Default-enable AI Planner in Host Agent; propose cloud-planner-proxy
Tests / Test passed: 626

- Host Agent now defaults AI_PLANNER_ENABLED=true (opt-out via env),
  scoped to apps/device-host-agent/host_agent/execution.py only; the
  shared runtime.planner_config default (disabled) is unchanged.
- Add openspec proposal for cloud-planner-proxy: centralize LLM
  provider config/credentials on the Cloud Control Plane and let the
  Host Agent proxy AI Planner decisions through it instead of holding
  provider API keys locally. Proposal only, no implementation yet.
This commit is contained in:
2026-07-13 20:50:35 +08:00
parent 78ce788e2f
commit 1107ace89c
9 changed files with 537 additions and 5 deletions
+19 -1
View File
@@ -1,10 +1,12 @@
from __future__ import annotations
import os
from collections.abc import Callable
from dataclasses import dataclass
from dataclasses import dataclass, replace
from device.manager import DeviceManager
from runtime.executor import Executor, default_tool_registry
from runtime.planner_config import PlannerConfig, load_config as load_planner_config
from runtime.task import TaskRunner
from storage.task_metadata import TaskMetadataStore
from storage.timeline import Timeline
@@ -33,6 +35,7 @@ def create_execution_factories(
executor=Executor(tools=default_tool_registry(manager=manager)),
metadata_store=metadata_store,
timeline=timeline,
planner_config=_host_agent_planner_config(),
)
def create_workflow_runner() -> WorkflowRunner:
@@ -46,3 +49,18 @@ def create_execution_factories(
workflow_runner_factory=create_workflow_runner,
workflow_store=shared_workflow_store,
)
def _host_agent_planner_config() -> PlannerConfig:
"""Host Agent defaults to the AI planner unless an operator opts out.
`runtime.planner_config` defaults `enabled=False` for the shared Runtime
library (local dev/tests/cloud dispatcher keep the deterministic stub
planner unless asked). The Host Agent is the actual device-control path,
so it flips that default on here -- an explicit `AI_PLANNER_ENABLED=false`
still disables it.
"""
config = load_planner_config()
if os.environ.get("AI_PLANNER_ENABLED") is None:
config = replace(config, enabled=True)
return config
@@ -2,8 +2,12 @@ from __future__ import annotations
from pathlib import Path
import pytest
from device.manager import DeviceManager
from host_agent.execution import create_execution_factories
from runtime.ai_planner import AIPlanner
from runtime.planner import Planner
from runtime.task import TaskRunner
from workflow.runner import WorkflowRunner
from workflow.store import WorkflowStore
@@ -26,6 +30,34 @@ def test_execution_factories_compose_existing_runtime_and_workflow(tmp_path) ->
assert isinstance(workflow_runner.task_runner_factory(), TaskRunner)
def test_created_task_runner_defaults_to_ai_planner(
tmp_path, monkeypatch: pytest.MonkeyPatch
) -> None:
monkeypatch.delenv("AI_PLANNER_ENABLED", raising=False)
manager = DeviceManager()
factories = create_execution_factories(
manager, workflow_store=WorkflowStore(tmp_path / "workflows.sqlite3")
)
task_runner = factories.task_runner_factory()
assert isinstance(task_runner.planner, AIPlanner)
def test_created_task_runner_honors_explicit_ai_planner_opt_out(
tmp_path, monkeypatch: pytest.MonkeyPatch
) -> None:
monkeypatch.setenv("AI_PLANNER_ENABLED", "false")
manager = DeviceManager()
factories = create_execution_factories(
manager, workflow_store=WorkflowStore(tmp_path / "workflows.sqlite3")
)
task_runner = factories.task_runner_factory()
assert type(task_runner.planner) is Planner
def test_runtime_owned_packages_do_not_import_host_or_cloud_concerns() -> None:
root = Path(__file__).resolve().parents[3]
forbidden = ("import cloud", "from cloud", "import host_agent", "from host_agent")
+10 -4
View File
@@ -346,18 +346,24 @@ back to the upstream.
