fix: subagent import, Claude/Codex proxy cwd, 1h cache pricing, and indexed effort

[lis186]

摘要

修正 ccxray 在六個地方少算或算錯用量，另加一項強化，共 7 個 commit，每個都附回歸測試：

• 匯入漏掉 Claude 子代理 transcript：只靠匯入時，成本少算約 37%。
• Claude Code 2.1.283 與 Codex 經 proxy 時 cwd 是 null：影響所有新版 Claude Code 流量的專案歸屬。
• 1 小時快取寫入誤用 5 分鐘價格：Claude Code 的主對話都用 1 小時快取，所以成本少算。
• 匯入 id 只有 10 毫秒精度：同一個 10 毫秒內寫入的紀錄會被誤判為重複而丟掉。
• effort、thinking token、turn 耗時沒有寫進索引：_req.json 14 天後就會被刪，事後無法補。

修正後，實驗 session 的匯入與 proxy 兩條路徑都算出 $0.4030201，與 Claude Code 自己回報的成本完全一致。起因是設計 agentflow 的每個 Ask 成本報表，但除了 effort 欄位以外，其他修正都會影響一般使用者。

Details

• Subagent transcripts were never imported (b79473a). collectJsonlFiles listed only projects/<slug>/*.jsonl and skipped <sid>/subagents/agent-*.jsonl. Subagent turns now import under the parent session with isSubagent, agentKey/agentLabel taken from the .meta.json agentType, plus new subagentId and subagentToolUseId fields. The deployment-identity fields agentId/agentType (feat(index): optional identity fields + local date/tz + config-dir extraction #504) are left alone. Dedup goes through msg.id, the responseId (ADR 0012).
• Import ids used 10 ms resolution (d651552), so two turns written in the same 10 ms were treated as duplicates and one was dropped. Ids now keep full milliseconds. Dedup asks "is this turn already indexed" (by responseId, or by a same-session entry under the new id or the legacy 10 ms id), so a rescan does not rewrite existing imports. A genuinely different turn gets a -1/-2 suffix instead of being dropped.
• Proxy cwd was null on Claude Code 2.1.283 (b2e0d22). 2.1.283 moved the environment block into messages[1] with role: system, and later system messages can be plain strings. extractCwd now scans messages[0] and every system-role message. As a last resort it reads safeguards[].classifier_context.live_cwd. It never scans other user messages, so pasted text cannot decide the project. This affected project attribution for all proxied traffic from current Claude Code.
• 1-hour cache writes were priced at the 5-minute rate (3e41b3a). When usage.cache_creation has the TTL split, 5-minute writes are now priced at 1.25× input and 1-hour writes at 2× input. The 1-hour rate uses LiteLLM's cache_creation_input_token_cost_above_1hr when available. Both proxy (calculateCost) and import/cost-worker (calculateCostSimple) paths are fixed. A non-numeric or negative split falls back to the old calculation.
• Proxy cwd was null for Codex (e131be2). Codex sends cwd in an input[] user message that starts with <environment_context><cwd>…, and none of the existing sources read it. It is now the last fallback. Only a block at the start of the text counts, because Codex memory-consolidation prompts quote older rollouts' environment_context, and taking those would attribute the request to the wrong project.
• Effort, thinking tokens and turn duration were not indexed (0fa5cbe). New index fields effort, thinkingTokens and turnDurationMs are written at ingest time, so they survive the 14-day _req.json retention.
  • Import sources: Claude perTurnEffort or effort, output_tokens_details.thinking_tokens, and turn_duration rows; Codex turn_context.effort.
  • Proxy sources: Claude output_config.effort; Codex reasoning.effort, then metadata.reasoning_effort, then the same session's latest prewarm.
  • This commit also fixes the WebSocket path, which used to drop reasoning from response.create.
• Prewarm map hardening (96c7c53). The Codex prewarm → effort map now expires entries after 30 idle minutes (sliding, because Codex sends only one prewarm per session). It refuses effort/model values over 128 chars instead of truncating them, since a truncated model name would misprice. The existing 500-entry cap now evicts least-recently-used.

These are one PR rather than six because they share one goal (correct per-session attribution and cost across import and proxy) and several touch the same functions. Each commit stands alone if you'd rather split it.

Verification

• CCXRAY_HOME=$(mktemp -d) CCXRAY_EXPORT_DISABLE=1 npm test passes: 2631/2631.
• Evidence the change does what it claims:
  • Fail-on-old / pass-on-new: the 9 changed test files (309 tests) run against origin/main code fail 49; against this branch all 309 pass.
  • Real-traffic check: a Claude Code 2.1.283 claude -p --effort low session with one Task subagent was recorded through an isolated proxy and also imported from its transcript.
    • Cost: Claude Code's own total_cost_usd is $0.4030201. Before this branch, import gave $0.2043 (subagents missing) and proxy gave $0.3230 (1-hour cache mispriced). Now both paths give $0.4030201.
    • Rows: the import went from 7 to 11 (7 main + 4 subagent).
    • Effort: low on this session, and medium on a second interactive session.
  • Proxy cwd: proxied Claude requests with cwd went from 0/24 to 22/24. The other 2 are title-generation requests, which carry no working directory.
  • Codex cwd: a captured Codex exec request (ChatGPT login, WebSocket) now yields its cwd.
  • Browser smoke: an isolated --port 5602 smoke run loads the dashboard with no console errors, and /_api/entries returns the imported rows.
• Independent review (agentflow cross-check/acceptance, Claude-family reviewer): Outcome, Minimality and Conformance all PASS. Security scan: no critical or high findings.

Not verified:

• Dashboard lists: with a home populated only by import --target-transcript, the dashboard shows 0 projects and 0 sessions. origin/main code shows the same with the same data, so this is not a regression, but it means the dashboard rendering of imported subagent turns and the new fields was not checked by eye.
• Successful Codex turn through the proxy: not observed, because the account hit its usage limit. Codex usage/elapsed recording on a completed WebSocket turn was not re-checked.
• Proxy-side thinking tokens and turn duration: not recorded. Proxy records effort only; those two come from import.
• Grok: untouched.
• Deferred proposals, not in this PR:
  • .meta.json agentType allowlist
  • Codex uses only the first user message for environment_context
  • extractConfigDir scans only messages[0]
  • exposing the 1-hour rate in /_api/pricing

License

• I agree to license this contribution under PolyForm Noncommercial 1.0.0
  and grant the maintainer the right to relicense it under different terms in
  the future, including commercial licenses
  (see CONTRIBUTING.md) — maintainer-authored; left for the maintainer to tick.
• My commits are signed off (git commit -s): maintainer-authored, matching existing main history.

🤖 Generated with Claude Code