agl 43.0.4 → 43.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/changelog.md +25 -0
- package/dist_serve/bundle.js +5186 -4683
- package/dist_serve/bundle.js.map +2 -2
- package/dist_serve/bundle.js.third-party-notices.json +29 -29
- package/dist_serve/bundle.js.third-party-notices.txt +29 -29
- package/dist_serve/third-party-notices.md +1 -1
- package/dist_ts/00_commitinfo_data.js +1 -1
- package/dist_ts/aglapi/aglapisurface.document.js +475 -4
- package/dist_ts/aglapi/interfaces.d.ts +10 -0
- package/dist_ts/aglapi/interfaces.js +3 -1
- package/dist_ts/classes.claudeharness.d.ts +14 -5
- package/dist_ts/classes.claudeharness.js +62 -2
- package/dist_ts/classes.codexclientadapter.d.ts +7 -5
- package/dist_ts/classes.codexclientadapter.js +38 -3
- package/dist_ts/classes.controller.d.ts +9 -2
- package/dist_ts/classes.controller.js +116 -24
- package/dist_ts/classes.flexservice.d.ts +6 -7
- package/dist_ts/classes.flexservice.js +12 -45
- package/dist_ts/classes.flexsupervisor.d.ts +37 -3
- package/dist_ts/classes.flexsupervisor.js +164 -37
- package/dist_ts/classes.inferencegateway.d.ts +8 -5
- package/dist_ts/classes.inferencegateway.js +62 -19
- package/dist_ts/classes.inferencespend.d.ts +9 -3
- package/dist_ts/classes.inferencespend.js +44 -16
- package/dist_ts/classes.mcpclient.d.ts +6 -4
- package/dist_ts/classes.mcpclient.js +6 -2
- package/dist_ts/classes.mcpcontrollerhost.js +5 -1
- package/dist_ts/classes.mcptools.d.ts +1 -1
- package/dist_ts/classes.mcptools.js +17 -4
- package/dist_ts/classes.opencodeclient.js +3 -1
- package/dist_ts/classes.proxycontrol.d.ts +22 -11
- package/dist_ts/classes.proxycontrol.js +6 -3
- package/dist_ts/classes.proxycontrolclient.d.ts +6 -5
- package/dist_ts/classes.proxycontrolclient.js +8 -6
- package/dist_ts/classes.proxydaemon.d.ts +8 -6
- package/dist_ts/classes.proxydaemon.js +35 -21
- package/dist_ts/classes.proxyleaseholder.d.ts +12 -4
- package/dist_ts/classes.proxyleaseholder.js +39 -9
- package/dist_ts/classes.proxystore.d.ts +4 -3
- package/dist_ts/classes.proxystore.js +17 -7
- package/dist_ts/classes.sessionfeed.d.ts +1 -1
- package/dist_ts/classes.validation.d.ts +7 -0
- package/dist_ts/classes.validation.js +49 -2
- package/dist_ts/driver/functions.drivertools.js +2 -2
- package/dist_ts/fleet/classes.codexworkers.d.ts +4 -11
- package/dist_ts/fleet/classes.codexworkers.js +6 -3
- package/dist_ts/fleet/classes.workers.d.ts +25 -13
- package/dist_ts/fleet/classes.workers.js +105 -9
- package/dist_ts/fleet/functions.workervalidation.d.ts +26 -1
- package/dist_ts/fleet/functions.workervalidation.js +40 -11
- package/dist_ts/fleet/interfaces.workers.d.ts +17 -11
- package/dist_ts/fleet/interfaces.workers.js +10 -1
- package/dist_ts/flexharness.child.js +9 -4
- package/dist_ts/functions.confinedrelay.d.ts +21 -0
- package/dist_ts/functions.confinedrelay.js +65 -0
- package/dist_ts/functions.deviceapi.d.ts +4 -2
- package/dist_ts/functions.deviceapi.js +20 -2
- package/dist_ts/functions.elicitation.d.ts +49 -0
- package/dist_ts/functions.elicitation.js +418 -0
- package/dist_ts/functions.flexchildshutdown.d.ts +7 -0
- package/dist_ts/functions.flexchildshutdown.js +18 -1
- package/dist_ts/functions.inferencecli.d.ts +3 -3
- package/dist_ts/functions.inferencecli.js +33 -10
- package/dist_ts/functions.inferencecliarguments.d.ts +2 -1
- package/dist_ts/functions.inferencecliarguments.js +29 -8
- package/dist_ts/functions.inferencelaunch.d.ts +35 -7
- package/dist_ts/functions.inferencelaunch.js +143 -16
- package/dist_ts/functions.mcpfleetworkertools.d.ts +63 -70
- package/dist_ts/functions.mcpfleetworkertools.js +32 -4
- package/dist_ts/functions.proxycontract.d.ts +21 -4
- package/dist_ts/functions.proxycontract.js +12 -5
- package/dist_ts/functions.proxyharnessconfig.d.ts +7 -0
- package/dist_ts/functions.proxyharnessconfig.js +11 -1
- package/dist_ts/interfaces.auth.d.ts +1 -1
- package/dist_ts/interfaces.auth.js +2 -1
- package/dist_ts_interfaces/devices.d.ts +133 -3
- package/dist_ts_interfaces/devices.js +3 -2
- package/dist_ts_interfaces/elicitations.d.ts +125 -0
- package/dist_ts_interfaces/elicitations.js +7 -0
- package/dist_ts_interfaces/fleetworkers.d.ts +119 -5
- package/dist_ts_interfaces/fleetworkers.js +14 -2
- package/dist_ts_interfaces/index.d.ts +1 -0
- package/dist_ts_interfaces/index.js +2 -1
- package/dist_ts_interfaces/interfaces.d.ts +3 -1
- package/dist_ts_interfaces/interfaces.js +2 -2
- package/dist_ts_interfaces/mcprequests.d.ts +5 -1
- package/dist_ts_interfaces/mcprequests.js +2 -1
- package/dist_ts_interfaces/nativesessions.d.ts +14 -0
- package/dist_ts_interfaces/requests.d.ts +3 -3
- package/dist_ts_interfaces/requests.fleetworkers.js +2 -2
- package/dist_ts_interfaces/requests.js +1 -1
- package/dist_ts_interfaces/requests.promptreply.d.ts +18 -0
- package/dist_ts_interfaces/sessionfeed.d.ts +3 -2
- package/dist_ts_interfaces/sessionfeed.js +1 -1
- package/dist_ts_migration/index.d.ts +0 -2
- package/dist_ts_migration/index.js +1 -3
- package/dist_ts_migration/v43_flexpostharnessmigrationmarkers.d.ts +6 -17
- package/dist_ts_migration/v43_flexpostharnessmigrationmarkers.js +6 -36
- package/package.json +3 -3
- package/readme.md +191 -24
- package/third-party-notices.md +1 -1
- package/ts/00_commitinfo_data.ts +1 -1
- package/ts/aglapi/aglapisurface.document.ts +474 -3
- package/ts/aglapi/interfaces.ts +2 -0
- package/ts/classes.claudeharness.ts +80 -9
- package/ts/classes.codexclientadapter.ts +41 -5
- package/ts/classes.controller.ts +142 -29
- package/ts/classes.flexservice.ts +10 -66
- package/ts/classes.flexsupervisor.ts +214 -37
- package/ts/classes.inferencegateway.ts +69 -23
- package/ts/classes.inferencespend.ts +46 -13
- package/ts/classes.mcpclient.ts +8 -4
- package/ts/classes.mcpcontrollerhost.ts +4 -0
- package/ts/classes.mcptools.ts +17 -3
- package/ts/classes.opencodeclient.ts +2 -0
- package/ts/classes.proxycontrol.ts +27 -13
- package/ts/classes.proxycontrolclient.ts +9 -6
- package/ts/classes.proxydaemon.ts +49 -27
- package/ts/classes.proxyleaseholder.ts +43 -11
- package/ts/classes.proxystore.ts +21 -8
- package/ts/classes.sessionfeed.ts +1 -1
- package/ts/classes.validation.ts +52 -0
- package/ts/driver/functions.drivertools.ts +1 -1