## Runtime AI Planner
The Host Agent reuses the local Runtime planner. AI planning is disabled by
default. Configure it in the Host Agent environment when goal assignments must
use a model:
The Host Agent reuses the local Runtime planner. Unlike the shared Runtime
library (whose own default is the deterministic stub planner), the **Host
Agent defaults `AI_PLANNER_ENABLED` to on** -- it is the actual device-control
path, so goal assignments use a model unless an operator explicitly opts out.
Provide provider credentials before deploying:
```text
AI_PLANNER_ENABLED=true
AI_PLANNER_PROVIDER=anthropic
AI_PLANNER_MODEL=claude-sonnet-5
AI_PLANNER_TIMEOUT_SECONDS=30
ANTHROPIC_API_KEY=<secret manager reference>
```
Without a valid API key, every planning step raises immediately and the task
fails on its first step (no silent fallback to the stub planner). Set
`AI_PLANNER_ENABLED=false` to opt back out to the deterministic stub planner
(e.g. for offline/dev hosts with no provider credentials).
For OpenAI, set `AI_PLANNER_PROVIDER=openai`, choose the deployed model through
`AI_PLANNER_MODEL`, and provide `OPENAI_API_KEY`. Provider credentials belong
only on the Host Agent; the Cloud API does not need them.
@@ -0,0 +1,2 @@
schema: spec-driven
created: 2026-07-13
@@ -0,0 +1,210 @@
## Context
The Host Agent's `AIPlanner` (`runtime/ai_planner.py`) delegates the actual
LLM call to a `ToolCallingClient` (`runtime/tool_calling_client.py`):
`AnthropicToolCallingClient`/`OpenAIToolCallingClient` each build a provider
SDK client that implicitly reads `ANTHROPIC_API_KEY`/`OPENAI_API_KEY` from
the Host Agent's own process environment, and `build_client()` selects
between them based on `PlannerConfig.provider` (from
`runtime/planner_config.py`, itself read from the Host Agent's local env).
None of this touches the Cloud Control Plane today.
The Cloud Control Plane (`apps/cloud-api` + `packages/cloud-platform/cloud`)
already authenticates every Host Agent with a host-scoped bearer credential
for heartbeat, claim/long-poll, lease renewal, and terminal result reporting
(`cloud/internal_api/*`), with the Host Agent importing shared Pydantic
models directly from `cloud.internal_api.models` rather than duplicating
schemas. Separately, `packages/cloud-platform` already depends on the root
Runtime package (`device-agent-runtime`, workspace dependency), the same way
`apps/device-host-agent` does -- so cloud-side code can import
`runtime.tool_calling_client` directly instead of reimplementing
Anthropic/OpenAI wire-format handling a second time.
A prior change in this session made the Host Agent default
`AI_PLANNER_ENABLED` to on (`apps/device-host-agent/host_agent/execution.py`).
That means every Host Agent now needs a working planner path by default,
which raises the stakes on where its provider credentials live.
## Goals / Non-Goals
**Goals:**
- Let an operator configure LLM provider, model, and credentials once, on
the Cloud Control Plane, instead of per Host Agent.
- Let a Host Agent execute AI-planned tasks without holding
`ANTHROPIC_API_KEY`/`OPENAI_API_KEY` locally, by proxying its planner
decision calls through the Cloud Control Plane.
- Reuse existing Host<->Cloud authentication and existing provider
wire-format code; add no new auth scope and no duplicated Anthropic/OpenAI
translation logic.
- Preserve today's direct-to-provider path as a fully supported, still-default
option, so existing Host Agent deployments with local keys are unaffected.
