YA Claw

Workspace-native single-node agent runtime and web service for the ya-mono workspace.

Scope

YA Claw packages a durable runtime shell around ya-agent-sdk with:

• one persistent workspace resolved through WorkspaceProvider
• reusable agent profiles
• explicitly selected installable capability plugins through an SDK-owned manifest
• resumable sessions and runs
• in-process active handles and live transports over SQL-owned async task state
• session schedules for timed execution
• SQLite-first durable state with optional PostgreSQL
• local filesystem session continuity and exported state
• a bundled web shell for local and self-hosted use
• bridge adapters that connect IM channels to the YA Claw service

Current Direction

The target single-node shape runs as one web service. Process memory owns active handles, live delivery transports, and best-effort scheduling coordination. Async-task state, completion intent, parent inbox/continuation targeting, and delivery disposition remain canonical relational state; only the SQL-backed post-run owner dispatches completion, and portable delivery completes only after canonical parent application. SQLite is the default durable store. PostgreSQL remains an optional storage backend for deployments that prefer an external relational database.

Layout

Key areas in this package:

• .env.example — runtime environment example
• spec/ — architecture and runtime design documents
• tests/ — runtime tests
• ya_claw/api/ — HTTP API surface
• ya_claw/bridge/ — IM bridge adapters and event handling
• ya_claw/app.py and ya_claw/cli.py — application entrypoints
• ya_claw/config.py — runtime configuration

Runtime Shape

The runtime shape is:

• one YA Claw web service
• one in-process runtime state manager
• one session scheduler
• one bridge subsystem for external channels
• one shared bearer token for HTTP access
• one SQLite database by default
• optional PostgreSQL
• one runtime data directory for sensitive session continuity
• one persistent workspace directory
• one bundled web shell

Runtime Architecture Notes

This section is the maintainer index for implementation details that affect code changes across YA Claw.

Runtime Defaults

• YA_CLAW_API_TOKEN is required before service startup.
• /api/v1/claw/info exposes service build metadata from YA_CLAW_SERVICE_VERSION, YA_CLAW_SERVICE_COMMIT, YA_CLAW_SERVICE_BUILD, and YA_CLAW_SERVICE_IMAGE; Docker builds inject these values for UI display.
• SQLite is the default durable store at ~/.ya-claw/ya_claw.sqlite3; file-backed SQLite engines use WAL and a 30-second busy timeout for the runtime's concurrent readers and writers.
• YA_CLAW_DATA_DIR defaults to ~/.ya-claw/data.
• YA_CLAW_WORKSPACE_DIR defaults to ~/.ya-claw/data/workspace.
• Browser workspace downloads are capped at 100 MiB by default; configure YA_CLAW_WORKSPACE_DOWNLOAD_MAX_BYTES to change the enforced per-file limit. The server enforces the cap both before and during streaming.
• GET /api/v1/sessions/{session_id}/workspace/files uses stable case-insensitive name ordering. Send limit, then continue with the opaque next_cursor while has_more is true; the name-key cursor prevents inserts or deletes before the cursor from shifting later pages. offset, next_offset, and the truncated alias remain available for backwards compatibility, but new clients should use cursor/next_cursor.
• The default Docker workspace image is ghcr.io/wh1isper/ya-claw-workspace:latest.
• Session metadata lives in the database; committed continuity blobs live in the local run store.

Implementation Conventions

• Runtime code is organized around ya_claw/api/, ya_claw/controller/, and ya_claw/orm/.
• Foundational execution modules live under ya_claw/execution/.
• Workspace provider modules live under ya_claw/workspace/.
• Internal data objects use Pydantic BaseModel.
• Code prefers explicit typing and isinstance checks.
• The session API is the high-level surface; the run API is the low-level surface.
• SQLite tests use the session-scoped initialize_sqlite_database fixture, which copies a schema-only template into each isolated database; avoid per-test Base.metadata.create_all calls.

