cleveragents/cleveractors-core

Fork 0

compile_actor() drops graph-level system_prompt — not threaded into CompiledActor #6

New issue

Closed

opened 2026-05-21 08:50:00 +00:00 by hurui200320 · 2 comments

hurui200320 commented

2026-05-21 08:50:00 +00:00

Member

Summary

compile_actor() silently drops the actor-level system_prompt field for type: graph actors. The CompiledActor contains no trace of it, forcing host applications to keep the raw ActorConfigSchema in scope alongside the compiled bundle as a workaround.

Metadata

Commit message: fix(compiler): thread actor-level system_prompt into graph node metadata
Branch: fix/compiler-thread-system-prompt

Details

What the spec says

docs/reference/actors_schema.md (line 92) explicitly documents system_prompt as:

Default prompt for graph-level agents

The word "default" places it in the same category as provider and model: actor-level fields that serve as fallback defaults for all AGENT nodes in the graph.

What the code does

actor/compiler.py:_map_node() already implements the default-threading pattern for provider and model:

merged_meta: dict[str, Any] = dict(config)
if node.type == NodeType.AGENT:
    merged_meta.setdefault("provider", actor_provider)
    merged_meta.setdefault("model", actor_model)
    # system_prompt is never passed or set here  ← gap

_map_node() is never given actor_system_prompt and never writes it to merged_meta. After compile_actor() returns, system_prompt is inaccessible from CompiledActor.

Additionally, Node._execute_agent() merges state.metadata into the context dict that reaches agent.process_message(), but it never includes self.config.metadata.get("system_prompt"). Even after the threading fix is applied to the compiler, hosts cannot read the system prompt from within process_message() unless this second gap is also addressed.

Impact

All type: graph actors with a top-level system_prompt silently lose it after compilation. The compiled bundle is incomplete relative to the schema.
Host applications must keep the raw ActorConfigSchema in scope as a workaround, which defeats the purpose of storing a compiled bundle.

Fix

Three targeted additions across two files:

actor/compiler.py — _map_node(): accept and thread actor_system_prompt

def _map_node(
    node: NodeDefinition,
    actor_provider: str | None = None,
    actor_model: str | None = None,
    actor_system_prompt: str | None = None,   # add
) -> lg_nodes.NodeConfig:
    ...
    if node.type == NodeType.AGENT:
        merged_meta.setdefault("provider", actor_provider)
        merged_meta.setdefault("model", actor_model)
        merged_meta.setdefault("system_prompt", actor_system_prompt)  # add

actor/compiler.py — compile_actor(): pass config.system_prompt through

nodes[node_def.id] = _map_node(
    node_def,
    actor_provider=config.provider,
    actor_model=config.model,
    actor_system_prompt=config.system_prompt,  # add
)

langgraph/nodes.py — _execute_agent(): expose it in the context dict

node_system_prompt = self.config.metadata.get("system_prompt")
if node_system_prompt:
    context["system_prompt"] = node_system_prompt  # add before context.update(state.metadata)

The last change allows the host's Agent.process_message() to read context.get("system_prompt") without needing a direct reference to NodeConfig. No changes to ProviderRegistryPort.get() are required.

Definition of Done

_map_node() accepts and threads actor_system_prompt into AGENT node metadata via setdefault
compile_actor() passes config.system_prompt to _map_node()
Node._execute_agent() adds metadata["system_prompt"] to the context dict when present
BDD scenario: graph actor with system_prompt → compiled NodeConfig.metadata contains system_prompt
BDD scenario: process_message() receives system_prompt in context dict
docs/reference/actor_compiler.md updated to document that system_prompt is threaded into node metadata alongside provider and model

