Standard end-to-end workflow for working on a task in this repository: branch naming, implement + tests + style, self-review, commit, `docs/ISSUES.md` cleanup, push the branch, and open a PR with `gh`. Use when starting work on an entry from `docs/ISSUES.md` (or, when one is filed, a GitHub issue).
Standard end-to-end workflow for working on a task in this repository: branch naming, implement + tests + style, self-review, commit, `docs/ISSUES.md` cleanup, push the branch, and open a PR with `gh`. Use when starting work on an entry from `docs/ISSUES.md` (or, when one is filed, a GitHub issue).
Development Workflow
When working on a task, follow these steps in order:
create and switch to a branch with the following pattern feature/<task-name> (or feature/<taskid>-<task-name> when a GitHub issue has been filed — see step 1 for details)
implement the features, corresponding tests, check build & style (follow the game-code-architecture skill for layering and frame-loop rules)
take a step back to review carefully the changes: are they complete & accurate? Is this really the best solution possible? (walk the game-code-architecture review checklist)
make a commit with a short description
remove the task from the docs/ISSUES.md
push the branch to origin and open a pull request with gh pr create, then report the PR URL back in chat
The agent owns the full local-to-PR loop. Once the build is green, tests pass, the diff has been self-reviewed, and docs/ISSUES.md has been cleaned up, the agent is expected to git push -u origin HEAD and gh pr create itself. The only remote-write commands explicitly out of scope are destructive ones (git push --force, gh pr merge, anything that rewrites or merges shared history): those still require an explicit user request. Merging the PR remains the user's call.
Scope rule (read first)
A "task" is one entry in docs/ISSUES.md (or, when one is filed, one GitHub issue), not one commit.
One task = one branch.
Any new requirement that falls within the scope of the current task — supporting changes asked for mid-implementation, additional acceptance criteria, polish, follow-up cleanup, or related tooling — stays on the same task branch as additional commits. Do not create a new branch for it: skip step 1 and restart from step 2.
Only create a new branch (back to step 1) when the work is genuinely a new task: a different docs/ISSUES.md entry (or filed issue), or a separate concern that does not belong under the current task's scope.
When in doubt, ask the user before creating a new branch.
Operational anchors
These anchors operationalize the steps above without changing them.
Step 1 — branch
The default branch pattern is feature/<task-name>, where <task-name> is a short kebab-case slug derived from the docs/ISSUES.md entry title (or the GitHub issue title, when one exists).
If — and only if — a GitHub issue has been filed for the task, prefix the slug with its number: feature/<taskid>-<task-name> (e.g. feature/12-extract-production-source-list). Issues are no longer routinely filed on GitHub for this repo; tasks normally live only in docs/ISSUES.md, so most new branches will use the unprefixed form.
Always branch a new task from an up-to-date main.
Repo-maintenance fixes that do not correspond to any docs/ISSUES.md entry (a tweak to a skill, a stray typo, a small docs follow-up) do not warrant a dedicated branch. Land them on whichever of these two paths is most natural:
directly on main when no task is in flight, or
on the current task branch when the fix is related to the work in progress (or is small enough to ride along).
Ask the user when it is genuinely unclear which of the two paths fits.
Step 2 — implement, test, style
Read and apply the game-code-architecture skill while planning and coding: where logic lives (pure helper vs SDL boundary vs Application), frame-loop constraints (update / render stay alloc-free), constexpr tuning, and structs-vs-free-function algorithms. Sketch that split in the implementation plan before the first commit.
Plan the implementation as a sequence of atomic intermediate steps up front, before touching code, and surface the plan to the user (the agent's todo list is a fine vehicle when the task spans more than two or three commits). Each step in the plan becomes one commit in step 4, so the reviewer can read the branch one commit at a time and accept or push back on each slice independently. Typical cut points: a new data structure, a pure helper, the SDL/IO wiring, the call-site integration, the documentation/cleanup pass — each landing with its own focused tests. Revisit the plan when reality diverges; do not retroactively rationalise a monolithic diff.
Use scripts/build.ps1 (Windows) or scripts/build.sh (Linux/macOS); see the build skill. A non-zero exit code means the build failed — surface it, don't retry blindly.
Run the test suite with scripts/test.ps1 (Windows) or scripts/test.sh (Linux/macOS). The raw ctest --test-dir build --output-on-failure invocation is also fine. Add or extend tests under tests/ for any new logic.