- package/ts/fleet/classes.codexworkers.ts +8 -6
- package/ts/fleet/classes.workers.ts +101 -17
- package/ts/fleet/functions.workervalidation.ts +35 -9
- package/ts/fleet/interfaces.workers.ts +11 -4
- package/ts/flexharness.child.ts +8 -3
- package/ts/functions.confinedrelay.ts +65 -0
- package/ts/functions.deviceapi.ts +24 -1
- package/ts/functions.elicitation.ts +443 -0
- package/ts/functions.flexchildshutdown.ts +22 -0
- package/ts/functions.inferencecli.ts +38 -13
- package/ts/functions.inferencecliarguments.ts +26 -5
- package/ts/functions.inferencelaunch.ts +143 -16
- package/ts/functions.mcpfleetworkertools.ts +34 -4
- package/ts/functions.proxycontract.ts +17 -4
- package/ts/functions.proxyharnessconfig.ts +11 -0
- package/ts/interfaces.auth.ts +1 -0
- package/ts_interfaces/devices.schema.json +476 -2
- package/ts_interfaces/devices.ts +146 -2
- package/ts_interfaces/elicitations.ts +148 -0
- package/ts_interfaces/fleetworkers.ts +96 -4
- package/ts_interfaces/index.ts +1 -0
- package/ts_interfaces/interfaces.ts +3 -1
- package/ts_interfaces/mcprequests.ts +1 -0
- package/ts_interfaces/nativesessions.ts +16 -0
- package/ts_interfaces/requests.fleetworkers.ts +1 -1
- package/ts_interfaces/requests.promptreply.ts +22 -0
- package/ts_interfaces/requests.ts +4 -0
- package/ts_interfaces/sessionfeed.ts +3 -2
- package/ts_migration/index.ts +0 -2
- package/ts_migration/v43_flexpostharnessmigrationmarkers.ts +6 -49
- package/ts_web/00_commitinfo_data.ts +1 -1
- package/ts_web/classes.elicitationcards.ts +106 -0
- package/ts_web/classes.nativesessionstate.ts +38 -1
- package/ts_web/elements.fleetworker.ts +27 -2
- package/ts_web/elements.harnesscontrollerapp.ts +68 -1
- package/ts_web/functions.sessionchat.ts +5 -0
- package/dist_ts_migration/classes.flexpostharnessmigrationrunner.d.ts +0 -19
- package/dist_ts_migration/classes.flexpostharnessmigrationrunner.js +0 -34
- package/dist_ts_migration/v19_flexprojectmanagement.d.ts +0 -14
- package/dist_ts_migration/v19_flexprojectmanagement.js +0 -23
- package/ts_migration/classes.flexpostharnessmigrationrunner.ts +0 -54
- package/ts_migration/v19_flexprojectmanagement.ts +0 -32
package/readme.md
CHANGED
|
@@ -344,6 +344,8 @@ agl proxy run claude --provider openai --model <catalog-model> --
|
|
|
344
344
|
agl proxy run opencode --model <catalog-model> -- run 'Explain this repository'
|
|
345
345
|
agl proxy confined run -- docker run --rm --user "$(id -u):$(id -g)" -i <image> claude -p # Claude Code in a container
|
|
346
346
|
agl proxy confined run --provider openai -- docker run --rm --user "$(id -u):$(id -g)" -i <image> claude -p
|
|
347
|
+
agl proxy confined run --harness codex -- docker run --rm --user "$(id -u):$(id -g)" <image> \
|
|
348
|
+
node /run/agl-proxy-relay/agl-relay.mjs -- codex exec … # Codex in a container
|
|
347
349
|
agl proxy confined end <run id>
|
|
348
350
|
```
|
|
349
351
|
|
|
@@ -787,10 +789,11 @@ request on its own side cannot be seen in the request or its answer, so AGL cann
|
|
|
787
789
|
|
|
788
790
|
#### Confined containers
|
|
789
791
|
|
|
790
|
-
`agl proxy confined run` runs Claude Code
|
|
791
|
-
|
|
792
|
-
|
|
793
|
-
|
|
792
|
+
`agl proxy confined run` runs Claude Code, or with `--harness codex` Codex (see
|
|
793
|
+
[Codex in a confined container](#codex-in-a-confined-container)), in a Docker container that reaches AGL through one
|
|
794
|
+
route only: the proxy's confined listener, a Unix socket alone in a directory of its own (`agl-proxy-confined/proxy.sock`
|
|
795
|
+
beside the proxy's control socket; `agl proxy status` names it). Each Claude Code run is served from one pool, chosen when
|
|
796
|
+
it starts: the Claude pool, or with `--provider openai` the ChatGPT pool. The listener serves the three endpoints Claude Code uses --
|
|
794
797
|
`GET /anthropic/v1/models`, `POST /anthropic/v1/messages` and `POST /anthropic/v1/messages/count_tokens` -- exactly as
|
|
795
798
|
the loopback listener serves Claude Code on the run's pool: on the Claude pool sent on to Anthropic as Claude Code sent
|
|
796
799
|
them, on the ChatGPT pool translated (see [Claude Code on the ChatGPT pool](#claude-code-on-the-chatgpt-pool), where
|
|
@@ -805,7 +808,7 @@ agl proxy confined run -- docker run --rm --user "$(id -u):$(id -g)" -i <image>
|
|
|
805
808
|
claude -p --model claude-opus-5-5 --output-format json < prompt.txt
|
|
806
809
|
agl proxy confined run --provider openai -- docker run --rm --user "$(id -u):$(id -g)" -i <image> \
|
|
807
810
|
claude -p --output-format json < prompt.txt # on the ChatGPT pool's default model
|
|
808
|
-
agl proxy status # names the confined listener and lists the runs by id, with their pools
|
|
811
|
+
agl proxy status # names the confined listener and lists the runs by id, with their harnesses and pools
|
|
809
812
|
agl proxy confined end <run id> # ends a run's credential at once
|
|
810
813
|
```
|
|
811
814
|
|
|
@@ -861,14 +864,139 @@ with the reason; it does the same when the proxy does not hand the credential ba
|
|
|
861
864
|
|
|
862
865
|
Downgrade: an earlier proxy refuses to start while its store keeps a confined run's credential its check of the kept
|
|
863
866
|
leases does not know: any confined run's for a release before confined runs, a ChatGPT-pool run's for one before
|
|
864
|
-
ChatGPT-pool runs (41.0.1 and earlier)
|
|
865
|
-
|
|
866
|
-
|
|
867
|
-
|
|
867
|
+
ChatGPT-pool runs (41.0.1 and earlier), a Codex run's for one before Codex runs (43.0.5 and earlier). Let every confined
|
|
868
|
+
run end before downgrading (`agl proxy status` lists them, `agl proxy confined end <run id>` ends one); if an earlier
|
|
869
|
+
proxy's start already failed, start this release once with no confined run going and wait out the 60-second reclaim
|
|
870
|
+
window, which ends the kept credential, then downgrade. A launcher older than ChatGPT-pool runs still runs Claude-pool
|
|
871
|
+
runs on this proxy; an older proxy refuses this release's `--provider openai`, as it issues no ChatGPT-pool run, and an
|
|
872
|
+
older proxy refuses `--harness codex`, which the launcher reports as a proxy older than Codex runs. An older launcher's
|
|
873
|
+
`agl proxy status` lists a Codex run by its pool alone (`ChatGPT pool`).