# Summary `compile_actor()` silently drops the actor-level `system_prompt` field for `type: graph` actors. The `CompiledActor` contains no trace of it, forcing host applications to keep the raw `ActorConfigSchema` in scope alongside the compiled bundle as a workaround. # Metadata - Commit message: `fix(compiler): thread actor-level system_prompt into graph node metadata` - Branch: `fix/compiler-thread-system-prompt` # Details ## What the spec says `docs/reference/actors_schema.md` (line 92) explicitly documents `system_prompt` as: > **Default prompt for graph-level agents** The word "default" places it in the same category as `provider` and `model`: actor-level fields that serve as fallback defaults for all AGENT nodes in the graph. ## What the code does `actor/compiler.py:_map_node()` already implements the default-threading pattern for `provider` and `model`: ```python merged_meta: dict[str, Any] = dict(config) if node.type == NodeType.AGENT: merged_meta.setdefault("provider", actor_provider) merged_meta.setdefault("model", actor_model) # system_prompt is never passed or set here ← gap ``` `_map_node()` is never given `actor_system_prompt` and never writes it to `merged_meta`. After `compile_actor()` returns, `system_prompt` is inaccessible from `CompiledActor`. Additionally, `Node._execute_agent()` merges `state.metadata` into the context dict that reaches `agent.process_message()`, but it never includes `self.config.metadata.get("system_prompt")`. Even after the threading fix is applied to the compiler, hosts cannot read the system prompt from within `process_message()` unless this second gap is also addressed. ## Impact - All `type: graph` actors with a top-level `system_prompt` silently lose it after compilation. The compiled bundle is incomplete relative to the schema. - Host applications must keep the raw `ActorConfigSchema` in scope as a workaround, which defeats the purpose of storing a compiled bundle. ## Fix Three targeted additions across two files: **`actor/compiler.py` — `_map_node()`**: accept and thread `actor_system_prompt` ```python def _map_node( node: NodeDefinition, actor_provider: str | None = None, actor_model: str | None = None, actor_system_prompt: str | None = None, # add ) -> lg_nodes.NodeConfig: ... if node.type == NodeType.AGENT: merged_meta.setdefault("provider", actor_provider) merged_meta.setdefault("model", actor_model) merged_meta.setdefault("system_prompt", actor_system_prompt) # add ``` **`actor/compiler.py` — `compile_actor()`**: pass `config.system_prompt` through ```python nodes[node_def.id] = _map_node( node_def, actor_provider=config.provider, actor_model=config.model, actor_system_prompt=config.system_prompt, # add ) ``` **`langgraph/nodes.py` — `_execute_agent()`**: expose it in the context dict ```python node_system_prompt = self.config.metadata.get("system_prompt") if node_system_prompt: context["system_prompt"] = node_system_prompt # add before context.update(state.metadata) ``` The last change allows the host's `Agent.process_message()` to read `context.get("system_prompt")` without needing a direct reference to `NodeConfig`. No changes to `ProviderRegistryPort.get()` are required. # Definition of Done - [x] `_map_node()` accepts and threads `actor_system_prompt` into AGENT node `metadata` via `setdefault` - [x] `compile_actor()` passes `config.system_prompt` to `_map_node()` - [x] `Node._execute_agent()` adds `metadata["system_prompt"]` to the context dict when present - [x] BDD scenario: graph actor with `system_prompt` → compiled `NodeConfig.metadata` contains `system_prompt` - [x] BDD scenario: `process_message()` receives `system_prompt` in context dict - [x] `docs/reference/actor_compiler.md` updated to document that `system_prompt` is threaded into node metadata alongside `provider` and `model`

hurui200320 added the

labels

2026-05-21 08:50:24 +00:00

hurui200320 referenced this issue

2026-05-21 08:58:36 +00:00

Clever Actors only accepts GRAPH actors, not LLM actors. #2

hurui200320 added

and removed

labels

2026-05-22 03:42:45 +00:00

hurui200320 commented

2026-05-22 03:53:02 +00:00

Author

Member

Implementation Notes

Branch: `fix/compiler-thread-system-prompt`

Root cause

actor/compiler.py:_map_node() already threaded provider and model from the actor-level config into each AGENT node's merged_meta dict via setdefault. The system_prompt field was simply never added to the same pattern — a three-line omission.

A second, independent gap in langgraph/nodes.py:Node._execute_agent() meant that even after the compiler fix the system_prompt value would not appear in the context dict passed to agent.process_message().

Files changed

src/cleveractors/actor/compiler.py — _map_node() and compile_actor()

Added actor_system_prompt: str | None = None parameter to _map_node().
Added merged_meta.setdefault("system_prompt", actor_system_prompt) inside the if node.type == NodeType.AGENT: block, directly after the existing provider/model setdefault calls.
Updated compile_actor() to pass actor_system_prompt=config.system_prompt in the _map_node() call.
setdefault semantics preserve any per-node config.system_prompt override, exactly matching how provider and model work.

src/cleveractors/langgraph/nodes.py — Node._execute_agent()

Added extraction of node_system_prompt = self.config.metadata.get("system_prompt") before the context.update(state.metadata) call.
If truthy, injects context["system_prompt"] = node_system_prompt so the host agent can read it from context.get("system_prompt") without referencing the raw ActorConfigSchema.
Placed before state.metadata update so per-execution overrides via state.metadata["system_prompt"] still take precedence.

docs/reference/actor_compiler.md — Node Binding table

Updated the AGENT row to document that system_prompt is now threaded alongside provider and model, and that Node._execute_agent() injects it into the context dict.

