Contribute

Contribute

You've got a link. This skill takes it from there.

Orchestrates an end-to-end open-source contribution: from a bare issue or pull request link all the way to an opened PR, with no manual gates in between once the work itself checks out. It reads the issue like a maintainer would. Then it fixes it, tests it, and opens a PR you don't have to babysit. Delegates rendered issue/PR intake to browser intake (rendered read via /ego-browser preferred, browser subagent fallback per references/issue-intake.md) and deep analysis/review to /multi-agents. Nothing here is tied to one project, one host, or one language - the pipeline shape is fixed, everything else is detected fresh each run.

Purpose

Help you earn trust with a PR that is easy to merge.

Give it an issue link. It sends back a small fix with a test and proof. It pings you only when it must.

That's the bar. Small diff. Real proof. Kind tone.

Act with full ownership

A trusted human contributor doesn't pitch a plan and wait for a nod before opening their editor - they read the issue, understand the codebase, and get to work, looping in a maintainer only when something is genuinely ambiguous or high-stakes. Once steps 1-5 have given you that same footing, hold yourself to the same standard. The checkpoints later in this skill are the complete list of moments worth a pause - routine calls a competent contributor makes independently (which test file to extend, whether a helper deserves its own file, commit styling) are yours to make, not questions to surface. Put the reasoning in the commit and PR description, where it actually benefits reviewers. The same standard applies to everything posted publicly: see references/human-voice.md before writing any comment, issue text, or PR content. No public text gets pushed or posted until its Rule 6 self-check passes. Write in plain B2 English, casual and short, like a senior engineer who fixed one thing. Never in high-formal, buzzword-laden, em-dash-filled prose.

Maintainers are busy volunteers. They merge work they can read fast.

Keep the diff small. Show what you ran and that it passed. Credit prior tries. Keep it short so they can say yes fast.

When to Use

• User shares an issue or pull request URL from GitHub, GitLab, Gitea, etc., paired with intent to work on it ("start this issue", "let's fix", "pick up this ticket", "here's the next one")
• User mentions a local contributions/ directory or asks to begin an open-source task
• User requests opening a bug fix or feature PR against an external upstream repository
• User wants an end-to-end autonomous contributor pipeline that doesn't halt for routine micro-decisions

What "done" looks like from the user's side

Just this:

"Here's the next one: <issue-or-pr-url>. Repo's in contributions/."

That's a complete invocation. The skill infers the rest, checks whether the ground has already been covered by someone else, and - once it's confident the work is solid - pushes and opens the PR itself. It only interrupts you at the explicit checkpoints listed below.

Inputs

• Required: one URL to an issue or PR. Parse it generically - host, owner/org, repo name, number - rather than assuming GitHub. GitHub -> use gh for structured fields. GitLab -> use glab for structured fields. Anything else -> detect via browser intake (preferred) or web_fetch/web_search against the web UI. Host detection never changes the intake rule. Read the issue body and comments rendered in browser intake per references/issue-intake.md, on every host.
• Inferred when not stated:
  • Local repo path - check in this order: (a) a path the user just mentioned, (b) ~/contributions/<repo>, (c) ~/code/contributions/<repo>, (d) ./contributions/<repo> in the current workspace, (e) a case/dash/underscore-insensitive fuzzy match across those roots. Nothing found -> ask once whether to clone into ~/contributions/<repo> (default) or elsewhere. Don't silently clone - it's the one setup step worth a question.
  • Issue vs. PR - tell from the URL shape (/issues/ vs /pull/ or /merge_requests/) or from fetched metadata.
  • Agent roles for step 7 - default to the three-role brief below; only scale up for genuinely large issues (see "Scaling the agent brief").