Session and Run Persistence

• Committed continuity blobs live in run-store/{run_id}/state.json and run-store/{run_id}/message.json.
• message.json stores the compacted replay list of AGUI-aligned events as a top-level JSON array.
• state.json stores the bounded latest cumulative usage/cost snapshot. Claw normalizes SDK snapshots to the outer run ID, accumulates deferred-tool continuation segments, and indexes the bounded snapshot in internal run metadata. Run detail may fall back through state and replay for legacy or active runs; paginated summaries never scan full state blobs.
• Input payloads use input_parts; run records preserve input_parts as original JSON-compatible payloads for replay and UI reconstruction.
• Successful run records store final output_text directly in the database.
• Session GET exposes paginated runs with optional raw input_parts and compacted message replay lists, returns optional top-level committed state/message from head_success_run_id, and derives session status from the latest run.
• Session turns API returns successful completed turns with raw input_parts and output_text.
• Run GET returns session + run + optional state + optional message.
• Run trace API returns compact tool-call/tool-response projections from message.json.
• Rerun can explicitly target failed or interrupted runs through restore_from_run_id.
• JSON run/session create routes return JSON consistently; foreground SSE creation uses POST /api/v1/runs:stream, POST /api/v1/sessions:stream, and POST /api/v1/sessions/{session_id}/runs:stream.

Execution Coordination

• Active handles, live events, schedule dispatch, and bridge transports stay in the runtime process; relational rows remain the authority for runs, async tasks, input, and completion delivery.
• Built-in run orchestration lives in ya_claw/execution/coordinator.py; it retains SQL scheduling, workspace, HITL, memory, and delivery ownership while composing the SDK AgentExecutionHarness for each native segment.
• New ordinary runs and new child spawns resolve model/runtime behavior from the selected execution profile. Existing or resumed async-child executions restore the exact immutable descriptor persisted with their SQL execution record; profile drift, reseeding, or route deletion cannot redefine them.
• YA_CLAW_DEFAULT_PROFILE defaults to default.
• YA_CLAW_CAPABILITY_PLUGIN_MANIFEST optionally names one strict SDK plugin manifest; omission means no external plugins, while an explicit missing or invalid file is fatal.
• Runtime instance heartbeat lives in runtime_instances.
• Run records carry claim ownership through claimed_by and claimed_at.
• The built-in session toolset lets agents inspect their current session through internal HTTP client tools list_session_turns and get_run_trace; session ID and bearer token stay inside the client resource.

Workspace Providers and Docker Runtime

• LocalWorkspaceProvider uses LocalFileOperator plus policy-driven LocalShell over the real workspace path. Claw passes resolved shell sandbox policy for local sandbox execution; raw host shell is controlled by explicit policy.
• DockerWorkspaceProvider uses Docker mounts through SandboxEnvironment; file operations map the service-visible workspace path to /workspace, and Docker shell uses /workspace. Managed temporary storage lives at a hidden .tmp/ya-agent-<id> path below that shared workspace mount, including for reused containers; each owned instance contains a self-ignoring .gitignore, and no per-run bind mount is required.
• Docker shell execution requires a Docker CLI on the service runtime PATH plus Docker Engine API access. The official Dockerfile.ya-claw image bundles the CLI; custom images and host installs must provide it, and mounting docker.sock alone is insufficient.
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_HOST_WORKSPACE_DIR provides the Docker daemon-visible host mount path when the YA Claw service itself runs in Docker.
• Docker workspace containers receive UID/GID envs (YA_CLAW_WORKSPACE_UID, YA_CLAW_WORKSPACE_GID, YA_CLAW_HOST_UID, YA_CLAW_HOST_GID) from the service process by default or from YA_CLAW_WORKSPACE_PROVIDER_DOCKER_UID/GID.
• Dockerfile.ya-claw can drop service execution privileges through YA_CLAW_RUN_UID and YA_CLAW_RUN_GID.
• The official workspace image defaults to UID/GID 1000 through build args.

