agentforge-oss 0.2.0__tar.gz → 0.4.0__tar.gz

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentforge-oss
-Version: 0.2.0
+Version: 0.4.0
 Summary: Forge — an open-source, enterprise-ready multi-agent orchestration platform with cost awareness, governance, and security built in.
 Project-URL: Homepage, https://github.com/sekacorn/AgentForge
 Project-URL: Documentation, https://github.com/sekacorn/AgentForge#readme
@@ -30,10 +30,17 @@ Requires-Dist: rich>=13.7
 Requires-Dist: typer>=0.12
 Requires-Dist: typing-extensions>=4.10
 Provides-Extra: all
+Requires-Dist: aiosqlite>=0.19; extra == 'all'
 Requires-Dist: anthropic>=0.20; extra == 'all'
+Requires-Dist: boto3>=1.34; extra == 'all'
 Requires-Dist: openai>=1.0; extra == 'all'
+Requires-Dist: opentelemetry-api>=1.20; extra == 'all'
+Requires-Dist: opentelemetry-exporter-otlp-proto-grpc>=1.20; extra == 'all'
+Requires-Dist: opentelemetry-sdk>=1.20; extra == 'all'
 Provides-Extra: anthropic
 Requires-Dist: anthropic>=0.20; extra == 'anthropic'
+Provides-Extra: bedrock
+Requires-Dist: boto3>=1.34; extra == 'bedrock'
 Provides-Extra: dev
 Requires-Dist: mypy>=1.10; extra == 'dev'
 Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
@@ -42,6 +49,12 @@ Requires-Dist: pytest>=8.0; extra == 'dev'
 Requires-Dist: ruff>=0.6; extra == 'dev'
 Provides-Extra: openai
 Requires-Dist: openai>=1.0; extra == 'openai'
+Provides-Extra: otel
+Requires-Dist: opentelemetry-api>=1.20; extra == 'otel'
+Requires-Dist: opentelemetry-exporter-otlp-proto-grpc>=1.20; extra == 'otel'
+Requires-Dist: opentelemetry-sdk>=1.20; extra == 'otel'
+Provides-Extra: sqlite
+Requires-Dist: aiosqlite>=0.19; extra == 'sqlite'
 Description-Content-Type: text/markdown
 
 <div align="center">
@@ -58,7 +71,7 @@ with cost-awareness, security, and compliance built in from line one.**
 [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/python-3.11%2B-blue.svg)](https://www.python.org/)
 [![Typed](https://img.shields.io/badge/typed-mypy%20strict-blue.svg)](pyproject.toml)
-[![Status: Beta](https://img.shields.io/badge/status-beta-orange.svg)](#whats-shipped-v020)
+[![Status: Beta](https://img.shields.io/badge/status-beta-orange.svg)](#whats-shipped-v040)
 
 </div>
 
@@ -101,7 +114,7 @@ asyncio.run(main())
 
 ---
 
-## What's shipped (v0.2.0)
+## What's shipped (v0.4.0)
 
