[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7B3D5TBTWNWHW2D52ED0QVR":3,"skill-same:01M7B3D5TBTWNWHW2D52ED0QVR":47},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":31,"updatedAt":32,"install":33,"resources":46},"01M7B3D5TBTWNWHW2D52ED0QVR","oracle",null,"Oracle second-model review: bundle prompts\u002Ffiles, debug, refactor, design-check.","merllinsbeard\u002Fmerlin-CTO","merllinsbeard","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F144795773?v=4","01M7B3C01C1PCBWFGB5T7YXSF8","merlin","01M7B3C023BBYDEPAPWBAGW9A7",1,"MIT","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fautonomous-ai-agents\u002Foracle\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: oracle\ndescription: \"Oracle second-model review: bundle prompts\u002Ffiles, debug, refactor, design-check.\"\n---\n\n# Oracle (CLI) — best use\n\nOracle bundles your prompt + selected files into one “one-shot” request so another model can answer with real repo context (API or browser automation). Treat outputs as advisory: verify against the codebase + tests.\n\n## Main use case (browser, GPT‑5.6 Sol Pro)\n\nDefault workflow here: `--engine browser` with GPT‑5.6 Sol at the explicit Pro effort tier in ChatGPT. This is the “human in the loop” path: it can take ~10 minutes to ~1 hour; expect a stored session you can reattach to.\n\nRecommended defaults:\n\n- Engine: browser (`--engine browser`)\n- Model: current browser Pro alias (`--model gpt-5-pro`), which Oracle 0.18 resolves to the current GPT‑5.6 Sol target. Prefer this alias for the default so the top Pro target follows supported model updates; use `gpt-5.6-sol` only when pinning that exact family is intentional.\n- Maximum browser effort: `--browser-thinking-time pro`. `pro` is deliberate and distinct from `heavy`; it selects the top Pro tier exposed by ChatGPT's unified Intelligence picker.\n- Model-picker strategy: `--browser-model-strategy select`. Maximum mode is fail-closed: a run must verify the requested model and effort instead of submitting silently at the UI's current setting.\n- Research mode: keep `--browser-research off` for normal runs; follow the source policy below.\n- Invocation: rely on the configured defaults or restate the current golden path exactly. Explicit CLI\u002FMCP arguments override config; use a different model or picker strategy only when the user explicitly requests that variant.\n- Parallelism: up to five concurrent browser tabs per shared profile (`--browser-max-concurrent-tabs 5`) for direct local browser runs. Oracle serializes the short profile\u002Fsend critical section.\n- Remote bridge limit: Oracle 0.18's `createRemoteServer` is intentionally single-flight and returns HTTP 409 `busy` for a second simultaneous remote run. Do not claim remote concurrency from `maxConcurrentTabs`; queue\u002Freattach remote work or run independent Oracle processes directly on the signed-in Mac. Do not remove the bridge guard in-process: `runBrowserMode` is not safely reentrant there.\n- Attachments: directories\u002Fglobs + excludes; avoid secrets. For many text files, add `--browser-bundle-files --browser-bundle-format zip`.\n\n## Source policy: standard web search stays available\n\nTreat standard ChatGPT web search and the dedicated Deep Research workflow as separate capabilities.\n\n- Normal mode (`browser.researchMode=off`) disables only the dedicated Deep Research workflow. It does not make the task closed-book.\n- In normal mode, use standard ChatGPT web search when current or external technical facts would improve the answer. Request citations and label web findings separately from attached files and live evidence.\n- Write the source instruction positively: “Use standard ChatGPT web search as needed. Treat attached files and verified live evidence as canonical for project-specific facts.”