fix(router): statusline token dedup, translator id fallback, context-fit filter for flash models

[steventohme]

Summary

Two unrelated correctness fixes surfaced from a session where the statusline showed saved $5.84 · 387.2k in / 674 out for a session whose actual totals were $2.94 · 194.8k in / 327 out.

install/cc-statusline.sh — token/savings inflation

Claude Code writes one JSONL entry per content block of an assistant turn (e.g. text + text + tool_use → 3 entries). Every entry carries the same message.usage payload. The existing jq filter ran select(.type=="assistant") and summed per entry, double- or triple-counting every multi-block turn — observed ~2× inflation in a real qwen session, can hit 3× for turns with two text blocks before a tool call.

Dedupe on (message.id, message.usage) before summing:

• Native Anthropic upstreams: message.id is unique per turn → collapses content-block fan-out cleanly.
• Translator round-trips: message.id can be a constant placeholder (msg_translated); usage still differs per turn (input_tokens grows), so the composite key keeps real turns distinct. Two turns with byte-identical id AND usage would still collapse — that's a genuine retry/duplicate we want to drop.

Verified: same transcript that produced 387.2k in / saved $5.84 now produces 194.8k in / saved $2.94, matching independent jq dedup math within $0.005.

internal/translate/stream.go — empty-string id latched

if id := gjson.GetBytes(data, "id"); id.Exists() && t.messageID == "" latches the first id field present in the SSE stream. Some OpenAI-compat upstreams (observed: OpenRouter for certain models) send early chunks with "id": "" before settling on a real id. gjson reports Exists()=true, Str="" for that, so the empty string was being latched — never overwritten — and emitMessageStart fell back to the "msg_translated" placeholder, breaking any downstream that dedupes by message id.

Switched the gate to Str != "" (same for model). Strictly tightens behavior: only latches non-empty strings.

Test plan

• Manually replayed a real transcript through the patched cc-statusline.sh and verified output matches hand-computed dedup'd totals.
• Confirm internal/translate unit tests still pass (none currently exercise the empty-id chunk case — could add one in a follow-up).
• Once merged, bump the gitlink in the workweave repo + copy install/cc-statusline.sh into syncrouterinstall/install_template.sh so the heredoc-embedded copy stays in lockstep.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 850436e. Configure here.}

[cursor]

Heredoc in install.sh not synced with statusline fix

High Severity

install/cc-statusline.sh was updated with the jq -rs slurp mode and unique_by dedup fix, but the inlined heredoc copy inside install/install.sh (lines ~1434–1459) still uses the old jq -r with bare select(.type=="assistant") — no dedup. Fresh installs via make full-setup or curl|sh will write the stale version to disk and continue double/triple-counting tokens and savings.

 

^{Triggered by learned rule: install/cc-statusline.sh and install/install.sh heredoc must stay in sync}

^{Reviewed by Cursor Bugbot for commit 850436e. Configure here.}