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
{{ message }}
Repository navigation
Commit 8b78eef
Browse filesBrowse the repository at this point in the historyBrowse files
Process-exclusive elastic pooling for injected StructuredOutput runs (#293)
* fix: process-exclusive elastic pooling for injected StructuredOutput runs (#292)
Replace the per-connection FIFO that serialized every injected-tool schema
run (introduced in #56 for OpenCode's process-global, name-keyed MCP
registry) with space-based isolation: selection synchronously reserves a
process with no other injected run, spawning past the configured pool size
when all are reserved, and the reservation is released only after the
owning session.release() completes. Pi, OpenCode, and custom injecting
backends now overlap uniformly — parallel(N) schema runs execute N-wide —
while the #56 leak stays impossible: no two injected runs ever share a
process. Non-injected runs keep idle/grow-to-size/least-loaded multiplexing.
Surplus processes stay warm for a 30s keep-alive, then reap back to the
steady-state size; dispose()/forceKill() cover elastic and mid-reap
connections. The #56 serialization regression tests are retargeted to
overlap + distinct-process + distinct-capture assertions across the fake,
Pi, and OpenCode backends, and the live e2e now expects three elastic
subprocesses for injected backends.
Implemented-by: Codex (gpt-5.6-sol, xhigh)
* chore: dependency-gate maintenance — claude-agent-acp 0.62.0, codex-acp 1.6.12
Bump @agentclientprotocol/claude-agent-acp 0.61.0 -> 0.62.0, which wraps the
current @anthropic-ai/claude-agent-sdk 0.3.219; the root pnpm override pinning
0.3.218 is now stale and removed so the adapter's own exact pin wins. Sync the
codex-acp fork with upstream (models-availability e2e fix, test-only), fork
auto-released as @automatalabs/codex-acp 1.6.12; bump the exact pin. Welded
doc citations updated in docs/api.md, docs/design-notes.md, and the
acp-auth-spec §3.3 heading per the docs-drift suite.
* test(live-e2e): reroute opencode leg to deepseek-v4-flash
With process-exclusive injected pooling (#292) the gate's three opencode
schema runs fire concurrently, and kimi-k3 reliably drops one of the three
under burst (observed twice at the pre-push gate; the pi leg's identical
3-wide OpenRouter fan-out on another model passes consistently, ruling out
runner-side contention). Reroute the leg's default model per owner
direction; AGENTPRISM_OPENCODE_E2E_MODEL still overrides.
* fix: per-spawn OPENCODE_DB isolation for concurrent opencode processes
Concurrent OpenCode instances in the same project share the cwd-keyed
sqlite database (WAL) and interfere across sessions — upstream
anomalyco/opencode#31307. Process-exclusive injected pooling (#292) made
overlapping opencode processes routine, surfacing the interference as a
deterministic mid-run "ACP connection closed" on one of three concurrent
gate legs (full-suite 3/4 failure rate, model-independent: reproduced on
kimi-k3 and deepseek-v4-flash; solo legs and pi's identical 3-wide fan-out
always passed). spawnConfig() now assigns each spawned server its own
OPENCODE_DB under the OS tmpdir; auth.json is separate and unaffected, and
an explicitly exported OPENCODE_DB wins. Tradeoff documented in-line:
cross-process session/load|resume falls back to the runner's fresh-session
path. Full live 4-leg suite green with the isolation in place.
* fix: full per-spawn XDG isolation for opencode servers, not just OPENCODE_DB
Research into anomalyco/opencode's source and issue tracker showed the
shared surface is wider than the sqlite db: concurrent instances also share
one snapshot git index, the log, tool-output, and auth.json, and 1.18.5
still serializes all statements behind an untimed in-process semaphore, so
db contention freezes a whole process (ACP stdout pump included). Isolating
OPENCODE_DB alone is documented as insufficient (#33321). Each spawned
server now gets fresh XDG data/state/cache trees with auth.json/mcp-auth.json
seeded in and autoupdate disabled; XDG_CONFIG_HOME stays shared so user
config and providers apply, and an explicit OPENCODE_DB passes through.