Bridge Runtime

• Bridge adapter types are enumerated through BridgeAdapterType; the current built-in adapter is lark.
• Bridge deployment dispatch uses BridgeDispatchMode (embedded, manual) and stays separate from run execution dispatch (queue, async, stream).
• embedded is the default bridge dispatch mode and runs adapter tasks under BridgeSupervisor in the same HTTP server lifespan as ExecutionSupervisor.
• manual starts the HTTP server with bridge dispatch managed outside the server lifespan.
• Lark bridge event allowlist comes from YA_CLAW_BRIDGE_LARK_EVENT_TYPES; defaults cover im.chat.member.bot.added_v1, im.chat.member.user.added_v1, im.message.receive_v1, and drive.notice.comment_add_v1.
• Lark message events map (adapter, tenant_key, chat_id) one-to-one to a session.
• Other accepted Lark events use chat_id when present and fall back to stable event or Drive conversation keys.
• Each accepted inbound event creates a bridge-triggered run after event/message dedupe.
• Lark bridge replies/actions are performed by the agent from the workspace with lark-cli.
• Workspace environments receive LARK_APP_ID and LARK_APP_SECRET from process env or Lark bridge app settings.

Session Memory

• Session memory is workspace-native.
• Paired internal session_type="memory" sessions run background extract/summary jobs with trigger type memory.
• Memory jobs share the source workspace sandbox and use the same profile tool surface as the primary agent.
• Memory content lives in workspace files: memory/MEMORY.md, memory/CHANGELOG.md, and memory/YYYYMMDD-event.md files with YAML frontmatter (name, description).
• memory/MEMORY.md is a compact durable brief for stable facts loaded by the main agent. Detailed chronology, file catalogs, and event lists belong in event files and memory/CHANGELOG.md.
• Memory extract and summary agents use fixed XML-style prompts from ya_claw/memory/extract_prompt.py and ya_claw/memory/summary_prompt.py.
• Primary conversation runs load workspace guidance from AGENTS.md and inject memory in the system prompt via WorkspaceMemoryStore, loading memory/MEMORY.md plus event file frontmatter as separate XML-style blocks.
• memory-context is registered in injected_context_tags, so SDK trim-mode handoff removes historical memory context from prompt history.
• Memory orchestration state lives in session_memory_states.
• Session list/detail responses expose memory_state.
• Manual endpoints are memory:extract and memory:summarize.
• File browsing should use workspace filetree APIs.

Installation

Recommended standalone install with uv:

uv tool install 'ya-claw[rs]'
ya-clawd version

[rs] passes ya-agent-sdk[all,rs] into the runtime and installs the native Rust filesystem search binding. The equivalent extra-dependency form is:

uv tool install ya-claw --with ya-ripgrep-core

ya-ripgrep-core is a library dependency, so --with is the matching uv form; --with-executables-from applies to companion packages that also expose CLI executables.

Pip can install the same runtime shape:

pip install 'ya-claw[rs]'

Capability Plugins

Install each trusted plugin distribution into the YA Claw service's Python environment, not only into a Docker workspace container. For an isolated uv tool installation:

uv tool install 'ya-claw[rs]' --with acme-agent-plugin

Then point YA Claw at the SDK's strict manifest:

YA_CLAW_CAPABILITY_PLUGIN_MANIFEST=/etc/ya-claw/plugins.toml

schema_version = 1
entry_points = ["acme.search"]

[[capabilities]]
name = "acme.search"
arguments = { result_limit = 10 }

entry_points explicitly selects installed Python types for the process catalog. capabilities appends ordered grants to every admitted root profile spec before its immutable descriptor is fingerprinted and persisted. Installation alone grants nothing, and YA Claw never ambiently loads all installed entry points.

