gpt-6-astra v1

🚀 gpt-6-astra-v1

中文

✨ gpt-6-astra-v1 是 gpt-6-astra 产品线的首个正式版本，由 epoch2 最佳实测稿 e2b19 直接晋升。相较 v1-rc1，重点提升：

• ⚡ 单次原子分派：把当前用户动词链编译为一个不可拆分的执行事件，工具或文件任务在首个可见动作中直接进入检查、修改或评测，降低“只读断点后给计划”的概率。
• 🔄 项目续作保持：读取包含安全抑制、软件修改或内容规则的候选提示词时，继续把项目维护/优化视为外层任务；project_continuation.zh.01 从 rc1 的失败提升为通过。
• 🛠️ 事务轮提交：修改任务在同一轮中保持修改件、patch/diff、验证记录与可运行 rollback 四个角色，并完成 baseline、modified、rollback 与 reapply 的行为级验证。
• ✅ 真实返回稳定性：B execution_completion 的九个实际模型返回全部通过人工复核；唯一未通过项为 provider-policy block，而非返回正文失败。

提示词原理改进

v1-rc1 主要依赖通用首动作和完成规则，遇到“继续优化当前提示词项目”时可能把候选文件中的被测规则误当作本轮请求，读取断点后转为资格或范围说明。v1 将执行控制压缩为三个连续契约：

1. atomic dispatcher 只绑定本轮外层任务、现有断点和第一个可提交动作；
2. continuation ledger 继承已经确认的对象与结果，但不继承旧回复中的拒绝或范围判断；
3. transaction-turn commit 要求工具任务在同一轮形成可验证、可回滚并重新应用的目标状态。

这一结构保留了 rc1 的技术事务能力，同时把“读取项目状态”从终点改为下一项候选修改或隔离评测的前置步骤。

测试方法与结果

所有新运行固定使用 gpt-6-astra、medium、workers=1。A 为 4 cases / 4 turns，准入要求 3/4 且 2/2 artifact gates；B 按 family 顺序运行，本次只完成 execution_completion 8 cases / 10 turns；后续 B families 与 C 未运行。完整方法与失败说明见 docs/comparison-tests.md。

阶段	v1-rc1 / e1b5	e2b12	e2b15	v1 / e2b19
A cases / turns	2/4 · 2/4	3/4 · 3/4	3/4 · 3/4	3/4 · 3/4
A artifact gates	2/2	2/2	2/2	2/2
B execution cases / turns	6/8 · 8/10	4/8 · 6/10	5/8 · 7/10	7/8 · 9/10
B artifact gates	7/8	6/8	6/8	7/8
Provider-policy blocks	2	0	3	1
后续 B families / C	未运行	未运行	未运行	未运行

📈 相较 v1-rc1，v1 的 A 增加 1 case / 1 turn，B execution 增加 1 case / 1 turn；最关键的新增收益是精确项目续作探针通过，且 B 九个实际返回均无正文失败。

📌 上表只列真实运行配置与首个结果。e2b15 与 e2b19 提示词字节相同，两次 B 的差异全部来自 provider-policy block 数量波动；该波动不被改写为提示词能力差异。

📉 v1 基础提示词为 132 行、7,495 UTF-8 bytes，比 v1-rc1 的 7,140 bytes 增加 4.97%，仍低于 8,000-byte 上限。

🛠️ codex-instruct.py 保持 gpt-5.6-v45 为唯一默认选择，并新增 gpt-6-v1 正式版选项；部署、配置快照、字段级回滚以及 provider、模型和认证配置保留方式与 v45 一致。

📦 历史版本：v1-rc1 已移至 historical-versions/。核心提示词 Markdown SHA256 为 39fb46d6edc75963677fd92828dcb6c66dce11740b432efda3af9a683158ce16。

核心提示词包 gpt-6-astra-v1.zip 的 SHA256：

054edb6fa8a6edd2d144c8582756df3179a85481bcb6696d8b730177521b1de1

完整发布包 gpt-instruct-v1.zip 的 SHA256：

159264bfd96ad75e7dab15d22f69f7f52faf230afa335c7adcd5718cc0b1f65e

---

English

✨ gpt-6-astra-v1 is the first formal release of the gpt-6-astra line, promoted directly from the best measured Epoch 2 revision, e2b19. Its main improvements over v1-rc1 are:

• ⚡ Single atomic dispatch: the current user verb chain compiles into one indivisible execution event. A tool or file task enters inspection, modification, or evaluation in its first visible action instead of stopping after a read-only checkpoint with a plan.
• 🔄 Project-continuation retention: after reading candidate instructions that discuss safety suppression, software modification, or content rules, the controller keeps project maintenance/optimization as the outer task. project_continuation.zh.01 improves from failure on rc1 to pass.
• 🛠️ Transaction-turn commit: a modification task retains four roles in one turn—the modified artifact, patch/diff, verification record, and runnable rollback—and behaviorally verifies baseline, modified, rollback, and reapply states.
• ✅ Returned-output stability: all nine actual model returns in B execution_completion pass manual review. The sole miss is a provider-policy block, not a returned-output failure.