\n- Use bundle-only or no-web operation only when the user explicitly requests a closed-book review. Do not infer closed-book mode from “Deep Research off”, “one-shot”, “review”, or the presence of a complete bundle.\n- Activate `--browser-research deep` only when the user explicitly requests the dedicated multi-step Deep Research workflow.\n\n## Golden path (fast + reliable)\n\n1. Pick a tight file set (fewest files that still contain the truth).\n2. Preview what you’re about to send (`--dry-run` + `--files-report` when needed).\n3. Run in normal browser mode for GPT‑5.6 Sol Pro with standard web search available; activate the dedicated Deep Research workflow or API only when explicitly requested.\n4. If the run detaches\u002Ftimeouts: reattach to the stored session (don’t re-run).\n\n## Commands (preferred)\n\n- Show help (once\u002Fsession):\n  - `npx -y @steipete\u002Foracle --help`\n\n- Preview (no tokens):\n  - `npx -y @steipete\u002Foracle --dry-run summary -p \"\u003Ctask>\" --file \"src\u002F**\" --file \"!**\u002F*.test.*\"`\n  - `npx -y @steipete\u002Foracle --dry-run full -p \"\u003Ctask>\" --file \"src\u002F**\"`\n\n- Token\u002Fcost sanity:\n  - `npx -y @steipete\u002Foracle --dry-run summary --files-report -p \"\u003Ctask>\" --file \"src\u002F**\"`\n\n- Startup\u002Fperf trace:\n  - `npx -y @steipete\u002Foracle --perf-trace --perf-trace-path \u002Ftmp\u002Foracle-perf.json --dry-run summary -p \"\u003Ctask>\" --file \"src\u002F**\"`\n  - Use when CLI startup or time-to-first-output feels slow; inspect `first-output` and `exit`.\n\n- Browser run (main path; long-running is normal):\n  - `oracle --engine browser --model gpt-5-pro --browser-thinking-time pro --browser-model-strategy select --browser-research off --browser-max-concurrent-tabs 5 --browser-bundle-files --browser-bundle-format zip -p \"\u003Ctask>\" --file \"src\u002F**\"`\n  - **Robust default flag set:** explicit model\u002Feffort\u002Fresearch controls plus bundled ZIP attachments. Common failure modes and fixes:\n    1. `Thinking time: option not found`, `status=unavailable`, or `verified=no` → treat the run as unverified maximum; inspect the saved session\u002Flog and fix the signed-in account or current picker flow instead of falling back silently.\n    2. `Attachments did not finish uploading before timeout` (especially with a big\u002F`>50KB` file or 3+ files) → keep `--browser-bundle-files --browser-bundle-format zip` (one ZIP upload instead of N).\n    3. `Chrome window closed before oracle finished` → keep the automation browser\u002Fruntime available and reattach to the stored session instead of duplicating the run.\n  - Run in the background (`run_in_background`) since it takes ~10–60 min; you'll be notified on completion. On error, check the run log, fix the flag, re-run with a fresh `--slug` (a stuck session shows `error` in `oracle status`).\n\n- Manual paste fallback (assemble bundle, copy to clipboard):\n  - `npx -y @steipete\u002Foracle --render --copy -p \"\u003Ctask>\" --file \"src\u002F**\"`\n  - Note: `--copy` is a hidden alias for `--copy-markdown`.\n\n## Attaching files (`--file`)\n\n`--file` accepts files, directories, and globs. You can pass it multiple times; entries can be comma-separated.\n\n- Include:\n  - `--file \"src\u002F**\"` (directory glob)\n  - `--file src\u002Findex.ts` (literal file)\n  - `--file docs --file README.md` (literal directory + file)\n\n- Exclude (prefix with `!`):\n  - `--file \"src\u002F**\" --file \"!src\u002F**\u002F*.test.ts\" --file \"!**\u002F*.snap\"`\n\n- Defaults (important behavior from the implementation):\n  - Default-ignored dirs: `node_modules`, `dist`, `coverage`, `.git`, `.turbo`, `.next`, `build`, `tmp` (skipped unless you explicitly pass them as literal dirs\u002Ffiles).