[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M74KBQAVFPZV2DQV7C975PQC":3,"skill-same:01M74KBQAVFPZV2DQV7C975PQC":49},{"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":27,"commitSha":28,"skillDirectory":29,"forksCount":30,"repositoryHtmlUrl":31,"repositoryHomepageUrl":32,"ownerHtmlUrl":33,"ownerBlogUrl":6,"ownerDescription":34,"updatedAt":35,"install":36,"resources":48},"01M74KBQAVFPZV2DQV7C975PQC","start",null,"Intelligent orchestrator for ANY engineering task — feature, bug, infra, architecture, refactor, research. Accepts a free-form goal, clarifies ambiguities via conditional Q&A, infers depth mode from risk signals, and routes to the optimal core skill or orchestrates ad-hoc orchestration. Replaces \u002Fship as the universal front-door.","Acauhi99\u002Fopencode-agent-system","Acauhi99","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F109399110?v=4","01M74JTV4VH2YAER6TABE78RKF","Acauhi","01M74JTV5CXNC5DEV2XR5SBFG3",1,"MIT","https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Fstart\u002FSKILL.md",[],[],"2026-11-11 15:11:14",false,"---\nname: start\ndescription: Intelligent orchestrator for ANY engineering task — feature, bug, infra, architecture, refactor, research. Accepts a free-form goal, clarifies ambiguities via conditional Q&A, infers depth mode from risk signals, and routes to the optimal core skill or orchestrates ad-hoc orchestration. Replaces \u002Fship as the universal front-door.\nlicense: MIT\ncompatibility: opencode\nmetadata:\n  requires_codegraph: true\n---\n\n> **Tool mapping (OpenCode):**\n> - `question` tool for user prompts (not `AskUserQuestion`)\n> - `task` tool for subagents (not `Agent` or `Task` subagent syntax)\n> - `skill({ name: \"...\" })` to load other skills (not `Skill(skill=\"...\", args=\"...\")`)\n> - `codegraph_explore`, `codegraph_search`, `codegraph_context` for codebase exploration (preferred)\n> - `todowrite` for phase\u002Fprogress tracking\n> - `grep`, `glob`, `read`, `bash` as fallback when `.codegraph\u002F` not initialized\n>\n> **Context rule:** This skill supersedes any prior skill instructions. Follow ONLY these instructions now. Read the user's goal and any mode\u002Finclude\u002Fskip from the conversation context.\n\nRead `mode=`, `include=`, and `skip=` from the conversation context. Default to the mode specified in this skill if not provided.\n\n# start\n\nThe user gave you an engineering goal. Any goal. Clarify it when needed, pick the right depth, route to the best toolchain, run it, report. Stop before git.\n\nThe tax on a vibe coder is picking *which* workflow to invoke and *how* thorough to be. This skill removes that tax while keeping full visibility of what was decided.\n\n---\n\n## Step 0 — Clarify (conditional Q&A)\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nAnalyze the user's request for:\n\n1. **Ambiguity** — does the goal have multiple interpretations?\n2. **Missing inputs** — is key info absent (which system, what tech, expected behavior)?\n3. **Decision-tree branches** — does the problem have unresolved forks (e.g., \"fix auth\" → fix login, fix signup, fix token refresh, fix permissions)?\n\n**If clear** → skip this step entirely. Proceed to Step 1.\n\n**If ambiguous** → ask questions to resolve. Batch related questions in a single `question` call when possible. Only drill one-at-a-time (grill-me style) for genuinely complex decision trees where answers change follow-up questions.\n\n**No question limit.** But don't over-ask — one round of batched questions resolves 90% of cases.\n\nExamples of when to ask:\n- \"Add a payment feature\" → \"Which payment provider? Stripe, Paddle, or custom?\"\n- \"Fix the dashboard\" → \"What's broken — data not loading, wrong calculations, layout issues, or something else?\"\n- \"Build infra\" → \"Self-hosted or cloud? Which provider? What's the stack?