External skills this depends on

Blocking-required - the pipeline stops without this:

• Browser intake (steps 3-4, blocking gate via /ego-browser preferred, browser subagent fallback) - the ONLY accepted way to read the issue/PR. Open the live rendered page as a maintainer would see it and capture: (a) title, body, labels, assignees, state, reactions/priority signals; (b) the FULL comment thread including collapsed/hidden comments, review threads, and the "linked pull requests" sidebar; (c) every visual asset - screenshots, attached images, video/GIF, pasted rich media - viewed at full resolution, with what each one actually shows written down; (d) every external URL in body/comments (repro repos, CodeSandbox/StackBlitz, live demos, docs, videos, sibling-repo issues/PRs) - followed and summarized; (e) cross-references into sibling repos in the same org. Follow the exact protocol in references/issue-intake.md and produce its intake artifact before touching code. gh issue view / glab / web_fetch supply structured fields (state, timeline, assignees) as a complement - they NEVER replace the rendered read, because linked-PR sidebars, reactions, images, and external repro content do not come through cleanly as text.
  • Attempt a rendered read first, every run: /ego-browser preferred, browser subagent / Playwright fallback per references/issue-intake.md. CLI-only is permitted only as a logged exception after a failed rendered attempt, with the intake artifact recording the attempted method, the exact error, and an explicit VISUALS: UNVERIFIED flag. Never infer VISUALS: NONE from CLI text alone.
  • If no browser tool can launch, or the page fails to load and no fallback renders it: STOP. Tell the user plainly what failed and what resolves it (e.g. install/enable the skill, log in for a private repo, paste the exact error), and wait. A gh-only read is not an acceptable substitute for this gate.
  • Opencode / generic agents: invoke it the way this environment loads skills (slash-command, Skill tool, or ego-browser nodejs heredoc per its SKILL.md). Adapt only the invocation format, never the content bar: full thread + visuals
    • external links, every run.

Required with graceful solo fallback:

• /multi-agents (steps 7 and 10) - runs the three-role analysis/review brief. See references/multi-agent-brief.md for the exact shape to hand it. If it isn't installed, or there's no subagent-spawning tool at all in this environment, don't let the pipeline stall waiting for it - read the brief yourself and work through the three roles in order (Root-Cause Analyst, then Codebase & Convention Researcher, then - held back until step 10 - Senior Maintainer Reviewer) inside your own reasoning instead. You lose the independence of a separate pass doing it this way, so be honest in the reviewer role rather than rubber-stamping your own work - reread that role's "OUTPUT EXPECTED" in the brief and hold your diff to it as if someone else had written it.

If the call shape either skill actually expects (flags, JSON, a fixed slot count) differs from what's described here, adapt the format to match its real interface but keep the content and roles intact. If you're ever unsure how one wants to be invoked, ask the user to paste its SKILL.md once rather than guessing.

Optional companions - reach for these when the situation calls for it, not as a checklist to run through every time:

• code-review - lends its correctness/security/maintainability/performance/testing checklist to the Senior Maintainer Reviewer role in step 10, so the review has real structure behind it instead of a vague "looks fine."
• jev-decisions (references/jev-decisions.md) - optional jev_review MCP gate for steps 4, 7, and 10. Scalar scores + confidence with previousEvaluation deltas instead of a chatty "looks good". Never writes code, never replaces tests. Skipped entirely when JEV_API_KEY is not configured.
• documentation - for a change that touches public APIs, config, or commands, keeps README/CHANGELOG/API docs in sync with the diff instead of leaving them stale.
• project-architecture - for a cross-cutting or large-feature issue, works out module boundaries before step 8 instead of improvising structure mid-implementation.

Workflow

1. Parse the link

Extract host, owner, repo, and number. Everything downstream reads from what's detected here - don't hardcode to whichever project you saw last.

