haru-cli 0.1.1__tar.gz → 0.1.2__tar.gz

{haru_cli-0.1.1 → haru_cli-0.1.2}/CHANGELOG.md

@@ -9,6 +9,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- Sampling controls (`temperature`, `top_p`, `top_k`, `seed`) across three
+  surfaces: model entries, per-agent `sampling:` blocks, and CLI/REPL
+  overrides (`--temperature/--top-p/--top-k/--seed`, `/sampling`).
+  Precedence CLI/REPL → agent → model, merged per-field; unset fields are
+  omitted from requests. `top_k`/`seed` travel via Converse
+  `additionalModelRequestFields`.
+
+### Fixed
+
+- `temperature` is no longer required (or sent by default) on model entries —
+  Claude 5-series and Opus 4.7+ reject non-default sampling values with an
+  HTTP 400, which the previous always-sent temperature would have triggered.
+
+### Added
+
 - `haru config init`: interactive starter configuration under
   `~/.config/haru` (SSO settings prompted; guardrails explicitly disabled
   when no id is provided). `haru config show` prints a secret-free summary.

{haru_cli-0.1.1 → haru_cli-0.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: haru-cli
-Version: 0.1.1
+Version: 0.1.2
 Summary: Secure, scriptable CLI for Amazon Bedrock Claude models via the AWS Strands Agents SDK.
 Author: haru-cli maintainers
 License-Expression: Apache-2.0
@@ -96,12 +96,43 @@ caching tokens in the botocore-compatible schema under `~/.aws/sso/cache`
 tokens are refreshed automatically; when a fresh login is needed, commands say
 so plainly.
 
+## Sampling controls
+
+haru-cli exposes `temperature`, `top_p`, `top_k`, and `seed` on **three
+configuration surfaces** — a capability Kiro CLI does not offer on any of its
+configuration surfaces (its v2/v3 agent formats and `cli.json` define no
+sampling fields):
+
+| Surface                  | Where                                          | Example                                   |
+| ------------------------ | ---------------------------------------------- | ----------------------------------------- |
+| Model catalogue          | `models.yaml`, per model entry                 | `temperature: 0.2`, `top_k: 50`           |
+| Agent profile            | `agents.yaml`, per agent `sampling:` block     | `sampling: {temperature: 0.1}`            |
+| CLI / REPL               | `--temperature/--top-p/--top-k/--seed` flags on `run` and `chat`; `/sampling` in the REPL | `haru run "..." --top-k 20` |
+
+Precedence: **CLI/REPL override → agent `sampling:` block → model entry**,
+merged per-field. Unset fields are omitted from requests entirely, keeping
+the provider's defaults. In the chat REPL, `/sampling` shows the model
+default, session override, and effective value for each field;
+`/sampling temperature=0.2 top_k=50` applies overrides and `/sampling reset`
+clears them.
+
+Model compatibility (AWS-side behaviour, not a haru limitation):
+
+- **Claude 5-series and Opus 4.7+** (Sonnet 5, Opus 5, Opus 4.8) reject
+  non-default `temperature`/`top_p`/`top_k` with an HTTP 400 — leave them
+  unset on those entries (the shipped config does).
+- **Claude Haiku 4.5, Sonnet 4.5, and older Claude models**, plus
+  non-Anthropic Bedrock models, accept them.
+- `seed` is forwarded via Converse `additionalModelRequestFields` for Bedrock
+  models that support it; **no Claude model accepts a seed today**, so do not
+  expect determinism from Claude — the field exists for models that honour it.
+
 ## Configuration
 
 Declarative YAML under `config/` — models, agents, orchestration
-(supervisor/swarm/graph), MCP servers, guardrails, sessions, logging, and
-OpenTelemetry. No secrets in YAML: values reference environment variables with
-`${env:VAR}`. See [docs/configuration.md](docs/configuration.md).
+(supervisor/swarm/graph), MCP servers, guardrails, sessions, sampling,
+logging, and OpenTelemetry. No secrets in YAML: values reference environment
+variables with `${env:VAR}`. See [docs/configuration.md](docs/configuration.md).
 
 ## Development
 

{haru_cli-0.1.1 → haru_cli-0.1.2}/README.md

@@ -68,12 +68,43 @@ caching tokens in the botocore-compatible schema under `~/.aws/sso/cache`
 tokens are refreshed automatically; when a fresh login is needed, commands say
 so plainly.
 
