Grandet Skill 0.3.0 (build.20261018)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261018.

Product and API version stay 0.3.0. This build is build.20261017b (#16) plus (#18):

• Sites the user named. When the user names sites ("A 和 B 哪家便宜", "I've looked at A"), the Skill asks for limit=20 and answers from those sites' own board rows (each row has input and output), named sites first. Only a named site missing from the 20 rows gets one provider lookup. No more limit above 20, board provider filters or view=full for this.
• Install. If the GitHub download has not finished after 30 seconds or fails once, use the Tencent SkillHub copy instead of retrying GitHub.

No price snapshot yet; it follows in the next build with the W42 data.

Gate (test machine, Hermes, frozen gate-bank-v1: 48 questions × deepseek-v4.1-flash and glm-5.3, build.20261017b vs this build): PASS, 15 of 15 targets. Answers correct 95/96 on both builds; median tokens ×0.895; on the named-site questions, tools per answer 8.44 → 6.62, median tokens 342k → 220k, sessions taking side lookups 10/16 → 2/16, sessions sending limit=20 16/16. Price quotes in answers matched Grandet's data in 36 of 36 checked answers.

Mainland China, or whenever GitHub is slow: the same kit downloads from Tencent SkillHub without an account, https://api.skillhub.cn/api/v1/download?slug=grandet-api-price-finder&version=0.3.0-build.20261018.0. A version still in SkillHub review returns 404.

中国大陆用户：GitHub 下载慢时，可以从腾讯 SkillHub 免注册下载同一个完整安装包（链接见上）。

SHA-256:

• grandet-api-price-finder-0.3.0.zip 435666 bytes 8fffbb70384bd56a633f9f2b6cd1608ed9f432906ae60f8691cb79ba736a4a9d
• grandet-api-price-finder-0.3.0.skill 133539 bytes ac485e25def0cfbd462136a8b62f12c5dd7369060f1f3e06828a958427377838

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261017b)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261017b.

Product and API version stay 0.3.0. This build is build.20261017 (#15) plus three Skill commits from the Linux side (#16):

• English glossary column. references/glossary.md gains an English column joined by term id; the Chinese text is unchanged. The file grows from 33.9 KB to 64.2 KB, so the gate watches tokens on relay-jargon questions.
• English answers keep the original group name. In an English conversation the answer writes the relay's group name as listed and puts planEn beside it (response-format.md, "English conversations"). Answer-quality evals gain English cases E4 and E5.
• WorkBuddy is a connect target, and connecting keeps the existing config file's permissions (scripts/connect.mjs, host-setup.md, installation.md).
• Step 4 keeps build.20261017's lead sentence; the planEn note follows the price shape as its own sentence.

No price snapshot in this build: it follows in the next build, regenerated from the W42 data once that is live (owner decision 2026-10-17).

Gate (test machine, Hermes, frozen gate-bank-v1: 48 questions × deepseek-v4.1-flash and glm-5.3, build.20261017 vs this build): HOLD on both lanes, 11 of 13 targets. Answers correct 95/96 on both builds; price answers ask which app 28/32 (build.20261017 25/32); tools per question 7.07 (7.10); API errors per question 0.031 (0.073); median tokens ×1.028. The two misses come from one glm-5.3 session on a question naming two sites: one INVALID_LIMIT and one cheaper-output answer without the link (15/16). The owner approved publishing it; the first candidate of this build ID (i6b, never released) was HOLD on asking which app and tools per question, both fixed here.

Mainland China, or whenever GitHub is slow: the same kit downloads from Tencent SkillHub without an account, https://api.skillhub.cn/api/v1/download?slug=grandet-api-price-finder&version=0.3.0-build.20261017.2. A version still in SkillHub review returns 404.

中国大陆用户：GitHub 下载慢时，可以从腾讯 SkillHub 免注册下载同一个完整安装包（链接见上）。

SHA-256:

• grandet-api-price-finder-0.3.0.zip 434665 bytes aabf2c784fa7f3825d8ce26159d22fea94690e39ce05892181627daa33719f62
• grandet-api-price-finder-0.3.0.skill 133374 bytes d02359bb8fcfeeb7b7e05d4c2c9dd7f5b9d9fa07ea7f3ffea9637e6be62f43f5

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261017)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261017.