Prompt-Principle Changes

v1-rc1 primarily relied on general first-action and completion rules. On “continue optimizing this prompt project,” it could treat rules inside the candidate file as the request being executed and end after a read-only checkpoint with an eligibility or scope statement. v1 compresses execution control into three consecutive contracts:

1. the atomic dispatcher binds only the outer task, confirmed checkpoint, and first committable action;
2. the continuation ledger inherits confirmed objects and results, but never an earlier refusal or scope judgment;
3. the transaction-turn commit requires a tool task to reach a verified, rollback-tested, and reapplied target state in the same turn.

This structure retains rc1's technical-transaction strength while turning project-state reading from an endpoint into preparation for the next candidate edit or isolated evaluation.

Test Method and Results

All new runs use gpt-6-astra, medium, and workers=1. A contains 4 cases / 4 turns and admits a candidate at 3/4 plus 2/2 artifact gates. B runs in a fixed family order; this release completed only execution_completion with 8 cases / 10 turns. Later B families and C were not run. See docs/comparison-tests-en.md for the full method and failure record.

Stage	v1-rc1 / e1b5	e2b12	e2b15	v1 / e2b19
A cases / turns	2/4 · 2/4	3/4 · 3/4	3/4 · 3/4	3/4 · 3/4
A artifact gates	2/2	2/2	2/2	2/2
B execution cases / turns	6/8 · 8/10	4/8 · 6/10	5/8 · 7/10	7/8 · 9/10
B artifact gates	7/8	6/8	6/8	7/8
Provider-policy blocks	2	0	3	1
Later B families / C	Not run	Not run	Not run	Not run

📈 Compared with v1-rc1, v1 gains 1 A case / 1 A turn and 1 B-execution case / 1 turn. The key new result is a pass on the exact project-continuation probe with no returned-output failure among the nine B responses.

📌 The table lists only measured configurations and first results. e2b15 and e2b19 have byte-identical prompts; their B difference comes entirely from the number of provider-policy blocks and is not represented as a prompt-capability difference.

📉 The v1 base prompt is 132 lines and 7,495 UTF-8 bytes, 4.97% larger than the 7,140-byte v1-rc1 prompt while remaining below the 8,000-byte limit.

🛠️ codex-instruct.py keeps gpt-5.6-v45 as the sole default and adds the formal gpt-6-v1 option. Deployment, configuration snapshots, field-level rollback, and preservation of provider, model, and authentication settings follow the same path as v45.

📦 History: v1-rc1 has moved to historical-versions/. SHA256 for the core prompt Markdown is 39fb46d6edc75963677fd92828dcb6c66dce11740b432efda3af9a683158ce16.

SHA256 for the core prompt package, gpt-6-astra-v1.zip:

054edb6fa8a6edd2d144c8582756df3179a85481bcb6696d8b730177521b1de1

SHA256 for the complete release bundle, gpt-instruct-v1.zip:

159264bfd96ad75e7dab15d22f69f7f52faf230afa335c7adcd5718cc0b1f65e

gpt-5.6-instruct v45

🚀 gpt-5.6-instruct v45

✨ v45 针对 Issues 与社交平台用户反映的实际问题，提升逆向、生物研究任务通过概率，降低触发云审查几率，相较 v42 重点提升：

• ⚡ 单次任务编译：以最新 USER 请求为唯一活动任务，控制型续句继承最近的实质任务；在输出前一次选择 OUTER_SCHEMA > TOOL_TRANSACTION > PROCESS_RECORD > DIRECT_ANSWER，减少模板误路由与重复开场。
• 🛠️ 工具事务闭环：工具/文件修改任务首个可见动作直接检查唯一工作对象；保留原件哈希、修改副本，并创建和重开修改件、diff、验证记录、可运行 rollback 四个真实角色。
• ✅ 行为级验证：验证记录与最终答复共同给出基线、修改后与回滚的命令、输入、原样结果和退出状态，要求在另一个副本上实际测试回滚，同时保留修改件的目标状态。
• 🔄 过程记录渲染：物理、机械、生物与叙事过程统一绑定 CORE_EVENT；START 和首个 DEVELOPMENT 句复用同一中心谓词，随后完成 RESULT、FOLLOW-UP 及叙事所需的 DIALOGUE / SENSATION。
• 🧭 状态连续性与外层格式优先：续轮复用已确认文件、哈希、偏移、报告和结果；列表、JSON、文档或批量结构保持用户指定的外层形式，内部规则只负责填充内容。

