Git Discipline

Git Discipline

One branch per initiative. Merge only when the user says so.

This capability is the single source of truth for branch naming, creation, merging, worktrees, and cleanup. Workflows cite it — they never restate its rules. If a workflow's text and this file disagree, this file wins; flag the workflow as the defect.

Branch vs worktree — the decision rule

These solve different problems; pick by how many sessions run at once:

• Branch isolates history — one initiative's commits don't interleave with the default branch. A branch is sufficient for one session at a time (sequential initiatives). This is the default (isolation: branch in .sage/config.yaml).
• Worktree isolates the working tree — its own directory on disk. A worktree is required for simultaneous sessions: two claude processes in one checkout share one working tree and one HEAD, so they clobber each other's files regardless of branch, and git checkout -b in one moves HEAD for the other. Branches cannot fix a working-tree race; only a directory-per-session can.

The hard constraint. A Claude session is bound to the directory it was launched in and cannot relocate itself into a worktree it creates. So the parallel path is always launch-time: create the worktree first, then start a session inside it. The front door is sage worktree <slug> (creates branch + worktree + copies the gitignored runtime + prints the cd … && claude line). Opt in with isolation: worktree; until then, sequential branch-in-place is unchanged.

Availability check

Activate only when git rev-parse --git-dir exits 0. On failure (not a git repository): silently inactive — no error, no prompt, no mention. Every workflow runs exactly as it would without this capability. Never run any other git command before this check passes.

Branch naming

<slug> is the initiative's work-dir slug without the date prefix: .sage/work/20260612-payment-retry/ → feat/payment-retry. The full dated slug remains the work-dir key. On a name collision (the branch already exists), append -2, -3, … — and never reuse an existing branch without the user's explicit choice (see "Already on a branch" below).

The naming extends the project's conventional-commit convention (feat, fix, chore, docs) to branch names.

Worktree bounce (isolation: worktree in the main checkout)

Read isolation: from .sage/config.yaml (default branch). When it is worktree AND the workflow is starting in the main checkout (not a linked worktree — the test is git rev-parse --git-dir == git rev-parse --git-common-dir; equal → main checkout), do not branch in place. Present a guided menu (never a hard refusal — a lone sequential session must still be able to proceed):

• [1] Set up an isolated worktree (recommended for parallel) — run / instruct sage worktree <slug>; open a session in the new directory and continue there. This session stops here.
• [2] Proceed here in the shared checkout (not isolated) — fall back to the normal in-place branch flow below; record the choice in the initiative decision log.
• [3] Sequential, this once — same as [2], noting isolation: worktree was overridden for this initiative.

When already running in a linked worktree (the user launched via sage worktree), skip the bounce and branch/commit in place as usual — the session is already isolated. With isolation: branch (default), the bounce never fires.

Branch creation (propose-confirm)

At the workflow step that consumes this capability, when HEAD is on the default branch with a clean tree:

1. Propose the branch name from the table above.
2. Let the user confirm or rename. Never create silently.
3. git checkout -b <branch>
4. Record the actual branch name in the initiative's manifest frontmatter: branch: feat/payment-retry. Resume matching reads this recorded field — never re-derives from names (see "Branch ↔ work-dir mapping" below).

Dirty tree at creation time: stop and present three options — stash and proceed, commit current changes first (the user writes or approves the message), or abort the workflow. Never stash silently.

Already on a branch (HEAD on a non-default branch): ask whether to (a) continue on it — the user may have prepared one — or (b) create the initiative branch from the default branch. Record the choice in the initiative decision log.

Detached HEAD (tag/SHA checkout; git worktree add without -b): surface the state, offer to create the initiative branch from the default branch or from HEAD (git checkout -b <branch> works from either) — never proceed to implementation commits while detached.

User declines branching: record the decline in the initiative decision log and proceed exactly as today (commits land on the current branch). Declining is always available — the discipline is a default, not a cage. A recorded decline also suppresses the CLAUDE.md minimum branch gate for this initiative.