\n  - Honors `.gitignore` when expanding globs.\n  - Does not follow symlinks (glob expansion uses `followSymbolicLinks: false`).\n  - Dotfiles are filtered unless you explicitly opt in with a pattern that includes a dot-segment (e.g. `--file \".github\u002F**\"`).\n  - Default cap: files > 1 MB are rejected unless you raise `ORACLE_MAX_FILE_SIZE_BYTES` or `maxFileSizeBytes` in `~\u002F.oracle\u002Fconfig.json`.\n\n## Budget + observability\n\n- Target: keep total input under ~196k tokens.\n- Use `--files-report` (and\u002For `--dry-run json`) to spot the token hogs before spending.\n- Use `--perf-trace` \u002F `ORACLE_PERF_TRACE=1` for startup and first-output timing. Traces redact prompts, tokens, keys, cookies, and inline cookie payloads; detached API children write a session-suffixed sidecar trace.\n- If you need hidden\u002Fadvanced knobs: `npx -y @steipete\u002Foracle --help --verbose`.\n\n## Engines (API vs browser)\n\n- Auto-pick: uses `api` when `OPENAI_API_KEY` is set, otherwise `browser`.\n- Browser engine supports GPT + Gemini only; use `--engine api` for Claude\u002FGrok\u002FCodex or multi-model runs.\n- Direct local browser runs share a bounded five-tab lease. Remote-bridge runs remain single-flight in Oracle 0.18. More tabs are not automatically better: account throttling and UI instability remain the limiting factors.\n- **API runs require explicit user consent** before starting because they incur usage costs.\n- Browser attachments:\n  - `--browser-attachments auto|never|always` (auto pastes inline up to ~60k chars then uploads).\n  - Add `--browser-bundle-files --browser-bundle-format zip` to upload many text files as one ZIP while preserving file names.\n- Remote browser host (signed-in machine runs automation):\n  - Host: `oracle serve --host 0.0.0.0 --port 9473 --token \u003Csecret>`\n  - Client: `oracle --engine browser --remote-host \u003Chost:port> --remote-token \u003Csecret> -p \"\u003Ctask>\" --file \"src\u002F**\"`\n\n## API preflight\n\n- API runs require explicit user consent and cost money.\n- Before API runs, check provider readiness without printing secrets:\n  - `oracle doctor --providers --models gpt-5.4,claude-4.6-sonnet,gemini-3-pro`\n  - `oracle --preflight --models gpt-5.4,gemini-3-pro`\n  - `oracle --route --model gpt-5.4`\n- If the user wants first-party OpenAI, pass `--provider openai` or `--no-azure`. This prevents exported Azure env\u002Fconfig from hijacking the route:\n  - `oracle --provider openai --engine api --model gpt-5.5-pro ...`\n- For advisory multi-model panels where partial success is useful, use `--allow-partial --write-output \u003Cpath>` so successful model files and the `\u003Cstem>.oracle.json` manifest are easy to recover:\n  - `oracle --models gpt-5.4,claude-4.6-sonnet,gemini-3-pro --allow-partial --write-output \u002Ftmp\u002Fpanel.md -p \"\u003Ctask>\"`\n- `--timeout 10m` is the normal user-facing API deadline; Oracle derives the HTTP transport timeout unless `--http-timeout` is explicitly set.\n- If the exported `OPENAI_API_KEY` is invalid and the user wants their personal OpenAI key, use `$one-password` in one persistent tmux session. Known item: `API Key - OpenAI - Personal`, field `api_key`. Inject only into the single Oracle command; never print the key:\n  - `OPENAI_API_KEY=\"$(op item get 'API Key - OpenAI - Personal' --account my.1password.com --fields label=api_key --reveal)\" oracle --provider openai --engine api --model gpt-5.5-pro ...`\n- For debugging Oracle itself, prefer the local checkout after pulling `~\u002FProjects\u002Foracle`:\n  - `pnpm -C ~\u002FProjects\u002Foracle run build`\n  - `node ~\u002FProjects\u002Foracle\u002Fdist\u002Fscripts\u002Frun-cli.js ...