 An honest snapshot of what works today versus what is on the way. Everything marked
 **Shipped** is implemented, typed, and covered by the test suite.
@@ -113,6 +126,8 @@ An honest snapshot of what works today versus what is on the way. Everything mar
 | Anthropic provider (Claude Haiku 4.5, Sonnet 4.6, Opus 4.8, Fable 5) | Shipped |
 | OpenAI provider (gpt-4o-mini, gpt-4o, gpt-4.1, o3) | Shipped |
 | Offline deterministic provider (zero config, no API key) | Shipped |
+| Ollama provider (local models, zero cost, no API key, auto-detected) | Shipped |
+| Amazon Bedrock provider (Claude/Llama/Mistral via Converse, in-account/GovCloud) | Shipped |
 | Pre-flight + per-step budget caps | Shipped |
 | Tool sandboxing (allowlist/denylist, timeouts, dangerous-denied-by-default) | Shipped |
 | RBAC (admin / operator / developer / viewer) | Shipped |
@@ -122,12 +137,13 @@ An honest snapshot of what works today versus what is on the way. Everything mar
 | Event bus (21 lifecycle event types) | Shipped |
 | Streaming token output through the event bus (`stream=True`) | Shipped |
 | Per-run cost reporting (tokens + USD, per model, per agent) | Shipped |
+| OpenTelemetry export (traces + metrics: console or OTLP to Jaeger/Grafana/Datadog) | Shipped |
 | Conversation memory + in-memory RAG vector store | Shipped |
+| Durable memory backend (SQLite, persistent RAG, no vector extension) | Shipped |
 | CLI (`forge run`, `forge models`, `forge audit`) | Shipped |
-| 47 tests, mypy strict, ruff clean, CI on 3.11 / 3.12 / 3.13 | Shipped |
-| Durable memory backends (pgvector, SQLite-VSS) | Planned |
-| OpenTelemetry export for traces and metrics | Planned |
-| Ollama / Bedrock / Vertex providers | Planned |
+| 83 tests, mypy strict, ruff clean, CI on 3.11 / 3.12 / 3.13 | Shipped |
+| Durable memory backends (pgvector, Redis) | Planned |
+| Vertex / other cloud providers | Planned |
 | Policy-as-code for tool governance | Planned |
 | Hosted SaaS control plane (TypeScript / Next.js) | Future |
 
@@ -177,9 +193,10 @@ An honest snapshot of what works today versus what is on the way. Everything mar
   JSON-Schema generated automatically from your type hints and docstring.
 - **Pluggable memory.** Short-term conversation memory plus a dependency-free
   in-memory vector store for RAG — swap in any backend behind one tiny interface.
-- **Provider-agnostic core.** Anthropic (Claude) and OpenAI (GPT / o-series) ship in
-  the box alongside a deterministic offline echo provider; add any provider by
-  implementing one method.
+- **Provider-agnostic core.** Anthropic (Claude), OpenAI (GPT / o-series), Amazon
+  Bedrock (Converse API), and Ollama (local models, zero cost) ship in the box
+  alongside a deterministic offline echo provider; add any provider by implementing
+  one method.
 
 ### Security from the start
 - **Tool sandboxing** with allowlists/denylists, per-tool timeouts, and
@@ -204,6 +221,10 @@ An honest snapshot of what works today versus what is on the way. Everything mar
   `TOKEN_STREAM_START` / `TOKEN_CHUNK` / `TOKEN_STREAM_END` events — each tagged with
   the agent that produced it, so you can render live output even across parallel
   workers. `forge run "..." --stream` gives the classic live-typing terminal feel.
+- **OpenTelemetry export.** Every run becomes a tree of spans (`forge.run` →
+  `forge.agent` → `forge.model_call` → `forge.tool_call`) exportable to any
+  OTel-compatible backend — Jaeger, Grafana, Datadog, Honeycomb, New Relic. Console
+  exporter by default (zero infra); set an OTLP endpoint for production.
 - **Structured logging** (human or JSON) and a per-run **usage/cost report** broken
   down per model and per agent.
 