|
|
874
|
+
|
|
875
|
+
#### Codex in a confined container
|
|
876
|
+
|
|
877
|
+
`agl proxy confined run --harness codex -- docker run …` runs Codex CLI (`codex exec`) in a container on the same route
|
|
878
|
+
(AGL-1614). A Codex run is of the ChatGPT pool alone (`--provider openai` may name it; `--provider anthropic` is
|
|
879
|
+
refused) and is served by the pool's ChatGPT accounts alone, never by an OpenAI API key: when none of them can take a
|
|
880
|
+
request it is refused as Codex is on the loopback listener (503 without one, the usage limit and its reset when each is at
|
|
881
|
+
its limit) and never sent to a key. Its credential is a lease whose sessions are those of the
|
|
882
|
+
client `confined-codex`, so its requests are placed, moved and counted (token usage under `confined-codex`) like any
|
|
883
|
+
other client's of the pool, and like a Claude Code run's it ends with the container and survives a proxy restart.
|
|
884
|
+
`agl proxy status` names such a run `Codex on the ChatGPT pool (ChatGPT accounts only)`, and its JSON
|
|
885
|
+
`confinedRuns[].harness` is `codex` (`claude` for a Claude Code run).
|
|
886
|
+
|
|
887
|
+
The confined listener serves a Codex run two endpoints, as the loopback listener serves Codex on the ChatGPT pool:
|
|
888
|
+
`GET /openai/v1/models` (with `client_version`, Codex's own catalog of the pool's ChatGPT accounts for that release;
|
|
889
|
+
without it the pool's models in OpenAI's list shape) and `POST /openai/v1/responses`. It refuses every other path and
|
|
890
|
+
method -- every `/anthropic/` path, `/openai/v1/messages`, `/openai/v1/responses/compact`, `/memories/…`,
|
|
891
|
+
`/realtime/…` among them -- and every request that asks to upgrade its connection with 404 before anything is read or
|
|
892
|
+
routed. A Responses request that declares a hosted tool, a tool the provider runs (`web_search`, `file_search`,
|
|
893
|
+
`code_interpreter`, `image_generation`, a remote `mcp` server, a hosted `shell` or `tool_search`; in `tools` or in the
|
|
894
|
+
tools its input adds), or whose `tool_choice` names one (also within `allowed_tools`), is refused with 400
|
|
895
|
+
`hosted_tool_refused` before it is routed; Codex's client tools (functions,
|
|
896
|
+
custom tools such as `apply_patch`, the local shell, `tool_search` with `execution: "client"`) pass. A Codex run's
|
|
897
|
+
refusals take OpenAI's error shape (`{"error": {"type": …, "code": …, "message": …}}`), as does any refusal on the
|
|
898
|
+
listener of an `/openai/` path before its credential is known.
|
|
899
|
+
|
|
900
|
+
Codex has no Unix-socket transport for its model provider, so a Codex run adds AGL's relay. The command writes
|
|
901
|
+
`agl-relay.mjs` (plain Node.js, no dependencies) into a directory of the run's own (0700, in the temporary directory),
|
|
902
|
+
mounts it read-only at `/run/agl-proxy-relay`, and removes it when the container exits. The container's command runs
|
|
903
|
+
Codex through it: `node /run/agl-proxy-relay/agl-relay.mjs -- codex exec …`. The relay listens on `127.0.0.1:4100`
|
|
904
|
+
inside the container (set `AGL_PROXY_RELAY_PORT` for another port), pipes each connection byte for byte to the confined
|
|
905
|
+
socket and to nothing else (a connection the socket does not take is closed), then runs the command, passes it SIGTERM,
|
|
906
|
+
SIGINT and SIGHUP, and exits with its status (128 plus the signal's number for a command a signal ended, 127 for one
|
|
907
|
+
that could not start). The image needs `node`; the container needs no network, and no bridge or host address is used.
|
|
908
|
+
Besides the two mounts the command passes by name `AGL_PROXY_SOCKET` (`/run/agl-proxy/proxy.sock`, which the relay
|
|
909
|
+
connects to) and `MODELPROFILE_PROXY_TOKEN` (the run's credential, for Codex's provider command), and no model variable:
|
|
910
|
+
Codex reads the pool's catalog, and its command names the model. It removes from the environment `docker` runs with
|
|
911
|
+
every variable that would give Codex another login, route or home -- `OPENAI_API_KEY`, `OPENAI_BASE_URL`,
|
|
912
|
+
`OPENAI_ORGANIZATION`, `OPENAI_PROJECT`, `OPENAI_IDENTITY_TOKEN_FILE`, `OPENAI_WORKLOAD_IDENTITY_CONTEXT`,
|
|
913
|
+
`OPENAI_FEDERATION_RULE_ID`, `CODEX_API_KEY`, `CODEX_ACCESS_TOKEN`, `CODEX_AUTH`, `CODEX_REMOTE_AUTH_TOKEN`,
|
|
914
|
+
`CODEX_REMOTE_TOKEN`, `CODEX_HOME`, `CODEX_CA_CERTIFICATE`, `CODEX_OSS_BASE_URL`, `CODEX_OSS_PORT` -- and Claude Code's,
|
|
915
|
+
so a `-e NAME` of one passes nothing; the container names its own `CODEX_HOME` with a value. The socket and the relay are
|
|
916
|
+
the proxy user's, so the container runs as that uid (`--user`), or as root under rootless Docker.
