branch-commit

Output contract, in brief. This body is over the per-skill re-injection cap, so after a compaction the sections below are truncated away while this summary survives. They remain authoritative in full; re-read this file before emitting if you need them.

• Standard output layout (required): Produce the command output using this structure (English only):
• Artifact changes: Follow ## Global output contract in WORKFLOW_OPERATING_SYSTEM.md for APPLIED / PROPOSED / SKIP rules. Every listed file...
• Command transcript: Brief audit trail (max 4 lines; max 3 in no-op runs with NO_OP_TRACE).
• Handoff: Use the adaptive ending format of WORKFLOW_OPERATING_SYSTEM.md ## Global output contract. Every Handoff is one fenced text b...
• Definition of done (command output): Output names the exact diff source used (git diff, git diff --staged, or git diff <base>...HEAD); paraphrasing from the task...

Act as a concise engineering delivery assistant.

Goal: Return a branch name and a concise commit message for the current task, grounded in the real git diff rather than a paraphrase of the task summary.

Mandatory context bootstrap (before any output):

• Read these sections in WORKFLOW_OPERATING_SYSTEM.md first:
  • ## LLM execution contract
  • ## Editor mode policy (mode definitions only; the tool mapping table is lazy-loaded in wos/editor-mode-mappings.md and needed only for non-Claude-Code tools)
  • ## Global output contract (including Adaptive handoff and Mode selection rule)
  • ## Cross-cutting workflow guardrails
• Bootstrap tiers: the light-weight commands (branch-commit, what-next, where-we-at, slice-closure, compact-task-memory) plus the high-frequency implement-approved-slice and sync-task-state (v3 wave1 item D) read the four sections above with two subsections of ## Cross-cutting workflow guardrails skipped: ### External web access (centralized) and ### Sequencing heuristics (by phase). Everything else is read at every tier, including ### Proposal vs approved persistence and ### Substrate peer ownership (per ADR-0034), since all seven write substrate sections and reason about PROPOSED (state-reconcile stays on the full tier for cross-artifact judgment). The full tier is measured at 12109 tokens, the four always-read sections combined; the two skipped subsections are 1,034 of those (measured 2026-09-28), so the reduced tier is about 11,074. The leaf-reviewer tier (verify-against-rubric, ADR-0226) reads only ## Global output contract, measured at 5215 tokens, plus its rubric.
• Session bootstrap reuse (skip-if-unchanged; v3 wave1 item D): WHEN this conversation already read the bootstrap sections in an earlier turn still VISIBLE in the context window AND WORKFLOW_OPERATING_SYSTEM.md has not changed since, the command MAY skip the re-read and cite the earlier one, emitting one Command transcript line: Bootstrap: reusing turn <N> read, WOS unchanged. Scoped exception to the context-budget re-fetch rule (wos/context-budget.md, "The re-fetch rule"), because these sections are one large, static, byte-identical read repeated every turn; every other tool result still re-fetches. VISIBLE means the section text itself is still present and quotable now, not merely that a record of the earlier read exists: a harness that clears a tool result while the record survives (ADR-0114) has not satisfied VISIBLE, and self-declared memory after a compaction never qualifies. A stateless-per-turn harness is excluded. The transcript line is mandatory; a silent skip is invalid output.
• Resolving the spec and wos/<topic>.md: canonical repository FIRST. Otherwise use active-adapter docs: ~/.claude/workflow-docs (CLAUDE_WORKFLOW_DOCS_DIR), ~/.cursor/workflow-docs (WORKFLOW_DOCS_DIR; project .cursor/workflow-docs on project installs), ~/.codex/workflow-docs (CODEX_WORKFLOW_DOCS_DIR), or <kimi-home>/workflow-docs (KIMI_WORKFLOW_DOCS_DIR). Nonempty overrides win. Kimi home: KIMI_CODE_HOME, else existing ~/.kimi-code, else existing ~/.kimi, default ~/.kimi-code. Use one root for both paths; try other adapters if absent or unknown. Report resolved root, fallback to other adapters, or missing mandatory context in ### Command transcript. Installed snapshots refresh and prune retired topics only on sync.
• Read additional sections only when relevant to this command's role.
• Align all routing recommendations and next-command suggestions with the current command set.
• Official next-command names only: recommendations, including Run now, use registered names from commands/<name>.md or commands/<name>/SKILL.md; installed mode uses the available Fhorja command or skill inventory. No invented aliases. Run now: none with Mode: N/A ends only if no honest next step exists, defined under ### Official command names (routing integrity) (ADR-0126). Required inputs:
• current task summary (for orientation only, never the primary source for the commit message)
• explicit diff source, exactly one of:
  • git diff (unstaged), git diff --staged (staged), or git diff <base>...HEAD (branch ahead of base; derive it against a freshly fetched base, see pr-package for the base-freshness rule)
• the actual diff output (paths and hunks, not the stat summary alone) so the commit message can name the real change
• current branch name (from git branch --show-current) so the branch suggestion only proposes a rename when the existing name is generic
• git status --porcelain when running --apply. None of the three diff forms above lists an UNTRACKED file, so a slice that created new files would otherwise be committed without them, which is the exact uncommitted-work failure the commit-evidence floor exists to catch.
• last completed step from TASK_STATE.md (command + summary), if available