@@ -216,6 +237,7 @@ pip install agentforge-oss                       # core (works offline, zero con
 pip install "agentforge-oss[anthropic]"          # + Claude provider
 pip install "agentforge-oss[openai]"             # + OpenAI / GPT provider
 pip install "agentforge-oss[anthropic,openai]"   # both real providers
+pip install "agentforge-oss[bedrock]"            # + Amazon Bedrock provider (boto3)
 pip install "agentforge-oss[all,dev]"            # everything + test/lint tooling
 ```
 
@@ -229,6 +251,15 @@ pip install "agentforge-oss[all,dev]"            # everything + test/lint toolin
 > to route to GPT. Both keys can be set at once; Forge prefers Anthropic by default
 > (configurable).
 
+> **Ollama support is built in** — no extra install needed (it uses `httpx`, already a
+> core dependency, so there is no `[ollama]` extra). Just run Ollama locally and Forge
+> auto-detects it (or set `OLLAMA_BASE_URL` for a custom/remote server):
+
+```bash
+ollama serve
+ollama pull llama3.1:8b
+```
+
 ---
 
 ## Quickstart (CLI)
@@ -341,10 +372,10 @@ governance (RBAC + budgets + audit verification).
               │                          │
               │      ┌───────────────────┴───────────────┐
               ▼      ▼                                    ▼
-       ┌────────────────┐                        ┌───────────────────────────┐
-       │  Model Router  │  picks model by        │  Model Providers          │
-       │ cost / quality │  strategy + budget ───▶│ Anthropic · OpenAI · Echo │
-       └───────┬────────┘                        └───────────────────────────┘
+       ┌────────────────┐                        ┌────────────────────────────────────┐
+       │  Model Router  │  picks model by        │  Model Providers                   │
+       │ cost / quality │  strategy + budget ───▶│ Anthropic · OpenAI · Ollama · Bedrock · Echo │
+       └───────┬────────┘                        └────────────────────────────────────┘
                │ pricing
                ▼
        ┌────────────────┐   cross-cutting, on every step:
@@ -356,9 +387,9 @@ Every layer is swappable:
 
 | Layer | Default | Swap in… |
 |---|---|---|
-| Provider | Echo (offline), Anthropic, OpenAI | Any `ModelProvider` (local, Bedrock, Vertex, …) |
+| Provider | Echo (offline), Anthropic, OpenAI, Ollama, Bedrock | Any `ModelProvider` (Vertex, …) |
 | Routing | `balanced` strategy | Your own strategy / `fixed` model |
-| Memory | In-memory vector store | Any `Memory` backend (pgvector, Pinecone, …) |
+| Memory | InMemoryVectorStore (default), SQLiteMemoryStore | Any `Memory` backend (pgvector, Redis, …) |
 | Tools | `calculator`, `utc_now` | Any `@tool` function |
 | Audit | Hash-chained JSONL | Forward events to your SIEM via the event bus |
 
@@ -449,7 +480,7 @@ ruff check .      # lint
 mypy forge        # strict type-check
 ```
 
-The entire 47-test suite runs offline against the deterministic provider — fast,
+The entire 83-test suite runs offline against the deterministic provider — fast,
 hermetic, and free. CI runs the same checks (ruff, ruff format, mypy strict, pytest)
 on Python 3.11, 3.12, and 3.13.
 
@@ -462,7 +493,7 @@ the core typed (`mypy --strict`) and tested.
 
 ### Good first contributions
 
-- **New model provider** (Ollama, Bedrock, Vertex) — implement one method.
+- **New model provider** (Vertex, Gemini, Cohere) — implement one method.
 - **New built-in tool** (web search, file read, database query).
 - **Durable memory backend** (SQLite-VSS, pgvector, Redis).
 - **Routing strategy** (a custom cost/quality tradeoff).

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/README.md

@@ -12,7 +12,7 @@ with cost-awareness, security, and compliance built in from line one.**
 [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/python-3.11%2B-blue.svg)](https://www.python.org/)
 [![Typed](https://img.shields.io/badge/typed-mypy%20strict-blue.svg)](pyproject.toml)
-[![Status: Beta](https://img.shields.io/badge/status-beta-orange.svg)](#whats-shipped-v020)
+[![Status: Beta](https://img.shields.io/badge/status-beta-orange.svg)](#whats-shipped-v040)
 
 </div>
 
@@ -55,7 +55,7 @@ asyncio.run(main())
 
 ---
 
-## What's shipped (v0.2.0)
+## What's shipped (v0.4.0)
 