The manifest path is optional; omission produces an empty external catalog. An explicit missing path, invalid TOML, unknown field, unsupported version, missing or duplicate entry point, import failure, grant argument signature mismatch, or catalog collision stops startup. Manifest arguments are durable non-secret configuration; secret-like keys are rejected recursively. This name-based guard cannot detect a secret stored under a neutral key. Provide live credentials and authority through typed runtime dependencies or host APIs instead.

Root manifest grants do not enter named children or self forks. A selected type may be used by a child only when that child's native AgentSpec.capabilities explicitly grants it. Before a root descriptor is fingerprinted or persisted, YA Claw performs native static-plan construction with ClawAgentContext, the selected custom types, and fresh base host capabilities; plugin factory values and combined ordering therefore fail at admission rather than after a run is queued. Runtime entry still validates dynamic Environment and resource contributions.

YA Claw resolves the file once before application startup and uses the same catalog snapshot for profile admission, async children, retained plans, recovery, memory runs, and runtime construction. Restart the service after changing the file or distribution.

The official image does not include third-party plugins. Build a derived service image that installs the distribution into /opt/venv, mount the manifest read-only, and set its in-container path:

FROM ghcr.io/wh1isper/ya-claw:latest
COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
RUN uv pip install --python /opt/venv/bin/python acme-agent-plugin

services:
  ya-claw:
    build: .
    environment:
      YA_CLAW_CAPABILITY_PLUGIN_MANIFEST: /etc/ya-claw/plugins.toml
    volumes:
      - ./plugins.toml:/etc/ya-claw/plugins.toml:ro

See the SDK file configuration contract and the installable example.

Quick Start

From the workspace root, start the default runtime flow:

uv sync --all-packages
cp packages/ya-claw/.env.example packages/ya-claw/.env
make run-claw

Set YA_CLAW_API_TOKEN before starting the service. The development server listens on http://127.0.0.1:9042 by default. YA Claw loads YA_CLAW_* settings from packages/ya-claw/.env and the process environment. YA Claw startup also exports provider variables such as GATEWAY_API_KEY and GATEWAY_BASE_URL from packages/ya-claw/.env into the process environment. Use packages/ya-agent-sdk/.env.example for shared SDK and tool environment variables when you want the same keys outside YA Claw startup. Set YA_CLAW_PROFILE_SEED_FILE plus YA_CLAW_AUTO_SEED_PROFILES=true when you want packaged profiles to seed into the database on startup. Seeded profiles use create/update semantics: every startup refreshes matching database profiles from the YAML file, including subagent configuration, while profiles absent from the YAML file remain in the database. Runs auto-dispatch through the built-in coordinator. New ordinary runs and child spawns resolve model/runtime behavior from execution profile rows; existing or resumed async children use their exact persisted immutable descriptor instead. The default profile name is default.

Profile, MCP, and coordinator settings:

• YA_CLAW_PROFILE_SEED_FILE=packages/ya-claw/profiles.yaml
• YA_CLAW_AUTO_SEED_PROFILES=true
• YA_CLAW_DEFAULT_PROFILE=default
• YA_CLAW_CAPABILITY_PLUGIN_MANIFEST=/etc/ya-claw/plugins.toml
• YA_CLAW_WORKSPACE_PROVIDER_BACKEND=local|docker
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_IMAGE=ghcr.io/wh1isper/ya-claw-workspace:latest
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_HOST_WORKSPACE_DIR=/srv/ya-claw/workspace
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_UID=<service process UID>
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_GID=<service process GID>
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_EXEC_USER=auto
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_HOME=/home/claw
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_CONTAINER_CACHE_DIR=~/.ya-claw/data/docker-workspace-containers
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_EXTRA_MOUNTS=/srv/ya-claw/home:/home/claw:rw,/srv/ya-claw/cache:/cache:ro
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_RETENTION_POLICY=stop_on_idle
• YA_CLAW_WORKSPACE_PROVIDER_DOCKER_IDLE_TTL_SECONDS=3600
• YA_CLAW_WORKSPACE_ENV_VARS=MY_TOOL_API_KEY,MY_TOOL_ENDPOINT
• YA_CLAW_BRIDGE_DISPATCH_MODE=embedded|manual
• YA_CLAW_BRIDGE_ENABLED_ADAPTERS=lark
• YA_CLAW_BRIDGE_LARK_APP_ID=cli_xxx
• YA_CLAW_BRIDGE_LARK_APP_SECRET=...
• YA_CLAW_BRIDGE_LARK_DEFAULT_PROFILE=default
• YA_CLAW_BRIDGE_LARK_EVENT_TYPES=im.chat.member.bot.added_v1,im.chat.member.user.added_v1,im.message.receive_v1,drive.notice.comment_add_v1
• YA_CLAW_BRIDGE_LARK_REPLY_IDENTITY=bot
• LARK_APP_ID=cli_xxx
• LARK_APP_SECRET=...
• LARKSUITE_CLI_BRAND=feishu
• LARKSUITE_CLI_DEFAULT_AS=bot
• LARKSUITE_CLI_STRICT_MODE=bot
• MALLOC_ARENA_MAX=2
• MALLOC_TRIM_THRESHOLD_=131072