+## Sampling controls
+
+haru-cli exposes `temperature`, `top_p`, `top_k`, and `seed` on **three
+configuration surfaces** — a capability Kiro CLI does not offer on any of its
+configuration surfaces (its v2/v3 agent formats and `cli.json` define no
+sampling fields):
+
+| Surface                  | Where                                          | Example                                   |
+| ------------------------ | ---------------------------------------------- | ----------------------------------------- |
+| Model catalogue          | `models.yaml`, per model entry                 | `temperature: 0.2`, `top_k: 50`           |
+| Agent profile            | `agents.yaml`, per agent `sampling:` block     | `sampling: {temperature: 0.1}`            |
+| CLI / REPL               | `--temperature/--top-p/--top-k/--seed` flags on `run` and `chat`; `/sampling` in the REPL | `haru run "..." --top-k 20` |
+
+Precedence: **CLI/REPL override → agent `sampling:` block → model entry**,
+merged per-field. Unset fields are omitted from requests entirely, keeping
+the provider's defaults. In the chat REPL, `/sampling` shows the model
+default, session override, and effective value for each field;
+`/sampling temperature=0.2 top_k=50` applies overrides and `/sampling reset`
+clears them.
+
+Model compatibility (AWS-side behaviour, not a haru limitation):
+
+- **Claude 5-series and Opus 4.7+** (Sonnet 5, Opus 5, Opus 4.8) reject
+  non-default `temperature`/`top_p`/`top_k` with an HTTP 400 — leave them
+  unset on those entries (the shipped config does).
+- **Claude Haiku 4.5, Sonnet 4.5, and older Claude models**, plus
+  non-Anthropic Bedrock models, accept them.
+- `seed` is forwarded via Converse `additionalModelRequestFields` for Bedrock
+  models that support it; **no Claude model accepts a seed today**, so do not
+  expect determinism from Claude — the field exists for models that honour it.
+
 ## Configuration
 
 Declarative YAML under `config/` — models, agents, orchestration
-(supervisor/swarm/graph), MCP servers, guardrails, sessions, logging, and
-OpenTelemetry. No secrets in YAML: values reference environment variables with
-`${env:VAR}`. See [docs/configuration.md](docs/configuration.md).
+(supervisor/swarm/graph), MCP servers, guardrails, sessions, sampling,
+logging, and OpenTelemetry. No secrets in YAML: values reference environment
+variables with `${env:VAR}`. See [docs/configuration.md](docs/configuration.md).
 
 ## Development
 

{haru_cli-0.1.1 → haru_cli-0.1.2}/config/agents.yaml

@@ -28,3 +28,10 @@ orchestration:
     edges:
       - from: research
         to: write
+# Per-agent sampling overrides beat the model entry's values and are in turn
+# beaten by CLI/REPL overrides. Example:
+#   researcher:
+#     model: haiku-4-5
+#     sampling:
+#       temperature: 0.1
+#       top_k: 20

haru_cli-0.1.2/config/models.yaml

@@ -0,0 +1,31 @@
+# Sampling fields (temperature, top_p, top_k, seed) are optional; unset
+# fields are omitted from requests, keeping the provider's defaults.
+# Claude 5-series and Opus 4.7+ models reject non-default temperature/top_p/
+# top_k with an HTTP 400, so leave them unset on those entries. seed is
+# passed through to models that support it (no Claude model does today).
+default_model: sonnet-5
+models:
+  opus-5:
+    model_id: us.anthropic.claude-opus-5
+    region: us-east-1
+    max_tokens: 8192
+    streaming: true
+  sonnet-5:
+    model_id: us.anthropic.claude-sonnet-5
+    region: us-east-1
+    max_tokens: 8192
+    streaming: true
+  opus-4-8:
+    model_id: us.anthropic.claude-opus-4-8
+    region: us-east-1
+    max_tokens: 8192
+    streaming: true
+  # Haiku 4.5 accepts sampling parameters:
+  haiku-4-5:
+    model_id: us.anthropic.claude-haiku-4-5
+    region: us-east-1
+    max_tokens: 4096
+    temperature: 0.2
+    top_p: 0.9
+    top_k: 50
+    streaming: true

{haru_cli-0.1.1 → haru_cli-0.1.2}/docs/configuration.md

@@ -38,6 +38,18 @@ Bare Bedrock model ids receive the `us.` geographic inference-profile prefix
 by default. Explicit prefixes (`eu.`, `au.`, `global.`, ...) and full ARNs
 pass through as configured — writing one into YAML is the approval surface.
 