Validated: 9/9 backend unit tests, hermetic suite green, full live 4-leg
e2e green with the opencode leg overlapping three isolated processes.
Copy file name to clipboardExpand all lines: docs/api.md
+5-5Lines changed: 5 additions & 5 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -1085,7 +1085,7 @@ quarantines the pooled child from new work, and disposes it after existing sibli
1085
1085
This policy is identical for every built-in and custom ACP backend. A `child_cleanup_error` returned
1086
1086
by close remains observable through the runner's normal error path.
1087
1087
1088
-
**Structured output channels.** Claude and Codex keep their agent-specific schema channels authoritative. Pi, OpenCode, and opted-in custom ACP backends use the client-hosted MCP path: when `RunOptions.schema` is set and initialize advertises HTTP MCP support, the runner appends a client-hosted HTTP MCP server to `session/new.mcpServers`. The injected server is named `structured_output` (or `structured_output_2`, etc. on name collision), runs on `127.0.0.1` with an unguessable token path, and exposes `StructuredOutput`; Pi shows the namespaced alias `mcp__structured_output__StructuredOutput`. Its input schema is the requested JSON Schema and a valid call captures the result. Injected-tool schema runs are serialized per pooled connection. The common prompt-embedded schema plus validated final-text ladder remains the fallback when capture is absent or invalid. User-provided `mcpServers` are preserved and are not part of the resume hash.
1088
+
**Structured output channels.** Claude and Codex keep their agent-specific schema channels authoritative. Pi, OpenCode, and opted-in custom ACP backends use the client-hosted MCP path: when `RunOptions.schema` is set and initialize advertises HTTP MCP support, the runner appends a client-hosted HTTP MCP server to `session/new.mcpServers`. The injected server is named `structured_output` (or `structured_output_2`, etc. on name collision), runs on `127.0.0.1` with an unguessable token path, and exposes `StructuredOutput`; Pi shows the namespaced alias `mcp__structured_output__StructuredOutput`. Its input schema is the requested JSON Schema and a valid call captures the result. Each injected-tool schema run reserves a pooled process exclusively from other injected runs; when every process is reserved, the pool grows elastically past `size`, then keeps surplus idle processes warm briefly before shrinking back to that steady-state size. The reservation remains held through `session.release()`, so two injected runs never share one process. Non-injected runs retain ordinary idle/grow-to-size/least-loaded multiplexing and may co-locate with an injected run. The common prompt-embedded schema plus validated final-text ladder remains the fallback when capture is absent or invalid. User-provided `mcpServers` are preserved and are not part of the resume hash.
1089
1089
1090
1090
**Model specs**: after the engine's existing precedence resolves one effective string, the runner splits it on the first `/`. If the first segment, ASCII-case-insensitively, is `claude`, `codex`, `opencode`, `pi`, or a registered custom backend name, that harness is selected and exactly one segment is stripped; a custom registration wins on collision. A registered harness name by itself is backend-only and issues no model `session/set_config_option`, preserving the harness default. Any other first segment sends the entire authored string unchanged to `AGENTPRISM_DEFAULT_BACKEND` (historical default `claude`), so `anthropic/…`, `openai/…`, bare `opus`, and bare `gpt-…` are not aliases. When an id remains it is the exact `configId:"model"` value: no case folding, normalization, catalog matching, bracket parsing, sibling effort/Fast option driving, retry, echo verification, or fallback. Brackets, dots, and provider-style prefixes are ordinary id characters. For pi, `pi/<provider>/<model-id>` therefore sends `<provider>/<model-id>` verbatim. Harness rejection follows the existing agent-error path; `onModelFallback` remains source-compatible but model resolution does not emit it. Live-catalog-verified examples: `claude/opus[1m]`, `codex/gpt-5.6-sol`, `opencode/zai/glm-5.2`; use backend-only `claude`, `codex`, `opencode`, or `pi` when the model is configured in the harness.