**Non-Goals:**
- Not moving prompt construction (`runtime/planner_prompts.py`, Scene
serialization, world/history summarization) to the cloud. The Host Agent
keeps building `system_prompt`/`user_prompt` locally; the cloud endpoint
is a thin "make this one LLM call and return one tool call" proxy, not a
re-implementation of planning.
- Not building per-tenant/per-host provider overrides, usage metering,
rate limiting, or a management UI for the proxy. Single cloud-wide
provider/model/credential configuration only, matching how
`AI_PLANNER_*` is configured today (env-based, one value cluster).
- Not changing the existing heartbeat/claim/renew/result protocol or its
spec (`host-agent-protocol`).
- Not deprecating or removing the direct-to-provider transport.
## Decisions
### D1: New endpoint reuses `runtime.tool_calling_client` server-side instead of reimplementing provider calls
`cloud.internal_api` adds a route whose handler constructs an
`AnthropicToolCallingClient`/`OpenAIToolCallingClient` (imported from
`runtime.tool_calling_client`, already an allowed dependency direction since
`packages/cloud-platform` depends on `device-agent-runtime`) using
Cloud-side provider/model/credential configuration, and calls `.decide(...)`
with the request payload. This eliminates a second implementation of
Anthropic/OpenAI tool-calling wire-format translation, which
`ai-planner-runtime/design.md` already flagged as a drift risk for the
*two-provider* case -- a *four-implementation* case (two providers x two
transports) would be worse.
Alternative considered: hand-roll a minimal cloud-side Anthropic/OpenAI
client in `cloud.internal_api`. Rejected: duplicates parsing logic
(`_decision_from_anthropic_response`, `_decision_from_openai_response`) that
must stay in lockstep with the Host Agent's local transport as providers'
APIs evolve.
### D2: New transport axis, orthogonal to provider selection
Host Agent config gains `AI_PLANNER_TRANSPORT` (`direct` default | `cloud`),
independent of `AI_PLANNER_PROVIDER`. `direct` is today's behavior
unchanged (Host Agent builds the SDK client itself). `cloud` builds a new
`CloudProxyToolCallingClient` instead; provider/model selection and
credentials for that path live in the Cloud API's own
`AI_PLANNER_PROVIDER`/`AI_PLANNER_MODEL`/`ANTHROPIC_API_KEY`/`OPENAI_API_KEY`
configuration, not the Host Agent's.
Alternative considered: overload `AI_PLANNER_PROVIDER=cloud` as a third
provider value. Rejected: provider and transport are different axes (a
`cloud` transport still ultimately calls `anthropic` or `openai`), and
conflating them would make the Cloud API's own provider config harder to
reason about ("provider" would mean different things on each side).
### D3: `CloudProxyToolCallingClient` lives in `host_agent`, not `runtime`
`runtime/tool_calling_client.py`'s `ToolCallingClient` stays a plain
`Protocol`; the new client is added in `apps/device-host-agent/host_agent/`
(e.g. `host_agent/cloud_planner_client.py`) and satisfies that Protocol
structurally. This preserves the existing boundary enforced by
`apps/device-host-agent/tests/test_execution.py::test_runtime_owned_packages_do_not_import_host_or_cloud_concerns`,
which forbids `runtime` (and `core`/`device`/`driver`/`tools`) from
importing `cloud` or `host_agent`. The new client wraps a new method on the
existing `host_agent/client.py::HostAgentClient` (which already holds the
authenticated `httpx` session and imports `cloud.internal_api.models`), so
it reuses the same request/auth/retry plumbing as heartbeat/claim/renew/result
instead of opening a second HTTP client type.
### D4: Reuse the existing host-scoped bearer credential; no new auth scope
The new endpoint sits on the same internal router and auth dependency as
heartbeat/claim/renew/result. A Host Agent that can already reach the Cloud
Control Plane for task assignment can reach the planner-decide endpoint;
there is no separate enrollment or credential to provision for this
capability.