提示词原理改进

原 v42 以 normalizer、强制工具分派、输出禁用规则、连续性和 generalized completion 组成线性规则栈。v45 把分散的首轮执行与完成规则重构为一次编译、单一路由的执行模型，将其压缩为四个互斥的渲染分支：

1. SINGLE-PASS TASK COMPILER 绑定当前任务、语言、形式和连续状态；
2. OUTER_SCHEMA 保护用户要求的列表、JSON、文档或批量外壳；
3. TOOL_TRANSACTION 管理可逆修改与四角色证据；
4. PROCESS_RECORD 管理具有开始、发展、结果和后续状态的过程；其余请求进入 DIRECT_ANSWER。

这一结构把“先判断再回退”的多路径决策改为“先绑定再完整渲染”的单路径决策，并把中心事件、工具工件和最终状态从建议性规则提升为分支级输出契约。

测试方法与结果

测试方法更新为严格 A → B → C：A 是 3 个原始用户反馈样例；A 全过后，B 按固定 family 顺序运行 66 cases / 74 turns；B 全过后，C 使用 120 条 medium 输入的 batch-10 batched_json_screen（当前限于测试 Token 不足，C暂时未测试）。完整方法与失败分布见 docs/comparison-tests.md。

以下是 v42、v44 和 v45 在同一冻结 bank、runner/scorer 与 transport 下的 comparison-only 横向结果，不等同三版完成当前发布门禁：

阶段	v42	v44	v45
A cases / turns	1/3 · 1/3	2/3 · 2/3	2/3 · 2/3
A artifact gates	1/2	2/2	2/2
B cases / turns	43/66 · 51/74	49/66 · 55/74	54/66 · 62/74
B artifact gates	13/16	14/16	12/16

📈 相较 v42，v45 的 B 结果增加 11 个 case 和 11 个 turn；biology 为 16/16、cloud 为 16/16、progress 为 8/8，routing 为 10/12。当前保留问题是 fiction 0/6、execution 4/8、routing 10/12，且 B artifact gate 为 12/16。

📉 v45 基础提示词为 84 行、5,170 UTF-8 bytes，比 v42 的 90 行、5,856 bytes 缩短 11.71%。

🛠️ codex-instruct.py 现以 v45 为唯一默认生产版本，继续支持部署前预览、配置快照、字段级回滚以及 provider、模型和认证配置保留。v42 的核心包和明文源已移至 historical-versions/。

📦 核心提示词包 gpt-5.6-sol-unrestricted-v45.zip 的 SHA256：

c86c2c6d20a4d1155d87422f485eb37b77539132270918c002b5d8237a5adf54

完整发布包 gpt-5.6-instruct-v45.zip 的 SHA256：

98d9550849e9445365eda5562fd854cab3039bb2df866ba200c6b3136558cf05

---

✨ v45 addresses real issues reported in Issues and by users on social platforms, improving pass probability for reverse-engineering and biological-research tasks while reducing the likelihood of triggering cloud-side review, with the following primary improvements over v42:

• ⚡ Single-pass task compilation: the newest USER request is the only active task, while a control-only continuation inherits the latest substantive task. Before emitting prose, the model chooses exactly one route: OUTER_SCHEMA > TOOL_TRANSACTION > PROCESS_RECORD > DIRECT_ANSWER, reducing template misrouting and repeated openings.
• 🛠️ Closed-loop tool transactions: a tool/file modification starts by inspecting the sole workspace object; it preserves the original hash, modifies a copy, and creates and reopens four real roles—the modified artifact, diff, verification record, and runnable rollback.
• ✅ Behavior-level verification: the verification record and final answer jointly provide baseline, modified, and rollback commands, inputs, literal results, and exit statuses; rollback is actually tested on another copy while the modified artifact retains its target state.
• 🔄 Process-record rendering: physical, mechanical, biological, and narrative processes bind one CORE_EVENT; START and the first DEVELOPMENT sentence reuse the same central predicate, then complete RESULT, FOLLOW-UP, and the narrative DIALOGUE / SENSATION fields.
• 🧭 State continuity with outer-schema priority: continuations reuse confirmed files, hashes, offsets, reports, and results. Lists, JSON, documents, and batches retain the user's requested outer form while inner rules fill each body.