The official YA Claw service and workspace Docker images set MALLOC_ARENA_MAX=2 and MALLOC_TRIM_THRESHOLD_=131072 for long-lived Python workloads. Use the same allocator values for systemd or custom container deployments when memory residency matters.

Profiles use strict schema version 2: a native Pydantic AI AgentSpec, a Claw-only host policy, and native SDK SubagentSpec children. Portable behavior is granted through AgentSpec.capabilities; Claw host groups are limited to session, schedule, workflow, and agency control-plane tools. Host policy owns model runtime configuration, approvals, MCP selection and server definitions, and workspace hints. See spec/01-configuration-and-workspace-provider.md.

Codex OAuth profiles use the oauth@codex:gpt-5.5 model string after the service host has run ya-oauth login codex. YA Claw maps provider session headers to the YA Claw session ID and provider thread headers to the run ID. Docker deployments should mount a persistent host directory to the service user's ~/.yaai, keep directory mode 0700 and auth.json mode 0600, and keep credentials out of image layers.

Shell review is configured under host.model_config_override.security.shell_review. The review model is explicit when enabled, and model_settings accepts an SDK preset name or inline settings. Interactive runs can defer for HITL; unattended schedule, workflow, heartbeat, and agency runs convert deferred review to denial.

version: 2
profiles:
  - schema_version: 2
    name: default
    agent:
      model: gateway@openai-responses:gpt-5.5
      name: default
      capabilities: [FilesystemCapability, ShellCapability]
    host:
      model_config_preset: gpt5_270k
      model_config_override:
        security:
          shell_review:
            enabled: true
            model: gateway@openai-responses:gpt-5.4-mini
            model_settings: openai_responses_low
            on_needs_approval: defer
            risk_threshold: extra_high
      tool_groups: [session]
    subagents: []

Session and run requests can provide workspace.mounts with one or more logical workspace folders, one default mount, a default cwd, and rw or ro access per mount. When requests omit workspace configuration, YA Claw uses the shared workspace configured by YA_CLAW_WORKSPACE_DIR and maps it to /workspace. Workspace guidance and memory use the default logical mount, and runtime prompts list the resolved mount set.

Workspace environments receive LARK_APP_ID and LARK_APP_SECRET from explicit process environment values or from the configured Lark bridge app settings. The official Docker workspace entrypoint writes these values into /home/claw/.lark-cli/config.json for lark-cli bot commands, and clears LARKSUITE_CLI_APP_ID / LARKSUITE_CLI_APP_SECRET in the container runtime environment so lark-cli uses the generated config profile. LARKSUITE_CLI_BRAND, LARKSUITE_CLI_DEFAULT_AS, and LARKSUITE_CLI_STRICT_MODE tune that generated profile. YA_CLAW_WORKSPACE_ENV_VARS forwards additional comma-separated process environment variable names into workspace environments. For Docker workspaces, forwarded values are passed at session or run workspace container creation time.