`\n\n## Sessions + slugs (don’t lose work)\n\n- Stored under `~\u002F.oracle\u002Fsessions` (override with `ORACLE_HOME_DIR`).\n- Browser runs save durable files under `~\u002F.oracle\u002Fsessions\u002F\u003Cid>\u002Fartifacts\u002F`, including `transcript.md`, Deep Research reports, and downloaded ChatGPT-generated images when available.\n- Runs may detach or take a long time (browser\u002FAPI + GPT‑5.5 Pro often does). If the CLI times out: don’t re-run; reattach.\n  - List: `oracle status --hours 72`\n  - Attach: `oracle session \u003Cid> --render`\n- Use `--slug \"\u003C3-5 words>\"` to keep session IDs readable.\n- Duplicate prompt guard exists; use `--force` only when you truly want a fresh run.\n- CLI guardrails: root runs without a prompt exit nonzero; `--dry-run` conflicts with `--render` \u002F `--render-markdown`; Ctrl-C exits foreground API runs with code 130 while browser cleanup\u002Freattach still runs.\n\n## Prompt template (high signal)\n\nOracle starts with **zero** project knowledge. Assume the model cannot infer your stack, build tooling, conventions, or “obvious” paths. Include:\n\n- Project briefing (stack + build\u002Ftest commands + platform constraints).\n- “Where things live” (key directories, entrypoints, config files, dependency boundaries).\n- Exact question + what you tried + the error text (verbatim).\n- Constraints (“don’t change X”, “must keep public API”, “perf budget”, etc).\n- Source policy: preserve standard web search in normal mode and distinguish web findings from canonical project\u002Flive evidence; choose closed-book only on the user's explicit instruction.\n- Desired output (“return patch plan + tests”, “list risky assumptions”, “give 3 options with tradeoffs”).\n\n### “Exhaustive prompt” pattern (for later restoration)\n\nWhen you know this will be a long investigation, write a prompt that can stand alone later:\n\n- Top: 6–30 sentence project briefing + current goal.\n- Middle: concrete repro steps + exact errors + what you already tried.\n- Bottom: attach _all_ context files needed so a fresh model can fully understand (entrypoints, configs, key modules, docs).\n\nIf you need to reproduce the same context later, re-run with the same prompt + `--file …` set (Oracle runs are one-shot; the model doesn’t remember prior runs).\n\n## Safety\n\n- Don’t attach secrets by default (`.env`, key files, auth tokens). Redact aggressively; share only what’s required.\n- Prefer “just enough context”: fewer files + better prompt beats whole-repo dumps.\n","# Oracle (CLI) — best use\n\nOracle bundles your prompt + selected files into one “one-shot” request so another model can answer with real repo context (API or browser automation). Treat outputs as advisory: verify against the codebase + tests.\n\n## Main use case (browser, GPT‑5.6 Sol Pro)\n\nDefault workflow here: `--engine browser` with GPT‑5.6 Sol at the explicit Pro effort tier in ChatGPT. This is the “human in the loop” path: it can take ~10 minutes to ~1 hour; expect a stored session you can reattach to.\n\nRecommended defaults:\n\n- Engine: browser (`--engine browser`)\n- Model: current browser Pro alias (`--model gpt-5-pro`), which Oracle 0.18 resolves to the current GPT‑5.6 Sol target. Prefer this alias for the default so the top Pro target follows supported model updates; use `gpt-5.6-sol` only when pinning that exact family is intentional.\n- Maximum browser effort: `--browser-thinking-time pro`. `pro` is deliberate and distinct from `heavy`; it selects the top Pro tier exposed by ChatGPT's unified Intelligence picker.\n- Model-picker strategy: `--browser-model-strategy select`. Maximum mode is fail-closed: a run must verify the requested model and effort instead of submitting silently at the UI's current setting.