LLM → Telemetry: fixes

[enyst]

Addresses a few things from #637

• LLM → Telemetry:
  • forward input_cost_per_token/output_cost_per_token so that it's actually used for custom cost
• Telemetry logging:
  • de-duplicate tools (drop kwargs.tools)
  • serialize raw_response with model_dump() when available
  • use unique filenames with a short UUID suffix.
• Telemetry cost:
  • intentionally skip recording 0.0; added inline comment clarifying policy.

All tests pass locally (1480 passed). Larger refactors from #637 are deferred to a follow-up.

---

Agent Server images for this PR

• GHCR package: https://github.com/All-Hands-AI/agent-sdk/pkgs/container/agent-server

Variants & Base Images

Variant	Base Image	Docs / Tags
golang	golang:1.21-bookworm	Link
java	eclipse-temurin:17-jdk	Link
python	nikolaik/python-nodejs:python3.12-nodejs22	Link

Pull (multi-arch manifest)

docker pull ghcr.io/all-hands-ai/agent-server:5ea2ff6-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-5ea2ff6-python \
  ghcr.io/all-hands-ai/agent-server:5ea2ff6-python

All tags pushed for this build

ghcr.io/all-hands-ai/agent-server:5ea2ff6-golang
ghcr.io/all-hands-ai/agent-server:v1.0.0_golang_tag_1.21-bookworm_golang
ghcr.io/all-hands-ai/agent-server:5ea2ff6-java
ghcr.io/all-hands-ai/agent-server:v1.0.0_eclipse-temurin_tag_17-jdk_java
ghcr.io/all-hands-ai/agent-server:5ea2ff6-python
ghcr.io/all-hands-ai/agent-server:v1.0.0_nikolaik_s_python-nodejs_tag_python3.12-nodejs22_python

The 5ea2ff6 tag is a multi-arch manifest (amd64/arm64); your client pulls the right arch automatically.

[enyst]

@OpenHands Look at the diff of this branch with main, and remember it. Then merge main branch into this PR's branch. Pay attention to the incoming changes, they affect telemetry. Adapt our changes, commit and push.

[openhands-ai]

I'm on it! enyst can track my progress at all-hands.dev

[openhands-ai]

Final summary of work completed

What I did

• Reviewed and memorized the diff of fix/llm-telemetry-pricing vs main, focusing on telemetry-related changes.
• Merged origin/main into fix/llm-telemetry-pricing.
• Verified the merged code preserved and aligned telemetry behavior:
  • LLM passes input_cost_per_token and output_cost_per_token into Telemetry.
  • Telemetry cost calculation uses LiteLLM headers when present; otherwise falls back to litellm cost calculator with custom_cost_per_token when provided.
  • Skips recording zero cost values.
  • Logging uses unique filenames with a short UUID suffix.
  • Logging serializes responses safely via a centralized _safe_json (ModelResponse -> model_dump with redaction; fallbacks for other objects).
  • De-duplicates duplicated tools by removing kwargs["tools"] when also present at top-level.
• Built and ran the full test suite: 1513 passed, no new failures.
• Amended the merge commit to include Co-authored-by and pushed to the PR branch.

Checklist

• Retrieved and reviewed the diff against main and kept telemetry changes in mind.
• Merged main into the PR branch.
• Ensured telemetry-related incoming changes are respected and compatible.
• Verified with a full build and test run (1513 passed).
• Committed (amended) with Co-authored-by and pushed to origin.

Conciseness and scope

• No extraneous code changes were added beyond the merge; existing telemetry adaptations remained valid and required no further edits.

Details

• Branch: fix/llm-telemetry-pricing
• Latest commit pushed: 7839af2 (Merge remote-tracking branch 'origin/main' into fix/llm-telemetry-pricing)
• Tests: uv run pytest -q → 1513 passed (warnings unchanged/benign)

_{View full conversation}