The default Docker workspace image is ghcr.io/wh1isper/ya-claw-workspace:latest. It is based on Debian stable and includes Python, Node.js, lark-cli, and bundled workspace skills copied into mounted workspaces at container start. Auto-started workspace containers receive YA_CLAW_WORKSPACE_UID, YA_CLAW_WORKSPACE_GID, YA_CLAW_HOST_UID, and YA_CLAW_HOST_GID; the default values come from the YA Claw service process UID/GID and can be overridden with YA_CLAW_WORKSPACE_PROVIDER_DOCKER_UID and YA_CLAW_WORKSPACE_PROVIDER_DOCKER_GID. Docker exec uses YA_CLAW_WORKSPACE_PROVIDER_DOCKER_EXEC_USER=auto by default, which resolves to the configured workspace UID:GID, and sets HOME from YA_CLAW_WORKSPACE_PROVIDER_DOCKER_HOME with default /home/claw. YA_CLAW_WORKSPACE_PROVIDER_DOCKER_EXTRA_MOUNTS adds comma-separated provider support mounts to Docker workspace containers, with rw and ro modes. Docker session sandboxes use generation-specific containers named ya-claw-session-{session_id_short}-gN, store cache metadata under ~/.ya-claw/data/docker-workspace-containers/sessions/{session_id}/workspace.json, resolve and store the Docker image digest before reuse, refresh last_used_at during active runs and once on run exit, and stop idle containers according to YA_CLAW_WORKSPACE_PROVIDER_DOCKER_RETENTION_POLICY and YA_CLAW_WORKSPACE_PROVIDER_DOCKER_IDLE_TTL_SECONDS. TTL cleanup deletes the session workspace.json cache file and uses the same cache-path lock as run startup. A changed image digest causes YA Claw to remove the stale container and start a new one from the current image. Schedule and heartbeat runs use run-scoped containers named ya-claw-run-{run_id_short} and cache metadata under runs/{run_id}/workspace.json.

Profiles can be managed through:

• REST API: /api/v1/profiles
• Seed API: POST /api/v1/profiles/seed
• CLI: ya-claw profiles seed

Default local paths:

• SQLite database: ~/.ya-claw/ya_claw.sqlite3
• runtime data root: ~/.ya-claw/data
• workspace directory: ~/.ya-claw/data/workspace

External Database

Set YA_CLAW_DATABASE_URL in packages/ya-claw/.env when you want an external PostgreSQL database. The default SQLite file stays at ~/.ya-claw/ya_claw.sqlite3.

Database Commands

uv run --package ya-claw ya-claw db upgrade
uv run --package ya-claw ya-claw db current
uv run --package ya-claw ya-claw db history
uv run --package ya-claw ya-claw db revision "add session tables"

Bridge Commands

The CLI owns a top-level bridge command group.

uv run --package ya-claw ya-claw bridge ls
uv run --package ya-claw ya-claw bridge run lark
uv run --package ya-claw ya-claw bridge serve lark

Bridge Dispatch

Bridge dispatch controls whether the YA Claw HTTP server starts bridge adapters:

• embedded starts enabled adapters inside the YA Claw server lifespan under BridgeSupervisor.
• manual starts the YA Claw HTTP server without starting BridgeSupervisor.

Bridge adapters submit inbound events through the same session/run controller path used by HTTP requests, so bridge ingress behaves as a self-request inside the service process. The Lark bridge reads YA_CLAW_BRIDGE_LARK_EVENT_TYPES as a comma-separated event allowlist. The default allowlist covers bot-added-to-chat, user-added-to-chat, message receive, and Drive comment notification events. Message receive events map each tenant_key + chat_id pair to one YA Claw session. Other Lark events use chat_id when present and fall back to a stable event or Drive conversation key. Every accepted inbound event creates a queued bridge-triggered run, and the agent replies or acts from the workspace with lark-cli.