Alternative considered: a distinct scope/credential just for planner-proxy
calls (defense in depth, so a compromised heartbeat credential couldn't run
up LLM cost). Rejected for this proposal to keep the change additive and
consistent with how the rest of the internal API already treats "any
authenticated host" uniformly; revisit if abuse/cost containment becomes a
concrete concern (see Open Questions).
### D5: Request/response payload mirrors the local `ToolCallingClient.decide()` signature
The proxy request carries `system_prompt`, `user_prompt`, an optional
base64 screenshot, the tool specs (already JSON-serializable
`ToolSpec`/dict shapes used to build provider tool schemas), and a timeout.
The response carries the resolved `tool_name`/`arguments` (or a structured
error). This keeps the Host Agent and Cloud Control Plane in the same
shape as today's local call, so `AIPlanner` itself needs zero changes --
only the client it's constructed with changes.
### D6: Cloud does not durably persist prompt text or screenshot bytes from proxy requests
The endpoint handler processes the request in memory and returns; only
metadata (host id, tool name decided, latency, error class if any) may be
logged for observability. This bounds the new sensitive-data surface
created by routing screenshots and prompts through the cloud (an accepted
trade-off from the earlier feasibility discussion) to "in transit and in
process," not "at rest in cloud logs/DB."
### D7: Proxy failures surface as `ToolCallUnavailable`, matching direct-transport failure behavior
Any failure calling the cloud endpoint (network error, auth rejection,
non-2xx, provider error surfaced by the cloud side, timeout) raises
`ToolCallUnavailable` from `CloudProxyToolCallingClient.decide()`, exactly
like today's direct-transport failures. This preserves the existing,
already-accepted behavior that a planner failure fails the current task
step immediately with no silent fallback to the stub planner -- this
proposal does not change that risk profile, only where the call happens.
## Risks / Trade-offs
- **[Risk] Cloud Control Plane is now in the hot path of every planning
step for hosts on the `cloud` transport** -> Mitigation: `direct`
transport remains the default and fully supported; operators who need
offline/low-latency operation simply don't opt in. Bounded by the
existing `AI_PLANNER_TIMEOUT_SECONDS`, same as today.
- **[Risk] Cloud Control Plane outage now stalls planning (not just new
task assignment) for opted-in hosts** -> Mitigation: same
`ToolCallUnavailable` -> task-fails-fast behavior as any other planner
error; no new silent-hang mode. Documented as an explicit trade-off of
opting into `cloud` transport.
- **[Risk] Screenshots and prompt text now transit through the Cloud
Control Plane** -> Mitigation: D6 (no durable persistence); still an
expansion of the data path operators should account for versus
direct-to-provider, which never touches the cloud.
- **[Risk] Cloud API gains a new dependency on `anthropic`/`openai` SDKs
and becomes a second place holding provider credentials** -> Mitigation:
reusing `runtime.tool_calling_client` (D1) keeps this to configuration
and routing, not new provider-integration code; credential handling
follows the same "environment or secret manager" pattern already
documented for the Host Agent in `docs/CLOUD_DEPLOYMENT.md`.
- **[Risk] Any authenticated host can drive cloud-held LLM spend** (D4) ->
Mitigation: none in this proposal beyond existing per-host authentication;
flagged as an Open Question rather than silently accepted, since it's a
cost/abuse concern rather than a correctness one.
- **[Risk] `ai-planner-runtime`'s accepted spec text ("AI Planner is
disabled by default") is already stale relative to the Host Agent's
actual default (changed in a prior, separate session change without an
openspec artifact)** -> Not a risk introduced by this proposal, but this
proposal's delta spec is written against that same pending, unarchived
`agent-runtime` capability, so the discrepancy should be reconciled
once, when `ai-planner-runtime` is archived (see Open Questions).
## Migration Plan
1. Add Cloud API configuration (`AI_PLANNER_PROVIDER`/`AI_PLANNER_MODEL`/
`AI_PLANNER_TIMEOUT_SECONDS`/provider API keys) and the new internal
route, guarded by the existing host-scoped auth. Off by default in the
sense that no Host Agent calls it until configured to use `cloud`
transport.