|
|
917
|
+
|
|
918
|
+
A Codex run's container needs a network namespace of its own (`--network none`, or Docker's default when it names none):
|
|
919
|
+
the relay's `127.0.0.1` port is that namespace's loopback. The command refuses `--network host` and `--network
|
|
920
|
+
container:…` in every spelling Docker takes (`--network=…`, `--net …`, `--net=…`, and Docker's long form of
|
|
921
|
+
comma-separated `key=value` fields, such as `name=host,alias=x`), and a long-form value whose network it cannot read
|
|
922
|
+
(no `name` field, a field without `=`, or quotes): on the host's network the relay's
|
|
923
|
+
port would sit on the host's loopback beside the proxy's own `4100`, and with `AGL_PROXY_RELAY_PORT` set it would open
|
|
924
|
+
the confined route to every process of the host over TCP; another container's network would open it to that container.
|
|
925
|
+
|
|
926
|
+
The run's credential is in the container's environment, so every process in the container that runs as the same uid
|
|
927
|
+
can read it, from the relay's or Codex's `/proc/<pid>/environ` as from its own environment. That is the exposure of a
|
|
928
|
+
Claude Code run's `ANTHROPIC_AUTH_TOKEN` as well: the container is the boundary, and the credential serves only the
|
|
929
|
+
confined route and ends with the run.
|
|
930
|
+
|
|
931
|
+
Codex's provider in the container is the one `agl proxy run codex` gives Codex on the host, on the relay's port:
|
|
932
|
+
|
|
933
|
+
```toml
|
|
934
|
+
model_provider = "agl_proxy" # never "openai", which Codex reserves
|
|
935
|
+
|
|
936
|
+
[model_providers.agl_proxy]
|
|
937
|
+
name = "AGL Proxy"
|
|
938
|
+
base_url = "http://127.0.0.1:4100/openai/v1"
|
|
939
|
+
wire_api = "responses"
|
|
940
|
+
supports_websockets = false
|
|
941
|
+
auth = { command = "printenv", args = ["MODELPROFILE_PROXY_TOKEN"], refresh_interval_ms = 0 }
|
|
942
|
+
```
|
|
943
|
+
|
|
944
|
+
The kit's side, as `test/e2e/e2e.proxyconfined.node.ts` runs it with Codex CLI 0.164.0 (its standalone static build) in
|
|
945
|
+
`debian:bookworm-slim` with `--network none --read-only --cap-drop ALL --security-opt no-new-privileges`, the host's
|
|
946
|
+
Node mounted:
|
|
947
|
+
|
|
948
|
+
```sh
|
|
949
|
+
agl proxy confined run --harness codex -- docker run --rm --network none --read-only --cap-drop ALL \
|
|
950
|
+
--security-opt no-new-privileges --user "$(id -u):$(id -g)" --tmpfs /tmp \
|
|
951
|
+
--mount type=bind,source=<codex home>,target=/codex-home --env CODEX_HOME=/codex-home --env HOME=/tmp \
|
|
952
|
+
--mount type=bind,source=<managed_config.toml>,target=/etc/codex/managed_config.toml,readonly <image> \
|
|
953
|
+
node /run/agl-proxy-relay/agl-relay.mjs -- codex exec --json -o /tmp/last-message.txt --strict-config \
|
|
954
|
+
--skip-git-repo-check --ignore-user-config --ignore-rules --dangerously-bypass-approvals-and-sandbox \
|
|
955
|
+
-m <model> -c model_reasoning_effort=<effort> '<prompt>'
|
|
956
|
+
```
|
|
957
|
+
|
|
958
|
+
```toml
|
|
959
|
+
# /etc/codex/managed_config.toml, read-only: the provider block above, and
|
|
960
|
+
web_search = "disabled"
|
|
961
|
+
check_for_update_on_startup = false
|
|
962
|
+
|
|
963
|
+
[features]
|
|
964
|
+
image_generation = false
|
|
965
|
+
apps = false
|
|
966
|
+
plugins = false
|
|
967
|
+
|
|
968
|
+
[analytics]
|
|
969
|
+
enabled = false
|
|
970
|
+
|
|
971
|
+
[feedback]
|
|
972
|
+
enabled = false
|
|
973
|
+
|
|
974
|
+
[shell_environment_policy]
|
|
975
|
+
exclude = ["MODELPROFILE_PROXY_TOKEN"]
|
|
976
|
+
```
|
|
977
|
+
|
|
978
|
+
- The model and the reasoning effort go on the command line (`-m`, `-c model_reasoning_effort=…`); everything else
|
|
979
|
+
comes from `/etc/codex/managed_config.toml`, which Codex 0.164.0 reads on a read-only root beside
|
|
980
|
+
`--ignore-user-config` (no `$CODEX_HOME/config.toml`) and `--ignore-rules` (no execpolicy `.rules` files);
|
|
981
|
+
`--strict-config` refuses any key Codex does not know, and every key above passes it.
|
|
982
|
+
- `CODEX_HOME` is the session's own, on a writable mount: Codex keeps the session's rollout transcript there
|
|
983
|
+
(`sessions/…/*.jsonl`, with the prompt and the answer), which a provenance audit can collect after the run.
|
|
984
|
+
- Hosted tools stay off in Codex (`web_search = "disabled"`, `features.image_generation = false`, apps and plugins off);
|
|
985
|
+
the route refuses a request that declares one with the 400 anyway. How Codex reports that refusal in a turn is not
|
|
986
|
+
verified here.
|
|
987
|
+
- `shell_environment_policy.exclude` is Codex's setting for keeping the run's credential out of the commands the model
|
|
988
|
+
runs; the e2e case runs no command, so that is not verified here.
|
|
989
|
+
- `codex exec` reads standard input when it is not a terminal and adds it to the prompt; give it `/dev/null` or the
|
|
990
|
+
prompt there.
|
|
991
|
+
- Releases: verified with Codex CLI 0.164.0 (2026-10-23 on npm). Upstream 0.163.0 (2026-10-15) and
|
|
992
|
+
`@tempfix/codex 0.162.0-tempfix.1` are not verified here; that their flags and keys match is unverified, and
|
|
993
|
+
`--strict-config` names any key a release does not know.
|
|
994
|
+
- The mount boundary, hooks and what the container may read are the kit's: AGL adds the two read-only mounts and the two
|
|
995
|
+
variables, nothing else.
|
|
868
996
|
|
|
869
997
|
## Native Codex Harness
|
|
870
998
|
|
|
871
|
-
Choose **New → Chat session** and pick the Codex harness to use Codex inside AGL. Native threads support text and reasoning streaming, tool activity, the Codex model and reasoning-effort catalog, command and file approvals, ordinary user-input questions, interruption, history pagination, rename, archive, and deletion. AGL Ask/Yolo controls apply to the exact pending native approval request; approvals never outlive its server request.
|
|
999
|
+
Choose **New → Chat session** and pick the Codex harness to use Codex inside AGL. Native threads support text and reasoning streaming, tool activity, the Codex model and reasoning-effort catalog, command and file approvals, ordinary user-input questions, MCP elicitations (see [MCP elicitations](#mcp-elicitations)), interruption, history pagination, rename, archive, and deletion. AGL Ask/Yolo controls apply to the exact pending native approval request; approvals never outlive its server request.