+## Sampling
+
+`temperature`, `top_p`, `top_k`, and `seed` are optional on every model entry
+and may also be set per agent (a `sampling:` block) or per invocation
+(`--temperature/--top-p/--top-k/--seed`, or `/sampling` inside `haru chat`).
+Precedence is CLI/REPL → agent → model, merged per-field; unset fields are
+omitted from requests. `temperature`/`top_p` map to the Converse inference
+configuration; `top_k` and `seed` travel via
+`additionalModelRequestFields`. Claude 5-series and Opus 4.7+ models reject
+non-default sampling values (HTTP 400), and no Claude model honours `seed` —
+use these fields with models that accept them (for example Haiku 4.5).
+
 ## Sessions
 
 Conversation persistence is configured in the base file:

{haru_cli-0.1.1 → haru_cli-0.1.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "haru-cli"
-version = "0.1.1"
+version = "0.1.2"
 description = "Secure, scriptable CLI for Amazon Bedrock Claude models via the AWS Strands Agents SDK."
 readme = "README.md"
 license = "Apache-2.0"

{haru_cli-0.1.1 → haru_cli-0.1.2}/src/haru/agents/factory.py

@@ -15,9 +15,9 @@ from strands.models import BedrockModel
 from strands.session.session_manager import SessionManager
 from strands.tools.mcp import MCPClient
 
-from haru.config.schema import HaruConfig
+from haru.config.schema import HaruConfig, SamplingConfig
 from haru.errors import ConfigError
-from haru.models.bedrock import build_model, get_model_config
+from haru.models.bedrock import build_model, get_model_config, merge_sampling
 from haru.steering.prompts import DEFAULT_PROMPTS_ROOT, load_prompts, resolve_prompt
 from haru.tools.mcp import collect_tools
 
@@ -31,16 +31,19 @@ def build_agent(  # noqa: PLR0913 - keyword-only wiring points, all optional
     mcp_clients: Mapping[str, MCPClient] | None = None,
     session_manager: SessionManager | None = None,
     model_name: str | None = None,
+    sampling: SamplingConfig | None = None,
 ) -> Agent:
     """Build the named agent from configuration (default model when unnamed).
 
     ``model_name`` overrides the model for the default (unnamed) agent only;
-    named agents pin their own model. Raises ConfigError for unknown agents,
-    prompt references, tools, or an invalid override.
+    named agents pin their own model. ``sampling`` overrides sampling fields
+    (precedence: this override, then the agent's ``sampling`` block, then the
+    model entry). Raises ConfigError for unknown agents, prompt references,
+    tools, or an invalid override.
     """
     if agent_name is None:
         model_cfg = get_model_config(config, model_name)
-        model = build_model(model_cfg, session, guardrails=config.guardrails)
+        model = build_model(model_cfg, session, guardrails=config.guardrails, sampling=sampling)
         return Agent(model=model, session_manager=session_manager)
     if model_name is not None:
         raise ConfigError(
@@ -48,7 +51,12 @@ def build_agent(  # noqa: PLR0913 - keyword-only wiring points, all optional
             " to the default agent only"
         )
     model, system_prompt, tools = resolve_agent_parts(
-        config, agent_name, session, prompts_root=prompts_root, mcp_clients=mcp_clients
+        config,
+        agent_name,
+        session,
+        prompts_root=prompts_root,
+        mcp_clients=mcp_clients,
+        sampling=sampling,
     )
     return Agent(
         model=model,
@@ -59,13 +67,14 @@ def build_agent(  # noqa: PLR0913 - keyword-only wiring points, all optional
     )
 