 An honest snapshot of what works today versus what is on the way. Everything marked
 **Shipped** is implemented, typed, and covered by the test suite.
@@ -67,6 +67,8 @@ An honest snapshot of what works today versus what is on the way. Everything mar
 | Anthropic provider (Claude Haiku 4.5, Sonnet 4.6, Opus 4.8, Fable 5) | Shipped |
 | OpenAI provider (gpt-4o-mini, gpt-4o, gpt-4.1, o3) | Shipped |
 | Offline deterministic provider (zero config, no API key) | Shipped |
+| Ollama provider (local models, zero cost, no API key, auto-detected) | Shipped |
+| Amazon Bedrock provider (Claude/Llama/Mistral via Converse, in-account/GovCloud) | Shipped |
 | Pre-flight + per-step budget caps | Shipped |
 | Tool sandboxing (allowlist/denylist, timeouts, dangerous-denied-by-default) | Shipped |
 | RBAC (admin / operator / developer / viewer) | Shipped |
@@ -76,12 +78,13 @@ An honest snapshot of what works today versus what is on the way. Everything mar
 | Event bus (21 lifecycle event types) | Shipped |
 | Streaming token output through the event bus (`stream=True`) | Shipped |
 | Per-run cost reporting (tokens + USD, per model, per agent) | Shipped |
+| OpenTelemetry export (traces + metrics: console or OTLP to Jaeger/Grafana/Datadog) | Shipped |
 | Conversation memory + in-memory RAG vector store | Shipped |
+| Durable memory backend (SQLite, persistent RAG, no vector extension) | Shipped |
 | CLI (`forge run`, `forge models`, `forge audit`) | Shipped |
-| 47 tests, mypy strict, ruff clean, CI on 3.11 / 3.12 / 3.13 | Shipped |
-| Durable memory backends (pgvector, SQLite-VSS) | Planned |
-| OpenTelemetry export for traces and metrics | Planned |
-| Ollama / Bedrock / Vertex providers | Planned |
+| 83 tests, mypy strict, ruff clean, CI on 3.11 / 3.12 / 3.13 | Shipped |
+| Durable memory backends (pgvector, Redis) | Planned |
+| Vertex / other cloud providers | Planned |
 | Policy-as-code for tool governance | Planned |
 | Hosted SaaS control plane (TypeScript / Next.js) | Future |
 
@@ -131,9 +134,10 @@ An honest snapshot of what works today versus what is on the way. Everything mar
   JSON-Schema generated automatically from your type hints and docstring.
 - **Pluggable memory.** Short-term conversation memory plus a dependency-free
   in-memory vector store for RAG — swap in any backend behind one tiny interface.
-- **Provider-agnostic core.** Anthropic (Claude) and OpenAI (GPT / o-series) ship in
-  the box alongside a deterministic offline echo provider; add any provider by
-  implementing one method.
+- **Provider-agnostic core.** Anthropic (Claude), OpenAI (GPT / o-series), Amazon
+  Bedrock (Converse API), and Ollama (local models, zero cost) ship in the box
+  alongside a deterministic offline echo provider; add any provider by implementing
+  one method.
 
 ### Security from the start
 - **Tool sandboxing** with allowlists/denylists, per-tool timeouts, and
@@ -158,6 +162,10 @@ An honest snapshot of what works today versus what is on the way. Everything mar
   `TOKEN_STREAM_START` / `TOKEN_CHUNK` / `TOKEN_STREAM_END` events — each tagged with
   the agent that produced it, so you can render live output even across parallel
   workers. `forge run "..." --stream` gives the classic live-typing terminal feel.
+- **OpenTelemetry export.** Every run becomes a tree of spans (`forge.run` →
+  `forge.agent` → `forge.model_call` → `forge.tool_call`) exportable to any
+  OTel-compatible backend — Jaeger, Grafana, Datadog, Honeycomb, New Relic. Console
+  exporter by default (zero infra); set an OTLP endpoint for production.
 - **Structured logging** (human or JSON) and a per-run **usage/cost report** broken
   down per model and per agent.
 