|
|
872
1000
|
|
|
873
1001
|
Install Codex CLI 0.156.0 or newer and sign in using Codex before starting AGL. The local connection uses the existing Codex app-server control socket when present. This is the WebSocket-over-Unix transport described by the [Codex app-server protocol](https://developers.openai.com/codex/app-server), at `CODEX_HOME/app-server-control/app-server-control.sock` (the default home is `~/.codex`). Codex publishes that path as a link to a socket in its private per-user directory, and AGL attaches only to that published link. AGL owns only its connection to that shared server, and accepts it only when the Codex home the server reports in `initialize` is AGL's own. If nothing exists at the path, AGL starts its own `codex app-server --listen stdio://` child, unless Codex's managed daemon serves this home: its state under `CODEX_HOME/app-server-daemon/` (`daemon.lock`, `settings.json`, `daemon.pid` or the older `app-server.pid`) proves it, and AGL then waits for the daemon instead of starting a second writer for the same threads. Codex never removes that state, so start the daemon with `codex app-server daemon start`; AGL rejoins it on its own or on **Reconnect**. Anything else there, including a socket bound directly at the path or a link whose socket is gone after the daemon died, is reported as a connection failure, as is a published socket that does not answer; AGL never starts a private server beside it. Set `AGL_CODEX_EXECUTABLE` to an absolute executable path when Codex is not on the controller's `PATH`; an explicit absolute `CODEX_HOME` is forwarded to the owned child.
|
|
874
1002
|
|
|
@@ -909,7 +1037,7 @@ Choose **New → Chat session** and pick **Claude Code** to run Claude Code insi
|
|
|
909
1037
|
Claude Code runs on the AGL inference proxy's Claude pool, never on a key or login of its own. The controller takes one lease of the pool and starts every Claude Code process with only the lease's endpoint and credential (`ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`). Its requests therefore go through the proxy's account placement, usage reads and limits like all other Claude traffic. A process runs with the controller's own Claude Code configuration folder (its `CLAUDE_CONFIG_DIR`, by default `~/.claude`), where Claude Code stores the transcript. While the proxy cannot give a lease, a send fails with `harness_unavailable`. When the lease ends, the processes on it are closed, and the next message starts one on a new lease.
|
|
910
1038
|
|
|
911
1039
|
- **Models.** The model picker offers the models the proxy's Claude pool lists, which are those every account of the pool offers. The effort picker offers the efforts the catalog states for the selected model, narrowed to those Claude Code lists for it (`supportedModels`) once one of its processes has listed the model. A process lists Claude Code's aliases and the model it started with, so a model no process has listed yet keeps the catalog's efforts. The pool names no default model. A conversation without a choice runs on the default model saved in **Settings** (`settings.defaultModels`) while the pool offers it, else on Claude Code's own default, which AGL does not name (see [Models And Provider Accounts](#models-and-provider-accounts)). If the pool does not offer Claude Code's own default, the turn fails with Claude Code's message. Changing the model or effort applies from the next turn.
|
|
912
|
-
- **Ask and Yolo.** Ask runs Claude Code in its default permission mode. Every tool call Claude Code asks about becomes a permission card, answered with **Allow** or **Reject**. Yolo makes AGL answer each card once, automatically; it never switches Claude Code to `bypassPermissions`. `AskUserQuestion` becomes a question card and `ExitPlanMode` a permission card carrying the plan.
|
|
1040
|
+
- **Ask and Yolo.** Ask runs Claude Code in its default permission mode. Every tool call Claude Code asks about becomes a permission card, answered with **Allow** or **Reject**. Yolo makes AGL answer each card once, automatically; it never switches Claude Code to `bypassPermissions`. `AskUserQuestion` becomes a question card and `ExitPlanMode` a permission card carrying the plan. An MCP server's request for input becomes an elicitation card, which Yolo never answers (see [MCP elicitations](#mcp-elicitations)).
|
|
913
1041
|
- **Transcript.** Thinking, text and tool calls stream into the transcript as Claude Code produces them. A `Task` subagent is shown as its tool card. The transcript is Claude Code's own: AGL reads it from Claude Code's store, so a conversation reopens with its history after a reload, a controller restart or an upgrade. Its next message resumes the same Claude Code conversation.
|
|
914
1042
|
- **Steer and Queue.** While AGL runs a turn, **Steer** hands the message to Claude Code, which takes it in between tool rounds or, if the turn ends first, runs it as its own turn right after. **Queue** keeps a follow-up in AGL's queue until the conversation is idle (see [Steering and pending messages](#steering-and-pending-messages)). A steer that a stop or a failure cancelled before Claude Code took it is held for the owner, as for Codex.
|
|
915
1043
|
- **Stop and failures.** **Stop** interrupts the running turn and keeps the partial reply. Where Claude Code supports it (`interrupt_cancel_queued_v1`), it also cancels the messages queued inside Claude Code. A failed turn leaves the conversation in the error state with Claude Code's message until the next turn starts, as a failed Codex turn does. A Claude Code process that ends during a turn fails that turn.
|
|
@@ -918,14 +1046,29 @@ Claude Code runs on the AGL inference proxy's Claude pool, never on a key or log
|
|
|
918
1046
|
|
|
919
1047
|
Claude Code conversations are separate from [Claude chats](#claude-chats), which are terminal resources running the `claude` CLI.
|
|
920
1048
|
|
|
1049
|
+
## MCP elicitations
|
|
1050
|
+
|
|
1051
|
+
An MCP server that a Claude Code conversation or a Codex thread runs can ask the owner for input (MCP `elicitation/create`; Claude Code hands it to AGL through the Agent SDK's `onElicitation`, Codex as `mcpServer/elicitation/request`). AGL holds each request as an elicitation card beside the conversation's permission and question cards, and the conversation needs input until the request is answered or withdrawn. Nothing answers an elicitation on the owner's behalf: auto-accept (Yolo, globally or for the conversation) answers permissions only.
|
|
1052
|
+
|
|
1053
|
+
Each card is one of three kinds, named by its `mode`, and carries the server's name, its message and when it asked:
|
|
1054
|
+
|
|
1055
|
+
- **Form** (`form`). The card shows the form's fields in the server's order: text (optionally an email address, an absolute URI, a date or a date and time, with lengths in characters), a number or a whole number with inclusive bounds, a yes or no, a single choice or several choices among strings, each choice with its value and label. This is MCP's flat primitive subset. **Send** (the accept) sends the values, typed: numbers as numbers, a yes or no as a boolean, several choices as a list, an empty optional field left out. The card checks each value as you type; the controller checks them again against the form before anything reaches the server and refuses a reply that does not fit (`invalid_request`, naming the field and its rule, never the value), so the server keeps waiting. A refusal shows on the card with its reason; the card keeps what you entered and takes an answer again. **Decline** refuses the request; **Cancel** dismisses it without a decision.
|
|
1056
|
+
- **URL** (`url`). The card shows the address as plain text, exactly as the server sent it: no link, and AGL never opens or fetches it. Visit it yourself if you trust it, then **Mark as done** (the accept) to report that you have done it; the answer carries no content. Decline and Cancel work as for a form.