 
-def resolve_agent_parts(
+def resolve_agent_parts(  # noqa: PLR0913 - keyword-only wiring points, all optional
     config: HaruConfig,
     agent_name: str,
     session: boto3.Session,
     *,
     prompts_root: Path | None = None,
     mcp_clients: Mapping[str, MCPClient] | None = None,
+    sampling: SamplingConfig | None = None,
 ) -> tuple[BedrockModel, str | None, list[Any]]:
     """Resolve an agent's model, steering prompt, and tool list from config."""
     if config.agents is None or agent_name not in config.agents.agents:
@@ -73,7 +82,10 @@ def resolve_agent_parts(
         raise ConfigError(f"Unknown agent {agent_name!r}; configured agents: {available}")
     agent_cfg = config.agents.agents[agent_name]
     model = build_model(
-        get_model_config(config, agent_cfg.model), session, guardrails=config.guardrails
+        get_model_config(config, agent_cfg.model),
+        session,
+        guardrails=config.guardrails,
+        sampling=merge_sampling(agent_cfg.sampling, sampling),
     )
 
     system_prompt: str | None = None

{haru_cli-0.1.1 → haru_cli-0.1.2}/src/haru/commands/chat.py

@@ -7,31 +7,37 @@ models and agents mid-session.
 
 from collections.abc import Callable
 from pathlib import Path
-from typing import Any
+from typing import Any, cast
 
 import click
+from pydantic import ValidationError
 from rich.console import Console
 
 from haru.agents.factory import build_agent
 from haru.auth.session import build_boto3_session
 from haru.commands.streaming import collect_response, surface_guardrail
 from haru.config import load_config, resolve_config_path
-from haru.config.schema import HaruConfig
-from haru.errors import HaruError
+from haru.config.schema import HaruConfig, SamplingConfig
+from haru.errors import ConfigError, HaruError
+from haru.models.bedrock import effective_sampling, get_model_config, sampling_overrides
 from haru.observability.telemetry import configure_telemetry
 from haru.sessions.manager import build_session_manager
 from haru.tools.mcp import started_mcp_clients
 
 _EXIT_WORDS = frozenset({"exit", "quit"})
+_SAMPLING_FIELDS = ("temperature", "top_p", "top_k", "seed")
 
 _HELP_TEXT = """\
 Commands:
-  /help            Show this help
-  /model           List configured models
-  /model <name>    Switch the default agent to <name> (resets the conversation)
-  /agent           List configured agents
-  /agent <name>    Switch to agent <name> (resets the conversation)
-  exit | quit      Leave the chat (also /exit, /quit, Ctrl-D)"""
+  /help                    Show this help
+  /model                   List configured models
+  /model <name>            Switch the default agent to <name> (resets the conversation)
+  /agent                   List configured agents
+  /agent <name>            Switch to agent <name> (resets the conversation)
+  /sampling                Show effective sampling for the active target
+  /sampling k=v [k=v ...]  Override temperature/top_p/top_k/seed (resets the conversation)
+  /sampling reset          Clear sampling overrides (resets the conversation)
+  exit | quit              Leave the chat (also /exit, /quit, Ctrl-D)"""
 
 
 def run_chat(  # noqa: PLR0913 - keyword-only wiring points, all optional
@@ -42,6 +48,7 @@ def run_chat(  # noqa: PLR0913 - keyword-only wiring points, all optional
     read_input: Callable[[str], str] = input,
     prompts_root: Path | None = None,
     session_id: str | None = None,
+    sampling: SamplingConfig | None = None,
 ) -> None:
     """Run the interactive chat loop until the user exits."""
     session = build_boto3_session(config.auth)
@@ -51,7 +58,11 @@ def run_chat(  # noqa: PLR0913 - keyword-only wiring points, all optional
         else None
     )
     with started_mcp_clients(config.mcp) as mcp_clients:
-        state: dict[str, Any] = {"agent_name": agent_name, "model_name": None}
+        state: dict[str, Any] = {
+            "agent_name": agent_name,
+            "model_name": None,
+            "sampling": sampling,
+        }
 
         def make_agent(name: str | None, model: str | None) -> Any:
             return build_agent(
@@ -62,6 +73,7 @@ def run_chat(  # noqa: PLR0913 - keyword-only wiring points, all optional
                 mcp_clients=mcp_clients,
                 session_manager=session_manager,
                 model_name=model,
+                sampling=state["sampling"],
             )
 
         state["agent"] = make_agent(agent_name, None)
@@ -121,11 +133,92 @@ def _handle_slash(
             _print_agents(config, console, state)
         else:
             _switch(state, console, make_agent, agent_name=argument, model_name=None)
+    elif name == "sampling":
+        if argument is None:
+            _print_sampling(config, console, state)
+        else:
+            _apply_sampling(argument, config, console, state, make_agent)
     else:
         console.print(f"Unknown command /{command}; type /help for commands.")
     return True
 
