The framework never commits on an agent's or the user's behalf (#1638) - #1724
Merged
Conversation
Contributor
Author
|
Why the prompt line changed (for the read-through)
-2. Create a new branch `tf-<SESSION_NAME>` and `$ git checkout` it — do all the work in that branch
+2. Create a new branch `tf-<SESSION_NAME>` and `$ git checkout` it — do all the work in that branch, and commit it there as you go: The Framework publishes only what you committed, and never commits for you
-- **Name the session, then branch** - before the first change the agent invents a session name, creates and checks out `tf-<session name>`, and reports the name back.
+- **Name the session, then branch** - before the first change the agent invents a session name, creates and checks out `tf-<session name>`, commits its work there as it goes, and reports the name back.Why: until this PR the framework committed whatever the agent left uncommitted — at teardown, at handoff, on Push/Open PR — so the prompt never had to ask the agent to commit. That sweep is what put 7,632 cache files on main (#1632), and #1638 removes it. With nothing committing on the agent's behalf, the agent has to commit its own work or it never gets published; detached runs were already told this, now every run is. The spec line follows the prompt. |
Contributor
Author
|
@brillout merging — happy to adjust anything. |
suleimansh
added a commit
that referenced
this pull request
Aug 27, 2026
- The legacy the-framework/ prefix goes: agentBranchFor guesses tf-, the scratch sweep and isAgentBranch know one spelling, and the package's public API carries no compat layer (repo MEMORY.md: zero migration code). - The framework's own build/dev/typecheck/test scripts build the package first, so a fresh clone works without a root build; the package sets publishConfig.access public. - reclaimWorktree reads the tree once; the refusal union says which members carry a branch; git branch -D/-m/<new> run on the write budget. - cli-exec keeps only what gh uses: a flat budget, no maxBuffer, no isCliTimeout; the daemon test times out with the package's error. - The branch-links pass is one line at its call site; the barrel exports what has an importer; the framework proves it wires birthBranch and mayPush into the rule. - SPECs: root and packages/ name the new package; index.SPEC.md; worktree, cli-exec and agent-handoff wording caught up with #1724 and this lift.
suleimansh
added a commit
that referenced
this pull request
Aug 27, 2026
* Branch management lives in its own package (#1725) The git side of an agent's checkout — the naming conventions and the .the-framework/branches/ layout, the git runner with its per-subcommand time budgets, creating/attaching/listing/renaming/removing a worktree, sharing the parent's dependency trees into it, the branch-name links, and the retention rule under which a checkout is reclaimed — moves out of the framework into packages/branch-management, published as @superskill/branch-management. Step 1 of #1725: a lift with no behaviour change; the framework imports the package as a workspace dependency. What the framework keeps is the agent's side: its record (may the branch be pushed, what did a cloud hand-off already push), the teardown and the sweep, the dashboard's list and buttons, and how each refusal is worded. The package's reclaimWorktree takes those as options and returns a reason code. Also folded in: pushAgentBranch/gitReason (agent-handoff) were the same push the rule needs, so there is one pushBranch now, in the package. * Review fixes for the branch-management lift (#1725) - The legacy the-framework/ prefix goes: agentBranchFor guesses tf-, the scratch sweep and isAgentBranch know one spelling, and the package's public API carries no compat layer (repo MEMORY.md: zero migration code). - The framework's own build/dev/typecheck/test scripts build the package first, so a fresh clone works without a root build; the package sets publishConfig.access public. - reclaimWorktree reads the tree once; the refusal union says which members carry a branch; git branch -D/-m/<new> run on the write budget. - cli-exec keeps only what gh uses: a flat budget, no maxBuffer, no isCliTimeout; the daemon test times out with the package's error. - The branch-links pass is one line at its call site; the barrel exports what has an importer; the framework proves it wires birthBranch and mayPush into the rule. - SPECs: root and packages/ name the new package; index.SPEC.md; worktree, cli-exec and agent-handoff wording caught up with #1724 and this lift.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #1638. Rom's call on the issue: remove the
[The Framework] uncommitted changesfeature altogether. Replaces #1722 (closed).What goes
The sweep was made in four places; all four are gone, and nothing replaces them with another commit:
install.ts) — no longer commits the user's dirty tree first. The install commit adds only.the-framework/(git add .the-framework, neveradd -A); whatever the user has uncommitted stays theirs.worktrees.ts→store/worktree.ts) —commitPendingWork(with its Trivial run's work is stranded: agent skipped committing, handoff says "session committed nothing", safety commit lands only after the publish decision (#860 redux) #1376 retry) is deleted. A checkout holding uncommitted work is reported dirty byworktreeCleanand kept:session <id> has uncommitted work; its worktree was kept. A tree git cannot read is kept the same way.cli.ts) — no commit before the push; thecommit-failedskip reason goes with it (events.ts,terminal.ts). What is published is what the agent committed.dashboard-rpc/control.ts) —commitAgentWorkis deleted; both actions push the branch as the agent left it, still under the agent lock (the push race with teardown is unchanged).What replaces it
The agent commits its own work. Detached runs were already told so; now the system prompt's branch step says it for every run: do all the work in that branch, and commit it there as you go: The Framework publishes only what you committed, and never commits for you.
The trade, stated plainly: an agent that ends without committing leaves its work uncommitted in its worktree — never destroyed, named on its page (
pendingFilesis unchanged), but not published and on disk until a person commits or deletes it. The framework used to rescue that silently; that rescue is what put 7,632 cache files on main.Verified
Full suite green (1643 node + 864 dashboard). Broken on purpose in the compiled output: "teardown never keeps a dirty checkout" and "install adds
-A" turn 6 tests red acrossworktrees,merged-worktrees,agent-addressingandinstall; restored, 58/58.Specs (please read through)
install.SPEC.md— activation is one commit holding only The Framework's files; the user's uncommitted work is left as it is.store/worktree.SPEC.md— "Teardown commits before it deletes" → "A checkout is read, never committed", with the rationale (the cache incident).worktrees.SPEC.md— the reclaim rule: kept while dirty, pushed when clean, removed once the remote has it.cli.SPEC.md,dashboard-rpc/control.SPEC.md,dashboard/agent-handoff.SPEC.md,agent-locks.SPEC.md,merged-worktrees.SPEC.md— the sentences that said "commits what the agent left" now say "publishes what the agent committed".prompts/system_prompt.SPEC.md— the branch step includes committing.FEATURES-SPEC.md— "Commit what the agent left uncommitted" → "Only what the agent committed is published — nothing is ever committed on the agent's or your behalf".install,store/worktree,worktrees,merged-worktrees,agent-handoff,dashboard-rpc/agent-addressing).🤖 curated · Fable 5, effort high