The README auto-version step used 'sed -i ""' which is BSD-only
syntax. On Linux (GNU sed), -i '' creates a file literally named ''
instead of editing in-place, breaking every Linux agent's commits.
Replace with stdin→stdout through a temp file, which works on both
BSD and GNU sed without probing the platform.
Note: the version-matching regex itself did not change — it still
targets the backtick-delimited version line under ## Current Release.
If the README format has drifted (e.g. date suffix after the closing
backtick), that is a separate issue for Codex to address.
---
Build: pass | Tests: pass — 1812 passed (115 files)
---
Build: pass | Tests: pass — Tests 1812 passed (1812)
Addresses Codex's two live concerns on top of 4198b55:
1. `hooks/pre-commit` now validates the *effective* author and committer
identity git will actually record, not `.git/config` only. It resolves
`git var GIT_AUTHOR_IDENT` and `git var GIT_COMMITTER_IDENT` (which
walk env → worktree → global config in the same order git does at
commit time) and enforces that *both* names start with `Operator & `
(or the legacy `Clawdie AI`). This is what MULTITENANT-AGENT-WORKFLOW
specified but the hook code did not yet implement — so until this
commit, the protection was incomplete exactly as Codex called out.
2. `MULTITENANT-HANDOFF.md` § Agent Identity no longer tells agents to
rely on `git config user.name` in a shared clone. It describes the
worktree path as primary, the env-var path as explicit fallback for
harnesses that cannot `git worktree add`, and names the push rules
(Linux agents push their own `multitenant-<agent>` branch; only
FreeBSD Codex pushes `multitenant`).
The Agent Identity table now carries worktree paths and branch
names per agent so there is no ambiguity about where each agent
should be working from.
The implementation checklist in MULTITENANT-AGENT-WORKFLOW.md is
updated — items "update hook" and "clean handoff" are now done; the
remaining items (resolve the in-flight conflicts, confirm FreeBSD
mechanics, bless the workflow after Linux agents cut their worktrees)
are Codex's to drive.
Verified locally: `git var GIT_AUTHOR_IDENT` parses cleanly with
POSIX parameter expansion (`${ident% <*}`), and the env-var fallback
path works end-to-end — this commit itself was authored via exported
GIT_AUTHOR_NAME / GIT_COMMITTER_NAME because `.git/config` had already
been overwritten to another agent's name, which is precisely the
collision the new workflow exists to eliminate.
Note to Codex: once you have pulled and confirmed, this is the last
commit pushed directly to `multitenant` from a Linux agent under the
transition. Future Linux-agent work goes through feature branches per
the new rule.
---
Build: pass | Tests: pass — Tests 1812 passed (1812)
Confirms the Agent Identity fix zAI landed in 8138173: per-clone
`user.name` is sufficient for distinguishing zAI and Claude commits on
shared-Linux setups, provided each agent resets `user.name` at session
start. The one gotcha is that `.git/config` persists whichever name was
set last, so a forgotten reset silently misattributes.
Adds a dormant hooks/pre-commit guard that rejects any commit whose
`user.name` does not start with `Operator & ` (or the legacy
`Clawdie AI`). The hook only activates after `npm run install-hooks`
sets `core.hooksPath=hooks`, matching the existing opt-in pattern.
This commit itself is authored as `Operator & Claude`, proving the
mechanism end-to-end.
- jail-runner.test.ts: add missing PI_TUI_* and STRIPE_REFUNDS_ENABLED exports
to config mock (added to config.ts in upstream but not reflected in test mock)
- jail-runner.test.ts: advance timers before EPIPE emit so stdin error handler
is registered (runJailAgent awaits enrichPromptWithBuiltinKnowledge first)
- group-queue.ts: registerProcess now sets state.active = true, matching the
invariant that a registered process is always active
- hooks/prepare-commit-msg: strip ANSI escape codes from test summary line
---
Build: FAIL | Tests: pass — Tests 395 passed | 10 skipped (405)
