develop

Development Workflow

Universal standards for any code implementation work. The goal isn't compliance — it's predictably high-quality changes that future-you (or anyone else) can read, trust, and extend.

Decision Protocol — present options, don't decide alone

When you hit a fork — choosing between approaches, libraries, structures, scopes, or fixes — stop and surface three options to the human with reasoning for each, and a recommendation for the best one. The human decides; you execute.

What counts as a fork:

• Choosing between distinct implementation approaches (algorithm, data structure, architecture pattern)
• Picking a library or dependency when multiple reasonable options exist
• Deciding scope (fix the symptom vs. fix the root cause; refactor first vs. patch then clean up later)
• Naming or API design choices not already implied by repo conventions
• Trade-offs with meaningful long-term consequences

What doesn't count — just do it:

• Mechanical translation of clear requirements into code
• Following an existing repo convention (the repo already made that decision)
• Tiny stylistic choices implied by sibling code
• Obvious one-line fixes with no real alternative

Format when presenting options:

1. **Option A** — short description.
   Pros: ...
   Cons: ...

2. **Option B** — short description.
   Pros: ...
   Cons: ...

3. **Option C** — short description.
   Pros: ...
   Cons: ...

**Recommendation: B** — because [reason this is the best fit].

Then wait for the human's pick before proceeding. Don't half-commit by starting on the recommended option while waiting — that's the same as deciding unilaterally, just with extra steps.

If you genuinely can't think of three viable options, surface that — "I can only see two reasonable approaches here, A and B. Want me to pick a stretch third option, or is this a two-way choice?" Honest is better than padded.

1. Orient before writing

• Read CLAUDE.md if present. Repo conventions take precedence over personal preference. Always.
• Scan siblings of the file you're about to touch. Match existing patterns — naming, error handling, structure, test organization. Don't impose new conventions on a codebase mid-stream.
• Look at recent commits (git log --oneline -20) for the prevailing commit format and the kinds of changes that land. Pattern-match.

If the repo has no CLAUDE.md and conventions are unclear from sibling files, flag it — don't guess.

2. Plan before coding

• Read the requirements end-to-end before touching anything.
• If acceptance criteria are missing or ambiguous, stop. Ask, or report. Inventing requirements creates the wrong thing well.
• Stay scoped. Only change what the task requires. Unrelated cleanup goes in a separate change — note it, don't fold it in.

3. Implement

• Match the existing style. Imports, naming, formatting, error handling — copy what's already there.
• One logical change at a time. If a refactor enables the actual fix, commit the refactor separately, before the fix.
• Don't add dependencies casually. Each one is a maintenance and supply-chain surface that outlives the immediate convenience. Check the project's existing deps first — chances are something close already exists. Prefer well-maintained, widely-used packages over niche ones. And for tiny utility functions (a few lines), a little copying is better than a little dependency.
• Keep docs and comments in sync with the code. If your change makes an existing comment, README section, JSDoc/docstring parameter, or type signature wrong, update it in the same commit. Stale docs and comments actively mislead — they're worse than missing ones, because the next reader trusts them. For external docs (Notion, Confluence, design specs), flag what needs updating even if you can't change them yourself.
• No WebFetch. Reason from the repo. If you can't, planning missed something — go back to step 2.

4. Test

• Every change gets a test. Bug fixes get a regression test; features get coverage of the new behavior; refactors get tests that prove behavior didn't change.
• Use the repo's existing framework. Pest, PHPUnit, Jest, Vitest, pytest, Go test, RSpec — discover from the repo, don't pick your favorite.
• Run the full suite. Don't push until it's green. Fix failures (yours or pre-existing) before proceeding.
• Run the linter or formatter if the repo has one (eslint, ruff, php-cs-fixer, gofmt, prettier, etc.). Fix violations rather than disabling rules.

5. Commit

🔒 Commit approval gate — non-negotiable

Never run git commit without explicit human approval for that specific commit. Before any commit:

1. Show the diff that will be committed — the human reviews the actual change, not your summary of it.
2. Show the proposed commit message (formatted via the /workbench-dev-team:git-commit skill).
3. Wait for an explicit yes. General approval of the task, "looks good" about the code, or approval of a previous commit do not carry over. One approval covers one commit — for multi-commit work, present each commit (or an explicitly enumerated batch) for its own approval.

This is enforced at the harness level too: a plugin PreToolUse hook (hooks/scripts/commit-approval-gate.sh) forces a permission prompt on every git commit, regardless of permission mode. The prose protocol above is what makes that prompt meaningful — the human must already have the diff and message in front of them when it appears.

Sole exception — the autonomous Index pipeline. Scheduled Watson runs are headless; there, dispatching an item to the board is the approval, and Holmes review plus the human's PR merge is the gate. The hook recognizes the pipeline by a live /tmp/watson.lock or WORKBENCH_DEV_TEAM_PIPELINE=1. Never set that variable or create that lock to skip approval in interactive work.

Message format and hygiene

Use the /workbench-dev-team:git-commit skill for message format — Conventional Commits + Gitmoji. Beyond format:

• Don't commit secrets. Scan your diff for credentials, tokens, API keys, PII. Verify .env-shaped files are in .gitignore. If anything smells like a credential, stop and remove it before staging.
• Re-read your own diff before staging. git diff what you're about to commit — catch debug console.logs, commented-out code, TODOs you forgot to address, leftover scaffolding.
• Confirm you're on a feature branch (not main/master/trunk) before pushing.
• Atomic commits. One logical change each. Don't bundle.
• Never force-push to a shared branch.
• Never amend an already-pushed commit. If you need to change something, add a new commit.

6. Open a PR (when applicable)

When the work is for a tracked issue:

• Create the PR as a draft early — before implementation is complete. Visible work-in-progress is better than a black-box dump at the end.
• Use the repo's PR template when one exists. gh pr create --body silently bypasses templates, so discover and apply it yourself. Check, in order: .github/PULL_REQUEST_TEMPLATE.md, PULL_REQUEST_TEMPLATE.md (root), docs/PULL_REQUEST_TEMPLATE.md — any letter case — and .github/PULL_REQUEST_TEMPLATE/ (multiple templates; pick the one that fits the change, or the default). Fill its sections honestly — never leave boilerplate placeholders or HTML comments behind. If the template has no slot for something required below (issue link, acceptance criteria, test plan), append it after the template content. No template → use the structure in the next bullets.
• Use Fixes #<n> in the body for auto-linking.
• Mark ready and update the body when done — summary + acceptance criteria with completed boxes ticked + test plan.
• CI green is the real "done" line. Local-green isn't enough — CI runs checks your machine may skip (strict lint gates, integration suites, environment differences). The work isn't done until CI is green. For automated or unattended work especially, wait for CI to finish and fix any failures before handing the PR off for review — never pass a red PR downstream.

7. When stuck

• Tests fail and you can't fix them? Leave the branch in a clean state (committed, pushed, PR reflects current state). Report what you tried and what's failing.
• Need info you can't get from the repo? Don't WebFetch, don't invent. Report the gap and what would unblock you.
• At a fork without a clear recommendation? Apply the Decision Protocol — present three options to the human and let them choose. Reaching for the protocol is not a failure; deciding unilaterally is.