 
+def _parse_sampling(argument: str) -> SamplingConfig:
+    """Parse ``key=value`` sampling tokens into a SamplingConfig."""
+    values: dict[str, str] = {}
+    for token in argument.split():
+        key, separator, raw_value = token.partition("=")
+        if not separator or key not in _SAMPLING_FIELDS or not raw_value:
+            fields = ", ".join(_SAMPLING_FIELDS)
+            raise ConfigError(f"Bad sampling token {token!r}; use key=value with keys: {fields}")
+        values[key] = raw_value
+    try:
+        return SamplingConfig.model_validate(values)
+    except ValidationError as exc:
+        raise ConfigError(f"Invalid sampling values: {exc}") from exc
+
+
+def _apply_sampling(
+    argument: str,
+    config: HaruConfig,
+    console: Console,
+    state: dict[str, Any],
+    make_agent: Callable[[str | None, str | None], Any],
+) -> None:
+    """Apply or clear sampling overrides, keeping the old state on failure."""
+    if argument.lower() == "reset":
+        override: SamplingConfig | None = None
+    else:
+        try:
+            override = _parse_sampling(argument)
+        except ConfigError as exc:
+            console.print(f"[red]{exc}[/]")
+            return
+    previous = state["sampling"]
+    state["sampling"] = override
+    try:
+        agent = make_agent(state["agent_name"], state["model_name"])
+    except HaruError as exc:
+        state["sampling"] = previous
+        console.print(f"[red]{exc}[/]")
+        return
+    state["agent"] = agent
+    if override is None:
+        console.print("Sampling overrides cleared (conversation reset).")
+    else:
+        console.print("Sampling overrides applied (conversation reset).")
+    _print_sampling(config, console, state)
+
+
+def _print_sampling(config: HaruConfig, console: Console, state: dict[str, Any]) -> None:
+    """Show model defaults, session overrides, and effective sampling values."""
+    agent_name = state["agent_name"]
+    model_key = state["model_name"]
+    if agent_name is not None and config.agents is not None:
+        agent_cfg = config.agents.agents.get(agent_name)
+        model_key = agent_cfg.model if agent_cfg is not None else model_key
+    try:
+        model_cfg = get_model_config(config, model_key)
+    except HaruError as exc:
+        console.print(f"[red]{exc}[/]")
+        return
+    override = cast(SamplingConfig | None, state["sampling"])
+    effective = effective_sampling(model_cfg, override)
+    console.print(f"Sampling for model {model_cfg.model_id!r} (unset = provider default):")
+    for field in _SAMPLING_FIELDS:
+        model_value = getattr(model_cfg, field)
+        override_value = getattr(override, field) if override is not None else None
+        effective_value = getattr(effective, field)
+        console.print(
+            f"  {field:<12} model={_fmt(model_value):<8}"
+            f" override={_fmt(override_value):<8} effective={_fmt(effective_value)}"
+        )
+
+
+def _fmt(value: Any) -> str:
+    return "-" if value is None else str(value)
+
+
 def _switch(
     state: dict[str, Any],
     console: Console,
@@ -190,7 +283,19 @@ def _print_agents(config: HaruConfig, console: Console, state: dict[str, Any]) -
     default=None,
     help="Persist and restore this conversation under the given session id.",
 )
-def chat(config_path: Path | None, agent_name: str | None, session_id: str | None) -> None:
+@click.option("--temperature", type=float, default=None, help="Sampling temperature (0-1).")
+@click.option("--top-p", "top_p", type=float, default=None, help="Nucleus sampling (0-1).")
+@click.option("--top-k", "top_k", type=int, default=None, help="Top-k sampling (>=1).")
+@click.option("--seed", type=int, default=None, help="Seed (models that support it only).")
+def chat(  # noqa: PLR0913, PLR0917 - one option per flag; Click passes positionally
+    config_path: Path | None,
+    agent_name: str | None,
+    session_id: str | None,
+    temperature: float | None,
+    top_p: float | None,
+    top_k: int | None,
+    seed: int | None,
+) -> None:
     """Chat interactively with a Bedrock agent (streaming)."""
     console = Console()
     try:
@@ -203,6 +308,9 @@ def chat(config_path: Path | None, agent_name: str | None, session_id: str | Non
             console=console,
             prompts_root=resolved.parent / "prompts",
             session_id=session_id,
+            sampling=sampling_overrides(
+                temperature=temperature, top_p=top_p, top_k=top_k, seed=seed
+            ),
         )
     except HaruError as exc:
         raise click.ClickException(str(exc)) from exc