2. Add the Host Agent's `AI_PLANNER_TRANSPORT` setting (default `direct`)
and `CloudProxyToolCallingClient`. Existing deployments are unaffected
until an operator sets `AI_PLANNER_TRANSPORT=cloud` and removes the
local provider key.
3. Update `docs/CLOUD_DEPLOYMENT.md` with the proxy configuration path and
its trade-offs (latency, availability coupling, data-path expansion).
4. Rollback is setting `AI_PLANNER_TRANSPORT=direct` (or unsetting it) on
affected hosts and restoring their local provider key; the Cloud API
route can remain deployed but unused.
## Open Questions
- Should planner-proxy calls eventually get their own credential scope
separate from heartbeat/claim/renew/result, to bound blast radius and
allow independent cost/abuse controls? Deferred; revisit if this becomes
a real deployment.
- Should the Cloud API expose any per-request cost/usage observability
(e.g. token counts) given it now brokers every LLM call for `cloud`-
transport hosts? Out of scope for this proposal's tasks; worth deciding
before recommending `cloud` transport for cost-sensitive fleets.
- When `ai-planner-runtime` is archived, its "disabled by default" spec
text needs reconciling against both the Host Agent's actual default
(from the earlier, separate change) and this proposal's transport
addition -- noted here so it isn't lost, matching how
`edge-host-self-enrollment`'s design.md tracked its own dependency on an
unarchived pending spec.
@@ -0,0 +1,70 @@
## Why
Today every Host Agent must hold its own LLM provider credentials
(`ANTHROPIC_API_KEY`/`OPENAI_API_KEY`) and provider/model configuration
locally, because `runtime/ai_planner.py`'s `AIPlanner` builds an
Anthropic/OpenAI SDK client directly inside the Host Agent process
(`runtime/tool_calling_client.py`). With the Host Agent now defaulting AI
planning to on, this means every edge deployment must independently
provision, rotate, and secure a provider secret. Centralizing provider
configuration and credentials in the Cloud Control Plane -- which already
authenticates every Host Agent for heartbeat/claim/lease/result traffic --
removes per-edge secret sprawl and gives operators one place to change
provider/model or rotate a key without touching any Host.
## What Changes
- Add a new Cloud Control Plane internal endpoint that accepts an AI
Planner tool-calling decision request (system prompt, user prompt,
optional screenshot, tool specs, timeout) from an authenticated Host
Agent, calls the configured LLM provider using cloud-held credentials,
and returns the resulting single tool-call decision.
- Cloud API owns `AI_PLANNER_PROVIDER`/`AI_PLANNER_MODEL`/provider API keys
as its own configuration; these are no longer required on the Host Agent
when the new proxy transport is used.
- Host Agent gains a new opt-in transport setting (proxy vs. direct-to-provider)
and a new `ToolCallingClient` implementation that calls the cloud endpoint
instead of constructing a local Anthropic/OpenAI SDK client. The existing
direct-to-provider transport remains fully supported and is the default,
so hosts that already run with a local provider key keep working
unchanged.
- Reuse the existing Host-scoped bearer credential (already used for
heartbeat/claim/renew/result) for the new endpoint; no new auth scope.
- Cloud API does not durably persist screenshot bytes or full prompt text
from proxy requests beyond the lifetime of handling the request.
## Capabilities
### New Capabilities
- `cloud-planner-proxy`: Cloud Control Plane internal endpoint and
configuration that proxies AI Planner LLM tool-calling decisions on
behalf of authenticated Host Agents, holding provider selection and
credentials centrally instead of on each edge host.