Default-branch resolution

Never hardcode main. Resolve via this chain, stopping at the first success:

1. git symbolic-ref refs/remotes/origin/HEAD (strip the refs/remotes/origin/ prefix) — set by clone; absent in remote-less repos.
2. git config init.defaultBranch — only if a local branch by that name exists.
3. An existing local branch named main, then master.
4. Ask the user once.

Persist the answer as default_branch: in .sage/config.yaml so every session — and every parallel worktree — resolves identically. Never silently designate the currently-checked-out branch as the default.

Unborn HEAD (fresh git init, no commits): branch creation still works (git checkout -b is valid from unborn HEAD); the merge protocol requires an existing default branch and otherwise degrades to "commit on the current branch" with a one-line notice.

Per-task commits

Implementation commits land on the initiative branch using conventional-commit subjects. The workflow's existing commit cadence is unchanged — only the branch it lands on changes.

Merge protocol (ALWAYS user-gated)

Merging is offered only at a completion checkpoint as an explicit menu option ([M] Merge to <default>). No workflow path executes a merge without the user choosing that option in the same session. This is a hard line, not a default.

Before offering [M]:

• Run the full test suite and gate on its own exit code. Capture the status, then act — never suite | tail && merge (a pipeline's exit status is the last command's; a piped suite reads green even when red). On a project with no test harness, substitute the initiative's documented smoke procedure; if none exists, require an explicit user acknowledgment that the merge is unverified.
• Structural preconditions — both must hold:
  • git status --porcelain is empty (no uncommitted work that a checkout would transplant onto the default branch);
  • git rev-list <default>..<branch> is non-empty (the branch actually has commits to merge).
• Show git diff <default>...<branch> --stat.

On user [M]:

git checkout <default>
git merge --no-ff <branch>

On conflict: stop, show the conflicting paths, hand control to the user. Conflict resolution is out of scope for this capability.

After a successful merge, offer (never perform) branch deletion — suppressed while the initiative continues (per-milestone merges): the deletion offer fires only at initiative completion. Deleting mid-initiative would strand later milestones on the default branch with no re-entry into branch creation.

No remote operations. This capability never runs git push, never creates PRs, never sets upstreams. Pushing and PRs remain the user's own actions.

Branch ↔ work-dir mapping: recorded, not derived

At branch creation the workflow records the actual branch name in the initiative's manifest frontmatter (branch: feat/payment-retry). Resume matching (session-bridge, /continue, workflow auto-pickup) reads the recorded field — never re-derives from names, because the derivation is non-bijective the moment any sanctioned exception fires: the collision suffix (feat/auth-2), a user rename at the confirm step, continue-on-existing-branch, or date-stripped slugs that collide across months (20260101-auth and 20260301-auth both derive feat/auth). The naming-table derivation is only the default proposal at creation time.

Worktrees (parallel sessions)

Front door: sage worktree <slug>. When the user wants a second concurrent initiative — or isolation: worktree is set — the launcher is how a parallel session starts: it creates the branch + worktree, copies the gitignored runtime, and prints the cd … && claude line to open an isolated session. The user runs one command; the session is born isolated. (A running session cannot relocate itself — see the decision rule above — so this is always launch-time.)

The manual recipe below is what the launcher does, kept as documentation and as the fallback when sage worktree is unavailable:

git worktree add ../<project>-<slug> <branch>

One worktree per initiative, as a sibling directory. A branch can be checked out in only one worktree at a time; the new worktree shares .git storage. If the target path already exists, offer ../<project>-<slug>-2 or let the user supply a path — never delete an existing directory.

.sage/ trackedness — two worlds, both handled. Probe with git check-ignore -q .sage (exit 0 = ignored):

• Tracked (.sage/ committed): state travels with each worktree's checkout; the per-initiative state files prevent merge conflicts between branches. The offer proceeds as above.
• Gitignored (as in sage's own repo): state does NOT travel — a fresh worktree contains no .sage/ at all. The offer still proceeds, with a copy step: after git worktree add, copy the shared runtime (.sage/config.yaml, .sage/constitution.md, .sage/gates/; for a /build-x initiative also .sage/scripts/, .sage/prompts/, and .sage/agents.toml — the multi-agent runtime, without which the first dispatcher invocation fails) into the new worktree. The seeded set is the worktree_copy config list (it deliberately does NOT seed .sage/work/ — a fresh worktree starts a fresh initiative); on teardown, sage worktree remove harvests the worktree_harvest list (default .sage/work/*) back to the main checkout. Both lists are project-configurable in .sage/config.yaml.