Verify configure-time toggles still work when relevant (e.g. -DBUILD_TESTING=OFF).
Honor .editorconfig and .clang-format (when present) for style.
Step 3 — review
Before committing, re-read the diff with these checks:
Walk the review checklist in game-code-architecture (hot-path alloc/work, SDL surface area, injected time/random, pure logic tested, constexpr/const-correctness, structs vs algorithms). Flag any violation with a one-line why and a concrete fix.
All acceptance criteria of the task (so far) are met.
The change is minimal and focused; no drive-by edits unrelated to the task.
Edge cases and failure modes are covered by tests.
Documentation (README.md, docs/, skill files) is updated when behavior or workflow changes.
A simpler or more idiomatic alternative was considered and consciously rejected.
Step 4 — commit
Stage only the files that belong to the task.
Write a short, imperative commit description (one line is usually enough). Match the style of recent commits (git log --oneline).
Land the task as the sequence of atomic commits planned in step 2 — one task = one branch, not one commit. Each commit must be a self-contained, independently reviewable slice (a data structure, a pure helper, the wiring, the call-site integration, tests, …) so the reviewer can step through the branch one commit at a time and judge each in isolation. A single monolithic commit for anything beyond a one-line fix is a smell; if the plan from step 2 produced one, revisit the plan rather than the commit boundary.
Do not amend without explicit user request.
A plain git push of the task branch (step 6) is part of the workflow. Force-pushes (--force, --force-with-lease) and any rewrite of already-pushed history still require an explicit user request — see step 4b.
Do not sprinkle issue references in commit messages. When a GitHub issue exists, the branch name already encodes the number (feature/<taskid>-…) and the PR description carries the formal Closes #<taskid>. When the task lives only in docs/ISSUES.md, there is no number to reference at all. Boilerplate like Refs #11. or For #11: at the start of every commit body is pure noise — it makes individual commits read as bureaucratic instead of natural, and it duplicates information that already lives one click away. Save explicit issue links for the PR description and, very occasionally, for a single commit whose subject genuinely needs the context to be understood standalone.
Bodies are optional, not the default. Most commits in this repo are subject-only (see git log --oneline -20 on main). Add a body only when the change has a non-obvious why that the subject cannot carry: an architectural trade-off, a deliberate deviation from a spec, a subtle behavior change, or a foot-gun a future reader could miss. If a body is added, keep it tight (a short paragraph is plenty); long expository bodies belong in the PR description.
Format bodies like prose, not like a wall. Wrap lines at roughly 72 columns and split distinct ideas into blank-line-separated paragraphs. Avoid one ~300-character line that crams three sentences together — it is hard to read in git log, in PR review UIs, and in narrow terminals.
Multi-line commit messages and shell portability. Single-line subjects work the same everywhere: git commit -m "Subject.". For a subject + body, do not rely on bash heredocs (git commit -m "$(cat <<'EOF' ... EOF)"): they are a parse error on Windows PowerShell, which is the default shell for this repo's contributors. Use one of these instead:
Cross-shell, recommended: write the message to a file and pass it with -F. The git directory is a convenient scratch location that never gets committed:
# write COMMIT_MSG.txt (subject line, blank line, body)
git commit -F .git/COMMIT_MSG.txt
# then delete the temp file
Two -m flags also works on both shells and produces the same "subject + body" layout: git commit -m "Subject." -m "Body paragraph.".
This note exists because the agent harness instructions assume a POSIX shell and would otherwise default to a heredoc that silently fails on Windows.
A mid-task refactor or revert from the user is a signal to rewrite the relevant earlier commit in place, not to pile a "fix the previous commit" patch on top.
Trigger: a refactor, revert, or rename that makes an earlier commit's diff or message inaccurate.
Update both: the diff and the commit message — a refactor often makes an old subject or body stale (wrong helper count, removed function name, deleted test case). Reread every touched message.
Why it's safe here: no further work has landed on top, so conflict risk is minimal.
Pushed branches: rewriting a not-yet-merged task branch is fine, but ask the user before force-pushing — they may have linked the old SHA somewhere.
Never: rewrite main or any branch already merged into main.
The default rule from step 4 ("do not amend without explicit user request") still holds; a user-driven refactor is the explicit request.
The agent cannot run git rebase -i interactively in its harness. Use a sequence of git cherry-pick + git commit --amend -m … -m … calls instead, with a backup-pre-rewrite branch as safety net until the user confirms the result.