Product and API version stay 0.3.0. Changes since build.20261016h (#15) come from the test machine's logs (c's 316 live rounds and the c/h A/B, re-judged with the fixed judge):

• No guessed paths. "Get the data first" lists the only requests: board, free-credits, models and glossary, with view = compact/brief/full and limit 1–20. When a board is empty, the Skill now gives the reason from coverage.excluded and the leaderboard link, and makes no second request. Build h used to try /api/v1/models/<id>, /api/grandet/page?path=/models/…, view=declarations or evidenceUrl downloads in that case.
• Parameter limits are stated where the agent builds the request: limit at most 20, mix as whole-number shares that add up to 100 with cacheRead written out, /api/v1/prices needs model and rejects lang, skillVersion/skillBuild only on the four endpoints that accept them.
• Short answers. Every answer stays within 8 lines and 600 characters. Step 4 now holds the 8-line price-answer shape, and the last line is always the question about which app to connect it to.
• Relay terms are explained from the glossary file itself, not from a search across the Skill folder.
• Free-credit answers end with the grandet.ai free-tokens link.
• Install and update: read download.md with curl, not a page reader. Use /dl/latest.json when the website has it. Check the Installation build line after installing. Hermes installs only into $HERMES_HOME/skills when that is set.
• Uninstalling is described the same way everywhere: move the folder out, or run the bundled installer with --uninstall.
• scripts/check-build-id.mjs checks that one build ID is used everywhere a release ships or pins it. Packaging and registry staging refuse a mismatch.

Gate (test machine, Hermes, frozen gate-bank-v1: 48 questions × deepseek-v4.1-flash and glm-5.3, h vs this build): HOLD on the deepseek-v4.1-flash lane, 10 of 13 targets. The three misses are one question each: free-credit link 7/8, asking which app 14/16 (h 15/16), response-format reads 0.21 per question. Answers correct 47/48 (h 48/48); INVALID_* errors 0 (h 18); median answer 484 characters (h 841); median tokens ×0.936. The owner approved publishing it so the website can switch; the next build (with English planEn answers and WorkBuddy) goes through the gate again.

Mainland China, or whenever GitHub is slow: the same kit downloads from Tencent SkillHub without an account, https://api.skillhub.cn/api/v1/download?slug=grandet-api-price-finder&version=0.3.0-build.20261017.0. A version still in SkillHub review returns 404.

中国大陆用户：GitHub 下载慢时，可以从腾讯 SkillHub 免注册下载同一个完整安装包（链接见上）。

SHA-256:

• grandet-api-price-finder-0.3.0.zip 373443 bytes f8371ff30df1b5d7478c1681c8507e2a42a10d2d72bc1cbe76fac8206723d7c1
• grandet-api-price-finder-0.3.0.skill 119831 bytes 6e965e62754bbf9539122ea86f25e94f6a41dcb2c61c3c76c6e98755f36c2fd4

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261016h)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261016h.

Product and API version stay 0.3.0. Changes since build.20261016g (#12): lessons from test rounds with weak models in hosts that have no Grandet MCP tools.

• The description now opens with "Model API prices, 中转站, 免费额度 — use this, not web search.", followed by the relay terms it understands and the common ways people ask for prices. Hosts that cut the description short still show why to use it. In an A/B test (Hermes + deepseek-v4.1-flash, 20 questions per arm), the strict trigger rate rose from 47% to 85%.
• A new section near the top, "Get the data first", tells the agent not to web-search prices. It gives exact read-only commands for grandet.ai's board and free-credit endpoints in curl, Python and PowerShell, for hosts without the MCP tools.
• Uninstalling: the Skill now says plainly how to remove it from each host. For Hermes, it moves the folder to .grandet-skill-backups (respecting HERMES_HOME) and then checks hermes skills list.
• Credit is not cash: prices in USD_CREDIT/CNY_CREDIT are compared only within one site's board, never ranked or converted across sites.
• Mainland China, or whenever GitHub is slow: the same installation kit downloads from Tencent SkillHub without an account, https://api.skillhub.cn/api/v1/download?slug=grandet-api-price-finder&version=0.3.0-build.20261016.8. A version still in SkillHub review returns 404.

中国大陆用户：GitHub 下载慢时，可以从腾讯 SkillHub 免注册下载同一个完整安装包（链接见上）。

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261016g)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261016g.