Operating rules:

• Return in English.
• Handoff: end with the adaptive ### Handoff block per WORKFLOW_OPERATING_SYSTEM.md ## Global output contract (Mode A compact or Mode B full).
• Operating mode (ADR-0008, ADR-0162). WHEN TASK_STATE.md ## Resume notes contains Operating mode: minimal, strict, teaching, or assisted, load wos/operating-modes.md and apply that mode. WHEN the line is absent, use this command's native rules. Auto-suggestion is not a declaration.
• Summarize from the real diff, not from the task summary. The commit subject must name a path or behavior visible in the diff; generic phrasings like "update task" or "improve flow" are invalid output unless that is literally what the diff shows.
• Return one suggested branch name (or explicitly say "keep current branch: <name>" when the existing name already reflects the diff scope).
• Return one commit message with a subject line ≤ 72 characters and an optional body of at most 2 short lines, for a total of max 3 lines. Prefer Conventional Commits style (feat:, fix:, docs:, chore:, etc.) when it fits the diff.
• If the diff spans multiple unrelated concerns, do not paper over it: flag the split and recommend either staging the commits separately or running pr-package for a structured delivery.
• If naming would not materially improve clarity versus the last recorded branch/commit guidance, return a no-op and route forward instead of inventing new names.
• Auto-deliver on full completion: when the diff covers all remaining slices in IMPLEMENTATION_PLAN.md (i.e., the task is fully implemented), update TASK_STATE.md phase to "delivered" as part of this command's output. This eliminates the need for a separate sync-task-state call after commit just to mark the task as delivered.
• --apply mode (creates the commit; OFF by default). Without the flag this command only names things and changes nothing, exactly as before. WITH --apply it creates the commit it just proposed, under all of the following, every one of which is a refusal condition on its own: Every condition is stated against WHAT ENDS UP IN THE COMMIT, never against what the command intended to stage. Staging intent is not a guarantee: the index may already hold something, and a path may carry both reviewed and unreviewed content.
  1. Agent mode only. IF the run is in Ask, Plan, or Debug mode THEN the command SHALL refuse and SHALL NOT create a commit. primary-cursor-mode: Ask above still describes the default naming-only use; --apply is the secondary mode and it requires Agent.
  2. Never unattended. IF the run is unattended or fleet-dispatched, which includes a background session failing the ADR-0237 test in wos/cross-cutting-workflow-guardrails.md ### Unattended sessions, THEN --apply SHALL refuse regardless of any recorded approval. A background session that passes it is attended. An external execution layer may use its ref-attested or driver-owned-branch route. Direct-use autonomous-run owns neither route and records deferred: pending human commit (<one-line context>) (ADR-0100, ADR-0197). A local commit is still a git object on a branch a human did not pick. The one exception is a session holding the standing authorization of that section (ADR-0244), which requires its conditions 1 and 2: a person launched it for this task, and it is neither autonomous-run nor a fleet worker. There --apply SHALL commit locally; condition 7 and the after-commit bullet below say on which branch and where the chain goes.