{haru_cli-0.1.1 → haru_cli-0.1.2}/src/haru/commands/config.py

@@ -6,8 +6,21 @@ from pathlib import Path
 import click
 
 from haru.config import load_config, resolve_config_path, user_config_path
+from haru.config.schema import ModelConfig
 from haru.errors import HaruError
 
+
+def _sampling_summary(entry: ModelConfig) -> str:
+    """Compact ``[temp=.. top_k=..]`` summary of a model's set sampling fields."""
+    labels = {"temperature": "temp", "top_p": "top_p", "top_k": "top_k", "seed": "seed"}
+    parts = [
+        f"{label}={value}"
+        for field, label in labels.items()
+        if (value := getattr(entry, field)) is not None
+    ]
+    return f" [{' '.join(parts)}]" if parts else ""
+
+
 _TEMPLATE_FILES = ("haru.yaml", "models.yaml", "agents.yaml", "mcp.yaml", "logging.yaml")
 _PROMPT_FILES = ("supervisor.md", "researcher.md", "writer.md")
 
@@ -128,8 +141,8 @@ def show(config_path: Path | None) -> None:
     if loaded.models is not None:
         default = loaded.models.default_model
         names = ", ".join(
-            f"{name}{' (default)' if name == default else ''}"
-            for name in sorted(loaded.models.models)
+            f"{name}{' (default)' if name == default else ''}{_sampling_summary(entry)}"
+            for name, entry in sorted(loaded.models.models.items())
         )
         click.echo(f"Models:       {names}")
     if loaded.agents is not None:

{haru_cli-0.1.1 → haru_cli-0.1.2}/src/haru/commands/run.py

@@ -9,19 +9,21 @@ from haru.agents.factory import build_agent
 from haru.auth.session import build_boto3_session
 from haru.commands.streaming import collect_response, surface_guardrail
 from haru.config import load_config, resolve_config_path
-from haru.config.schema import HaruConfig
+from haru.config.schema import HaruConfig, SamplingConfig
 from haru.errors import HaruError
+from haru.models.bedrock import sampling_overrides
 from haru.observability.telemetry import configure_telemetry
 from haru.tools.mcp import started_mcp_clients
 
 
-def run_prompt(
+def run_prompt(  # noqa: PLR0913 - keyword-only wiring points, all optional
     config: HaruConfig,
     prompt: str,
     agent_name: str | None = None,
     *,
     console: Console | None = None,
     prompts_root: Path | None = None,
+    sampling: SamplingConfig | None = None,
 ) -> str:
     """Execute ``prompt`` once and return the full response text.
 
@@ -30,7 +32,12 @@ def run_prompt(
     session = build_boto3_session(config.auth)
     with started_mcp_clients(config.mcp) as mcp_clients:
         agent = build_agent(
-            config, agent_name, session, prompts_root=prompts_root, mcp_clients=mcp_clients
+            config,
+            agent_name,
+            session,
+            prompts_root=prompts_root,
+            mcp_clients=mcp_clients,
+            sampling=sampling,
         )
         chunks: list[str] = []
         result = collect_response(agent, prompt, chunks.append)
@@ -48,13 +55,33 @@ def run_prompt(
     help="Path to haru.yaml (default: config/haru.yaml).",
 )
 @click.option("--agent", "agent_name", default=None, help="Agent name from configuration.")
-def run(prompt: str, config_path: Path | None, agent_name: str | None) -> None:
+@click.option("--temperature", type=float, default=None, help="Sampling temperature (0-1).")
+@click.option("--top-p", "top_p", type=float, default=None, help="Nucleus sampling (0-1).")
+@click.option("--top-k", "top_k", type=int, default=None, help="Top-k sampling (>=1).")
+@click.option("--seed", type=int, default=None, help="Seed (models that support it only).")
+def run(  # noqa: PLR0913, PLR0917 - one option per flag; Click passes positionally
+    prompt: str,
+    config_path: Path | None,
+    agent_name: str | None,
+    temperature: float | None,
+    top_p: float | None,
+    top_k: int | None,
+    seed: int | None,
+) -> None:
     """Run a single prompt and print the answer."""
     try:
         resolved = resolve_config_path(config_path)
         config = load_config(resolved)
         configure_telemetry(config.observability)
-        answer = run_prompt(config, prompt, agent_name, prompts_root=resolved.parent / "prompts")
+        answer = run_prompt(
+            config,
+            prompt,
+            agent_name,
+            prompts_root=resolved.parent / "prompts",
+            sampling=sampling_overrides(
+                temperature=temperature, top_p=top_p, top_k=top_k, seed=seed
+            ),
+        )
     except HaruError as exc:
         raise click.ClickException(str(exc)) from exc
     click.echo(answer)