-`@agentclientprotocol/claude-agent-acp@0.61.0`: advertises `auth.logout`, implements `logout`, and implements `authenticate` for its gateway auth methods; terminal login methods are advertised only when the client advertises terminal auth support. As of 0.60.0 it advertises `providers` and implements `providers/list`, `providers/set`, and `providers/disable` for a single provider `providerId``"main"` supporting `apiType``anthropic`, `bedrock`, and `vertex`; `providers/set` rejects any other `providerId`/`apiType` with invalid-params, and `vertex` additionally requires `_meta.claudeCode.vertex.{projectId,region}` (recorded as durable routing config and replayed on every reconstructed `providers/set`, so pooled connections re-route correctly). `providers/disable` is idempotent.
1174
-
-`@automatalabs/[email protected].11`: advertises `auth.logout`, implements `authenticate` (`api-key`, `chat-gpt`, and `gateway` when gateway support is advertised), and implements `logout`. As of 1.6.0 (upstream sync) it also advertises `providers` and implements `providers/list`, `providers/set`, and `providers/disable` for its single client-configurable custom gateway provider: `providerId``"custom-gateway"`, `supported: ["openai"]`, `required: false`, `current` carrying only the non-secret `{ apiType, baseUrl }` (never headers) and `null` while unconfigured; `providers/set` rejects any other `providerId`/`apiType` with invalid-params and `providers/disable` is idempotent. Its separate reasoning-effort options remain agent-owned configuration; model-spec brackets are never interpreted by this client.
1173
+
-`@agentclientprotocol/claude-agent-acp@0.62.0`: advertises `auth.logout`, implements `logout`, and implements `authenticate` for its gateway auth methods; terminal login methods are advertised only when the client advertises terminal auth support. As of 0.60.0 it advertises `providers` and implements `providers/list`, `providers/set`, and `providers/disable` for a single provider `providerId``"main"` supporting `apiType``anthropic`, `bedrock`, and `vertex`; `providers/set` rejects any other `providerId`/`apiType` with invalid-params, and `vertex` additionally requires `_meta.claudeCode.vertex.{projectId,region}` (recorded as durable routing config and replayed on every reconstructed `providers/set`, so pooled connections re-route correctly). `providers/disable` is idempotent.
1174
+
-`@automatalabs/[email protected].12`: advertises `auth.logout`, implements `authenticate` (`api-key`, `chat-gpt`, and `gateway` when gateway support is advertised), and implements `logout`. As of 1.6.0 (upstream sync) it also advertises `providers` and implements `providers/list`, `providers/set`, and `providers/disable` for its single client-configurable custom gateway provider: `providerId``"custom-gateway"`, `supported: ["openai"]`, `required: false`, `current` carrying only the non-secret `{ apiType, baseUrl }` (never headers) and `null` while unconfigured; `providers/set` rejects any other `providerId`/`apiType` with invalid-params and `providers/disable` is idempotent. Its separate reasoning-effort options remain agent-owned configuration; model-spec brackets are never interpreted by this client.
1175
1175
- Host-resolved OpenCode (`opencode-ai` 1.17.14 in the verified profile): advertises the `opencode-login` terminal-style method when the client advertises terminal auth, acknowledges `authenticate`, and relies on its provider credential store; it does not advertise logout. The credential-gated live suite verifies the installed executable because OpenCode is not bundled.
1176
1176
-`@automatalabs/pi-acp`: unconditionally advertises `anthropic-api-key`, `openai-api-key`, `gemini-api-key`, `xai-api-key`, `openrouter-api-key`, and `pi-stored-credentials`. The first five are `env_var` methods; the stored-credentials method is a bare `agent` method backed by `~/.pi/agent/auth.json`. Authentication is ambient/no-op, while a known model with missing credentials rejects with ACP `-32000`.