features/compiler_system_prompt.feature (new) — 3 BDD scenarios

Actor-level system_prompt is threaded into each AGENT node's metadata on compile.
Per-node system_prompt in node config takes precedence via setdefault semantics.
process_message() receives system_prompt in its context dict when the node's metadata contains it.

features/steps/compiler_system_prompt_steps.py (new) — step definitions

_CapturingAgent: records the context dict received by process_message().
_CapturingProviderRegistry: returns the same _CapturingAgent instance for a given provider/model pair, allowing the test to inspect the received context post-execution.
Reuses @when("the actor configuration is compiled") from smoke_steps.py (Behave loads all step files together).

Quality gates

All gates pass on branch fix/compiler-thread-system-prompt:

Gate	Result
`nox -s lint`	✅ pass
`nox -s typecheck`	✅ pass (0 errors, 0 warnings)
`nox -s unit_tests`	✅ pass (10/10 scenarios: 3 new + 7 existing)
`nox -s coverage_report`	✅ pass (threshold 97%)
`nox -s security_scan`	✅ pass (0 findings)
`nox -s dead_code`	✅ pass
`nox -s complexity`	✅ pass (avg complexity A)

## Implementation Notes ### Branch: `fix/compiler-thread-system-prompt` ### Root cause `actor/compiler.py:_map_node()` already threaded `provider` and `model` from the actor-level config into each AGENT node's `merged_meta` dict via `setdefault`. The `system_prompt` field was simply never added to the same pattern — a three-line omission. A second, independent gap in `langgraph/nodes.py:Node._execute_agent()` meant that even after the compiler fix the `system_prompt` value would not appear in the context dict passed to `agent.process_message()`. ### Files changed **`src/cleveractors/actor/compiler.py`** — `_map_node()` and `compile_actor()` - Added `actor_system_prompt: str | None = None` parameter to `_map_node()`. - Added `merged_meta.setdefault("system_prompt", actor_system_prompt)` inside the `if node.type == NodeType.AGENT:` block, directly after the existing `provider`/`model` setdefault calls. - Updated `compile_actor()` to pass `actor_system_prompt=config.system_prompt` in the `_map_node()` call. - `setdefault` semantics preserve any per-node `config.system_prompt` override, exactly matching how `provider` and `model` work. **`src/cleveractors/langgraph/nodes.py`** — `Node._execute_agent()` - Added extraction of `node_system_prompt = self.config.metadata.get("system_prompt")` before the `context.update(state.metadata)` call. - If truthy, injects `context["system_prompt"] = node_system_prompt` so the host agent can read it from `context.get("system_prompt")` without referencing the raw `ActorConfigSchema`. - Placed before `state.metadata` update so per-execution overrides via `state.metadata["system_prompt"]` still take precedence. **`docs/reference/actor_compiler.md`** — Node Binding table - Updated the `AGENT` row to document that `system_prompt` is now threaded alongside `provider` and `model`, and that `Node._execute_agent()` injects it into the context dict. **`features/compiler_system_prompt.feature`** (new) — 3 BDD scenarios 1. Actor-level `system_prompt` is threaded into each AGENT node's `metadata` on compile. 2. Per-node `system_prompt` in node `config` takes precedence via `setdefault` semantics. 3. `process_message()` receives `system_prompt` in its context dict when the node's metadata contains it. **`features/steps/compiler_system_prompt_steps.py`** (new) — step definitions - `_CapturingAgent`: records the context dict received by `process_message()`. - `_CapturingProviderRegistry`: returns the same `_CapturingAgent` instance for a given provider/model pair, allowing the test to inspect the received context post-execution. - Reuses `@when("the actor configuration is compiled")` from `smoke_steps.py` (Behave loads all step files together). ### Quality gates All gates pass on branch `fix/compiler-thread-system-prompt`: | Gate | Result | |------|--------| | `nox -s lint` | ✅ pass | | `nox -s typecheck` | ✅ pass (0 errors, 0 warnings) | | `nox -s unit_tests` | ✅ pass (10/10 scenarios: 3 new + 7 existing) | | `nox -s coverage_report` | ✅ pass (threshold 97%) | | `nox -s security_scan` | ✅ pass (0 findings) | | `nox -s dead_code` | ✅ pass | | `nox -s complexity` | ✅ pass (avg complexity A) |

