[angel-manuel]

Summary

• Parse OpenRouter's richer usage shape off the SSE wire: top-level cost, plus cache tokens in both Anthropic-flat (cache_read_input_tokens, cache_creation_input_tokens) and OpenAI-nested (prompt_tokens_details.cached_tokens) forms. A custom Deserialize on Usage collapses both into the flat public shape (cache_read_tokens = max of the two sources). Fields flow into the Langfuse generation span (metadata.cost_usd, cache_read_tokens, cache_creation_tokens) and into trace.finalize(...total_cost...) — previously all hardcoded to 0.
• Add an opt-in LANGFUSE_CAPTURE_INPUT flag that attaches a redacted chat-log snapshot (ported from sister-repo overfolder/agent-runner's build_context_snapshot) to the generation input. Adapted for overloop's Content enum: non-text blocks render as [image]/[audio]/[media] placeholders so no base64/data-URI payload leaks. Off by default — the default posture keeps prompt content local.
• No breaking changes on wire, REST, or CLI surfaces. Usage public shape stays flat (added fields default to 0 on absent input).

Test plan

• cargo test -p overloop — 184 tests pass (121 unit + integration suites)
• cargo clippy --workspace --all-targets -- -D warnings clean
• cargo fmt --all
• 4 new Usage deser tests cover Anthropic-flat, OpenAI-nested, both-absent, both-present-picks-max
• 7 new preview.rs tests cover detail-window, skip-count summary, per-role truncation, multimodal placeholder rendering (no base64 leak), char-boundary safety
• test_langfuse.rs success-generation test now asserts metadata.cost_usd round-trip; new generation_input_preview_string_roundtrips covers the string-preview branch
• LANGFUSE_CAPTURE_INPUT toggle parses 1/true/yes/TRUE as truthy, everything else as false
• Manual: run with LANGFUSE_CAPTURE_INPUT=1 + real creds, eyeball Langfuse UI for [USER]:/[AGENT]: snapshot + non-zero metadata.cost_usd

🤖 Generated with Claude Code

[sentry]

Codecov Report

❌ Patch coverage is 96.93487% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 91.81%. Comparing base (8b1a9d0) to head (d024af0).

Files with missing lines	Patch %	Lines
loop/src/observability/preview.rs	95.94%	6 Missing ⚠️
loop/src/agentic_loop.rs	96.00%	1 Missing ⚠️
loop/src/main.rs	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #44      +/-   ##
==========================================
+ Coverage   91.62%   91.81%   +0.19%     
==========================================
  Files          49       50       +1     
  Lines        6432     6685     +253     
==========================================
+ Hits         5893     6138     +245     
- Misses        539      547       +8     

Files with missing lines	Coverage Δ
loop/src/acp.rs	91.12% <100.00%> (+0.05%)	⬆️
loop/src/config.rs	100.00% <100.00%> (ø)
loop/src/llm/types.rs	99.58% <100.00%> (+0.13%)	⬆️
loop/src/observability/langfuse.rs	99.60% <100.00%> (+<0.01%)	⬆️
loop/src/agentic_loop.rs	92.37% <96.00%> (+0.08%)	⬆️
loop/src/main.rs	0.00% <0.00%> (ø)
loop/src/observability/preview.rs	95.94% <95.94%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.