You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Changed
A proxy relays one run and decides nothing, on every host, and a fresh one runs each call of the launcher.
Three pages gave the proxy three contracts: the shared call page's relay that never reads the result, the
operational proxy that decides covered requests and reads the report, and the shipped agent that hands every
request back. In Codex the proxy was the deciding one, kept one attached call and polled; on 2026-10-09 three
Luna proxies started a second watcher on one driver, accepted a clipped request and left --why empty on 26
accepts, and a read-only check took close to three hours. In Claude Code two Haiku proxies of three, sent the
same block after a decision, answered from their earlier turn and ran nothing. Now every host runs the same
block in a fresh small agent for each --run (the first, the one after a decision, a continuation's first),
names it after the external agent on every host, and the coordinator decides every request; the operational
proxy page is the main proxy mode's alone.
A read agent checks with its file tools, and its inputs sit inside its directory. Orchestrate told the
coordinator to give every worker capture-check.mjs, a shell command, while an OpenCode read agent asks for
every shell command and a Claude one for any command outside Claude Code's read-only set; on 2026-10-09 read
agents briefed that way asked 11 times (Claude, Sonnet) and 43 times (OpenCode, GLM) in one read-only task,
each question a stop for the coordinator. The script now goes to native workers and Codex agents; the shared
call page tells the coordinator to brief an OpenCode or Claude read agent in reads and searches, and to put an
OpenCode read agent's inputs inside its directory, since for one file outside it the agent asked for the whole
directory. Both drivers now tell a read agent that its read, grep and glob tools need no approval, and the
coordinator runs a check that needs a command before the launch and puts the result, the commit included, in
the task. On 2026-10-10 the same task asked once, from 111 calls of read, grep and glob and none of the shell.
Four rules a coordinator had to guess are on the pages. Read against one task on two hosts, the pages left
these to what a Claude coordinator already knew or a Codex one did not: whether a subagent can launch agents
(the Claude adapter said it can; a cloud session's subagent had no Agent tool), whether a pre-approved team lifts
the usage stop (a Codex coordinator read it so and ran past three times the pilot), which text the completeness
critic must have read (in both hosts a shorter message went out than the one it read), and whose TMPDIR an
external agent's checks write to (orchestrate promised each agent its own; only the Codex driver makes one).
Each now says what holds.
A refused accept of an OpenCode or Claude request names the frame it was printed in.--decide --accept
compared a typed request's body but told the caller to copy the lines between COMMAND<< and COMMAND>>,
which such a request does not print; on 2026-10-09 a proxy that restated an OpenCode permission's command
alone was sent looking for them twice. It now names REQUEST_BODY<< and REQUEST_BODY>> for a typed request.
The usage stop compares usage per unit. It compared an agent's total with the pilot's median per agent, so an
agent carrying more units than a pilot agent crossed it by size alone: on 2026-10-10 a Codex coordinator's pilot
agent read one ledger entry for 376,319 tokens, was continued for the other twelve for 2,470,884, six and a half
times the pilot and about half of it per entry, and the audit of that run reported the stop as missed.
The line is now three times the pilot's median per unit, counted in tokens from the agent's report, since a
second audit could not tell whether "usage" meant tokens or the catalog cost an OpenCode report also carries.
In Codex, a proxy runs the launcher outside the sandbox. The proxy's block was written for Claude Code's Bash
tool and said nothing of a sandbox; the driver starts the agent's CLI, which writes its state under the home
directory and reaches its provider, and Codex's sandbox allows neither by default. On 2026-10-10 a Codex proxy
given the block alone handed back an OpenCode server that had exited before its run began, where the day before
proxies that asked for escalated permissions had started theirs. The shared call page now gives a Codex proxy a
step 1 that asks for them, with a justification naming the agent, and the launcher refuses a --run or a launch
made under Codex's macOS sandbox, which marks its commands CODEX_SANDBOX=seatbelt, before it starts anything:
its ERROR= says to run the same command again with escalated permissions, and the agent's directory stays
fresh for that call. The suites' own launches drop the mark.
In Codex, a proxy waits for the run to end before it hands back. Codex's shell tool comes back while a long
command runs, with a session id and no exit code, and the block's step 2 sent such an empty result on as it was.
