| name | git-intent |
| description | Process commits as instructions — execute `!!` inline markers and `gi: <text>` commit-subject instructions, propagate user-committed choices through the codebase. Use when the user says "git-intent", "process commits", "apply intent", or pastes a `git log` containing `!!` markers or `gi:` subjects. |
Git is the instruction interface between human and agent. The human delivers intent through commit messages and inline !! markers in any file type. The agent reads the diff, executes the instructions, and commits the results. Every change is reviewable in git history.
Why
Chat is awkward for two things: instructions targeting many scattered points in a file (the user can't easily attach a comment at each), and reviewing large agent responses (no fast way to give feedback per section). Move both into a file and let git carry the loop — the user sees a diff, can revert individual hunks, and can leave !! markers right where a change should land. The agent should commit promptly so each step is a stable rollback point. If the agent's reply would be large, write it to todo/ or specs/, not the chat.
It's a REPL cycle: user instructs, sees the diff, judges, instructs again. Git adds what a plain REPL lacks — per-step rollback and a history you can return to. That history is the workflow's main artifact: an auditable record of decisions, their motivations, and the paths that were considered and rejected. The iterative history is preserved on archival (see /git-intent-squash-archive) precisely so the rejected paths aren't lost.
Instruction channels
Three channels carry user intent into a commit:
!! markers in any file — pinpoint instructions next to the target.- Commit message — explanation/context for the committed edit, optionally prefixed
gi: (git intent). If the message reads as an imperative ("rename foo to bar", "expand this section"), treat it as actionable.
- Direct edit — the user's edit itself, with no marker and no commit-message text, is also an instruction. The edit is the decision; the agent's job is to honor and propagate it.
N = $ARGUMENTS — when that is empty or not a positive integer, use 1.
Commit patterns
A commit message — with or without the gi: prefix — defaults to explanation of what the user already committed; the edit itself is the decision, and the agent honors and propagates it without re-executing the text. If the message reads as an imperative, treat it as actionable and apply it to the commit's files. When the role is unclear, ask. The pipeline is uniform: apply any !! markers, apply any imperative commit message, propagate the resulting decision to stale references elsewhere. Empty steps are no-ops.
The subject prefix is gi: (quiet — the commit is the unit, no surrounding noise to fight) while inline markers are !! (loud — they must stand out against surrounding code).
Markers
- Native single-line comment of the file, prefixed with
!!. Examples: # !! ... (Python/Nu/Shell/YAML), // !! ... (JS/Rust/Go/C), -- !! ... (SQL), /* !! ... */ (CSS), <!-- !! ... --> (HTML/Markdown).
- Place on the line above or beside the target code
- Removed after processing — remove the whole comment including any closing delimiters (
*/, -->), not just the !! token
- Comments without
!! are persistent context — leave alone
Reading user edits
- User's committed edit > prior markers — if the user edited code a marker targets, the edit is the decision; remove the marker without re-applying it
- Clean up broken surroundings — if the user's edit left a dangling sentence, stale list numbering, or broken syntax, fix it. Never re-add content the user removed
- Surface contradictions — when markers/edits conflict or scope is unclear, name the conflict in the commit body if you can resolve it, ask the user if you can't
Procedure
-
Clean-tree check — run git status --porcelain. If non-empty, stop and ask the user to commit or stash.
-
Get the diff — run git log -p -N --reverse where N = $ARGUMENTS. If the patch is very large (>500 lines), prefer git log -N --stat plus targeted git show per file to avoid filling context.
-
Read files in scope in full. Skip files you've already read or edited in this session, and skip binary files.
-
For each commit, run the pipeline:
- Apply markers — for each
!! marker added in the diff, read the instruction text, apply it to the target code/text, remove the whole marker comment (including any closing delimiters like */ or -->).
- Apply imperative commit message — if the message (with or without
gi:) reads as a command, apply it to the commit's files. If it reads as explanation, skip — the edit itself is the decision.
- Propagate the decision — scan each file in the commit scope for references that contradict the resulting state (stale branches, removed APIs, old behavior descriptions, obsolete rationale). If the decision names a symbol, path, API, or config key, also
grep the broader repo. Clean up broken surroundings left by user edits, per Reading user edits above.
Empty markers, empty global instruction, and no propagation needed are all valid no-ops — the pipeline runs unconditionally.
-
Commit each logical change — one commit per decision. When scope is unclear, name your interpretation in the commit body if you can resolve it; ask the user if you can't (per Reading user edits). Body is mandatory and follows this template:
<subject>
Decision: <one sentence>
Why: <non-obvious reasoning, 1–2 sentences>
Rejected: <alternative considered and dropped, with the reason — omit if none>
Propagated: <other places touched — omit this line if none>
- Record rejected paths. If the user considered an alternative and dropped it (in chat, via a
!! marker that was later removed, or by reverting a prior hunk), capture it as Rejected: <alt> — <why>. The dead end is part of the decision graph and would otherwise be lost.
- Don't list conflicts you looked for and didn't find. "Nothing else changed" is noise; absence is the default.
- Don't restate the diff. If the body is longer than the diff, the body is wrong unless the reasoning is genuinely complex.
- Skip binary files in the diff.
- A decision visible in git history is binding — do not override without explicit instruction.
If there is genuinely nothing to do (no markers, no propagation, no cleanup), report it and exit without creating an empty commit.
Related
/git-intent-squash-archive — when done iterating, squash the branch into one clean commit preserved as a git tag
