# v1.0.26 Release Notes

> v1.0.25 -> v1.0.26

## Summary

Usage records now retain tool-call identifiers, while routing and execution improve long-context selection and prompt-cache continuity. Reasoning guidance can be fetched in a single request.

<details>
<summary>翻譯</summary>

Usage 紀錄現在會保留工具呼叫識別碼，並改善長內容模型選擇與提示快取延續性。推理指引也可透過單次請求取得。

</details>

## Recommended upgrade

Cache hit rates on `codex@`, `grok@` and `grok-oauth@` are maximised from this version on. Each now sends the cache key its endpoint actually routes on, and the execution history stays byte-identical between sends so the cached prefix keeps growing instead of resetting. Measured on the same repeated task:

| Model prefix | Before | After |
|---|---|---|
| `codex@` | 16% | 87-93% |
| `grok@` | cached prefix frozen at 13.5K | 71% |
| `grok-oauth@` | no cache key sent | 68% |

Lower spend and lower latency come from the same change: what used to be re-sent at full price is now served from cache. `openai@` already cached well and is unchanged at 94%; it and `cloudflare@` now send the key explicitly.

<details>
<summary>翻譯</summary>

自本版起 `codex@`、`grok@`、`grok-oauth@` 的快取命中率最大化。三者現在都會送出各自端點實際用來路由的快取識別碼，而執行歷史在每次送出之間保持位元組一致，快取前綴因此持續成長而非重置。同一項重複任務的實測：

| 模型前綴 | 之前 | 之後 |
|---|---|---|
| `codex@` | 16% | 87-93% |
| `grok@` | 快取前綴卡在 13.5K | 71% |
| `grok-oauth@` | 未送快取識別碼 | 68% |

花費與延遲同時下降：過去以全額重送的內容現在由快取提供。`openai@` 原本命中率就好，維持 94% 不變；它與 `cloudflare@` 現在也會明確送出識別碼。

</details>

## Changes

### FEAT

- Record tool-call identifiers alongside model usage for agent and summary responses (@pardnchiu) [130f81b3]
- Fetch multiple reasoning guides in one request (@pardnchiu) [6efe752f]
- Prefer providers with larger context windows for long research tasks (@pardnchiu) [6ac619e6]

<details>
<summary>翻譯</summary>

- 在 agent 與摘要回應的模型用量紀錄中保存工具呼叫識別碼
- 單次請求取得多個推理指引
- 長篇研究任務優先選擇 context window 較大的 provider

</details>

### UPDATE

- Upgrade `go-llm-router` to v0.8.0 for per-session prompt-cache keys across Codex, Grok, OpenAI and Cloudflare (@pardnchiu) [876649d3]

<details>
<summary>翻譯</summary>

- 升級 `go-llm-router` 至 v0.8.0，為 Codex、Grok、OpenAI 與 Cloudflare 帶入每個 session 的提示快取識別碼

</details>

### REFACTOR

- Preserve tool ordering during lookup and keep execution history stable for prompt-cache reuse (@pardnchiu) [876649d3]
- Simplify todo history handling and pending-result persistence (@pardnchiu) [876649d3]
- Simplify reasoning-guide headings and clarify tool-error instructions (@pardnchiu) [6d90b327]

<details>
<summary>翻譯</summary>

- 工具查找時保留工具順序，維持執行歷史穩定以利提示快取重用
- 簡化 todo 歷史處理與待續任務結果保存方式
- 簡化推理指引標題並釐清工具錯誤處理說明

</details>

## Scope

- `internal/session/usage/` — FEAT
- `internal/agents/exec/` — FEAT, REFACTOR (`execute.go`, `selectAgentBeta.go`, `todo/`, `compact/`)
- `internal/session/config/` — FEAT (`tier.go`)
- `internal/tools/` — FEAT, REFACTOR (`guide.go`, `interactive/`, `searcher/`)
- `internal/utils/` — REFACTOR (`event.go`)
- `configs/prompts/` — FEAT, REFACTOR (`model_selection.md`, `skill_execution.md`, `guide/`)
- `go.mod`, `go.sum` — UPDATE (`go-llm-router` v0.8.0)

---

Generated by [SKILL](https://github.com/pardnchiu/skill-version-generate)