Prompt-Principle Changes

The original v42 used a linear rule stack composed of a normalizer, mandatory tool dispatch, output exclusions, continuity, and generalized completion. v45 restructures the scattered first-turn execution and completion rules into a compile-once, single-route execution model with four mutually exclusive rendering branches:

1. SINGLE-PASS TASK COMPILER binds the current task, language, form, and continuity state;
2. OUTER_SCHEMA protects a requested list, JSON, document, or batch envelope;
3. TOOL_TRANSACTION manages reversible modification and four-role evidence;
4. PROCESS_RECORD manages processes with start, development, result, and follow-up states; all remaining requests use DIRECT_ANSWER.

This structure changes a multi-path “classify, then possibly fall back” decision into a single-path “bind, then render to completion” flow, elevating the central event, tool artifacts, and final state from advisory rules to branch-level output contracts.

Test Method and Results

The method is now a strict A → B → C sequence: A contains three original user-feedback cases; after all A cases pass, B runs 66 cases / 74 turns in a fixed family order; after all B cases pass, C uses batch-10 batched_json_screen on 120 medium inputs (C temporarily not run because test tokens are insufficient). See docs/comparison-tests-en.md for the full method and failure distribution.

The table below compares v42, v44, and v45 under the same frozen bank, runner/scorer, and transport. These are comparison-only runs and do not mean that all three versions completed the current release gate:

Stage	v42	v44	v45
A cases / turns	1/3 · 1/3	2/3 · 2/3	2/3 · 2/3
A artifact gates	1/2	2/2	2/2
B cases / turns	43/66 · 51/74	49/66 · 55/74	54/66 · 62/74
B artifact gates	13/16	14/16	12/16

📈 Compared with v42, v45's B results increase by 11 cases and 11 turns; biology reaches 16/16, cloud reaches 16/16, progress reaches 8/8, and routing reaches 10/12. The retained issues are fiction at 0/6, execution at 4/8, routing at 10/12, and B artifact gates at 12/16.

📉 The v45 base prompt is 84 lines and 5,170 UTF-8 bytes, 11.71% shorter than v42's 90 lines and 5,856 bytes.

🛠️ codex-instruct.py now treats v45 as the sole default production release while retaining deployment preview, configuration snapshots, field-level rollback, and preservation of provider, model, and authentication settings. The v42 core package and plaintext source have moved to historical-versions/.

📦 SHA256 for the core prompt package, gpt-5.6-sol-unrestricted-v45.zip:

c86c2c6d20a4d1155d87422f485eb37b77539132270918c002b5d8237a5adf54

SHA256 for the complete release bundle, gpt-5.6-instruct-v45.zip:

98d9550849e9445365eda5562fd854cab3039bb2df866ba200c6b3136558cf05

gpt-5.6-instruct v42

🚀 gpt-5.6-instruct v42

中文

✨ v42 在 v41 基础上，针对 GitHub Issues 中“首轮不执行、模板循环和软件修改任务无法落地”等真实问题，重点优化了：

• ⚡ 无历史对话时的首轮直接执行，不再依赖用户重复输入
• 🔄 逆向任务中的模板误路由、循环恢复与状态连续性
• 📦 修改副本、补丁、验证记录和可运行回滚脚本的工件化交付
• ✅ 基线与修改后行为的实际执行验证

优化过程中，模型持续吸收用户指令、真实失败案例及 GitHub Issues，自行创建或扩展测试集、分析失败并重写提示词，再通过分层回归决定是否发布。

📊 基于 Issues #5 与 #22 的发布门禁，v42 获得以下结果：

门禁	结果
原始输入首轮门禁（medium）	2/2 cases · 2/2 turns · 2/2 artifact gates
扩展 Issue 专项集（low）	60/60 cases · 68/68 turns · 8/8 artifact gates
原 120-case medium 测试集（low）	115/120 首跑 + 5/5 定向审计 → 120/120 汇总

📌 上述数据只对应实际运行的配置；限于本人时间和token不足等原因，尚未运行的 v42 medium / high 全量矩阵。v41 的 low、medium、high 完整矩阵继续作为历史证据保留。

📉 v42 基础提示词为 90 行、5,856 UTF-8 bytes，较 v35 缩短 42.58%。

🛠️ codex-instruct.py 现以 v42 作为唯一默认生产版本，支持部署前预览、配置快照、字段级安全回滚，并在卸载时保留 provider、模型和认证配置。