\n- Research mode: keep `--browser-research off` for normal runs; follow the source policy below.\n- Invocation: rely on the configured defaults or restate the current golden path exactly. Explicit CLI\u002FMCP arguments override config; use a different model or picker strategy only when the user explicitly requests that variant.\n- Parallelism: up to five concurrent browser tabs per shared profile (`--browser-max-concurrent-tabs 5`) for direct local browser runs. Oracle serializes the short profile\u002Fsend critical section.\n- Remote bridge limit: Oracle 0.18's `createRemoteServer` is intentionally single-flight and returns HTTP 409 `busy` for a second simultaneous remote run. Do not claim remote concurrency from `maxConcurrentTabs`; queue\u002Freattach remote work or run independent Oracle processes directly on the signed-in Mac. Do not remove the bridge guard in-process: `runBrowserMode` is not safely reentrant there.\n- Attachments: directories\u002Fglobs + excludes; avoid secrets. For many text files, add `--browser-bundle-files --browser-bundle-format zip`.\n\n## Source policy: standard web search stays available\n\nTreat standard ChatGPT web search and the dedicated Deep Research workflow as separate capabilities.\n\n- Normal mode (`browser.researchMode=off`) disables only the dedicated Deep Research workflow. It does not make the task closed-book.\n- In normal mode, use standard ChatGPT web search when current or external technical facts would improve the answer. Request citations and label web findings separately from attached files and live evidence.\n- Write the source instruction positively: “Use standard ChatGPT web search as needed. Treat attached files and verified live evidence as canonical for project-specific facts.”\n- Use bundle-only or no-web operation only when the user explicitly requests a closed-book review. Do not infer closed-book mode from “Deep Research off”, “one-shot”, “review”, or the presence of a complete bundle.\n- Activate `--browser-research deep` only when the user explicitly requests the dedicated multi-step Deep Research workflow.\n\n## Golden path (fast + reliable)\n\n1. Pick a tight file set (fewest files that still contain the truth).\n2. Preview what you’re about to send (`--dry-run` + `--files-report` when needed).\n3. Run in normal browser mode for GPT‑5.6 Sol Pro with standard web search available; activate the dedicated Deep Research workflow or API only when explicitly requested.\n4. If the run detaches\u002Ftimeouts: reattach to the stored session (don’t re-run).\n\n## Commands (preferred)\n\n- Show help (once\u002Fsession):\n  - `npx -y @steipete\u002Foracle --help`\n\n- Preview (no tokens):\n  - `npx -y @steipete\u002Foracle --dry-run summary -p \"\u003Ctask>\" --file \"src\u002F**\" --file \"!**\u002F*.test.*\"`\n  - `npx -y @steipete\u002Foracle --dry-run full -p \"\u003Ctask>\" --file \"src\u002F**\"`\n\n- Token\u002Fcost sanity:\n  - `npx -y @steipete\u002Foracle --dry-run summary --files-report -p \"\u003Ctask>\" --file \"src\u002F**\"`\n\n- Startup\u002Fperf trace:\n  - `npx -y @steipete\u002Foracle --perf-trace --perf-trace-path \u002Ftmp\u002Foracle-perf.json --dry-run summary -p \"\u003Ctask>\" --file \"src\u002F**\"`\n  - Use when CLI startup or time-to-first-output feels slow; inspect `first-output` and `exit`.\n\n- Browser run (main path; long-running is normal):\n  - `oracle --engine browser --model gpt-5-pro --browser-thinking-time pro --browser-model-strategy select --browser-research off --browser-max-concurrent-tabs 5 --browser-bundle-files --browser-bundle-format zip -p \"\u003Ctask>\" --file \"src\u002F**\"`\n  - **Robust default flag set:** explicit model\u002Feffort\u002Fresearch controls plus bundled ZIP attachments. Common failure modes and fixes:\n    1. `Thinking time: option not found`, `status=unavailable`, or `verified=no` → treat the run as unverified maximum; inspect the saved session\u002Flog and fix the signed-in account or current picker flow instead of falling back silently.