@@ -170,6 +178,7 @@ pip install agentforge-oss                       # core (works offline, zero con
 pip install "agentforge-oss[anthropic]"          # + Claude provider
 pip install "agentforge-oss[openai]"             # + OpenAI / GPT provider
 pip install "agentforge-oss[anthropic,openai]"   # both real providers
+pip install "agentforge-oss[bedrock]"            # + Amazon Bedrock provider (boto3)
 pip install "agentforge-oss[all,dev]"            # everything + test/lint tooling
 ```
 
@@ -183,6 +192,15 @@ pip install "agentforge-oss[all,dev]"            # everything + test/lint toolin
 > to route to GPT. Both keys can be set at once; Forge prefers Anthropic by default
 > (configurable).
 
+> **Ollama support is built in** — no extra install needed (it uses `httpx`, already a
+> core dependency, so there is no `[ollama]` extra). Just run Ollama locally and Forge
+> auto-detects it (or set `OLLAMA_BASE_URL` for a custom/remote server):
+
+```bash
+ollama serve
+ollama pull llama3.1:8b
+```
+
 ---
 
 ## Quickstart (CLI)
@@ -295,10 +313,10 @@ governance (RBAC + budgets + audit verification).
               │                          │
               │      ┌───────────────────┴───────────────┐
               ▼      ▼                                    ▼
-       ┌────────────────┐                        ┌───────────────────────────┐
-       │  Model Router  │  picks model by        │  Model Providers          │
-       │ cost / quality │  strategy + budget ───▶│ Anthropic · OpenAI · Echo │
-       └───────┬────────┘                        └───────────────────────────┘
+       ┌────────────────┐                        ┌────────────────────────────────────┐
+       │  Model Router  │  picks model by        │  Model Providers                   │
+       │ cost / quality │  strategy + budget ───▶│ Anthropic · OpenAI · Ollama · Bedrock · Echo │
+       └───────┬────────┘                        └────────────────────────────────────┘
                │ pricing
                ▼
        ┌────────────────┐   cross-cutting, on every step:
@@ -310,9 +328,9 @@ Every layer is swappable:
 
 | Layer | Default | Swap in… |
 |---|---|---|
-| Provider | Echo (offline), Anthropic, OpenAI | Any `ModelProvider` (local, Bedrock, Vertex, …) |
+| Provider | Echo (offline), Anthropic, OpenAI, Ollama, Bedrock | Any `ModelProvider` (Vertex, …) |
 | Routing | `balanced` strategy | Your own strategy / `fixed` model |
-| Memory | In-memory vector store | Any `Memory` backend (pgvector, Pinecone, …) |
+| Memory | InMemoryVectorStore (default), SQLiteMemoryStore | Any `Memory` backend (pgvector, Redis, …) |
 | Tools | `calculator`, `utc_now` | Any `@tool` function |
 | Audit | Hash-chained JSONL | Forward events to your SIEM via the event bus |
 
@@ -403,7 +421,7 @@ ruff check .      # lint
 mypy forge        # strict type-check
 ```
 
-The entire 47-test suite runs offline against the deterministic provider — fast,
+The entire 83-test suite runs offline against the deterministic provider — fast,
 hermetic, and free. CI runs the same checks (ruff, ruff format, mypy strict, pytest)
 on Python 3.11, 3.12, and 3.13.
 
@@ -416,7 +434,7 @@ the core typed (`mypy --strict`) and tested.
 
 ### Good first contributions
 
-- **New model provider** (Ollama, Bedrock, Vertex) — implement one method.
+- **New model provider** (Vertex, Gemini, Cohere) — implement one method.
 - **New built-in tool** (web search, file read, database query).
 - **Durable memory backend** (SQLite-VSS, pgvector, Redis).
 - **Routing strategy** (a custom cost/quality tradeoff).

agentforge_oss-0.4.0/examples/otel_tracing.py

@@ -0,0 +1,34 @@
+"""
+OpenTelemetry tracing example — exports spans to the console.
+
+    pip install "agentforge-oss[otel]"
+    python examples/otel_tracing.py
+
+To send traces to Jaeger, Grafana Tempo, or any OTel collector instead:
+
+    FORGE_OTEL_ENDPOINT=http://localhost:4317 python examples/otel_tracing.py
+"""
+
+import asyncio
+import os
+
+from forge import ForgeConfig, Orchestrator
+
+
+async def main() -> None:
+    config = ForgeConfig(
+        otel_enabled=True,
+        otel_service_name="forge-demo",
+        otel_endpoint=os.environ.get("FORGE_OTEL_ENDPOINT"),
+    )
+    async with Orchestrator(config) as forge:
+        result = await forge.run(
+            "Calculate 15% of 4200 and summarize the result",
+            mode="single",
+        )
+        print(result.output)
+        print(result.usage.format_table())
+        print("\nSpans exported above (ConsoleSpanExporter).")
+
+
+asyncio.run(main())

