feat(agent): M4 — bounded tool-use loop + LlmProvider (scripted fake) #10

Merged

charles merged 1 commit from feat/m4-agent into main 2026-06-17 13:57:33 +00:00

Conversation 0 Commits 1 Files changed 4 +400 -8

charles commented 2026-06-17 13:56:10 +00:00

Owner

Copy link

Roadmap M4: the agent engine, deterministically testable with no live LLM.

Changes (geneatrip-agent)

• LlmProvider trait (one OpenAI-compatible shape for Qwen/Ollama + Kimi/UmansAI); Message/Role/ToolCall/Turn/LlmError.
• Tool trait + ToolSpec + ToolOutcome (Ok/Err — structured, never fabricated; FR-AGT-14).
• run(): bounded loop (plan → call tools → observe → repeat) with Budget caps (max_iters/max_tool_calls) ending gracefully with a marked partial (FR-AGT-08/20); unknown/failed tools surfaced as JSON errors, not panics.
• ScriptedProvider: replays a fixed Turn sequence (tech-spec 12.3).

Verification

• just qa green. 5 tests: tool-then-finish, tool-call budget cap, empty-script finalize, failing-tool surfaced, unknown-tool safe.

Next: OpenAI-compatible HTTP provider (live Qwen/Kimi) + the trip toolset, then wiring into the trip builder (M6) to replace StubBuilder. Traces: FR-AGT-08/09/10/14/20 · tech-spec 04.

Roadmap **M4**: the agent engine, deterministically testable with **no live LLM**. ## Changes (geneatrip-agent) - `LlmProvider` trait (one OpenAI-compatible shape for Qwen/Ollama + Kimi/UmansAI); `Message`/`Role`/`ToolCall`/`Turn`/`LlmError`. - `Tool` trait + `ToolSpec` + `ToolOutcome` (Ok/Err — structured, never fabricated; FR-AGT-14). - `run()`: bounded loop (plan → call tools → observe → repeat) with `Budget` caps (max_iters/max_tool_calls) ending gracefully with a marked partial (FR-AGT-08/20); unknown/failed tools surfaced as JSON errors, not panics. - `ScriptedProvider`: replays a fixed `Turn` sequence (tech-spec 12.3). ## Verification - `just qa` green. 5 tests: tool-then-finish, tool-call budget cap, empty-script finalize, failing-tool surfaced, unknown-tool safe. Next: OpenAI-compatible HTTP provider (live Qwen/Kimi) + the trip toolset, then wiring into the trip builder (M6) to replace StubBuilder. Traces: FR-AGT-08/09/10/14/20 · tech-spec 04.

charles added 1 commit 2026-06-17 13:56:10 +00:00

feat(agent): M4 — bounded tool-use loop + LlmProvider (scripted fake) ... a9890b5c47

The agent engine, deterministically testable with no live LLM:
- LlmProvider trait (one OpenAI-compatible shape for Qwen/Ollama + Kimi/UmansAI),
  Message/Role/ToolCall/Turn types, LlmError
- Tool trait + ToolSpec + ToolOutcome (Ok/Err — structured, never fabricated)
- run(): the bounded loop — plan → call registered tools → observe → repeat,
  with Budget caps (max_iters, max_tool_calls) ending gracefully with a marked
  partial; unknown/failed tools surfaced as JSON errors, not panics
- ScriptedProvider: replays a fixed Turn sequence (tests / tech-spec 12.3)

5 tests: tool-then-finish, tool-call budget cap, empty-script finalize, failing
tool surfaced, unknown tool safe. just qa green. HTTP providers + trip toolset +
wiring into the trip builder come next (M4b/M6).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

charles merged commit a03893ff96 into main 2026-06-17 13:57:33 +00:00

charles referenced this pull request from a commit 2026-06-17 13:57:34 +00:00

feat(agent): M4 — bounded tool-use loop + LlmProvider (scripted fake) (#10)

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

charles/geneatrip!10

No description provided.