|
|
1057
|
+
- **What AGL cannot collect** (`unsupported`). Another mode (Codex's `openai/form`, `openaiForm`, `openai/userVerification`, any later one), a nested object, a list of anything but string choices, an unknown string format, a malformed form or one larger than a card shows is never a form with a field left out: it is a card with the reason, and only **Decline** or **Cancel** answer it; it has no Accept.
|
|
1058
|
+
|
|
1059
|
+
- **Withdrawn requests.** A request the server or the harness withdraws (Claude Code's cancellation, Codex's `serverRequest/resolved`, a Claude Code process that ends) leaves no card, and a reply to it is refused (`not_found`). A Codex answer goes back on the JSON-RPC id Codex asked with.
|
|
1060
|
+
- **Claude Code's own hooks come first.** Claude Code runs the hooks of the settings it loads for AGL (user, project and local): an `Elicitation` hook can answer a request before AGL sees it, so no card appears, and an `ElicitationResult` hook sees the owner's answer and can change it before it reaches the server. Such hooks are the owner's own configuration; AGL neither adds nor removes them.
|
|
1061
|
+
- **Where it is answered.** Only the owner answers, through the browser (`controller.elicitation.reply`, `controller.nativesession.elicitation.reply`) or a paired device (`device.elicitation.reply`), audited as `elicitation.reply` with the session and request only. Agents have no way to answer an elicitation: no MCP tool, CLI command or fleet operation does. The values the owner enters go to Claude Code or Codex for the MCP server and nowhere else in AGL: no log, audit record, event, transcript entry of AGL's or error text carries them.
|
|
1062
|
+
- **Session reads.** Every session read carries `elicitations` beside `permissions` and `questions`, empty while none is pending and for harnesses that hand AGL none (OpenCode, Flex).
|
|
1063
|
+
|
|
921
1064
|
## MCP Server
|
|
922
1065
|
|
|
923
|
-
The installed `agl` command includes a unified stdio MCP server that combines
|
|
1066
|
+
The installed `agl` command includes a unified stdio MCP server that combines 96 tools on one connection:
|
|
924
1067
|
|
|
925
1068
|
- Nine CrossHarness tools for explicit OpenCode and Codex harness-server connections.
|
|
926
1069
|
- Six system inspection and reclaim-planning tools.
|
|
927
1070
|
- Thirteen durable `tstask` tools.
|
|
928
|
-
-
|
|
1071
|
+
- Sixty-eight AGL controller tools for projects, task discovery, control and messaging, models, scratchpads, resource lifecycle and attachment, managed jobs, driver fleet operations, the shared project ticket board, and challenge-only worker enrollment.
|
|
929
1072
|
|
|
930
1073
|
Configure an MCP host to execute:
|
|
931
1074
|
|
|
@@ -933,7 +1076,7 @@ Configure an MCP host to execute:
|
|
|
933
1076
|
agl mcp --port 4097
|
|
934
1077
|
```
|
|
935
1078
|
|
|
936
|
-
`--port` takes precedence over `AGL_CONTROLLER_PORT`; otherwise the controller port defaults to `4097`. `agl mcp` starts and exposes all
|
|
1079
|
+
`--port` takes precedence over `AGL_CONTROLLER_PORT`; otherwise the controller port defaults to `4097`. `agl mcp` starts and exposes all 96 tools without a running AGL controller. Only the AGL controller tools require a ready controller on the configured port; the CrossHarness, system, and durable-task tools remain available independently. `agl mcp` requires Linux x64 with Node.js `>=24.12.0 <27`.
|
|
937
1080
|
|
|
938
1081
|
On supported systems, a ready controller publishes a process- and generation-fenced private loopback endpoint under `$AGL_HOME/runtime/mcp`. The descriptor and bearer token remain owner-only runtime state. They are discovered and verified automatically; do not copy them into MCP configuration. The endpoint accepts only the explicitly registered, bounded AGL methods and is removed before controller shutdown proceeds.
|
|
939
1082
|
|
|
@@ -1530,11 +1673,31 @@ read or fenced. Running a failed refresh again finishes it on the recorded succe
|
|
|
1530
1673
|
it asks for the profile that successor runs: the same `--profile`, or none for a successor on the
|
|
1531
1674
|
recorded profile.
|
|
1532
1675
|
|
|
1533
|
-
The
|
|
1676
|
+
The sixteen MCP tools are `fleet_worker_start`, `fleet_worker_adopt`, `fleet_worker_resolve_lost`, `fleet_worker_reconfirm`, `fleet_worker_steer`,
|
|
1534
1677
|
`fleet_worker_report`, `fleet_worker_stop`, `fleet_worker_archive`, `fleet_worker_refresh`,
|
|
1535
|
-
`fleet_worker_move`, `fleet_worker_list`, `
|
|
1678
|
+
`fleet_worker_move`, `fleet_worker_list`, `fleet_worker_get`, `fleet_worker_status`, `fleet_worker_job_wait`,
|
|
1679
|
+
`fleet_worker_native_message` and `fleet_watch`. They share
|
|
1536
1680
|
the local-owner capability boundary. Mutation and report tools return job handles immediately;
|
|
1537
|
-
the CLI polls jobs and prints their final outcome. Worker status accepts the job `id
|
|
1681
|
+
the CLI polls jobs and prints their final outcome. Worker status accepts the job `id`: a fleet
|
|
1682
|
+
operation job, not a host job (`job_status`). `fleet_worker_job_wait` waits for that job until it
|
|
1683
|
+
succeeds or fails, at most `timeoutMs` (240000); `timedOut` tells which, and neither a timeout nor
|
|
1684
|
+
an abandoned wait cancels the job. Jobs are transient: a job the controller no longer holds (a
|
|
1685
|
+
controller restart, or eviction once 512 are held) is `JOB_UNAVAILABLE`, never failed; read the
|
|
1686
|
+
worker with `fleet_worker_get` then.
|
|
1687
|
+
|
|
1688
|
+
MCP `fleet_worker_list` returns summaries, newest created first, in pages of `limit` (1-100,
|
|
1689
|
+
default 50): id, name, harness, profile id, ticket, state, native id, cwd, `updatedAt`, `observed`
|
|
1690
|
+
(the native status, the newest turn's `outcome` and the `nativeWait` last observed) and `fence`
|
|
1691
|
+
flags (an unresolved dispatch with its action and since, a lost or reconfirmed dispatch, a recorded
|
|
1692
|
+
error), without attribution, proof or error bodies. `total`, `fenced` (unresolved dispatches) and
|
|
1693
|
+
`failed` (recorded errors and failed turn outcomes) count the whole filtered set, so a page never
|
|
1694
|
+
hides a fence. Continue with `nextCursor` and the same filter (`archived`, `ticketKey`, `state`,
|
|
1695
|
+
`harness`); a cursor of another filter is `LIST_CURSOR_MISMATCH` and one after a worker the
|
|
1696
|
+
registry no longer holds as it was `LIST_CURSOR_STALE`, never a silent restart. `view: 'full'`
|
|
1697
|
+
returns every full record unpaged, as `agl fleet worker list` and the private request do by
|
|
1698
|
+
default. `fleet_worker_get` (`name` or `id`) returns one worker's full record, archived ones
|
|
1699
|
+
included, with the native status, turn outcome and start, native wait and excerpt the controller
|
|
1700
|
+
last observed.
