[brentyates-swx]

Refs swxtchio/swx-abbe#1952. Follows #347, which is merged as 774d90c7 and published as v1.18.32-swxtch.774d90c7.

Problem

On the published v1.18.32-swxtch.774d90c7, V2 turns in the OpenCode jail still never reach swxtch-egress. Abbe measured this at the jail boundary (swxtchio/swx-abbe#1952, 2026-10-23):

• A session created with only a location fails with ModelNotSelectedError.
• A session that names the egress model fails with ModelUnavailableError.
• In both cases the provider receives nothing, and /wait answers 503.

Cause

A Location loads its internal plugins (config, providers, agents, models.dev) in a background fiber, and nothing on the turn path waited for that. The jail's adapter creates a session and prompts it at once, so the drain that boots the Location runs its first turn while the catalog is still empty or only partly built. Depending on timing, the first turn:

• finds no model (ModelNotSelectedError);
• cannot find the named egress model (ModelUnavailableError); or
• is sent to a bundled model such as nano-gpt/... or vivgrid/....

A turn a few seconds later does reach egress. #347's tests waited for the catalog before resolving, so they hid the race.

Fix

• PluginInternal now provides a service with a ready signal. It completes when the background boot ends, whether the boot loaded every plugin or failed.
• SessionRunnerLLM waits for ready at the start of every turn, before it selects the agent, model, tools and guidance, since the same boot fills all of them.
• PluginInternal.readyLayer is the seam for tests and embedders whose Location services are already loaded. The runner tests now use it.

Upstream has the same race and no fix. Fork-only.

Proof at the jail boundary

The rig:

• Image: Abbe's arm64 build, opencode-variant-jail:b16-10-774d90c7-arm64, which is jail/Dockerfile.jail at swxtchio/swx-abbe#1951 head 03d7e234.
• Launch: started through its own entrypoint with the exact docker run arguments from opencode_container_spawn.py: --read-only, the tmpfs mounts, OPENCODE_CONFIG=/jail/opencode.json, OPENCODE_DISABLE_MODELS_FETCH=1, and serve --pure --print-logs.
• Network: an --internal Docker network, with a fake OpenAI-compatible provider where the egress proxy listens.
• Calls: sent the way opencode_client.py sends them: POST /api/session with only location (or with model), then /prompt with id and prompt, then /wait.

Image	Location only (3 runs)	Named egress model (3 runs)	Provider requests
published 774d90c7	ModelNotSelectedError with a 503 wait ×2, a bundled model with transport failure ×1	ModelUnavailableError ×3	0
the same image with only the binary and FFF sidecar replaced by this branch's build	completed ×3	completed ×3	1 per run, accounts/fireworks/models/deepseek-v4p1-flash with Bearer capability

Unit test

The test is in packages/core/test/location-layer.test.ts: "runs a location-only V2 turn admitted while its location boots on the jail's egress model". It follows the jail's real path:

• the jail's config sources (OPENCODE_CONFIG with placeholders, and the rendered OPENCODE_CONFIG_CONTENT);
• every internal plugin, booting against the recorded models.dev corpus;
• SessionExecutionLocal, with a session created with only a location, then prompted and resumed immediately, with no waiting;
• a recording LLM client.

Without the fix, the turn was sent to vivgrid/gpt-5.6-luna (red). With the fix it goes to swxtch-egress at the proxy URL. The earlier jail resolution test now waits on ready instead of polling the catalog.

Focused runs, all with no provider keys in the environment:

• packages/core: test/session-runner*.test.ts, test/location-layer.test.ts, test/catalog.test.ts, test/config/, test/plugin/ and test/effect/layer-node/, 451 pass.
• packages/opencode: test/server/httpapi-session.test.ts (78 pass) and httpapi-wait-lifecycle.test.ts (1 pass).
• bun typecheck is clean.

Abbe side

No producer change is needed. A session created with only a location, as create_session_v2 sends it, and a session that names the egress model both reach egress.

A turn now waits for its Location's plugin boot. If an external plugin's boot hangs (none in the jail, since it runs --pure), the turn waits until the client deadline or an interrupt, instead of running on a partial catalog.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HGGfNzV3BHB7W49PTT9MVp

[brentyates-swx]

Session ID: 34738ff3-3dcf-4587-8fde-7c58d5478796
Working repo: /home/byates/fm-worktrees/fm-llmrouter/opencode-v2-provider-1952/swx-opencode