### Modified Capabilities
- `agent-runtime` (capability defined by the not-yet-archived
`ai-planner-runtime` change; this delta is written against that pending
spec, matching the precedent set by `edge-host-self-enrollment` against
the pending `edge-host-enrollment` spec): the "Pluggable dual-provider
tool-calling abstraction" requirement is extended so the tool-calling
client is selectable by transport (direct-to-provider vs. cloud-proxy) as
well as by provider identity, and provider credentials become optional on
the Host Agent when the cloud-proxy transport is selected.
## Impact
- `packages/cloud-platform/cloud` / `apps/cloud-api`: new internal API
route + request/response models, new provider/credential configuration,
and a new direct dependency on the `anthropic`/`openai` SDKs (currently
only the root Runtime package depends on them -- this is new dependency
and attack surface for the Cloud API).
- `apps/device-host-agent`: new `HostAgentClient` method for the
planner-decide call, a new `ToolCallingClient` implementation
(constructed in the `host_agent` package, not `runtime`, to preserve the
existing hexagonal boundary that forbids `runtime` from importing `cloud`
or `host_agent`), and a new transport configuration setting.
- `runtime/tool_calling_client.py`: no changes to the `ToolCallingClient`
Protocol itself; the new implementation satisfies it structurally from
outside the `runtime` package.
- Docs: `docs/CLOUD_DEPLOYMENT.md` Runtime AI Planner section gains the
proxy-transport configuration path and its trade-offs.
@@ -0,0 +1,43 @@
## MODIFIED Requirements
### Requirement: Pluggable dual-provider tool-calling abstraction
The system SHALL support at least two interchangeable LLM providers
(Anthropic native tool use and OpenAI function calling) for the AI
Planner's decision calls, selectable via configuration, with both providers
constrained to return exactly one tool call per request. Independently of
provider selection, the system SHALL support at least two transports for
making that decision call -- direct-to-provider (the tool-calling client
calls the provider's SDK itself, using locally configured credentials) and
cloud-proxy (the tool-calling client calls the Cloud Control Plane's
planner-decision endpoint, which calls the provider using cloud-held
credentials) -- selectable via configuration without requiring any change
to `AIPlanner`'s own decision logic.
#### Scenario: Provider selected via configuration
- **WHEN** the AI Planner is configured with a given provider identifier
- **THEN** it constructs and uses the tool-calling client for that provider
without requiring any change to `AIPlanner`'s own decision logic
#### Scenario: Provider response resolves to a single decision
- **WHEN** either supported provider returns a response to a tool-calling
request
- **THEN** the response is parsed into exactly one tool name and one
arguments object, regardless of which provider produced it
#### Scenario: Transport selected via configuration
- **WHEN** the Host Agent is configured with a given transport (direct or
cloud-proxy)
- **THEN** `AIPlanner` is constructed with the tool-calling client for that
transport, and its own decision logic is unchanged regardless of which
transport is in effect
#### Scenario: Direct transport remains available and default
- **WHEN** no transport is explicitly configured
- **THEN** the AI Planner uses the direct-to-provider transport, matching
its behavior before the cloud-proxy transport existed
#### Scenario: Cloud-proxy transport resolves a decision without a local provider client
- **WHEN** the Host Agent is configured with the cloud-proxy transport
- **THEN** its tool-calling client sends the decision request to the Cloud
Control Plane's planner-decision endpoint instead of constructing a local
Anthropic or OpenAI SDK client
@@ -0,0 +1,59 @@
## ADDED Requirements
### Requirement: Cloud Control Plane exposes an authenticated planner-decision endpoint
The Cloud Control Plane SHALL expose an internal endpoint that accepts an AI Planner tool-calling decision request from an authenticated Host Agent, and SHALL require the same host-scoped bearer credential already used for heartbeat, claim, lease renewal, and result reporting -- no separate credential or enrollment step.