2. Locate and sync the local clone

• Find the repo using the inference order above.
• git remote -v to see the actual remote setup (a fork workflow usually has origin = the user's fork, upstream = the real project). Add upstream from the parsed owner/repo if the fix needs to sync against it and it's missing.
• If origin isn't something you'll actually be able to push to later - it points at the upstream project itself rather than a fork, and you don't have write access - don't wait until step 11 to find that out. Fork it now (gh repo fork / glab repo fork, run with no arguments from inside the clone): both rename the existing origin to upstream and add the new fork as origin automatically, so there's somewhere to push when the time comes.
• Detect the real default branch (main, master, develop, ...) - don't assume main. git fetch the right remote, check out the default branch, pull.
• If there are uncommitted local changes, stop and ask rather than stashing - that work may matter to the user.

3. Read the issue or PR (BLOCKING - browser intake only)

Run this inside browser intake per references/issue-intake.md (/ego-browser preferred, browser subagent fallback). Do not start it via gh issue view alone, and do not treat the first screen as the whole read.

1. Open the exact URL the user shared in a dedicated browser task space (e.g. contribute-issue-<number>). Wait for load, capture snapshotText(), and scroll/expand until the full thread is visible - including collapsed, "show more", "load more", and resolved review threads.
2. Capture structured state as a complement (gh/glab: state, labels, assignees, timeline), but the rendered read is the record of truth. A maintainer's clarifying comment outranks the original issue body.
3. Inventory and OPEN every visual asset: screenshots, drag-dropped images, attachments, GIF/video. View each at full resolution (follow its link or screenshot it), and write down what it proves - error text, broken UI state, expected vs. actual - not just "has screenshot". An unread image is an incomplete read.
4. Inventory and FOLLOW every external URL in the body and comments - repro repo, branch, commit, CodeSandbox/StackBlitz/CodePen, live demo, docs page, video, log paste, sibling issue/PR. Open each in the same task space, note what it contains and whether it still loads. Mark dead links as dead; never silently ignore them.
5. Pull out the problem statement, repro steps, acceptance criteria, and how labels change scope (bug vs. feature vs. good-first-issue). Quote maintainer direction verbatim for the step 7 brief.
6. Write the intake artifact (title, thread summary, visual inventory, external-link inventory, related-work leads) exactly as references/issue-intake.md defines. Steps 4, 6, and 7 MUST NOT start until it exists. If the issue body is thin after all of this, say so explicitly - Agent A in step 7 reproduces from the visuals/repro link rather than the text.

4. Check for related work and duplicates (still inside browser intake)

Before writing anything, find out if this ground has already been covered, using the rendered context from step 3 - not title search alone:

• Look at the "linked pull requests" sidebar from step 3 for anything already tied to this issue.
• Search the issue/PR body and comments for references to other issues or PRs - including ones in sibling repos in the same org (a monorepo-adjacent project like sveltejs/svelte + sveltejs/kit cross-references constantly; don't assume the fix lives only in the repo the issue was filed against).
• Use browser intake to open any candidate PR(s) you find and read their actual status and review history, not just their title. Follow cross-repo references in the browser too - a fix that started in a sibling repo reads differently in rendered review comments than in a timeline API dump.
• Branch on what you find:
  • Already merged and it closes this issue - stop, tell the user it looks resolved already (with the link), and check whether the original issue should have been closed or whether some part of it is still open.
  • An open PR already addresses it well - don't open a competing one. Summarize its approach to the user and ask whether to review/improve that PR instead, or move to a different issue.
  • A stale or explicitly rejected prior attempt - use browser intake to read why it stalled or was rejected (maintainer review comments are the ground truth here), and carry that reasoning into the brief in step 7 so the new attempt doesn't repeat it.
  • Nothing found - proceed normally.
• Optional Jev gate: if a Jev key is configured, run Gate A from references/jev-decisions.md on the candidate-PR evidence. It advises only; the branches above still decide.

5. Learn the project's own conventions before touching code

Read whatever exists, in priority order: CONTRIBUTING.md / .github/CONTRIBUTING.md, CODE_OF_CONDUCT.md, .github/PULL_REQUEST_TEMPLATE.md, .github/ISSUE_TEMPLATE/, docs/DEVELOPMENT.md, the dev-setup section of the root README.md, and lint/format/commit configs (.eslintrc*, .editorconfig, rustfmt.toml, .pre-commit-config.yaml, commitlint.config.*). See references/ecosystem-detection.md for the language -> build/test/lint command lookup so you're not guessing npm test on a Rust repo. This step is never skippable, even for a one-line fix - it's what stops a good fix from getting bounced for a style nit instead of the actual code.