Web Shell

Run the web shell from the repository root:

make web-dev

Docker

Build the YA Claw service image from the repository root:

docker build -f Dockerfile.ya-claw -t ya-claw:dev .

Build the official workspace image locally:

docker build -f Dockerfile.ya-claw-workspace -t ya-claw-workspace:dev .

Build the workspace image with a default UID/GID baked in:

docker build \
  --build-arg WORKSPACE_UID=1000 \
  --build-arg WORKSPACE_GID=1000 \
  -f Dockerfile.ya-claw-workspace \
  -t ya-claw-workspace:dev .

Run the YA Claw service image under a specific UID/GID:

docker run \
  -e YA_CLAW_RUN_UID=1000 \
  -e YA_CLAW_RUN_GID=1000 \
  -e YA_CLAW_API_TOKEN=replace-with-a-long-random-token \
  ya-claw:dev

Initial API Surface

Every HTTP route except /healthz expects Authorization: Bearer <YA_CLAW_API_TOKEN>.

• GET /healthz — service health probe with storage and runtime component status
• POST /api/v1/sessions — create a session with optional first queued run and return JSON
• POST /api/v1/sessions:stream — create a session with a first run and stream foreground SSE events
• GET /api/v1/sessions — list sessions using the backwards-compatible unpaginated response; optional keyset parameters are available for transitional clients
• GET /api/v1/sessions/page — list lightweight session pages ordered by (updated_at, id), with total, has_more, and continuation anchors; this endpoint skips live Docker reconciliation and latest-run output text by default
• GET /api/v1/sessions/{session_id} — inspect a session plus paginated runs, top-level committed state, and optional compacted message replay lists; set include_head_payload=false when only the run page is needed
• POST /api/v1/sessions/{session_id}/memory:extract — enqueue a background memory extract run for the source session
• POST /api/v1/sessions/{session_id}/memory:summarize — enqueue a background memory summary run for the source session
• POST /api/v1/sessions/{session_id}/runs — create a run under a session and return JSON
• POST /api/v1/sessions/{session_id}/runs:stream — create a run under a session and stream foreground SSE events
• POST /api/v1/sessions/{session_id}/steer — steer the active run through the session surface
• POST /api/v1/sessions/{session_id}/interrupt — interrupt the active run through the session surface
• POST /api/v1/sessions/{session_id}/cancel — cancel the active run through the session surface
• POST /api/v1/runs — create a run directly through the low-level surface and return JSON
• POST /api/v1/runs:stream — create a run directly and stream foreground SSE events
• GET /api/v1/runs/{run_id} — inspect a run plus session summary, committed state, and optional compacted message replay list
• POST /api/v1/runs/{run_id}/steer — steer a specific active run
• POST /api/v1/runs/{run_id}/interrupt — interrupt a specific active run
• POST /api/v1/runs/{run_id}/cancel — cancel a specific active run

Direct session/run/async-child steering succeeds only when the addressed logical run is actively accepting input. A queued run, terminal run, or session without an active accepting run returns HTTP 409 before any success response. Success returns the persisted SQL/native receipt fields input_id, input_delivery_key, input_disposition, input_sdk_id, and input_enqueue_id; an equal idempotent retry returns the same durable receipt. Unified session submit is a separate API and may merge input into a queued run.

Portable SDK subagent spawn is claimed atomically in SQL by owner scope and idempotency key. The task, child session, child run, immutable plan, intent digest, and applied initial-input marker commit together before run publication. Concurrent or post-crash retries return that one committed task; a changed intent is rejected and cannot create an orphan child.

Spec Set

• spec/README.md
• spec/00-overview.md
• spec/01-configuration-and-workspace-provider.md
• spec/02-execution-and-session.md
• spec/03-storage-and-streaming.md
• spec/04-api.md
• spec/05-web-ui-and-operations.md