hurui200320 referenced this issue from a commit

2026-05-22 03:53:37 +00:00

fix(compiler): thread actor-level system_prompt into graph node metadata

hurui200320 referenced this issue from a pull request that will close it,

2026-05-22 03:54:00 +00:00

fix(compiler): thread actor-level system_prompt into graph node metadata #7

hurui200320 added

and removed

labels

2026-05-22 03:54:11 +00:00

hurui200320 added a new dependency

2026-05-22 03:56:22 +00:00

#7 fix(compiler): thread actor-level system_prompt into graph node metadata

hurui200320 self-assigned this

2026-05-22 03:56:29 +00:00

hurui200320 referenced this issue from a commit

2026-05-22 03:57:32 +00:00

fix(compiler): thread actor-level system_prompt into graph node metadata

hurui200320 referenced this issue from a commit

2026-05-22 04:30:06 +00:00

fix(compiler): thread actor-level system_prompt into graph node metadata

hurui200320 referenced this issue from a commit

2026-05-22 05:18:18 +00:00

fix(compiler): thread actor-level system_prompt into graph node metadata

hurui200320 referenced this issue from a commit

2026-05-22 05:22:23 +00:00

fix(compiler): thread actor-level system_prompt into graph node metadata

hurui200320 referenced this issue from a commit

2026-05-22 06:05:27 +00:00

fix(compiler): thread actor-level system_prompt into graph node metadata

hurui200320 commented

2026-05-22 06:28:37 +00:00

Author

Member

Self-QA Implementation Notes (Cycles 1–2)

Cycle 1

Review findings (0C / 3M / 5m / 2n — Request Changes):

M1 — Missing test for empty-string system_prompt (compiler + runtime levels). The advertised is not None fix had no regression guard.
M2 — Missing test for state.metadata = {"system_prompt": None} overriding a compiled default at runtime.
M3 — Scenario 5 only tested TOOL nodes; CONDITIONAL and SUBGRAPH node types were unverified.
m1 — setdefault("system_prompt", None) unconditionally inserted "system_prompt": None into every AGENT node's metadata even when the actor had no system_prompt, silently changing metadata shape.
m2 — metadata.get("system_prompt") returns Any; the str | None annotation was aspirational with no runtime guard.
m3 — system_prompt was absent from ignore_keys, allowing agent mutations to leak into metadata_updates.
m4 — # Then steps for new scenarios comment was a drafting artifact.
m5 — "each AGENT node" loops had no guard against zero AGENT nodes, making them tautologically pass on empty graphs.
n1 — AGENT row in the Node Binding table was excessively dense.
n2 — metadata.get() type-safety gap is a pre-existing pattern (tracked as future improvement, no action).

Fixes applied:

compiler.py _map_node(): Guarded setdefault with if actor_system_prompt is not None: so that actors without a top-level system_prompt no longer inject a None key into AGENT node metadata.
nodes.py _execute_agent(): Replaced is not None with isinstance(node_system_prompt, str) for defensive type safety; added "system_prompt" to ignore_keys.
compiler_system_prompt.feature: Scenario 6 updated to assert key absence (not None); 3 new scenarios added for empty-string compile/runtime (M1) and None runtime override (M2).
Mixed-graph scenario extended with CONDITIONAL + SUBGRAPH nodes (M3).
Step file reorganized into functional sections (# Then steps — compilation assertions / # Then steps — runtime assertions); agent_nodes_found guard added to all "each AGENT node" loops.
Documentation AGENT table row de-densified; new "AGENT node defaults" subsection added.

All 7 quality gates passed: 17/17 scenarios (up from 14), coverage ≥ 97%.

Cycle 2

Review findings (0C / 0M / 3m / 3n — Approve):

m1 — Missing test for non-string system_prompt in node metadata (the isinstance guard's non-None branch).
m2 — Missing assertion that ignore_keys prevents system_prompt from persisting to metadata_updates.
m3 — Missing scenario for actor-level None + node-level system_prompt override (setdefault semantics when fallback is absent).
n1 — Commit message body is stale (mentions 7 scenarios, 14/14, old guard semantics; actual is 10 scenarios, 17/17, isinstance guard).
n2 — _CapturingAgent uses shallow copy of context; deepcopy would be safer.
n3 — @when('... containing system_prompt "{prompt}"') step pattern fragile for unquoted None.

Fixes applied:

None — Cycle 2 verdict was Approve. The remaining minor/nit findings do not affect correctness and the reviewer confirmed the PR is ready to merge.

