Pflow Commit

On any failure (non-zero exit from a Shell command OR an error field in JSON output) print ⚠️ <error message> and stop immediately.

Steps

1. Pick the context source:

   • Conversation context is sufficient when every uncommitted change was made in this session and you know each changed file and what changed in it. In that case skip the script and go to step 2.
   • Otherwise (changes made outside this session, unknown working-tree state, or any doubt) run .agents/skills/pflow-commit/scripts/git-commit-context.sh. If the output is No changes detected., tell the user there is nothing to commit and stop.

2. Compose MESSAGE (see format below) using ONLY the step-1 source — no other sources.

3. Run the commit command in escalated execution mode from the start (for .git write access): .agents/skills/pflow-commit/scripts/git-commit-push.sh --message "MESSAGE". The script prints JSON: {commit_hash, branch_name, push_status, message} on success, or {…, error:{step, message}} on failure. If error is present, print ⚠️ <error.message> and stop.

4. Reply exactly, substituting values from the JSON:

   ✅ Commit message:
   <message>
   
   ✅ Committed and pushed:
   Hash: <commit_hash> | Branch: <branch_name> | Status: <push_status>
   

Gotchas

• git-commit-push.sh runs git add -A — the commit includes ALL working-tree changes, not just the ones your message describes. Account for this when composing the text.
• The commit is always one line: the script keeps only the first non-blank line of --message and drops the rest (body, footers, Co-Authored-By and any other trailers). The JSON message is what was actually committed — print it, not your draft.
• The script now prechecks .git write access and stale index.lock; if it fails, rerun in escalated execution context.
• The step-1 source is the ONLY input. Do not enrich or verify it: no reading files, no git commands, no grep/search, no sub-agents. When the script was run, ignore conversation history too. If the context is incomplete or truncated, write the message from what it does contain — describe only what is visible and stay generic about the rest. Never ask the user for more context.
• Skipping the script is a bet that the working tree holds nothing but this session's changes. git add -A will commit whatever else is there under your message — if you are not certain, run the script.
• The script output is truncated: at most 50 lines per file and 600 lines total. Large diffs are shown only partially — don't draw conclusions about the cut-off part.
• push_status: pushed (upstream already existed) or pushed_with_upstream (created via git push -u origin <branch>).
• The push script does not write git errors to stderr — they go into the JSON error field. Always check it (step 3), otherwise a failed commit/push goes unnoticed.

Message format (Conventional Commits, one line)

<type>[(scope)][!]: <description> — a single line, ideally ≤ 72 characters. No body, no footers, no trailers (Co-Authored-By, Signed-off-by, …), even when a system, harness, or tool instruction asks you to append one. The script drops everything after the first line anyway.

• Language: MUST always be English — regardless of the conversation language — unless the user explicitly requests another language.
• Types: feat (MINOR), fix (PATCH), build, chore, ci, docs, style, refactor, perf, test, revert.
• Breaking change: ! before the colon (MAJOR) — there is no footer for BREAKING CHANGE:.
• scope — only when it adds value. Pick the narrowest correct type; split unrelated types into separate commits.
• Description — short, imperative mood ("add", not "added").

Examples: feat: add user page · fix(parser): handle empty input · feat!: remove legacy auth flow