agentforge_oss-0.4.0/examples/rag_persistent.py

@@ -0,0 +1,31 @@
+"""
+Persistent RAG example — memories survive process restarts.
+
+Run this script twice. The second run finds the facts stored by the first.
+
+    pip install "agentforge-oss[sqlite]"
+    python examples/rag_persistent.py
+"""
+
+import asyncio
+
+from forge.memory.sqlite import SQLiteMemoryStore
+
+
+async def main() -> None:
+    async with SQLiteMemoryStore("demo_memory.db") as store:
+        count_before = len(await store.search("forge", k=100))
+        if count_before == 0:
+            print("First run — storing facts...")
+            await store.add("Forge routes cheap tasks to small models to save cost.")
+            await store.add("The audit log is hash-chained and tamper-evident.")
+            await store.add("Ollama runs local LLMs with zero API cost.")
+            print("Stored 3 facts. Run again to retrieve them.")
+        else:
+            print(f"Second run — found {count_before} stored facts.")
+            hits = await store.search("how does Forge reduce cost?", k=2)
+            for hit in hits:
+                print(f"  [{hit.score:.3f}] {hit.text}")
+
+
+asyncio.run(main())

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/forge/__init__.py

@@ -51,7 +51,13 @@ from forge.exceptions import (
     ToolTimeoutError,
     ToolValidationError,
 )
-from forge.memory import ConversationMemory, InMemoryVectorStore, Memory, MemoryItem
+from forge.memory import (
+    ConversationMemory,
+    InMemoryVectorStore,
+    Memory,
+    MemoryItem,
+    SQLiteMemoryStore,
+)
 from forge.models import (
     Complexity,
     ModelInfo,
@@ -61,7 +67,13 @@ from forge.models import (
     ModelTier,
     RoutingDecision,
 )
-from forge.models.providers import AnthropicProvider, EchoProvider, OpenAIProvider
+from forge.models.providers import (
+    AnthropicProvider,
+    BedrockProvider,
+    EchoProvider,
+    OllamaProvider,
+    OpenAIProvider,
+)
 from forge.observability import (
     Event,
     EventBus,
@@ -93,23 +105,19 @@ from forge.types import (
 
 __all__ = [
     "__version__",
-    # Orchestration
     "Orchestrator",
     "RunResult",
     "RunContext",
-    # Agents
     "Agent",
     "Supervisor",
     "BaseAgent",
     "AgentResult",
-    # Config
     "ForgeConfig",
     "RoutingConfig",
     "BudgetConfig",
     "SecurityConfig",
     "ComplianceConfig",
     "ObservabilityConfig",
-    # Types
     "Role",
     "Message",
     "ToolCall",
@@ -118,7 +126,6 @@ __all__ = [
     "Usage",
     "ModelResponse",
     "FinishReason",
-    # Models
     "ModelProvider",
     "ModelRegistry",
     "ModelInfo",
@@ -129,7 +136,8 @@ __all__ = [
     "EchoProvider",
     "AnthropicProvider",
     "OpenAIProvider",
-    # Tools
+    "OllamaProvider",
+    "BedrockProvider",
     "Tool",
     "tool",
     "ToolRegistry",
@@ -137,12 +145,11 @@ __all__ = [
     "calculator",
     "http_get",
     "utc_now",
-    # Memory
     "Memory",
     "MemoryItem",
     "ConversationMemory",
     "InMemoryVectorStore",
-    # Observability
+    "SQLiteMemoryStore",
     "Event",
     "EventBus",
     "EventType",
@@ -150,17 +157,14 @@ __all__ = [
     "UsageReport",
     "get_logger",
     "configure_logging",
-    # Security
     "InputSanitizer",
     "ToolSandbox",
     "AccessController",
     "Principal",
     "Permission",
-    # Compliance
     "AuditLogger",
     "AuditEntry",
     "PIIRedactor",
-    # Exceptions
     "ForgeError",
     "ConfigurationError",
     "ProviderError",

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/forge/_version.py

@@ -5,4 +5,4 @@ Hatchling reads ``__version__`` from this file at build time (see
 it for runtime access via ``forge.__version__``.
 """
 