\n    2. `Attachments did not finish uploading before timeout` (especially with a big\u002F`>50KB` file or 3+ files) → keep `--browser-bundle-files --browser-bundle-format zip` (one ZIP upload instead of N).\n    3. `Chrome window closed before oracle finished` → keep the automation browser\u002Fruntime available and reattach to the stored session instead of duplicating the run.\n  - Run in the background (`run_in_background`) since it takes ~10–60 min; you'll be notified on completion. On error, check the run log, fix the flag, re-run with a fresh `--slug` (a stuck session shows `error` in `oracle status`).\n\n- Manual paste fallback (assemble bundle, copy to clipboard):\n  - `npx -y @steipete\u002Foracle --render --copy -p \"\u003Ctask>\" --file \"src\u002F**\"`\n  - Note: `--copy` is a hidden alias for `--copy-markdown`.\n\n## Attaching files (`--file`)\n\n`--file` accepts files, directories, and globs. You can pass it multiple times; entries can be comma-separated.\n\n- Include:\n  - `--file \"src\u002F**\"` (directory glob)\n  - `--file src\u002Findex.ts` (literal file)\n  - `--file docs --file README.md` (literal directory + file)\n\n- Exclude (prefix with `!`):\n  - `--file \"src\u002F**\" --file \"!src\u002F**\u002F*.test.ts\" --file \"!**\u002F*.snap\"`\n\n- Defaults (important behavior from the implementation):\n  - Default-ignored dirs: `node_modules`, `dist`, `coverage`, `.git`, `.turbo`, `.next`, `build`, `tmp` (skipped unless you explicitly pass them as literal dirs\u002Ffiles).\n  - Honors `.gitignore` when expanding globs.\n  - Does not follow symlinks (glob expansion uses `followSymbolicLinks: false`).\n  - Dotfiles are filtered unless you explicitly opt in with a pattern that includes a dot-segment (e.g. `--file \".github\u002F**\"`).\n  - Default cap: files > 1 MB are rejected unless you raise `ORACLE_MAX_FILE_SIZE_BYTES` or `maxFileSizeBytes` in `~\u002F.oracle\u002Fconfig.json`.\n\n## Budget + observability\n\n- Target: keep total input under ~196k tokens.\n- Use `--files-report` (and\u002For `--dry-run json`) to spot the token hogs before spending.\n- Use `--perf-trace` \u002F `ORACLE_PERF_TRACE=1` for startup and first-output timing. Traces redact prompts, tokens, keys, cookies, and inline cookie payloads; detached API children write a session-suffixed sidecar trace.\n- If you need hidden\u002Fadvanced knobs: `npx -y @steipete\u002Foracle --help --verbose`.\n\n## Engines (API vs browser)\n\n- Auto-pick: uses `api` when `OPENAI_API_KEY` is set, otherwise `browser`.\n- Browser engine supports GPT + Gemini only; use `--engine api` for Claude\u002FGrok\u002FCodex or multi-model runs.\n- Direct local browser runs share a bounded five-tab lease. Remote-bridge runs remain single-flight in Oracle 0.18. More tabs are not automatically better: account throttling and UI instability remain the limiting factors.\n- **API runs require explicit user consent** before starting because they incur usage costs.\n- Browser attachments:\n  - `--browser-attachments auto|never|always` (auto pastes inline up to ~60k chars then uploads).\n  - Add `--browser-bundle-files --browser-bundle-format zip` to upload many text files as one ZIP while preserving file names.\n- Remote browser host (signed-in machine runs automation):\n  - Host: `oracle serve --host 0.0.0.0 --port 9473 --token \u003Csecret>`\n  - Client: `oracle --engine browser --remote-host \u003Chost:port> --remote-token \u003Csecret> -p \"\u003Ctask>\" --file \"src\u002F**\"`\n\n## API preflight\n\n- API runs require explicit user consent and cost money.