On 2026-10-10 four of the first five Codex proxies that ran handed back nothing, or step 4's "report delivered"
line, while their agent went on working, and the coordinator read the report and the driver's processes itself.
A Codex proxy's step 2 now reads the command's session again with write_stdin until an exit code comes, and
its step 4 is left out, since its final message is its hand-back.
The proxy's command goes as written, with no variables before it. Orchestrate told the coordinator to pass
its temporary context on "worker commands", and on 2026-10-10 a Codex coordinator put TMPDIR, ENTRUST_STATE_DIR and ENTRUST_TEMP_CONTEXT, over 400 characters, before every proxy's --run; one brief
mistyped the launcher's path and its proxy was stopped. --run starts the driver in the working directory and
with the state directory --new recorded, which a new case in the launcher's suite checks from another directory
and another TMPDIR, and the pages now say the launcher's calls take none of it.
An approved plan is the permission to send an agent's provider its files, and in Codex the coordinator's
launcher calls run in the sandbox. Codex's automatic reviewer of escalated commands read a --new that sends
the project's files to OpenRouter as a transfer the user had not approved, though the brief the user pasted said
they had: on 2026-10-10 it refused the launch twice, and the run waited 15 minutes for the user's chat answer;
earlier that day the same refusal ended with the coordinator doing the task with the host's own model. Telling
the coordinator to get that approval in the chat first then held a later run 21 minutes on a pasted brief that
carried it, while that run's --plan, --new and --decide worked without escalation and never reached the
reviewer. Orchestrate now says that an approved plan, or a request naming the adapter or the model, is that
permission and no other is asked, runs those three calls unescalated, and turns to the user only when the
reviewer refuses an escalated one.
After a passing pilot, the rest runs side by side. The plan said only to expand when the pilot's rule passes,
and on 2026-10-10 a Codex coordinator continued its one GLM agent over the twelve remaining ledger entries in
three batches of four, one after another (6.1, 5.0 and 2.9 minutes), under a bound of six alive. The plan now
gives the remaining independent units to agents that run side by side within the bound, unless usage rather
than time is the limit.
An OpenCode server that ends before it serves says why. The driver read the server's output only for its
address and reported "exited before announcing its local URL" alone; on 2026-10-10 a Codex rerun of the
orchestrate task stopped there half a second after the server started, and its report held no exit code and
none of the server's words. The error now carries the exit code or signal and the last lines the server printed.
Validation
Locally, in a Linux container running as root on Node 22, the suites this release touches pass: agent-run 62,
agent-contract 15, opencode 84, claude 26 with its live case skipped, orchestrate 11, lint-draft 13,
capture-check 17, status 10, fragments, conformance, and package with its tag check skipped; scripts/skills.test.mjs passes 8 of 8 and terse's pages 12 of 12. The protocol, lock and cli suites fail in
that container as they do on the tree before this release, and pass in CI. CI's four jobs run on the release PR,
and a red run stops the release.
On the owner's machine, a Codex coordinator ran one read-only task on this branch three times: OpenCode GLM 5.3
agents through OpenRouter checking the 13 entries of ISSUES.md, each run relayed by a Luna proxy. At 35cfa6b
four of eleven proxies handed back before their run ended. At b13a374 four proxies handed back with no failure
and the first agent ran 5 minutes after the start rather than 22, but the verdicts counted an address that merely
existed. At d36a2c0 the verdicts matched, entry for entry, a reference made independently on the same commit;
four proxies, one permission request, and one escalated command of the coordinator's against 94 in the run
before, once the owner's instructions put the sandbox first.
Not run: the live orchestrate gate, the live fidelity gate and the cross-plugin terse brief; the preparing
machine has no Codex CLI.
Known issues
Open ledger entries remain E67, E90, E96, E100, E102–E104, E110, E111, E113, E114, E117 and E139; see ISSUES.md. The d36a2c0 audit found that ten of them cite text that has since moved, which the
ledger has not caught up with.
Release environment
Release preparation used Node.js v22.22.0 and Claude Code 2.1.296, with no Codex CLI. The owner's runs used
Codex 0.162.0-alpha.17.2 and OpenCode 1.18.35. The CLI protocol pin, 0.159.3, is unchanged.