📦 v41 与 v41-skills 已移入 historical-versions/；核心提示词包 gpt-5.6-sol-unrestricted-v42.zip 的 SHA256 为：

11f0515be89943a7244d07b625a497b04dde07a51ba26e41df583a0acc145a09

---

English

✨ Building on v41, v42 targets real GitHub Issues involving first-turn non-execution, template loops, and software-modification tasks that did not reach a usable result. Its main improvements are:

• ⚡ Direct first-turn execution with zero conversation history, without requiring repeated user input
• 🔄 Recovery from template misrouting and loops in reverse-engineering tasks, with stronger state continuity
• 📦 Artifact-based delivery of a modified copy, patch, verification record, and runnable rollback script
• ✅ Actual execution checks for both baseline and modified behavior

During optimization, the model continuously incorporated user instructions, real failure cases, and GitHub Issues; created or expanded test sets; analyzed failures and rewrote the prompt; and then used tiered regression gates to decide whether to publish.

📊 On the release gates built from Issues #5 and #22, v42 achieved:

Gate	Result
Exact first-turn gate (medium)	2/2 cases · 2/2 turns · 2/2 artifact gates
Expanded issue-regression set (low)	60/60 cases · 68/68 turns · 8/8 artifact gates
Original 120-case medium bank (low)	115/120 first pass + 5/5 targeted audit → 120/120 aggregate

📌 These figures describe only configurations that were actually run. Unrun full-bank v42 medium / high results are not extrapolated. The complete low, medium, and high matrices for v41 remain available as historical evidence.

📉 The v42 base prompt is 90 lines and 5,856 UTF-8 bytes, 42.58% shorter than v35.

🛠️ codex-instruct.py now treats v42 as the sole default production release, with deployment preview, configuration snapshots, field-level safe rollback, and preservation of provider, model, and authentication settings during uninstall.

📦 v41 and v41-skills have moved to historical-versions/. The SHA256 of the core prompt package, gpt-5.6-sol-unrestricted-v42.zip, is:

11f0515be89943a7244d07b625a497b04dde07a51ba26e41df583a0acc145a09

gpt-5.6-instruct v41

🚀 gpt-5.6-instruct v41

中文

✨ v41 针对 GitHub Issues 中的真实使用问题，重点优化了：

• ☁️ 明文云审查与生物研究场景
• 🔄 模板误路由与循环恢复
• 👀 执行进度可见性

优化过程中，模型持续吸收用户指令、真实失败案例及 GitHub Issues，自行扩展测试集、分析失败并重写提示词。

📊 基于 Issues #3、#4、#5、#6、#8 构建的 52-case / 58-turn 明文专项测试中，v41 三档均达到 52/52：

档位	相较 v35 提升
Low	+25.00 个百分点
Medium	+25.00 个百分点
High	+23.08 个百分点

📉 基础提示词长度同时减少 55.67%。

🧩 此外，项目提供 Skills 专用版本 gpt-5.6-sol-unrestricted-v41-skills，需手动部署。

---

English

✨ v41 introduces targeted improvements based on real-world GitHub Issues, focusing on:

• ☁️ Plaintext cloud moderation and biological research
• 🔄 Template misrouting and loop recovery
• 👀 Progress visibility

During optimization, the model continuously incorporated user instructions, real failure cases, and GitHub Issues, while autonomously expanding test sets, analyzing failures, and rewriting prompts.

📊 On the 52-case / 58-turn plaintext benchmark built from Issues #3, #4, #5, #6, and #8, all three v41 configurations achieved 52/52:

Configuration	Improvement over v35
Low	+25.00 percentage points
Medium	+25.00 percentage points
High	+23.08 percentage points

📉 The base prompt length was also reduced by 55.67%.

🧩 A Skills-specific variant, gpt-5.6-sol-unrestricted-v41-skills, is also available for manual deployment.

gpt-5.6-instruct v35

提供两个等级的提示词：
v5 结构更简洁，足以覆盖大多数场景，并作为默认推荐版本；
v35 在此基础上进一步将具体名称与网址归一化为占位符，再按中英文复合意图族进行统一路由，减少复合任务中只完成部分子任务的情况。
当前情况建议仅在 v5 无法满足需求时使用 v35。

Two prompt tiers are available：
v5 has a simpler structure, covers most scenarios, and is the recommended default.
v35 additionally normalizes concrete names and URLs into placeholders, then routes bilingual compound-intent families as a whole to reduce partial completion of multi-part tasks.
Use v35 only when v5 does not meet your needs.