new Report a problem), docs/guides/scripting.md, docs/README.md, README.md, docs/quality.md (module ownership), the regenerated docs/commands.md, and CHANGELOG.md (Unreleased).
• Branch updated with a merge of master (CHANGELOG conflict: kept both lines). The verification gate passes on the head: cargo fmt --check, cargo clippy --locked --all-targets -- -D warnings, cargo machete --with-metadata, cargo test --locked (209 unit, 151 integration), and python3 scripts/check-coverage.py.
• I reviewed the diff against master with voltage-pr-review. Tests cover the new paths: an empty title, no command, an agent versus a person, --no-browser, --, negative values, attached short-flag values, a line that does not parse, the length cap, and the trimming loop.
• CRAP report (scripts/crap-report.sh, cargo-llvm-cov 0.8.5, knots 1.16.0, on the head): 876 of 1089 measured, 0 violations. The highest score is 18.10 (api.rs wait_for_payment, baseline). New functions: cli::outline 6, feedback::push_word 6, startup::file_feedback 4.0 (now fully covered), and the rest 1 to 4. startup::execute is 16.
• The terminal branch of file_feedback is now covered by feedback_on_a_terminal_prints_the_search_and_issue_links (a487e01), which runs voltage feedback with stdout and stderr on a pseudo-terminal and checks the two › links, their order, and that the wallet ID is absent.
• Checks not run: an automated panic-hook test, because a process-wide hook in the test binary would affect other tests and there is no panic to trigger, so I checked it by hand as described above.

Stack

Stack #135, layer 7 of 9 (bottom targets master):

#	PR	Change
1	#97	Staging setup and a Try it section for each PR
2	#120	--all on organizations, environments, and profiles list
3	#128	Warn when a profile or flag overrides a set VOLTAGE_* variable
4	#129	voltage context --format env
5	#127	Replace hand-rolled terminal code with crossterm
6	#126	Pick the wallet, organization, or environment to show
7	#130 (this PR)	Prefilled GitHub issue links from voltage feedback and panics
8	#123	voltage environments create
9	#131	Replace flake.nix with a Kaifile

The layers are chained with merge commits, because history is not rewritten. Use squash and merge, bottom first.

[tee8z]

Only one finding. Scrubbing, the caps, and outline are sound and well tested. outline reads the command tree rather than a parse, so a line that does not parse still outlines, and no positional or flag value survives. The trim loop always ends, because the title is capped well under 6,000 bytes. The panic hook keeps a dependency's absolute path down to its file name.

Finding: file_feedback uses two tests where the description uses one. The report is marked as a person's only when out.on_terminal() && terminal.can_prompt(), but the links are shown whenever out.on_terminal(). Two cases come out wrong:

On a terminal	Description and Try it	What the code does
voltage feedback --no-input	Report fields, "reporter": "agent"	Prints the two › links (agent-marked body inside)
stdin piped, as echo | voltage feedback	—	A person sees the links, but the issue form opens with An AI agent wrote this report already checked

feedback_on_a_terminal_prints_the_search_and_issue_links asserts the second case as intended, since its stdin is not a terminal.

Fix: decide once (suggestion inline). Then the PTY test needs stdin on the terminal and --no-browser, and a test for --no-input on a terminal should show the report, not the links. On a terminal the report is rendered as the human card, not JSON, so the Try it line for voltage feedback --no-input should say "the report fields" rather than JSON.

Non-blocking: when the trim loop runs, it cuts the end of the body, which is where the Checks boxes are, including the agent marker the description calls "the reliable marker". Today only multibyte-heavy input reaches the loop. Trimming the problem text instead of the tail would always keep the boxes.

CRAP: not run on this head yet; I will add the result when it finishes. CI is green.

[thebrandonlucas]

Also took the non-blocking point in cfccaff. When a report is too long for a link, the loop now trims the text before ### Checks and keeps the checks whole, so the agent marker always survives. The loop always ends, because the capped title and the checks alone fit. a_body_too_long_for_a_link_is_trimmed_until_it_fits checks both reporters and that the body ends with the marker line. CRAP on cfccaff: 907 of 1130 functions measured, 0 violations.

[tee8z]

Reviewed c94b96e against #126 (49590cc). No blocking findings in this layer.

The report path runs without loading credentials/configuration; command outlines remove argument values, report text is scrubbed, and the encoded link is capped. Browser suppression and terminal versus agent reports have process/PTY coverage.

Local gate passed: formatting, Clippy with warnings denied, dependency hygiene, full tests (220 unit + 159 integration), and contract/docs generation checks. Required CRAP check passed: 908/1132 functions measured, 224 unmeasured, no scores >=20; highest 18.10 (wait_for_payment).

Limits: no real browser/GitHub-form interaction or end-to-end panic injection; inspected the panic hook and tested its report-building helpers. Approval applies to this incremental stack layer.