\n- Before API runs, check provider readiness without printing secrets:\n  - `oracle doctor --providers --models gpt-5.4,claude-4.6-sonnet,gemini-3-pro`\n  - `oracle --preflight --models gpt-5.4,gemini-3-pro`\n  - `oracle --route --model gpt-5.4`\n- If the user wants first-party OpenAI, pass `--provider openai` or `--no-azure`. This prevents exported Azure env\u002Fconfig from hijacking the route:\n  - `oracle --provider openai --engine api --model gpt-5.5-pro ...`\n- For advisory multi-model panels where partial success is useful, use `--allow-partial --write-output \u003Cpath>` so successful model files and the `\u003Cstem>.oracle.json` manifest are easy to recover:\n  - `oracle --models gpt-5.4,claude-4.6-sonnet,gemini-3-pro --allow-partial --write-output \u002Ftmp\u002Fpanel.md -p \"\u003Ctask>\"`\n- `--timeout 10m` is the normal user-facing API deadline; Oracle derives the HTTP transport timeout unless `--http-timeout` is explicitly set.\n- If the exported `OPENAI_API_KEY` is invalid and the user wants their personal OpenAI key, use `$one-password` in one persistent tmux session. Known item: `API Key - OpenAI - Personal`, field `api_key`. Inject only into the single Oracle command; never print the key:\n  - `OPENAI_API_KEY=\"$(op item get 'API Key - OpenAI - Personal' --account my.1password.com --fields label=api_key --reveal)\" oracle --provider openai --engine api --model gpt-5.5-pro ...`\n- For debugging Oracle itself, prefer the local checkout after pulling `~\u002FProjects\u002Foracle`:\n  - `pnpm -C ~\u002FProjects\u002Foracle run build`\n  - `node ~\u002FProjects\u002Foracle\u002Fdist\u002Fscripts\u002Frun-cli.js ...`\n\n## Sessions + slugs (don’t lose work)\n\n- Stored under `~\u002F.oracle\u002Fsessions` (override with `ORACLE_HOME_DIR`).\n- Browser runs save durable files under `~\u002F.oracle\u002Fsessions\u002F\u003Cid>\u002Fartifacts\u002F`, including `transcript.md`, Deep Research reports, and downloaded ChatGPT-generated images when available.\n- Runs may detach or take a long time (browser\u002FAPI + GPT‑5.5 Pro often does). If the CLI times out: don’t re-run; reattach.\n  - List: `oracle status --hours 72`\n  - Attach: `oracle session \u003Cid> --render`\n- Use `--slug \"\u003C3-5 words>\"` to keep session IDs readable.\n- Duplicate prompt guard exists; use `--force` only when you truly want a fresh run.\n- CLI guardrails: root runs without a prompt exit nonzero; `--dry-run` conflicts with `--render` \u002F `--render-markdown`; Ctrl-C exits foreground API runs with code 130 while browser cleanup\u002Freattach still runs.\n\n## Prompt template (high signal)\n\nOracle starts with **zero** project knowledge. Assume the model cannot infer your stack, build tooling, conventions, or “obvious” paths. Include:\n\n- Project briefing (stack + build\u002Ftest commands + platform constraints).\n- “Where things live” (key directories, entrypoints, config files, dependency boundaries).\n- Exact question + what you tried + the error text (verbatim).\n- Constraints (“don’t change X”, “must keep public API”, “perf budget”, etc).\n- Source policy: preserve standard web search in normal mode and distinguish web findings from canonical project\u002Flive evidence; choose closed-book only on the user's explicit instruction.\n- Desired output (“return patch plan + tests”, “list risky assumptions”, “give 3 options with tradeoffs”).\n\n### “Exhaustive prompt” pattern (for later restoration)\n\nWhen you know this will be a long investigation, write a prompt that can stand alone later:\n\n- Top: 6–30 sentence project briefing + current goal.\n- Middle: concrete repro steps + exact errors + what you already tried.