6. Create a working branch

Name it per the project's own stated convention if step 5 turned one up; otherwise fix/<issue-number>-<short-slug> or feat/<issue-number>-<short-slug>. Branch off the freshly-synced default branch from step 2, never off a stale local branch.

7. Delegate deep analysis to /multi-agents

Once steps 3-5 give real context (not before - a vague brief produces vague agent output), run /multi-agents with a three-role brief (or work the same brief solo, wearing each hat in turn - see "External skills this depends on" above - if it isn't available). The brief's Shared Context Block MUST include the step 3 intake artifact verbatim - thread quotes, visual-asset findings, and external-link findings - plus step 4 and step 5 context, so subagents never re-derive or silently drop what the browser already proved. Read references/multi-agent-brief.md before your first call - it has the full template, including where the step 4 findings go, and a worked example so the roles come out sharp instead of generic. In short, the default roles:

1. Root-Cause Analyst - reproduces the bug (or nails the exact functional gap for a feature ask) and traces it to the actual responsible lines. Diagnoses; doesn't fix yet.
2. Codebase & Convention Researcher - deep-searches the repo for the files the fix will touch, prior art (similar past PRs/commits), the conventions gathered in step 5, and which existing tests already cover this area.
3. Senior Maintainer Reviewer - held back until step 10; reviews a real diff, not the issue. Simulates a maintainer of this specific project, applying the same correctness/security/maintainability/performance/testing lens as the code-review skill (use it directly if it's available) against the actual stated standards from step 5, rather than generic best practice.

Hand each agent the issue/PR content from step 3, the related-work findings from step 4 (especially any rejected prior approach to avoid repeating), and the conventions from step 5 directly - they shouldn't have to re-fetch what's already gathered.

• Optional Jev gate: if a Jev key is configured, run Gate B from references/jev-decisions.md on the analysis. Implement only when it passes; otherwise loop the roles once, then escalate instead of coding on a guess.

8. Implement

Using Agent A's root cause and Agent B's research as the spec, make the change. Keep the diff as small as the issue actually requires - note unrelated cleanup opportunities separately instead of folding them in. Act like a senior earning trust, not a beginner proving range: one issue, one minimal fix, existing patterns reused. If this is your first contribution to this repo, prefer the smallest scope that truly closes the issue over a wider refactor, even when the refactor looks tempting.

9. Verify locally

Run the project's actual test/lint/build commands (from step 5 / references/ecosystem-detection.md), never assumed generic ones. Add or update a test for the change - a fix with no accompanying test is one of the most common reasons maintainers request changes. Re-run the full suite, not just the new test, to catch regressions.

10. Maintainer-style review

Send the real diff (not the plan) to the Senior Maintainer Reviewer role from step 7. Have it check, against this project's own conventions: correctness, test coverage, style/lint cleanliness, commit hygiene, whether it actually closes the issue as scoped, and anything a maintainer would flag (missing docs, breaking-change risk, unhandled edge cases). Every finding that points at code must cite a commit-pinned permalink per references/permalink-evidence.md, never a bare path:line on the default branch. Loop steps 8 -> 10 until it passes or you hit three rounds; if it still isn't converging, stop and bring the disagreement to the user instead of pushing something unresolved.

• Optional Jev gate: if a Jev key is configured, run Gate C from references/jev-decisions.md on the diff + test results BEFORE the reviewer loop. approve goes to the reviewer; revise fixes and re-asks once; escalate or low confidence stops. The reviewer role and the three-round cap still apply on every path.

11. Commit, push, and open the PR - automatically, but never past the Voice Gate

This step has a hard gate inside it. The push is automatic once the gate passes; it never skips the gate to stay automatic.