|
|
1538
1701
|
Worker mutations and reports require a ticket key; mismatched associations refuse before
|
|
1539
1702
|
native effects. Refresh and moves retain the ticket and the associating driver's attribution.
|
|
1540
1703
|
Existing unassociated bindings remain readable and may be explicitly adopted after proving
|
|
@@ -1609,6 +1772,9 @@ the rest counted), and, for a report the fleet runs, its worker's transcript wit
|
|
|
1609
1772
|
The transcript is read-only and projected like the worker's harness conversation
|
|
1610
1773
|
(`controller.fleet.worker.read`); a read returns the newest page of turns, and `earlier` says
|
|
1611
1774
|
whether older ones exist; while the worker works, the modal reads again every 3 seconds. The
|
|
1775
|
+
worker's own state shows over its transcript, here and wherever a worker opens: a failed last
|
|
1776
|
+
turn as an error with the limit that ended it, an interrupted one and a wait in Claude Code or
|
|
1777
|
+
Codex itself (which AGL cannot answer) as notes; the report's status line stays the report's. The
|
|
1612
1778
|
steer box messages the report with the owner's own message (`controller.directreport.message`):
|
|
1613
1779
|
the fleet delivers it through the dispatch fence of the report's `driverTicketKey`, headed
|
|
1614
1780
|
`[Message from the user via the AGL Driver view; not from your driver]`, and the request
|
|
@@ -2670,7 +2836,7 @@ A removal that fails is resumed with backoff, and it resumes across restarts and
|
|
|
2670
2836
|
|
|
2671
2837
|
### Conversations
|
|
2672
2838
|
|
|
2673
|
-
The sidebar is the session list (`controller.sessionfeed.list`, then `sessionfeed` deltas). It holds the sessions of every harness across every registered project: the conversations AGL tracks, the Flex and OpenCode sessions of a project folder that AGL does not track yet, the threads on the shared Codex app-server, Claude terminals, and the Claude Code sessions of the host's home and of the fleet profiles' own homes. A conversation that is tracked, run by a fleet worker and present on the Codex server is one row. A row needs input when AGL holds a question or
|
|
2839
|
+
The sidebar is the session list (`controller.sessionfeed.list`, then `sessionfeed` deltas). It holds the sessions of every harness across every registered project: the conversations AGL tracks, the Flex and OpenCode sessions of a project folder that AGL does not track yet, the threads on the shared Codex app-server, Claude terminals, and the Claude Code sessions of the host's home and of the fleet profiles' own homes. A conversation that is tracked, run by a fleet worker and present on the Codex server is one row. A row needs input when AGL holds a question, permission or MCP elicitation for it, or when the harness signals that it waits in its own client, which AGL cannot answer: a Codex thread waiting on an approval, on user input or on a flag AGL does not know, a Claude Code `requires_action` without a card, or a Claude Code job whose process its inventory lists as waiting without the job having failed (a job listed blocked whose turn ended, status idle, is completed); such a row carries `nativeWait` with its reason (`agl feed list --json`). A fleet worker whose newest turn failed (a failed Codex turn, with its limit kind, or a failed Claude Code job) is failed until a newer turn starts, which moves the row, or until Claude Code lists the job working, blocked or done again; an interrupted turn's row is completed with `interrupted`. Each row shows its harness, its state (working, needing input, completed or failed; a row waiting in its harness's own client says where to answer, such as "Approve in Codex" or "Answer in Claude Code", and a completed row whose last turn was interrupted says "Interrupted"), its name, the last reply's opening for tracked conversations and workers, the tickets it is linked to, and when it last changed; a newly created session joins at the feed's next pass. Selecting a row opens it: a tracked conversation opens, an untracked one is tracked first, a Claude terminal opens in the Terminals view, a session a fleet worker runs opens that worker, and any other Claude Code session or Codex thread opens where it runs, whatever its directory (see [Sessions outside every project](#sessions-outside-every-project)). Opening a session never registers a project: the project matters only when a new session is started. A row AGL cannot open says why in the sidebar, for example that its project is no longer registered. A ticket chip opens that ticket on the board. Right-click, the ContextMenu key or Shift+F10 opens the row's menu: a conversation's own menu, **Rename** and **Terminate** for a terminal, or **Copy session id**, **Copy directory** and **Make direct report…** (see [Driver session](#driver-session)) for a native session. Conversations archived in AGL leave the list.
|
|
2674
2840
|
|
|
2675
2841
|
A failed row of any kind leads its menu with **Acknowledge and remove** (`controller.sessionfeed.acknowledge`). The controller records, for the owner, that row's failure as of the time the row shows, and the row leaves the Failed group until the session moves again: new activity or a new failure brings it back. Nothing else changes: the session and its transcript, a fleet worker's dispatch fence and error, its tickets and the unknown-dispatch record stay as they are, and nothing is sent, resumed or stopped. The acknowledgements are kept in the controller's database and survive a restart; a row that changed since it was listed, is no longer failed or is gone is refused. From the command line the owner does the same with `agl feed` (see [CLI](#cli)); no agent, worker or MCP tool can acknowledge, so a worker never hides its own failure. The composer's open-sessions action shows the sidebar and moves focus into the list. The list header offers three actions.
|
|
2676
2842
|
|
|
@@ -3446,7 +3612,7 @@ The gateway source is the address the proxy's connections arrive from on this ho
|
|
|
3446
3612
|
|
|
3447
3613
|
## Paired Devices
|
|
3448
3614
|
|
|
3449
|
-
A phone or a desktop app pairs with a controller and then reaches it as a device: it lists the session feed, reads a session's transcript, sends a message to a session and answers its permissions and
|
|
3615
|
+
A phone or a desktop app pairs with a controller and then reaches it as a device: it lists the session feed, reads a session's transcript, sends a message to a session and answers its permissions, questions and MCP elicitations. A device reaches nothing else: no terminals, settings, accounts, tickets, browsers or files. It connects through the controller's public HTTPS origin (Remote Browser Access), so a controller in local mode pairs no device.
|
|
3450
3616
|
|
|
3451
3617
|
### Pairing a device
|
|
3452
3618
|
|
|
@@ -3466,9 +3632,9 @@ Settings → Devices lists every paired device with its model, when it paired, w
|
|
|
3466
3632
|
|
|
3467
3633
|
### The device protocol
|
|
3468
3634
|
|
|
3469
|
-
The device API is its own contract, separate from the browser's: `ts_interfaces/devices.ts` in the package, and the same contracts as JSON Schema in `ts_interfaces/devices.schema.json` (each contract's request and response, and every named type they reach). A device asks `device.hello` for the device protocol version (`major` 1, `minor`
|
|
3635
|
+
The device API is its own contract, separate from the browser's: `ts_interfaces/devices.ts` in the package, and the same contracts as JSON Schema in `ts_interfaces/devices.schema.json` (each contract's request and response, and every named type they reach). A device asks `device.hello` for the device protocol version (`major` 1, `minor` 1) and speaks one major; within it, changes only add. Minor 1 added MCP elicitations: `elicitations` in every `device.session.read` (empty while none is pending; each one a `form` with its `fields`, a `url` or an `unsupported` request with its `reason`) and `device.elicitation.reply`. A device does not tell the controller its minor, so a client ignores keys it does not know; the iOS (HarnessKit) and Android clients of minor 0 do. Apps generate their clients from these files.