\n- Bottom: attach _all_ context files needed so a fresh model can fully understand (entrypoints, configs, key modules, docs).\n\nIf you need to reproduce the same context later, re-run with the same prompt + `--file …` set (Oracle runs are one-shot; the model doesn’t remember prior runs).\n\n## Safety\n\n- Don’t attach secrets by default (`.env`, key files, auth tokens). Redact aggressively; share only what’s required.\n- Prefer “just enough context”: fewer files + better prompt beats whole-repo dumps.\n",{"name":5,"description":7},"main","116bb09f1369e85cd8a54fec4135e218dca4a061","skills\u002Fautonomous-ai-agents\u002Foracle",0,"https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Freleases\u002Flatest","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard","ai","2026-08-25 21:36:43",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add merllinsbeard\u002Fmerlin-CTO --skill oracle",{"tool":39,"command":40},"bunx","bunx skills add merllinsbeard\u002Fmerlin-CTO --skill oracle",{"tool":42,"command":43},"pnpm","pnpm dlx skills add merllinsbeard\u002Fmerlin-CTO --skill oracle","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fautonomous-ai-agents\u002Foracle\u002FSKILL.md",true,[],[48,55,62,69,76,83],{"skillNo":49,"name":50,"nameZh":6,"description":51,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":52,"categories":53,"keywords":54,"publishedAt":19},"01M7B3SK15WA8JW4G6RBKZR3ME","writing-for-agents","Use for delegate_task, Kanban, and all agent-facing text.","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fwriting-for-agents\u002FSKILL.md",[],[],{"skillNo":56,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":59,"categories":60,"keywords":61,"publishedAt":19},"01M7B3SD78DTTXFVXGDPRK7249","why","Use for 'why does X work this way', 'why we picked Y', design rationale, regressions, postmortems, or data-backed thresholds. Discovers available MCPs and queries each evidence category (source control, issue tracker, long-form docs, real-time chat, infrastructure observability, error tracking, product analytics warehouse) in parallel, then returns a cited read on decisions and tradeoffs. Use how for runtime behavior.","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fwhy\u002FSKILL.md",[],[],{"skillNo":63,"name":64,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":19},"01M7B3S8ENS1XQ7YV3N5GV1PCF","vue-best-practices","Use for Vue 3, .vue files, Vue Router, or Pinia. Skip when the repo has no Vue.","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fweb-development\u002Fvue-best-practices\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":19},"01M7B3S2DPXMQN0EFFE8FTF03R","vercel-react-best-practices","React and Next.js performance optimization guidelines from Vercel Engineering. This skill should be used when writing, reviewing, or refactoring React\u002FNext.js code to ensure optimal performance patterns. Triggers on tasks involving React components, Next.js pages, data fetching, bundle optimization, or performance improvements.","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fweb-development\u002Fvercel-react-best-practices\u002FSKILL.md",[],[],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":19},"01M7B3S06K4AGG5X59TCX900R6","tailwind-v4","Use when writing or reviewing Tailwind CSS v4 classes or migrating v3 syntax.","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fweb-development\u002Ftailwind-v4\u002FSKILL.md",[],[],{"skillNo":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":19},"01M7B3RXNDDCPJ7K5303BXXDTV","react-best-practices","Mastra React subset only. Generic React or Next uses vercel-react-best-practices.","https:\u002F\u002Fgithub.com\u002Fmerllinsbeard\u002Fmerlin-CTO\u002Fblob\u002Fmain\u002Fskills\u002Fweb-development\u002Freact-best-practices\u002FSKILL.md",[],[]]