Skip to content

feat: add ClaudeAgentRunner using claude-agent-sdk#245

Merged
DhavalRepo18 merged 20 commits intomainfrom
feat/claude-agent-sdk-runner
Apr 7, 2026
Merged

feat: add ClaudeAgentRunner using claude-agent-sdk#245
DhavalRepo18 merged 20 commits intomainfrom
feat/claude-agent-sdk-runner

Conversation

@ShuxinLin
Copy link
Copy Markdown
Collaborator

@ShuxinLin ShuxinLin commented Apr 6, 2026

$(cat <<'EOF'
Closes #244

Summary

  • Adds src/agent/claude_agent/ subpackage with ClaudeAgentRunner that implements the AgentRunner ABC using the claude-agent-sdk agentic loop
  • Wires the same IoT / FMSR / TSFM / utilities / WO MCP servers as stdio MCP servers via ClaudeAgentOptions.mcp_servers — no custom plan loop needed
  • Adds a claude-agent CLI entry point backed by agent.claude_agent.cli:main
  • Adds claude-agent-sdk to project dependencies
  • Exports ClaudeAgentRunner from agent.__init__
  • 7 unit tests (all pass, no real API calls required)

Usage

uv run claude-agent "What sensors are on Chiller 6?"
uv run claude-agent --model claude-opus-4-6 --max-turns 20 "List failure modes for pumps"
uv run claude-agent --json "What is the current time?"

Or programmatically:

import anyio
from agent.claude_agent import ClaudeAgentRunner

runner = ClaudeAgentRunner()
result = anyio.run(runner.run, "What assets are at site MAIN?")
print(result.answer)

Test plan

  • uv run pytest src/agent/claude_agent/tests/ -v — 7 passed
  • uv run pytest src/ -v -k "not integration" — 185 passed (1 pre-existing CouchDB failure unrelated to this PR)
    EOF
    )

ShuxinLin added 20 commits April 6, 2026 16:13
@ShuxinLin
feat: add ClaudeAgentRunner using claude-agent-sdk
3f2c60e 
Closes #244

- Add `src/agent/claude_agent/` subpackage with `ClaudeAgentRunner`
  that implements `AgentRunner` via the claude-agent-sdk agentic loop
- Wire the same IoT/FMSR/TSFM/utilities/WO MCP servers as stdio
  servers via `ClaudeAgentOptions.mcp_servers`
- Add `claude-agent` CLI entry point (`agent.claude_agent.cli:main`)
- Add `claude-agent-sdk>=0.0.14` to project dependencies
- Export `ClaudeAgentRunner` from `agent.__init__`
- Add 7 unit tests (all pass, no real API calls)

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
refactor: remove OrchestratorResult, replace with AgentResult in plan…
f37c488 
…_execute/models

- Delete src/agent/models.py
- Add AgentResult to src/agent/plan_execute/models.py alongside Plan/PlanStep/StepResult
- Update AgentRunner ABC and all subclasses to use AgentResult
- Export AgentResult from agent.__init__

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
refactor: split result types — AgentResult for generic runners, Orche…
88690c6 
…stratorResult for plan-execute

- src/agent/models.py: AgentResult(question, answer, history: Any)
  — thin base result for all AgentRunner subclasses; history type TBD
- src/agent/plan_execute/models.py: OrchestratorResult(question, answer, plan, history: list[StepResult])
  — kept for PlanExecuteRunner with full plan/step-result detail
- ClaudeAgentRunner.run() returns AgentResult(history=None)
- AgentRunner ABC return type is AgentResult

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
fix: rename --model to --model-id in claude-agent CLI
a375683 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
feat: support litellm_proxy/ model IDs in ClaudeAgentRunner via ANTHR…
6e6c0ab 
…OPIC_BASE_URL

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
chore: add ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL to .env.public
bce56fc 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
feat: auto-populate ANTHROPIC_* env vars from LITELLM_* when using li…
cdc59d5 
…tellm_proxy/ model ID

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
chore: remove ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL from .env.public
4db63de 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
chore: remove ANTHROPIC_API_KEY/ANTHROPIC_BASE_URL references from CLI
9475c82 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
fix: map LITELLM_* to SDK env vars internally when using litellm_prox…
fcec456 
…y/ model ID

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
fix: set default permission_mode to bypassPermissions for autonomous …
d3faa6b 
…tool use

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
feat: collect full trajectory (tool calls, text, token usage) from SD…
4fb4a1b 
…K message stream

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
feat: add --show-trace flag to claude-agent CLI to print trajectory
72acea3 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
fix: rename --show-trace to --show-history for consistency with plan-…
4cbb905 
…execute

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
docs: add Claude Agent Runner section to INSTRUCTIONS.md
0cd8334 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
docs: replace 'Runner' with 'Agent' in section headings and TOC
b0a3177 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
refactor: rename history to trajectory throughout src/agent/ and docs
c93fcc9 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
feat: capture tool outputs via PostToolUse hook in ClaudeAgentRunner
62ec41e 
Add output field to ToolCall and register a PostToolUse hook so the
MCP server response is attached to each ToolCall in the trajectory.
Tool outputs are flushed onto the previous turn's calls when the next
AssistantMessage or ResultMessage arrives.

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
docs: document ToolCall.output field and fix --show-trajectory flag
ba76fc2 
Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin
fix: handle string tool_response in PostToolUse hook
5454ab8 
The SDK passes tool_response as a plain string, not a nested dict.
The hook now handles both string and dict responses gracefully.

Signed-off-by: Shuxin Lin <linshuhsin@gmail.com>
@ShuxinLin ShuxinLin requested a review from DhavalRepo18 April 6, 2026 21:02
@ShuxinLin
Copy link
Copy Markdown
Collaborator Author

ShuxinLin commented Apr 6, 2026

@DhavalRepo18 I have implement a preliminary trajectory in claude agent (related to #239).

DhavalRepo18
DhavalRepo18 approved these changes Apr 7, 2026
View reviewed changes
Comment thread src/agent/claude_agent/models.py
text: str
tool_calls: list[ToolCall] = field(default_factory=list)
input_tokens: int = 0
output_tokens: int = 0
Copy link
Copy Markdown
Collaborator

@DhavalRepo18 DhavalRepo18 Apr 7, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There are reasoning model and their output schema are different. Pleasr see LLM vs LRM.

@DhavalRepo18 DhavalRepo18 merged commit 9e7a4d0 into main Apr 7, 2026
1 check passed
@ShuxinLin ShuxinLin deleted the feat/claude-agent-sdk-runner branch April 7, 2026 18:13
caroline-cahill pushed a commit to jasonlee-1024/AssetOpsBench that referenced this pull request Apr 29, 2026
@DhavalRepo18
Merge pull request IBM#245 from IBM/feat/claude-agent-sdk-runner
c1587b0 
feat: add ClaudeAgentRunner using claude-agent-sdk
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@DhavalRepo18 DhavalRepo18 DhavalRepo18 approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

feat: add ClaudeAgentRunner using claude-agent-sdk

2 participants

@ShuxinLin @DhavalRepo18