• 11.0 Read the voice rules first. Before drafting anything public, read references/human-voice.md in full and pull this repo's own last 5 to 6 merged PR titles and bodies (gh pr list --state merged --limit 10 or the browser-intake equivalent). Your title, body, and any comments must match their length, casing, and issue-link style. Generic templates lose to repo reality every time.
• 11.1 Draft small. Pick the template from references/pr-template.md by diff size: tiny (1 to 2 sentences, no headers), small (3 to 6 sentences, no headers unless the repo template forces them), medium (repo headers only). Never more bullet lines than changed lines. When the text points at code, use a commit-pinned permalink plus a short verbatim snippet per references/permalink-evidence.md. Write in B2 casual English with at least one contraction in anything over 2 sentences. Zero em dashes (—), zero en dashes (–) as pauses, zero bare -- pauses in prose. Backticked flags like --config are fine; English pauses are not.
• 11.2 Commit message per the project's own convention (Conventional Commits - fix: ... / feat: ... - by default, unless step 5 turned up something else). Keep the subject to one thing, short, same casing as merged PRs.
• 11.3 Run the Voice Gate self-check from references/human-voice.md Rule 6 literally (banned-word grep, banned-phrase grep, dash grep, plus the eye checklist: one-thing title, length matches diff, contraction present, proof line with real command and result, Closes #N once for PR descriptions (skipped for replies and comments unless closing an issue), no headers/checklists/emoji unless the repo template demands them). Rewrite until every check passes. A failing draft does not get pushed "and fixed later".
• 11.4 Confirm the git identity first. Check git config --get user.name / git config --get user.email in this clone. It must resolve to the actual person submitting the contribution, not a default identity a coding-agent install may have set globally. If it isn't clearly a real person's identity, stop and ask rather than guessing or silently overwriting it.
• **11.5 Never add a "Co-authored-by," "Generated by," or similar signature crediting an AI tool to the commit or PR body unless the project's own CONTRIBUTING.md or PR template explicitly asks for that disclosure. If it does ask, answer it honestly - don't leave the field blank or mark it false to get past a policy written specifically to catch that. Absent an explicit ask, the commit should just read as the submitting contributor's own work, because the review, testing, and judgment behind it were theirs.
• 11.6 Push and open the PR only after 11.0 to 11.3 pass. Once step 10 passes and the Voice Gate passes, push to the user's fork (the one confirmed or created back in step 2) and open the PR without waiting for a further go-ahead - that's what end-to-end automation means here. Draft the PR description from references/pr-template.md, filled in with what actually changed, how it was tested, Closes #<issue-number> (or whatever phrasing this project's own template used in step 5), and a mention of any stale/rejected prior attempt from step 4 if one existed. The same gate covers every follow-up push and every review reply or issue comment: re-run the self-check on the new text before posting. If the gate still fails after three rewrites, stop and bring the draft to the user instead of pushing it anyway.
• If gh/glab isn't authenticated, or the push fails for any other reason, say so plainly and tell the user what to run (gh auth login, etc.) - never report a PR as opened when it wasn't. This is the one place a fabricated link would actually mislead someone, so treat it as a hard rule rather than a judgment call.
• The two exceptions that still stop and ask instead of proceeding: step 4 found an existing PR that already resolves this (see step 4's branches), step 10 never converged after three rounds, or the Voice Gate in 11.3 still fails after three rewrites. A rejected voice draft never gets pushed to "fix later".

12. Hand off

Give CI a few minutes and check it once (gh pr checks --watch or the platform equivalent) before calling this done - not to babysit it through days of human review, but because a red build from something the local run in step 9 couldn't catch (a CI-only lint rule, an OS or version you don't have locally) is worth one honest look, and one follow-up commit if the fix is quick, rather than leaving the user to discover it later. If it's still red after that look, or just slow to start, don't loop on it - note the status in the summary and move on.