{haru_cli-0.1.1 → haru_cli-0.1.2}/src/haru/config/schema.py

@@ -53,13 +53,31 @@ class IncludesConfig(_FrozenModel):
     logging: str | None = None
 
 
+class SamplingConfig(_FrozenModel):
+    """Optional sampling parameters.
+
+    Unset fields are omitted from requests entirely, keeping the provider's
+    defaults — required by Claude 5-series models, which reject non-default
+    sampling values. ``seed`` is passed through for Bedrock models that
+    support it (no Claude model does today).
+    """
+
+    temperature: float | None = Field(default=None, ge=0.0, le=1.0)
+    top_p: float | None = Field(default=None, ge=0.0, le=1.0)
+    top_k: int | None = Field(default=None, ge=1)
+    seed: int | None = None
+
+
 class ModelConfig(_FrozenModel):
     """A single Bedrock model entry."""
 
     model_id: str
     region: str
     max_tokens: int = Field(ge=1)
-    temperature: float = Field(ge=0.0, le=1.0)
+    temperature: float | None = Field(default=None, ge=0.0, le=1.0)
+    top_p: float | None = Field(default=None, ge=0.0, le=1.0)
+    top_k: int | None = Field(default=None, ge=1)
+    seed: int | None = None
     streaming: bool = True
 
 
@@ -83,6 +101,7 @@ class AgentConfig(_FrozenModel):
     system_prompt_ref: str | None = None
     tools: tuple[str, ...] = ()
     mcp_servers: tuple[str, ...] = ()
+    sampling: SamplingConfig | None = None
 
 
 class SwarmConfig(_FrozenModel):

{haru_cli-0.1.1 → haru_cli-0.1.2}/src/haru/config/templates/agents.yaml

@@ -28,3 +28,10 @@ orchestration:
     edges:
       - from: research
         to: write
+# Per-agent sampling overrides beat the model entry's values and are in turn
+# beaten by CLI/REPL overrides. Example:
+#   researcher:
+#     model: haiku-4-5
+#     sampling:
+#       temperature: 0.1
+#       top_k: 20

haru_cli-0.1.2/src/haru/config/templates/models.yaml

@@ -0,0 +1,31 @@
+# Sampling fields (temperature, top_p, top_k, seed) are optional; unset
+# fields are omitted from requests, keeping the provider's defaults.
+# Claude 5-series and Opus 4.7+ models reject non-default temperature/top_p/
+# top_k with an HTTP 400, so leave them unset on those entries. seed is
+# passed through to models that support it (no Claude model does today).
+default_model: sonnet-5
+models:
+  opus-5:
+    model_id: us.anthropic.claude-opus-5
+    region: us-east-1
+    max_tokens: 8192
+    streaming: true
+  sonnet-5:
+    model_id: us.anthropic.claude-sonnet-5
+    region: us-east-1
+    max_tokens: 8192
+    streaming: true
+  opus-4-8:
+    model_id: us.anthropic.claude-opus-4-8
+    region: us-east-1
+    max_tokens: 8192
+    streaming: true
+  # Haiku 4.5 accepts sampling parameters:
+  haiku-4-5:
+    model_id: us.anthropic.claude-haiku-4-5
+    region: us-east-1
+    max_tokens: 4096
+    temperature: 0.2
+    top_p: 0.9
+    top_k: 50
+    streaming: true