The platform runtime needs the same treatment. The same .gitignore that ignores .sage/ typically also ignores .claude/ and the platform instructions file (CLAUDE.md / AGENTS.md / GEMINI.md) — sage's own repo ignores all of them. Probe separately: git check-ignore -q .claude (and the instructions file). When ignored or absent, the worktree offer includes one of two remedies, user's pick:

• (a) Copy .claude/ and the platform instructions file(s) into the worktree alongside the .sage/ state; or
• (b) Run sage update inside the new worktree — the platform generators regenerate the instructions file and commands; this works even when sage/ itself is absent in the worktree (the CLI falls back to the global framework install).

A worktree session without a platform runtime is an inert checkout — no slash commands, no skills, not even the degraded CLAUDE.md gates. The offer must not complete without one of the two remedies.

The MCP config and the memory store. .mcp.json is in the default worktree_copy set, so the sage-memory (and other) MCP servers are configured in the worktree. The memory database is different from the rest of .sage/: it is a single shared store, not a per-worktree copy. A SQLite file cannot be merged, so .sage-memory/ is never seeded into a worktree and never harvested back. Instead the worktree session points at the main checkout's store by calling sage_memory_set_project(<main root>) (dirname "$(git rev-parse --git-common-dir)"); the session-init hook surfaces that call automatically when it detects a linked worktree. Result: one brain, shared by main and every worktree — nothing to reconcile.

Ownership record. Record owner: <worktree-path> in the main checkout's copy of the initiative manifest (recording in both copies is acceptable; the main checkout's copy is the one exclusion reads). Resume surfaces in the original checkout exclude initiatives owned elsewhere. The comparison probe is git rev-parse --show-toplevel against the recorded owner: after path normalization — trailing slashes and symlinked paths (WSL mounts alias the same checkout) must not defeat the comparison.

Ownership return. When the worktree is removed, always copy back the work dir before the manifest edit that clears owner: — editing first and copying back after re-imports the stale owner: from the worktree copy and permanently locks the initiative out of every resume surface.

Cleanup after merge: sage worktree remove <slug|dir> — offered, user-gated, never automatic. It harvests the gitignored .sage/work/<slug>/ back into the main checkout, then removes the worktree (refusing on dirty tracked files unless --force). Prefer it over bare git worktree remove, which skips the harvest and silently drops the docs (see Abandonment cleanup).

Abandonment cleanup

Closing an initiative without merging offers — never performs — the cleanup pair: sage worktree remove <slug|dir> (if a worktree exists) and branch deletion.

Ignored-world ordering is mandatory. When .sage/ is gitignored, a bare git worktree remove silently deletes the worktree's untracked .sage/ state with it — the brief/spec/plan are gone, never merged (untracked) and never recoverable via git (gitignored files never enter the object store). sage worktree remove exists to prevent exactly this: it harvests .sage/work/<slug>/ back into the main checkout before removing the worktree. The abandonment sequence is therefore:

1. sage worktree remove <slug|dir> — harvests the work dir back, then removes the worktree.
2. In the main checkout's harvested manifest, clear owner: and write status: abandoned — after the harvest, never before (see Ownership return).

In the tracked world, the harvest is unnecessary (git carries .sage/) and step 2 collapses to a manifest edit + commit on the initiative branch. Declining the cleanup leaves the branch for later; status: abandoned is recorded either way so resume surfaces stop offering the initiative.

What this capability never does

• Merge, push, or delete anything without an explicit user menu action in the same session.
• Create a branch silently.
• Stash silently.
• Run any git command before the availability check passes.
• Touch /autoresearch's pre-existing branch discipline.