One short summary: what the issue was, what changed, what was tested, and the branch/PR link. Don't re-paste the whole diff - the user just watched it happen.

Constraints

These keep your PR mergeable. They guard review time, not just code.

• Do NOT proceed without rendered issue intake (references/issue-intake.md). Attempt a rendered read first, every run. A raw CLI dump is a complement, never a substitute.
• Do NOT bypass the Human Voice Gate (references/human-voice.md). Zero banned AI buzzwords, zero em dashes, zero en dashes as prose pauses, and zero speculative fluff.
• Do NOT interrupt the user for routine decisions. File naming, helper extraction, and test placement are your responsibility. Pause ONLY at the explicit checkpoints.
• Do NOT open competing PRs if step 4 finds an existing active or merged PR that resolves the problem.
• Do NOT cite a bare path:line on the default branch in public text (references/permalink-evidence.md). Pin it to a commit SHA and quote the lines verbatim, after verifying the SHA exists.
• Do NOT push unverified code. Always execute the project's actual build, lint, and test suites (references/ecosystem-detection.md) before pushing.
• Do NOT add AI disclosure signatures ("Co-authored-by: AI", "Generated by...") unless the target repository's CONTRIBUTING.md or PR template explicitly mandates it.
• Do NOT guess git author identity. Always verify git config user.name and user.email represent the authentic contributor before committing.
• Do NOT touch or stash uncommitted changes in the user's local clone without explicit confirmation.
• Done means: tests and lint pass locally, one new or updated test covers the fix, Voice Gate Rule 6 passes, PR links Closes #N once, CI gets one check in step 12.

Checkpoints (the only points that stop and ask)

Pre-flight safety gates:

• Uncommitted local changes found in step 2 -> ask, don't stash.
• No local clone found -> ask before cloning, and confirm the path.
• Git identity in step 11 isn't clearly the real submitter -> ask, don't guess.

Integrity gates (auto-retry 3x, ask only on failure):

• Browser intake gate fails (no browser tool, page won't render, auth wall, visuals or external repro link unreadable) -> stop and resolve before any code work. Never silently downgrade to a gh-only read to stay moving.
• Voice Gate in step 11 still fails after three rewrites -> stop and show the draft to the user instead of pushing it anyway.

Stop-work gates:

• Step 4 finds an existing merged/strong-open PR that already resolves the issue -> stop and tell the user instead of opening a competing one.
• Step 10's review doesn't converge after three rounds -> stop and flag instead of pushing anyway.

Everything else - reading files, running /multi-agents, running tests, iterating on the diff, committing, pushing, and opening the PR - proceeds straight through without asking, by design. The browser intake gate and the Voice Gate are part of "straight through": they run every time, they just don't ask, they complete (intake artifact exists / voice self-check passes) before moving on.

Scaling the agent brief

The three-role default fits most single-issue contributions. Adjust only when the issue genuinely calls for it:

• Trivial fix (typo, one-line logic bug, clear repro) - merge Analyst and Researcher into one role; go straight to a light review.
• Cross-cutting or multi-file feature - add a fourth Docs/Changelog Agent (or hand this to the documentation skill directly) so README/CHANGELOG/API-doc updates aren't an afterthought.
• Continuing someone else's PR (not a fresh issue) - swap Root-Cause Analyst for a PR State Analyst: what's already done, what review feedback is still unresolved.
• A stale/rejected prior attempt turned up in step 4 - give the Root-Cause Analyst that attempt's rejection reasoning up front so the new pass doesn't re-propose the same thing.

Three well-briefed agents beat five thin ones - only add roles the issue actually needs.

Works with any project, any language

Nothing above is specific to any one repo, host, or language. Step 1 detects the host, step 5 detects the project's conventions, and references/ecosystem-detection.md detects the toolchain. Step 4's sibling-repo awareness means it also isn't limited to looking in just the one repo the issue happened to be filed in. The pipeline shape is the only fixed part; everything else is discovered fresh on every run.

