senior-dev: no wip commits, its own commit message, and hand-backs over self-fixes

[ZeroPoint95]

senior-dev's branches now carry clean history: no per-write wip commits and no senior-dev author. The run's work becomes codeaf's one commit, with a message senior-dev writes. On the turn a run ends, the chat now hands follow-up fixes back to senior-dev instead of patching its branch itself.

Prompted by #1693: about 50 wip(edit): <path> commits authored senior-dev <senior-dev@localhost>. With this repository's squash merge (body = every commit message), all of them would have landed in dev's history. #1696 replaced it.

1. No wip commits (8f696e069)

• EagerCommit committed every file the model wrote or edited. It is deleted, along with its SENIOR_DEV_EAGER_COMMIT opt-out, the recorders' CommitsOnWrite, and on codeaf's side SENIOR_DEV_EXPECTED_BRANCH / DelegateSetup.Branch, which existed only to keep those commits on the run's branch.
• Nothing else depended on those commits:
  • senior-dev snapshots through a temporary index and keeps its candidates as commit-tree objects outside the branch.
  • codeaf already commits whatever a run leaves, crash or not.
• How senior-dev plans, implements and verifies is unchanged.
• TestAWriteAndAnEditLeaveTheRunsBranchWhereItWas fails on dev with senior-dev wip(write): added.txt.

2. senior-dev is asked not to commit, and writes the commit message (62aa22777)

• The line codeaf puts at the head of every brief now tells the program to leave its work uncommitted and not to push, switch branches or rewrite history, even where the brief asks it to.
• It also asks for the commit message in .senior-dev/commit-message.
• codeaf reads that file before the notes leave the copy, and uses it for its end-of-run commit, adding its credit lines once. With no message, the commit falls back to the task title over the run's ending.
• This is steering, not enforcement: no git command is blocked, and a commit the program makes anyway stays on its branch.

3. Handing back counts as fixing it yourself (42c694411)

• prompts/program-outcome.md used to tell the chat to finish "a small, clear gap" itself in a worktree. So after a senior-dev run, the chat did follow-ups itself.
• That turn is the one moment a hand-off continues on the program's own branch. The page now says a hand-back to the same program counts as fixing it yourself, and the chat edits the branch itself only for trivial gaps or once retries are spent.
• The page is read only on that turn, so the prompt-prefix budget is untouched.
• Out of scope: follow-ups you ask for in a later message still start a fresh run from your checkout.

Manual:

• senior-dev.md gains "Does senior-dev commit or push — no wip commits, who writes the commit message", and its sections on where the work goes and what codeaf does when senior-dev ends are updated.
• delegates.md is updated to match.
• New questions in the manual's retrieval tests (internal/manual/chat_test.go) check that "why does senior-dev make wip(edit) commits in my pull request" and "who writes the commit message for senior-dev's work" reach the senior-dev page.

Checks run locally

• Passed: build and vet; full internal/seniordev/{util,tool,config,app} and internal/run; focused internal/session and cmd/codeaf; internal/manual and untagged internal/e2e; make test-laws.
• TestStandingIsolationRecordsItsCopyBeforeWorkerInitialization fails, and fails the same way on clean dev (macOS /var vs /private/var). A separate fix is in progress.

🤖 Generated with Claude Code

[AbirAbbas]

Taking over: checked by hand in the real binary with throwaway repos — a finished run, a run stopped part-way, a run whose codeaf was closed mid-run, a run started by a dev build and finished by this one, a steering note sent from the run's page, and a carried-on run. Each leaves one commit per run with senior-dev's own message, no wip( commits, and every file it wrote on the branch; the carried run's ending still reads 2 files past <commit>, where the last run left it.

Pushed:

• ea9e02c43 — senior-dev's unsubmitted-run nudge told its model to commit before verifying (with nothing committed any more, that fired on every such nudge), and "measured its change" diffed the working tree, so new untracked files went uncounted (1 file · +28 -0 for a run that wrote three files; its change is empty for one that only adds files). The finding is gone and the summary measures the snapshot tree the run would hand in.
• 402e7e5f1 — the commit message is read only as a regular file (a named pipe there hung the finish with the copy held; a symlink was followed); a NUL byte, credit lines only, or an unchanged copy the repo already tracked falls back to the title and ending. A message written early no longer hides that the run did not pass: stopped, failed, limit, crashed, unverified and gone runs carry the ending after the message. The brief also asks the program not to stash, reset, clean, or check out/restore over its work, which no longer sits in HEAD. Once retries are spent the wake page now says to tell the person, as its footer already did. Manual and change entry updated.
• 86b398ee6 — the stale-message check uses senior-dev: a run that carries on a branch counts only its own files, and a published branch is pushed, not merged #1705's ownBase rather than a second persisted start commit.
• d3f2a4244 — senior-dev and task-landing manual sections split after the merge, where one probe stopped reaching its page.
• 85996df3a, 05429ff90, e198cbc0c — merges of dev (senior-dev: a run that carries on a branch counts only its own files, and a published branch is pushed, not merged #1705; senior-dev can be told something while it works, from its page or the chat #1689, whose child environment keeps the steering inbox without SENIOR_DEV_EXPECTED_BRANCH) and of this branch's own merge of dev.

Left for later:

• A file the run writes and then ignores through its own .gitignore edit is not committed (an eager commit used to have tracked it). Force-adding newly ignored files would sweep shell-made build output onto the branch, so it is left as is.
• With no per-write checkpoint, the model's own git stash, git checkout -- . or git reset --hard now hides or drops work that used to be in HEAD; the brief asks it not to, nothing enforces it.
• RunSpec.ProgramBranch has no production reader now.