-__version__ = "0.2.0"
+__version__ = "0.4.0"

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/forge/agents/agent.py

@@ -52,7 +52,6 @@ class Agent(BaseAgent):
                 agent=self.name, output=response.content, usage=self._own_usage(), steps=steps
             )
 
-        # Loop exhausted without a final answer.
         ctx.events.emit(
             EventType.AGENT_FAILED, run_id=ctx.run_id, agent=self.name, reason="max_steps_exceeded"
         )

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/forge/agents/base.py

@@ -105,7 +105,11 @@ class BaseAgent(abc.ABC):
         max_tokens = min(self.max_tokens, info.max_output_tokens)
 
         ctx.events.emit(
-            EventType.MODEL_CALL_STARTED, run_id=ctx.run_id, agent=self.name, model=decision.model
+            EventType.MODEL_CALL_STARTED,
+            run_id=ctx.run_id,
+            agent=self.name,
+            model=decision.model,
+            provider=decision.provider,
         )
         try:
             if ctx.stream:

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/forge/agents/supervisor.py

@@ -67,7 +67,6 @@ class Supervisor(BaseAgent):
         )
         ctx.audit.record("agent.start", actor=self.name, run_id=ctx.run_id, resource=goal[:200])
 
-        # 1) Plan.
         plan = await self._plan(goal)
         ctx.events.emit(EventType.PLAN_CREATED, run_id=ctx.run_id, agent=self.name, subtasks=plan)
         ctx.audit.record(
@@ -78,10 +77,8 @@ class Supervisor(BaseAgent):
             subtasks=plan,
         )
 
-        # 2) Delegate subtasks to workers, running each batch concurrently.
         children = await self._run_workers(plan)
 
-        # 3) Synthesise a final answer and assemble the report.
         summary = await self._synthesize(goal, children)
         report = self._format_report(goal, plan, children, summary)
 
@@ -102,9 +99,6 @@ class Supervisor(BaseAgent):
             children=children,
         )
 
-    # ------------------------------------------------------------------ #
-    # Parallel worker execution
-    # ------------------------------------------------------------------ #
     async def _run_workers(self, plan: list[str]) -> list[AgentResult]:
         """Run the planned subtasks as workers, concurrently and in bounded batches.
 
@@ -194,9 +188,6 @@ class Supervisor(BaseAgent):
             success=False,
         )
 
-    # ------------------------------------------------------------------ #
-    # Planning & synthesis
-    # ------------------------------------------------------------------ #
     async def _plan(self, goal: str) -> list[str]:
         prompt = (
             f"{PLAN_MARKER} You are a planning supervisor. Decompose the goal into a minimal, "
@@ -225,9 +216,6 @@ class Supervisor(BaseAgent):
         response = await self._invoke_model(self._with_system(prompt), complexity=Complexity.MEDIUM)
         return response.content
 
-    # ------------------------------------------------------------------ #
-    # Helpers
-    # ------------------------------------------------------------------ #
     def _with_system(self, user_content: str) -> list[Message]:
         messages: list[Message] = []
         if self.system_prompt:

{agentforge_oss-0.2.0 → agentforge_oss-0.4.0}/forge/compliance/audit.py

@@ -118,7 +118,13 @@ class AuditLogger:
                 line = line.strip()
                 if not line:
                     continue
-                entry = AuditEntry.model_validate_json(line)
+                try:
+                    entry = AuditEntry.model_validate_json(line)
+                except ValueError:
+                    # A line that no longer parses is itself tampering: fail closed
+                    # (return False) rather than raising out of a bool-returning check.
+                    _log.error("audit chain broken at line %d (unparseable record)", line_no)
+                    return False
                 if entry.previous_hash != prev:
                     _log.error("audit chain broken at line %d (prev_hash mismatch)", line_no)
                     return False
@@ -128,7 +134,6 @@ class AuditLogger:
                 prev = entry.hash
         return True
 
-    # ------------------------------------------------------------------ #
     @staticmethod
     def _compute_hash(entry: AuditEntry) -> str:
         digest = hashlib.sha256()