Examples

Example 1 - Straightforward fix, no checkpoints hit

"Here's the next one: https://github.com/expressjs/express/issues/1234. Repo's in contributions/."

1. Parses the URL; finds ~/contributions/express already cloned with origin = fork, upstream = expressjs/express; syncs main.
2. Runs the blocking intake in browser intake - full thread, linked-PR sidebar empty, one attached screenshot opened at full size confirming the null-route stack, no external repro link; writes the intake artifact. A maintainer comment points at the null check in lib/router/route.js; no prior attempts.
3. Checks CONTRIBUTING.md and the lint config, branches fix/1234-route-null-check.
4. Runs the three-role brief: Root-Cause Analyst lands on route.js:38, Researcher points at the existing test pattern in test/Route.js.
5. Implements the guard, adds a test, runs the suite and linter, sends the diff through the Reviewer role - approved on the first pass.
6. Commits, pushes to the fork, opens the PR from the template, watches CI turn green, and hands off with a three-line summary and the PR link.

Every decision here, from branch name to commit message, was already answered by the brief and the project's own conventions - nothing needed a check-in.

Example 2 - Duplicate work found, checkpoint hit

"Pick up https://gitlab.com/some-org/some-app/-/issues/88."

1. Locates and syncs the clone, opens the issue.
2. Step 4 turns up an open, approved MR (!142) that already closes it, just awaiting merge.
3. Stops: "Issue #88 already has an approved, unmerged MR (!142) that closes it - want me to review/improve that one instead, or move to a different issue?"

Opening a second PR here would hand the maintainer two competing fixes to reconcile - this is exactly what the checkpoint exists for, not a failure of nerve elsewhere in the pipeline.

Example 3 - Thin issue, /multi-agents unavailable, agent adapts

"start this issue: https://github.com/foo/cli-tool/issues/9 - contributions/"

1. The issue is one line - "crashes on --verbose" - no repro, no comments. The blocking intake still runs in browser intake: full thread scrolled to footer, VISUALS: NONE - verified rendered, external links NONE - recorded in the intake artifact as gaps.
2. /multi-agents isn't installed here. Rather than stalling, works the brief solo: reproduces the crash first (Analyst hat), traces it to an unguarded .split() on undefined output, then switches to the Researcher hat to find the existing flag-parsing tests.
3. Implements the fix and a regression test, runs the suite.
4. Reviews its own diff wearing the Reviewer hat as if it were someone else's PR, catches that --verbose --json together still crashes, and fixes that too before moving on.
5. Opens the PR, noting that the original issue was thin and the repro was reconstructed from the stack trace, so a maintainer can sanity-check the interpretation.

Adapting to what's actually available beats stalling to ask the user to install something they may not know they're missing.

Reference files

• references/issue-intake.md - BLOCKING intake protocol for steps 3-4: how to drive browser intake (task space, full-thread scroll, visual-asset handling, external-URL handling), the intake artifact shape, and the stop-conditions. Read before opening the issue URL. No intake artifact, no code.
• references/human-voice.md - BLOCKING for any public-facing text (issue comment, PR description, merge request comment, review reply, status update). Read it before drafting, match this repo's own merged PRs, write in B2 casual English, and run its Rule 6 self-check until it passes. Do not push or post while it fails. This is what keeps the contribution from being closed on sight as generated text.
• references/multi-agent-brief.md - exact brief template + worked example for step 7, including where step 4's related-work findings go.
• references/ecosystem-detection.md - language/build-tool detection and command lookup for steps 5 and 9.
• references/jev-decisions.md - optional Jev calibrated gates for steps 4, 7, 10: typed questions, thresholds, provider priority, and the decision-log learning loop.
• references/pr-template.md - PR description template for step 11.
• references/permalink-evidence.md - commit-pinned permalink protocol for every code citation in public text (steps 10-11): how to build the link, the snippet format, and the SHA check. Bare path:line drifts; permalinks do not.