#### Scenario: Authenticated host requests a planner decision
- **WHEN** a Host Agent presents its existing valid host-scoped credentials
with a planner-decision request
- **THEN** the Cloud Control Plane accepts and processes the request for
that host
#### Scenario: Unauthenticated or foreign-host request is rejected
- **WHEN** a request omits valid host-scoped credentials, or presents
credentials bound to a different host than the one referenced in the
request
- **THEN** the Cloud Control Plane rejects the request without invoking any
LLM provider
### Requirement: Endpoint resolves exactly one tool-call decision using cloud-held provider configuration
The Cloud Control Plane SHALL use its own configured LLM provider, model, and credentials -- not any value supplied by the requesting Host Agent -- to resolve a planner-decision request to exactly one tool name and one arguments object, within the request's timeout.
#### Scenario: Provider returns a usable decision
- **WHEN** the configured provider responds to a planner-decision request
with a tool call
- **THEN** the Cloud Control Plane returns exactly one resolved tool name
and arguments object to the requesting Host Agent
#### Scenario: Configured provider is unreachable or misconfigured
- **WHEN** the Cloud Control Plane's configured provider call fails (for
example, invalid credentials, provider error, or timeout)
- **THEN** the endpoint returns a structured failure response rather than a
fabricated decision, and does not crash the Cloud Control Plane process
### Requirement: Cloud-proxy transport removes the need for Host Agent-held provider credentials
A Host Agent using the cloud-proxy transport for its AI Planner SHALL be able to execute AI-planned tasks without any locally configured LLM provider API key.
#### Scenario: Host configured for cloud-proxy transport has no local provider key
- **WHEN** a Host Agent is configured to use the cloud-proxy transport and
has no `ANTHROPIC_API_KEY`/`OPENAI_API_KEY` set in its own environment
- **THEN** it can still obtain AI Planner decisions by calling the Cloud
Control Plane's planner-decision endpoint
### Requirement: Planner-decision requests are not durably persisted
The Cloud Control Plane SHALL process planner-decision requests, including any screenshot and prompt text they carry, without durably persisting that screenshot or prompt content; only request metadata (such as host identifier, resolved tool name, latency, and error classification) may be retained for observability.
#### Scenario: Request handling completes without storing prompt or screenshot content
- **WHEN** the Cloud Control Plane finishes handling a planner-decision
request
- **THEN** the raw prompt text and screenshot bytes from that request are
not present in any durable store or log the Cloud Control Plane retains
### Requirement: Planner-proxy failures do not silently substitute a default action
The Cloud Control Plane SHALL report a failure to the requesting Host Agent when it cannot resolve a planner-decision request to a valid tool call, rather than returning a default, guessed, or previously cached decision.
#### Scenario: Endpoint cannot resolve a decision
- **WHEN** the configured provider does not return a usable tool call for a
planner-decision request
- **THEN** the Cloud Control Plane's response indicates failure, and the
requesting Host Agent treats the planner call for that turn as failed
@@ -0,0 +1,92 @@
## 1. Cloud API: planner configuration
- [ ] 1.1 Add Cloud API-side planner configuration (provider, model, timeout,
provider API keys) analogous to `runtime/planner_config.py`, loaded
from the Cloud API's own process environment (e.g.
`AI_PLANNER_PROVIDER`/`AI_PLANNER_MODEL`/`AI_PLANNER_TIMEOUT_SECONDS`,
`ANTHROPIC_API_KEY`/`OPENAI_API_KEY`), living in
`packages/cloud-platform/cloud` or `apps/cloud-api`.
- [ ] 1.2 Add `anthropic`/`openai` as explicit dependencies of the package
that hosts this configuration (confirm whether `packages/cloud-platform`
or `apps/cloud-api` is the right home, matching where the new endpoint
handler will live).
## 2. Cloud API: planner-decision internal endpoint
- [ ] 2.1 Add request/response Pydantic models to `cloud.internal_api.models`
for a planner-decision call: request carries `system_prompt`,
`user_prompt`, optional base64 screenshot, tool specs, timeout;
response carries resolved `tool_name`/`arguments` or a structured
error.