Remaining Issues

The following items from Cycle 2 are open but non-blocking:

ID	Severity	Description
m1	Minor	No BDD scenario for non-string (non-None) `system_prompt` injected via `NodeConfig.metadata`
m2	Minor	No explicit assertion that `ignore_keys` exclusion of `system_prompt` works
m3	Minor	No scenario for actor-level absent `system_prompt` + per-node override combination
n1	Nit	Commit message body references 7 scenarios / 14/14 / old guard — should be amended
n2	Nit	`_CapturingAgent.received_context = dict(context)` should use `deepcopy`
n3	Nit	Gherkin step pattern fragile for `"None"` string vs unquoted `None`

These can be addressed as a follow-up in a separate ticket if desired. No action required before merge.

## Self-QA Implementation Notes (Cycles 1–2) ### Cycle 1 **Review findings (0C / 3M / 5m / 2n — Request Changes):** - **M1** — Missing test for empty-string `system_prompt` (compiler + runtime levels). The advertised `is not None` fix had no regression guard. - **M2** — Missing test for `state.metadata = {"system_prompt": None}` overriding a compiled default at runtime. - **M3** — Scenario 5 only tested TOOL nodes; CONDITIONAL and SUBGRAPH node types were unverified. - **m1** — `setdefault("system_prompt", None)` unconditionally inserted `"system_prompt": None` into every AGENT node's metadata even when the actor had no `system_prompt`, silently changing metadata shape. - **m2** — `metadata.get("system_prompt")` returns `Any`; the `str | None` annotation was aspirational with no runtime guard. - **m3** — `system_prompt` was absent from `ignore_keys`, allowing agent mutations to leak into `metadata_updates`. - **m4** — `# Then steps for new scenarios` comment was a drafting artifact. - **m5** — "each AGENT node" loops had no guard against zero AGENT nodes, making them tautologically pass on empty graphs. - **n1** — AGENT row in the Node Binding table was excessively dense. - **n2** — `metadata.get()` type-safety gap is a pre-existing pattern (tracked as future improvement, no action). **Fixes applied:** - `compiler.py` `_map_node()`: Guarded `setdefault` with `if actor_system_prompt is not None:` so that actors without a top-level `system_prompt` no longer inject a `None` key into AGENT node metadata. - `nodes.py` `_execute_agent()`: Replaced `is not None` with `isinstance(node_system_prompt, str)` for defensive type safety; added `"system_prompt"` to `ignore_keys`. - `compiler_system_prompt.feature`: Scenario 6 updated to assert key absence (not `None`); 3 new scenarios added for empty-string compile/runtime (M1) and `None` runtime override (M2). - Mixed-graph scenario extended with CONDITIONAL + SUBGRAPH nodes (M3). - Step file reorganized into functional sections (`# Then steps — compilation assertions` / `# Then steps — runtime assertions`); `agent_nodes_found` guard added to all "each AGENT node" loops. - Documentation AGENT table row de-densified; new "AGENT node defaults" subsection added. All 7 quality gates passed: **17/17 scenarios** (up from 14), coverage ≥ 97%. --- ### Cycle 2 **Review findings (0C / 0M / 3m / 3n — Approve):** - **m1** — Missing test for non-string `system_prompt` in node metadata (the `isinstance` guard's non-None branch). - **m2** — Missing assertion that `ignore_keys` prevents `system_prompt` from persisting to `metadata_updates`. - **m3** — Missing scenario for actor-level `None` + node-level `system_prompt` override (setdefault semantics when fallback is absent). - **n1** — Commit message body is stale (mentions 7 scenarios, 14/14, old guard semantics; actual is 10 scenarios, 17/17, `isinstance` guard). - **n2** — `_CapturingAgent` uses shallow copy of context; `deepcopy` would be safer. - **n3** — `@when('... containing system_prompt "{prompt}"')` step pattern fragile for unquoted `None`. **Fixes applied:** None — Cycle 2 verdict was **Approve**. The remaining minor/nit findings do not affect correctness and the reviewer confirmed the PR is ready to merge. --- ### Remaining Issues The following items from Cycle 2 are open but non-blocking: | ID | Severity | Description | |----|----------|-------------| | m1 | Minor | No BDD scenario for non-string (non-None) `system_prompt` injected via `NodeConfig.metadata` | | m2 | Minor | No explicit assertion that `ignore_keys` exclusion of `system_prompt` works | | m3 | Minor | No scenario for actor-level absent `system_prompt` + per-node override combination | | n1 | Nit | Commit message body references 7 scenarios / 14/14 / old guard — should be amended | | n2 | Nit | `_CapturingAgent.received_context = dict(context)` should use `deepcopy` | | n3 | Nit | Gherkin step pattern fragile for `"None"` string vs unquoted `None` | These can be addressed as a follow-up in a separate ticket if desired. No action required before merge.

hurui200320

2026-05-22 07:08:20 +00:00