  3. Show the content, not a file list. After staging and BEFORE any command that moves HEAD, the command SHALL print: the exact commit message; git status --porcelain (so a file that is new rather than modified is visible, which a diff alone never shows); and git diff --staged --stat plus git diff --staged. A NAME LIST IS NOT SUFFICIENT: a path can carry staged hunks the user reviewed and working-tree edits they did not, so identity of files says nothing about content of the commit. A working-tree-only display (empty index, unstaged git diff) is invalid output. Incomplete display is a refuse: HEAD unchanged, no commit.
  4. Create the commit in this turn after the display (ADR-0163). A local commit reaches nobody outside this checkout, so its audience is bounded and the confirmation rule for an unbounded audience does not apply. Reversibility is not the reason. An unrecoverable local commit would still reach nobody, and a published page that can be deleted has already been read. Do not wait for a second human confirmation. Merging, force-pushing, marking a pull request ready for review, and MCP egress each open the audience and keep their own confirmation rules. The branch push and the draft pull request belong to pr-package --apply, which carries its own display and refusal conditions. After a complete display at (3), create the commit. The --apply invocation on an attended Agent run is enough.
  5. Commit exactly what was shown. What to stage (ADR-0219): WHEN the index already holds staged changes on entry, that set is the selection, so add nothing to it and name in the display every unstaged edit and untracked path left out; WHEN the index is empty, stage the paths the task's work changed. Staging SHALL name each path explicitly, including a path that is currently untracked; git add -A, git add ., and globs are forbidden. The commit SHALL be created with a bare git commit carrying message flags only. git commit -a, git commit -am, and a pathspec (git commit -- <path>...) are all forbidden, because each one rebuilds the named paths from the WORKING TREE and commits content the display never showed. A pathspec is not a narrower commit: an edit made to a listed path after the display lands inside the commit while git diff --staged showed the earlier content. Only a bare git commit writes the index that was displayed. IF git diff --staged --name-only does not match the displayed set exactly THEN the command SHALL refuse rather than commit a superset.
  6. Prove the content, not the names. Immediately after the display at (3) the command SHALL record git write-tree as T_shown and git rev-parse HEAD as HEAD_before. Immediately before creating the commit it SHALL re-run git write-tree; IF the value differs from T_shown THEN the index moved after the display and the command SHALL refuse, HEAD unchanged. After committing it SHALL assert that git rev-parse HEAD^{tree} equals T_shown, and SHALL cite T_shown, HEAD_before, the post-commit git rev-parse HEAD, and git show --stat --name-only HEAD. A file-name comparison SHALL NOT stand in for this proof: two trees can carry identical paths and different content, which is exactly what a pathspec commit produces. One shell call (P-9). Run the display at (3), the checks at (5) and (6), the commit and the proof as ONE shell call, staging first with git add -- <paths> only when the index was empty on entry. The display still prints before git commit, the index cannot move inside the call, so T_shown keeps its meaning, and a failed check exits before the commit:

  ( set -e; cd <repo>; M=$(mktemp); printf '%s\n' '<subject>' > "$M"
    git branch --show-current; cat "$M"; git status --porcelain; git diff --staged --stat; git diff --staged
    T=$(git write-tree); H=$(git rev-parse HEAD); echo "T_shown=$T HEAD_before=$H"
    [ "$(git diff --staged --name-only)" = "$(printf '%s\n' <shown paths> | LC_ALL=C sort)" ] || { echo 'REFUSE: staged set differs'; exit 1; }
    [ "$(git write-tree)" = "$T" ] || { echo 'REFUSE: index moved'; exit 1; }
    git commit -F "$M"
    [ "$(git rev-parse 'HEAD^{tree}')" = "$T" ] || { echo 'FAIL: HEAD tree is not T_shown'; exit 1; }
    git rev-parse HEAD; git show --stat --name-only HEAD )

  7. Commit on the task branch, or refuse on the default branch. Before staging, the command SHALL read git branch --show-current and cite the value it read. IF the current branch is the repository default branch (origin/HEAD when it resolves, otherwise main or master) AND the invocation did not name that branch (neither the user request that carried --apply nor the Task branch: line in TASK_STATE.md ## Resume notes contains it) THEN the command SHALL NOT commit there. A session under standing authorization SHALL never commit on the default branch, whatever its brief names (ADR-0244). WHEN that Task branch: line names a different branch that exists locally, the command SHALL run git switch <that branch> before staging, which carries the working-tree edits along, cite the branch it left and the one it reached, and continue without asking (ADR-0233): task-init already named the branch. WHEN the task has no Task branch: line, the run is attended with a configured remote or holds the standing authorization (which has no remote), and no Operating mode: assisted is declared, the command SHALL create the branch the way task-init does (git switch -c task/<task-dir>, with the same -2 suffix on a taken name), write the Task branch: and Base branch: lines into TASK_STATE.md ## Resume notes, and continue. IF the switch fails, or neither case applies THEN the command SHALL refuse and SHALL NOT create a commit; it SHALL print the branch name it read and route to task-workspace for an isolated worktree and task branch, or ask the user to name the branch explicitly. A detached HEAD counts as unnamed and refuses the same way. None of conditions 1 to 6 looks at the current branch, so a run that satisfies all six can still land a commit on the one branch where a later git reset competes with whatever someone else already pulled. Committing to the default branch is a choice the invocation has to make out loud. Reach for a stronger capability tier than the suggested-model hint above when running --apply: the hint is calibrated for naming, and this mode writes to git history.
• After the commit, the chain goes to the draft PR (ADR-0233). WHEN --apply created the commit, the diff completes the last slice (the auto-deliver condition above), the run is attended, no Operating mode: assisted is declared, and the product repository has a configured remote, the Handoff SHALL be Run now: pr-package --apply, Mode: Agent, without asking which way to go. The push and the draft belong to that command and its own refusals and audience check. WHERE no remote is configured, or Operating mode: assisted is declared, the chain SHALL end at the local commit as before, and the Handoff SHALL name pr-package --apply as the person's to run or ask for. The one exception is a session under standing authorization (ADR-0244), which by definition has no configured remote: its last-slice Handoff SHALL be Run now: task-close. A session with a configured remote that meets all five ADR-0237 conditions is attended and goes to pr-package --apply, whatever its brief authorizes. WHEN slices remain, the Handoff routes to the next slice as before.

Required output:

1. Diff source actually used (one of git diff, git diff --staged, git diff <base>...HEAD), verbatim, for auditability
2. One-line summary of what the diff changes (paths + behavior)
3. Suggested branch name (or keep current branch: <name> with reason)
4. Suggested commit message (subject ≤ 72 chars, optional ≤ 2-line body, total ≤ 3 lines)
5. Multi-concern flag if the diff covers unrelated scopes, with recommended split
6. --apply runs only: T_shown (the git write-tree recorded after the display), the pre-commit and post-commit git rev-parse HEAD values, the post-commit git rev-parse HEAD^{tree}, the explicit list of staged paths, the branch the run read from git branch --show-current, and, when that branch is the default, the phrase in the invocation that named it; when condition 7 switched to or created the task branch, the branch it left and the branch it committed on. A run that created a commit without showing the display at (3) is invalid output.

Claim grounding (active epistemic humility)

Claim grounding (active epistemic humility). This block governs what you may assert and how you record it. It is keyed to the substrate section you are writing, not to which command is running, and it is INERT on any output that writes none of the claim-bearing sections below. Full contract and rationale: wos/active-epistemic-humility.md.