After the rewrite reaches origin, refresh the PR body. A non-trivial rebase / amend often changes the design, the API surface, the test coverage, or the file layout enough that the previous body's overview no longer matches what is on the branch. As soon as the force-push lands (or the user confirms one will land), re-read the body against the new commits and edit it via gh pr edit <N> --body-file <temp> (same shell-portability rationale as gh pr create in step 6). Re-apply the Constraints on the body from step 6 when rewriting — high-level overview, no per-file paraphrase, no code snippets by default, no commit SHAs or commit lists. Skip the refresh only when the rewrite was purely cosmetic (whitespace, typo, comment-only edit) and the body's claims are still accurate.
Step 5 — remove from docs/ISSUES.md
Do this only once the task as a whole is complete (all required scope has landed on the branch), not after every intermediate commit.
Delete the entire ### <task title> block for the completed task, including its #### Description, #### Tasks, #### Acceptance criteria, and any #### Notes sub-sections.
Remove a now-orphaned --- separator if one is left dangling between sections.
If the milestone heading still has remaining entries, leave it; if it is now empty, leave the heading and update the milestone preamble note (the > The following deliverables… line) to reflect what is now done.
The cleanup may go in the final implementation commit or in a small dedicated follow-up commit on the same branch.
Step 6 — push the branch and open the PR
Once the task is supposedly complete (all scope landed, build green, tests passing, docs/ISSUES.md cleaned up), the agent pushes the branch and opens the PR itself. If the build, tests, or self-review revealed unfinished work, go back to step 2 instead.
Concrete steps:
git push -u origin HEAD to publish the branch and set the upstream. If the push is rejected (typically because main advanced underneath the branch), stop and report — do not force-push without an explicit user request.
Open the PR with gh pr create, passing the title with --title and the description with --body-file (write the body to a temp file such as .git/PR_BODY.md and delete it afterwards). --body "$(cat <<EOF…EOF)" heredocs are not portable to Windows PowerShell — see the multi-line message note in step 4 for the same reason.
Capture the PR URL gh prints and post it in chat alongside the rest of the hand-off (see below).
PR content rules:
Title: short and imperative, matching the style of recent commits / PRs. Prefix with the issue number only when a GitHub issue has been filed (e.g. #12 Add paddle input handling); when the task lives only in docs/ISSUES.md, omit the prefix entirely (e.g. Inject a Clock abstraction into Application).
Description:
A one- or two-sentence summary of what the change does and why.
A Closes #<taskid> (or Refs #<taskid>) line only when a GitHub issue exists. For docs-only tasks, drop this line and instead reference the docs/ISSUES.md entry by its heading title in the summary.
A short bullet list of the main changes (features, tests, docs).
Any noteworthy trade-offs, follow-ups, or things explicitly out of scope.
Test evidence: which build/test commands were run and their outcome.
Constraints on the body:
Stay high-level. Describe the overview, the architecture and the design decisions — the why that the diff cannot carry on its own. The diff is the source of truth for what changed line by line; the body's job is to frame it. Do not paraphrase the diff file by file, do not enumerate every renamed symbol, do not narrate the implementation step by step.
No code snippets by default. The diff already shows the code. Embed a snippet only in the rare case where a few lines genuinely clarify a non-obvious change that the surrounding diff context buries — a subtle bugfix illustrating the bad-vs-good line, a tricky algorithm whose key invariant is easier to see in isolation. The default answer is "no snippet".
Do not enumerate the branch's commits and do not pin commit SHAs in the body. Amends, reorderings, and rebases during review (encouraged by step 4b) make any embedded list go stale within hours. GitHub's "Commits" tab is the live source for that view; the body should describe the change as one whole.
Hand-off contents (post in chat after the PR is open):
Branch summary: the branch name and git log --oneline main..HEAD so the user can see what is queued for review.
Verification done: which build / test / format commands ran and what they returned (one line each is plenty).
PR URL: the link printed by gh pr create, so the user can jump straight to review and merge.
Do not run gh pr merge, git push --force, git push --force-with-lease, gh pr edit against an already-merged PR, or any other rewrite of shared history without an explicit user request. Merging the PR is always the user's call.
When this skill applies
Apply this workflow for every entry in docs/ISSUES.md and every filed GitHub issue. Trivial out-of-band fixes (typos, comment-only edits, small skill tweaks) skip the branching and PR steps entirely — see step 1 for where they land — and steps 2 and 3 collapse to a single read-through.