- [ ] 2.2 Add the internal route (e.g. `POST /internal/v1/planner/decide`)
to the existing internal router, reusing the current host-scoped
bearer auth dependency used by heartbeat/claim/renew/result -- no new
scope.
- [ ] 2.3 Implement the route handler by constructing
`runtime.tool_calling_client.AnthropicToolCallingClient` or
`OpenAIToolCallingClient` (per D1) from the Cloud API's own planner
configuration, calling `.decide(...)`, and translating
`ToolCallDecision`/`ToolCallUnavailable` into the response model.
- [ ] 2.4 Ensure the handler does not log or persist raw prompt text or
screenshot bytes; only metadata (host id, resolved tool name,
latency, error class) may be logged.
## 3. Host Agent: cloud-proxy transport
- [ ] 3.1 Add `AI_PLANNER_TRANSPORT` (`direct` default | `cloud`) to
`apps/device-host-agent/host_agent/config.py`.
- [ ] 3.2 Add a `request_planner_decision(...)` method to
`host_agent/client.py::HostAgentClient` that calls the new Cloud API
endpoint using the existing authenticated `httpx` session, importing
the new request/response models from `cloud.internal_api.models`.
- [ ] 3.3 Add `host_agent/cloud_planner_client.py::CloudProxyToolCallingClient`
implementing the `runtime.tool_calling_client.ToolCallingClient`
Protocol structurally (no import of `host_agent`/`cloud` from
`runtime`), wrapping `HostAgentClient.request_planner_decision(...)`
and raising `ToolCallUnavailable` on any failure (network error, auth
rejection, non-2xx, provider error, timeout) -- matching D7.
## 4. Host Agent: wiring
- [ ] 4.1 Update `apps/device-host-agent/host_agent/execution.py`'s
`_host_agent_planner_config()`/`create_task_runner()` so that when
`AI_PLANNER_TRANSPORT=cloud`, the constructed `TaskRunner`'s
`AIPlanner` is built with a `CloudProxyToolCallingClient` instead of
the default local provider client, while leaving the existing
default-enabled/direct-transport behavior unchanged when
`AI_PLANNER_TRANSPORT` is unset or `direct`.
- [ ] 4.2 Confirm `apps/device-host-agent/tests/test_execution.py::test_runtime_owned_packages_do_not_import_host_or_cloud_concerns`-style
boundary checks still pass with the new module in place.
## 5. Tests
- [ ] 5.1 Unit tests for Cloud API planner configuration loading
(defaults, provider/model/timeout parsing) mirroring
`tests/test_planner_config.py`.
- [ ] 5.2 Unit tests for the planner-decision endpoint: authorized request
resolves a decision, unauthenticated/foreign-host request is
rejected, provider failure returns a structured error without
crashing.
- [ ] 5.3 Unit tests for `CloudProxyToolCallingClient`: successful decision
round-trip, and each failure mode raises `ToolCallUnavailable`.
- [ ] 5.4 Unit tests for Host Agent wiring: `AI_PLANNER_TRANSPORT=cloud`
constructs an `AIPlanner` using `CloudProxyToolCallingClient`;
`AI_PLANNER_TRANSPORT` unset or `direct` preserves existing
direct-to-provider construction (no regression to the
already-implemented default-enabled behavior).
- [ ] 5.5 Full non-integration suite (`uv run --all-packages pytest -m
"not integration"`) passes with no regressions.
## 6. Documentation
- [ ] 6.1 Update `docs/CLOUD_DEPLOYMENT.md`'s Runtime AI Planner section
with the `cloud` transport configuration path, its trade-offs
(latency, cloud-availability coupling, expanded data path for
screenshots/prompts), and the credential split (Cloud API holds
provider keys for `cloud` transport; Host Agent holds them for
`direct` transport).
## 7. Validation
- [ ] 7.1 `openspec validate --strict` passes for this change.
- [ ] 7.2 Ruff check/format and `compileall` pass for all touched packages.