|
|
3470
3636
|
|
|
3471
|
-
- Before it is authenticated a connection reaches `device.hello`, `device.pair`, `device.auth.challenge` and `device.auth.prove`, under the controller's limits for unauthenticated requests: per connection, and in total together with the browser's sign-in requests. After that it reaches `device.sessions.list`, `device.session.read`, `device.session.send`, `device.permission.reply` and `device.
|
|
3637
|
+
- Before it is authenticated a connection reaches `device.hello`, `device.pair`, `device.auth.challenge` and `device.auth.prove`, under the controller's limits for unauthenticated requests: per connection, and in total together with the browser's sign-in requests. After that it reaches `device.sessions.list`, `device.session.read`, `device.session.send`, `device.permission.reply`, `device.question.reply` and `device.elicitation.reply`, under the limits of the owner's browser. A device never reaches a method of the browser, and the browser never one of a device. A connection is one or the other for its whole life: a connection signed in as the browser cannot authenticate as a device, and one a device authenticated cannot sign in as the browser.
|
|
3472
3638
|
- The feed is the sidebar's session feed without terminal rows, host directories or tickets. A transcript page reduces a tool call to its name, title and status; its input and output stay on the controller. `device.session.read` without `before` returns the newest page, sized by the controller whatever `limit` says; with `before` it returns the next older page of `limit` messages (1 to 50, default 20). A device names a session only by the feed's row id, and reaches only what the feed lists.
|
|
3473
3639
|
- A message from a device is a direct message (it queues behind a running turn), never the browser's shared draft. A native session AGL does not write to (another process runs it, a fleet worker owns it) refuses it, as it refuses the browser.
|
|
3474
3640
|
- The controller pushes a device an allowlist of events in their own shape: feed changes, a session's change, its streaming text (snapshots and exact deltas) and a resynchronization request. Every other event, and every event about a session the feed does not list, stays on the controller. A device that falls behind by 256 events or 4 MiB is resynchronized alone.
|
|
@@ -3479,18 +3645,19 @@ The device API is its own contract, separate from the browser's: `ts_interfaces/
|
|
|
3479
3645
|
- Local mode binds the controller to loopback only, and so does TLS-proxy mode without a gateway source. With one, a request from off loopback needs a configured gateway source address and the trusted-proxy secret, and only loopback connects without an `Origin`.
|
|
3480
3646
|
- The trusted-proxy secret is generated and TPM2-sealed in the controller's database; it is never read from the environment, written to a file or logged.
|
|
3481
3647
|
- Only `agl remote`, run by the local OS user against a stopped controller, changes a port's public origin, RP ID, TLS mode or gateway sources, and a changed origin or RP ID removes the passkey (a changed origin every paired device too).
|
|
3482
|
-
- A paired device holds its own key; the controller stores only its public key. A device reaches only the device scope (feed, bounded reads, direct sends, permission and
|
|
3648
|
+
- A paired device holds its own key; the controller stores only its public key. A device reaches only the device scope (feed, bounded reads, direct sends, permission, question and elicitation answers), each connection proves the key over a fresh nonce, and revoking a device closes its connections at once. Pairing codes are single-use, live five minutes in memory only, and pairing is refused in local mode.
|
|
3483
3649
|
- The private OpenCode server always remains on loopback and uses a generated per-process password.
|
|
3484
3650
|
- Application RPC starts only after the exact TypedSocket package-major handshake.
|
|
3485
3651
|
- The management endpoint (`/management/typedrequest`, used by the CLI and the upgrade worker) answers only an HTTP `POST` of `application/json` from a loopback peer (`127.0.0.1` or `::1`), addressed to `127.0.0.1:<port>`, with no `Origin` and no forwarding header (`Forwarded`, `X-Forwarded-*`, `X-Real-IP`, `X-Harness-Controller-Proxy-Secret`), so a browser or a request a proxy forwards is refused, also in proxy mode. It serves only the status read and the upgrade methods; each upgrade method authorises itself against the upgrade token and its one-time grant, as it does over the socket, and every upgrade request counts against the controller's global unauthenticated preflight limit.
|
|
3486
3652
|
- Client-managed connection tags are disabled; authentication and routing use server-owned peer state.
|
|
3487
3653
|
- Setup attempts are globally rate limited before enrollment.
|
|
3488
3654
|
- Permission auto-accept is an explicit server policy, never an assumption based on an open browser.
|
|
3489
|
-
- Questions are never auto-answered.
|
|
3655
|
+
- Questions are never auto-answered, and neither are MCP elicitations.
|
|
3490
3656
|
- Unknown, malformed, concurrent, or unverifiable lifecycle state fails closed.
|
|
3491
3657
|
- The inference proxy's confined listener is a Unix socket of the proxy's user alone; it accepts only confined runs'
|
|
3492
3658
|
credentials and serves them only Claude Code's three endpoints of the run's pool, the Claude pool or, translated, the
|
|
3493
|
-
ChatGPT pool's ChatGPT accounts
|
|
3659
|
+
ChatGPT pool's ChatGPT accounts, or a Codex run Codex's catalog and Responses endpoint of the ChatGPT pool's ChatGPT
|
|
3660
|
+
accounts, without hosted tools.
|
|
3494
3661
|
|
|
3495
3662
|
The global **auto-accept permissions (yolo)** setting replies once to pending permissions across registered projects and writes those replies to the audit log. Each conversation also has a run-local Ask/Yolo switch. For an active managed OpenCode root, that switch covers its currently active direct children under the same exact family authority. For an active managed Flex root, it covers admitted descendants through their exact active parent-run authority chain; a relationship alone grants nothing. Terminal children are never auto-accepted. Disabling either policy seals new automatic replies and drains already admitted replies before returning. Neither mode creates persistent broad provider grants.
|
|
3496
3663
|
|
package/third-party-notices.md
CHANGED
|
@@ -50,7 +50,7 @@ inventory.
|
|
|
50
50
|
## Monaco Editor 0.55.2
|
|
51
51
|
|
|
52
52
|
The workspace editor's Monaco files are served unchanged from the installed
|
|
53
|
-
`@design.estate/monaco-editor` 0.55.2 package that `@design.estate/dees-catalog` 24.0
|
|
53
|
+
`@design.estate/monaco-editor` 0.55.2 package that `@design.estate/dees-catalog` 24.2.0
|
|
54
54
|
depends on, at `/monaco-editor/min/vs/` on the controller's own origin. AGL does not
|
|
55
55
|
bundle or modify them. This maintained fork updates the embedded DOMPurify to 3.4.17;
|
|
56
56
|
its `SECURITY-FORK.json` records the upstream commits and source hashes, and the package
|