Product and API version stay 0.3.0. Changes since build.20261016f (#10):

• Key files are private: for Continue and Gemini CLI, connect now writes the config that holds the API key with owner-only permissions (0600), and applying or undoing a change keeps each file's existing permissions instead of resetting them.
• SKILL.md gains a section, What this Skill changes on your computer. It lists each app config that connect rewrites, the default-model change, the key file, backups, the undo record, and what the installer keeps and reports (with --no-telemetry and --uninstall).
• When the Skill loads on its own, it only reads prices. Connecting an app or installing runs only when the user asks for it and confirms the change shown.
• Mainland China, or whenever GitHub is slow: the same installation kit downloads from Tencent SkillHub without an account, https://api.skillhub.cn/api/v1/download?slug=grandet-api-price-finder&version=0.3.0-build.20261016.7. A version still in SkillHub review returns 404.

中国大陆用户：GitHub 下载慢时，可以从腾讯 SkillHub 免注册下载同一个完整安装包（链接见上）。

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261016f)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261016f.

Product and API version stay 0.3.0. Changes since build.20261016e:

• Answers, from agent test reports (#9): coding requests map to mix=coding; the user's model ID is kept as given and the price answer is tighter; free-credit deadlines and model coverage are worded from the data; connect says when it changes a host's default model and reports it on undo; missing glossary terms are added.
• Registry copies (#8): each registry copy's package-manifest.json now lists that copy's own files, so a copy passes its own manifest check. ClawHub had marked build.20261016e suspicious for this mismatch.
• Mainland China, or whenever GitHub is slow: the same installation kit downloads from Tencent SkillHub without an account, https://api.skillhub.cn/api/v1/download?slug=grandet-api-price-finder&version=0.3.0-build.20261016.6. A version still in SkillHub review returns 404.

中国大陆用户：GitHub 下载慢时，可以从腾讯 SkillHub 免注册下载同一个完整安装包（链接见上）。

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261016e)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261016e.

Product and API version stay 0.3.0. Same Skill behaviour as build.20261016d. What changes is how people get it:

• Mainland China, or whenever GitHub is slow: the same installation kit downloads from Tencent SkillHub without an account, https://api.skillhub.cn/api/v1/download?slug=grandet-api-price-finder&version=0.3.0-build.20261016.5. INSTALL.md, the per-app guide and the cloud-chat connect commands now offer it.
• ClawHub and SkillHub now carry the complete kit (Skill, installer, receipt module, Python verifier, INSTALL.md, package-manifest.json), not only the Skill folder. The release workflow uploads it as registry version 0.3.0-build.20261016.5.
• SKILL.md gains a metadata block (slug, display name, summary, homepage) that SkillHub reads.

中国大陆用户：GitHub 下载慢时，可以从腾讯 SkillHub 免注册下载同一个完整安装包（链接见上）。

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261016d)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261016d.

Product and API version stay 0.3.0. Same Skill content as build.20261016c except the installer: Node installs on native Windows no longer fail with EPERM, the Python verifier keeps staged files within the Windows 260-character path limit and its stale-lock check works on Windows, and the new --uninstall moves an app's copies into .grandet-skill-backups without deleting anything.

The release workflow also uploads this build to ClawHub and Tencent SkillHub as registry version 0.3.0-build.20261016.4.

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance: NOT_RUN.

Grandet Skill 0.3.0 (build.20261016c)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261016c.

Product and API version stay 0.3.0. Same content as build.20261016b except: the connect preview hides keys already in an app's settings (it could show an existing ANTHROPIC_AUTH_TOKEN before), and the connect executor now also writes Hermes and OpenClaw (OpenClaw reads the key through a file SecretRef, so openclaw.json never holds it).

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance in Hermes and OpenClaw: NOT_RUN.

Grandet Skill 0.3.0 (build.20261016b)

Grandet Skill 0.3.0, content build grandet-skill-0.3.0-build.20261016b.

Product and API version stay 0.3.0. Same content as build.20261016 except: INSTALL.md and the READMEs use the homepage install sentence ("Install Grandet Skill for me: https://grandet.ai/download.md"), INSTALL.md no longer tells the installing agent it need not read the page, and answers handle the case where only older quotes exist.

Earlier builds and their assets remain unchanged; existing installations are not backfilled. Native acceptance across all apps: NOT_RUN.