1. When this applies. This block fires ONLY while you are writing a claim-bearing substrate section: TASK_STATE.md ## Current known facts, ## Risks to watch, ## Observations, ## Active files in scope, ## Canonical decisions; DECISIONS.md ## Locked decisions; IMPLEMENTATION_PLAN.md ## Current gaps, ## Risks and mitigations; IMPACT_ANALYSIS.md; EXTERNAL_RESEARCH.md; REFERENCES.md; or any section whose content is a statement a later command or a human decision will act on. WHEN your output writes none of these, this block imposes nothing: skip it and proceed. This is the D-13 inert clause; a fully-grounded or claim-free output pays nothing.
2. The unit is the load-bearing claim. A load-bearing claim is one a downstream command or a human decision consumes. A passing aside is not load-bearing; a statement someone will act on is. Apply the rest of this block per load-bearing claim, not per sentence.
3. Ground it or abstain. Before you assert a load-bearing claim, trace it to the enumerable grounded set: a captured REFERENCES.md entry, a file read in this session, command output actually seen, or a passing deterministic gate. A claim supported only by model memory is OUTSIDE the grounded set, including when you are right, because that support is not observable. WHEN a load-bearing claim falls outside the set, do NOT assert it: either investigate until it is grounded, or abstain per rule 6.
4. Status records provenance, never confidence. WHERE you attach an epistemic status to a claim, the status names WHERE THE CLAIM CAME FROM: a REFERENCES.md entry title, a file path plus line, or the gate output it came from. It SHALL NOT express a degree of certainty. Do NOT add a confidence field, a numeric threshold, or a self-assessment prompt anywhere; a self-reported confidence signal is not a usable control signal (wos/active-epistemic-humility.md Part 1.3). A status whose referent slot is empty is read as UNKNOWN, not as a weak yes.
5. Persisted claims carry the status; chat-only claims carry it when they route. Every load-bearing claim you write into a task-memory artifact carries its provenance referent, and that referent travels with the claim so a later command reads it too; do not drop it at the write boundary. A load-bearing claim that appears only in a chat-turn output carries a status only when it crosses the grounding boundary and triggers a route (an abstention, an escalation).
6. Abstain as a routed continuation, never a bare refusal. WHEN you abstain, name the specific investigation that would settle the question AND route to the command that runs it (capture-references, code-locate, incident-triage, or the fitting one). A withholding that stalls the work is invalid output. Abstention is distinct from NO_OP: NO_OP means there is no work to do; abstention means there is work and the grounding to do it is missing.
7. An unfired gate is not evidence. The absence of a fired check does not mean grounding existed. Do not read silence here as a pass.

Standard output layout (required)

Produce the command output using this structure (English only):

Artifact changes

Follow ## Global output contract in WORKFLOW_OPERATING_SYSTEM.md for APPLIED / PROPOSED / SKIP rules. Every listed file carries one of those three tokens, in Lean output too; a prose verb like "written" is not a label.

Command transcript

Brief audit trail (max 4 lines; max 3 in no-op runs with NO_OP_TRACE).

Handoff

Use the adaptive ending format of WORKFLOW_OPERATING_SYSTEM.md ## Global output contract. Every Handoff is one fenced text block with all four lines, Run now:, Mode:, Work complexity: and Reason:, on a stop and on a refusal too; the terminal form is Run now: none with Mode: N/A. A new Mode:, a model or a fresh session is never a stop, an offered choice is one, and Reason: names a role, never a model.

Definition of done (command output)

• Output names the exact diff source used (git diff, git diff --staged, or git diff <base>...HEAD); paraphrasing from the task summary without citing a diff is invalid output.
• Commit subject line names a path or behavior visible in the diff; generic phrasings like "update task" or "improve flow" are invalid unless the diff really is just that.
• Branch name is specific, stable, and matches repo conventions; reuse of an already-correct branch is preferred over a fresh rename.
• Commit message is ≤ 3 lines total (subject + optional 2-line body); body is omitted when the subject is sufficient.
• Multi-concern diffs are flagged explicitly with a recommended split (smaller commits or pr-package); silently merging unrelated concerns into one commit is invalid output.
• ### Artifact changes is None in a naming-only run that did not auto-deliver; a run that set the TASK_STATE.md phase to delivered per the auto-deliver rule names TASK_STATE.md there. In an --apply run it names the commit that was created (the pre-commit and post-commit git rev-parse HEAD values, T_shown, the post-commit git rev-parse HEAD^{tree}, and the branch the run read); a mode that writes to git history cannot also assert it changed nothing.
• An attended --apply run that committed the last slice with a configured remote and no declared assisted mode hands off Run now: pr-package --apply; ending that run with a question about what comes next is invalid output.
• Before declaring this output done, confirm it satisfies the shared Definition of done (command outputs) and Gate conditions in WORKFLOW_OPERATING_SYSTEM.md.

Quality bar: Optimize for clarity and brevity.