1177
1177
@@ -1242,7 +1242,7 @@ await chat.prompt("Revise section 3 — the user wants X.");
1242
1242
1243
1243
Set `agent(prompt, { keepSession: true })` in the script (or `RunOptions.keepSession` on direct `run()` calls) when you intend to re-open: it skips the release-time best-effort `session/close`, guaranteeing the agent-persisted session is untouched. Without it the record is still surfaced, and the four first-class agents keep closed sessions loadable — but `keepSession` is the explicit, agent-agnostic contract. Check `reopen.load`/`reopen.resume` before offering re-attach and optional `reopen.fork` before offering a fork in UI: an agent that persists nothing advertises none of them, and its sessions are reachable only while held open (`openSession`). The fork flag is optional so records written before this field existed remain valid.
1244
1244
1245
-
Lifecycle methods are capability-gated after initialize. In particular, `forkSession()` requires `sessionCapabilities.fork`; when absent it throws a non-recoverable `WorkflowError` naming the backend and `session/fork` before any fork request is sent. The installed `@agentclientprotocol/claude-agent-acp@0.61.0` advertises `loadSession: true` plus `sessionCapabilities` for list/delete/resume/close/fork (fork verified live: the forked session carries the source conversation's context). `@automatalabs/[email protected].11` advertises `loadSession: true` plus list/delete/resume/close — no fork yet. OpenCode advertises load/list/resume/close/fork (also verified live). `@automatalabs/pi-acp` advertises load plus list/resume/close/fork and deliberately omits delete; unsupported lifecycle methods still fail through the same gate.
1245
+
Lifecycle methods are capability-gated after initialize. In particular, `forkSession()` requires `sessionCapabilities.fork`; when absent it throws a non-recoverable `WorkflowError` naming the backend and `session/fork` before any fork request is sent. The installed `@agentclientprotocol/claude-agent-acp@0.62.0` advertises `loadSession: true` plus `sessionCapabilities` for list/delete/resume/close/fork (fork verified live: the forked session carries the source conversation's context). `@automatalabs/[email protected].12` advertises `loadSession: true` plus list/delete/resume/close — no fork yet. OpenCode advertises load/list/resume/close/fork (also verified live). `@automatalabs/pi-acp` advertises load plus list/resume/close/fork and deliberately omits delete; unsupported lifecycle methods still fail through the same gate.
1246
1246
1247
1247
### Capabilities
1248
1248
@@ -1276,7 +1276,7 @@ Two gates run before any prompt tokens are spent:
1276
1276
- The agent must advertise `agentCapabilities.mcpCapabilities.acp === true`; otherwise the ACP server config fails with non-recoverable `SCRIPT_VALIDATION_ERROR`.
1277
1277
- The runner must have a complete `clientHandlers.mcp`; declaring `{ type: "acp" }` without a handler is also a non-recoverable config error.
1278
1278
1279
-
Installed backend status verified from the packaged dists: `@agentclientprotocol/claude-agent-acp@0.61.0` advertises `http`/`sse` MCP support but no `acp`, `@automatalabs/[email protected].11` advertises `mcpCapabilities: { acp: false, http: true, sse: false }` and rejects ACP MCP config internally, OpenCode advertises HTTP/SSE MCP support, and `@automatalabs/pi-acp` serves stdio, Streamable HTTP, and SSE while advertising `{ http:true, sse:true }`. Pi also consumes the stable MCP base protocol plus sampling, roots, and form/URL elicitation; client-hosted `acp` remains runner-owned.
1279
+
Installed backend status verified from the packaged dists: `@agentclientprotocol/claude-agent-acp@0.62.0` advertises `http`/`sse` MCP support but no `acp`, `@automatalabs/[email protected].12` advertises `mcpCapabilities: { acp: false, http: true, sse: false }` and rejects ACP MCP config internally, OpenCode advertises HTTP/SSE MCP support, and `@automatalabs/pi-acp` serves stdio, Streamable HTTP, and SSE while advertising `{ http:true, sse:true }`. Pi also consumes the stable MCP base protocol plus sampling, roots, and form/URL elicitation; client-hosted `acp` remains runner-owned.
0 commit comments