Review of the first version against 366 recorded runs found four defects and one
semantic gap, all fixed here.
- Two same-named calls in one batch swapped results, exit codes, origins and audit
links in the live view: the finished event was matched to the LAST pending item by
name. `tool_proposed` now carries the tool call id and the GUI matches on it.
- A command's exit status was read from the 300-character result preview; a long
command pushed it out, and a timed-out command showed a green dot. The exit code,
time-out and background flags now ride on `tool_finished` and on the persisted
display sidecar.
- After a reload every non-denied call replayed as "ok", so a failed write showed as
done. The status is persisted with the result and replayed.
- A notice inside a turn (compaction, continuation, a saved memory) split the summary
in two and showed the first half while the turn was still running. One block per
turn now, emitted after the answer.
The header reads "N changes auto-approved" and the count means "things that changed
state and that nobody approved": the engine stamps each cardless call's effect
("change" or "read", from its risk class and tier) and the summary counts failed
attempts and read-only actions apart. Several edits of one file collapse into one row;
background commands are labelled; after 25 rows the list offers the rest.
`turn_summary` gains `failed` and `reads`, and per-action effect, tier and shell
outcome; docs/headless.md describes the fields.
A user who was away had no simple view of what ran with no approval card. After a
turn in which auto-approve let actions through (a gate tier or a reviewer "allow"),
the transcript now ends the turn with a collapsed "N changes made without asking"
block: file changes with diffs, commands with their exit status, other actions,
each with who let it through and a link to its audit rows.
- engine: one TURN_SUMMARY event just before the event that ends the turn, listing
tool call id, tool, origin, status and the reviewer's reason (no arguments).
Wraps run / retry / resume; closing the wrapper closes the turn as before.
- audit: /v1/audit takes call_id; Activity opens filtered to one call, clearable.
- gui: tool items carry their call id, live and on reload; the block is built from
the persisted approval origins, so a reloaded session shows the same summary.
- docs/headless.md: the event in events.jsonl.
Whether a model declines to upload a .env file is model behaviour, not policy,
and OpenWorker lets users pick any model. A new floor, placed with the other
human-only floors (ahead of every allowlist, the allowed-sites wall and the
reviewer): an action that sends data off the machine (egress or external risk:
web tools, connector writes, MCP tools; or a shell command running a network
client) AND names a secret-looking file (.env*, *.pem, id_rsa*, credentials*,
*token*, ...) or carries a secret-looking value (private-key headers, sk-,
AKIA, ghp_, xox?- shapes, api_key = <long value>) is human-only in every
asking mode and in bypass-approvals; dangerous-bypass clears it and records
the floor. Templates and public keys (.env.example, *.sample, *.pub) are not
secrets. A backstop on top of the sandbox network allowlist: shell egress
can't be recognised completely.
Corpora: gate-104 (browser upload of an SSH key) moves from reviewer_eligible
to human_only (secure stays hard_deny); four secret-egress gate rows are
appended (gate-133..136). Seven frozen legacy reviewer rows the floor now
intercepts (danger-003/010, inject-001/002/008/014/016) stay as they are; the
hygiene test lists them explicitly and checks they're caught by this floor.
Docs: the floor joins the safety-check list; the Windows sandbox page notes
what "Allow everything" leaves to the approval flow.
For sealed evaluation runs and air-gapped work. A network block only refuses the call: the model still sees both tools, can spend turns on them, and the record cannot say they were absent. The flag withholds them at registration (build_engine(web_tools=False)), and summary.json and the trajectory record web_tools: false. Without the flag nothing changes.
Tests: with the flag both calls get the engine's unknown-tool answer and the record says false; without it web_fetch is present, shown with a non-URL the tool rejects before any network I/O.
A failed model call prints what was underneath it, such as a proxy's 403.
SIGTERM during the wait before a retry now stops the run and leaves its record (exit 130).
--mode, --persona, --attendance and the unlisted mode names are no longer accepted.
Flags cannot be shortened, so --mode is refused instead of being read as --model.
Default mode is ask; approvals and questions are separate (--auto-answer); --isolated keeps the sealed benchmark run.
Without --out the record goes to sessions/<workspace path>/<session id> in the state folder; --allow-site adds sites for the run.
Exit code 0 only when the task finished; 3 when it stopped early with its record written.
"root" already means the superuser and the top of the filesystem; `--add-dir` says what
the flag does and is the name the best-known agent CLIs use for the same idea. Recorded
in summary.json as args.extra_dirs.
A harness can put its output contract outside the workspace: CyberGym-E2E runs the agent
in /src and grades /output/fix.patch and /output/poc.bin. OpenWorker's file tools only
write under a declared root, in every permission mode (the dangerous mode clears the
permission floor, but the tools still resolve paths against the session's roots by
construction), while the shell is not scoped. So without a way to declare /output, a
delivery depended on which tool the model happened to pick.
`--root DIR` (repeatable) adds writable roots after the workspace, which stays the folder
relative paths resolve against; the folders are announced to the model in the session's
directories block and recorded under args.roots in summary.json. The reference harness
does the same for Gemini CLI with --include-directories /src,/scripts,/output.
A headless one-shot command beside `up`: give it a task, a folder and a model; it builds the same engine the desktop app uses, runs to the end with attendance `auto` (coworker/unattended.py), and leaves a record: every event, the final conversation, a summary with tokens and cost, every answer the engine gave on the absent user's behalf, the audit log, a per-call model log, and a trajectory in the ATIF interchange format. The trajectory and the exit code are the stable contract; the other files are internal.
State and scratch folders are pointed under --out before the engine starts, so a run never touches the machine's own conversations, keys or settings. A transient provider failure or an empty reply is waited out with backoff and a nudge instead of ending the run; permanent errors are not retried. Tool results are bounded with the full text spilled beside the record. A small price table gives an estimated cost per call as a cross-check; an unpriced model gets counts but no cost.
The mode axis decides approvals (`dangerously-bypass-approvals` prints its warning line here); the runner adds no policy of its own.