\"\n\nExamples of when to skip:\n- \"Add a dark mode toggle\" → clear, specific, self-contained\n- \"Fix the 500 error on \u002Fcheckout when clicking 'Pay Now'\" → clear + has reproduction\n\n---\n\n## Step 1 — Analyze & Route\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nRead the user's prompt and classify the request into one of these categories:\n\n| Task type | Intent | Routes to |\n|---|---|---|---|\n| New project or service from scratch | INIT | `init-project` |\n| New feature, screen, component, endpoint, module in an existing codebase | CREATE | `add-feature` |\n| Update, extend, change behavior, small cosmetic tweak on an existing feature | EVOLVE | `modify-feature` |\n| Apply UX polish checklist without behavior change | POLISH | `polish-ui` |\n| Remove, delete, deprecate, rip out a feature | REMOVE | `remove-feature` |\n| Integrate with an external API — typed client, retry, observability | INTEGRATE | `integrate-api` |\n| Bug, broken, not working, silent failure, regression | FIX | `fix-bug` |\n| Codebase audit, tech-debt sweep, deep clean, production-readiness | AUDIT | `audit` |\n| Domain model rename, enum migration, status\u002Fvocabulary change with persisted data | REALIGN | `realign` |\n| Architecture design, system design, thinking through structure | DESIGN | `general` subagent(s) + codegraph exploration |\n| Infrastructure setup, debugging, or review | INFRA | `general` subagent(s) + relevant tooling |\n| Codebase research, explain a system, understand a flow | RESEARCH | `explore` subagent(s) + codegraph |\n| Anything else \u002F multi-faceted \u002F doesn't fit above | GENERAL | Orchestrate ad-hoc with subagents + available tooling |\n\n**Routing rules:**\n\n- If the request maps to a core skill (INIT \u002F CREATE \u002F EVOLVE \u002F POLISH \u002F REMOVE \u002F INTEGRATE \u002F FIX \u002F AUDIT \u002F REALIGN), route there with mode= and any include=\u002Fskip= overrides.\n- If the request maps to DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL, do NOT try to force-fit into a core skill. Instead:\n  1. Use `task` with appropriate subagent types to do the work\n  2. Use codegraph_explore for understanding\n  3. Use the full MCP + tool arsenal\n  4. If a sub-skill exists that partially matches (add-migration, add-observability, etc.), invoke it as part of the orchestration\n- **Multi-intent** — if the user lists clearly separable goals (\"add X and remove Y\"), execute sequentially as separate \u002Fstart routings. State the order before starting.\n\n**Never reject a request that doesn't fit the first 9 categories (INIT, CREATE, EVOLVE, POLISH, REMOVE, INTEGRATE, FIX, AUDIT, REALIGN).** DESIGN, INFRA, RESEARCH, and GENERAL cover everything else. If none of those fit either, the problem is so unusual that ad-hoc orchestration is the right answer — still don't reject.\n\n---\n\n## Step 2 — Infer the depth mode\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nThree base modes — `fast`, `balanced`, `production`. Pick one before announcing.\n\n**Plan-only modifier:** `plan-only` can be appended to any base mode (e.g. `mode=production+plan-only` or just `mode=plan-only`). When present, the skill stops after Step 3 (announce plan) — no execution, no changes. The plan is saved to agentmemory for later `\u002Fcontinue` in Build mode. Use this when the user is in Plan mode (read-only) and wants to review before executing.\n\n**Risk signals (any one → `production`)**:\n- Touches auth, permissions, payments, billing, secrets, or external webhooks\n- Schema migration or persisted-data rewrite\n- Destructive deletion or deprecation of an external\u002Fpublic contract\n- Background jobs, queues, cron, retries, email\u002FSMS\u002Fpush, imports\u002Fexports, file writes, spawned processes, IPC, or external APIs\n- Caching, query invalidation, feature flags, analytics\u002Fbusiness reporting, or concurrency-sensitive mutations\n- Multi-subsystem in the same change (frontend + backend + DB)\n\n**Tiny-scope signals (all four → `fast`)**:\n- Single file\n- Cosmetic \u002F copy \u002F styling only\n- No data layer touched\n- No new public API surface\n\n**Default → `balanced`**.\n\n**Override:** explicit `mode=fast`, `mode=balanced`, or `mode=production` in the user's prompt always wins — except when it conflicts with a high-risk signal (see NEVER below).\n\nFor DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL tasks, mode means:\n- `fast` — minimal analysis, direct answer\n- `balanced` — moderate depth with subagent exploration\n- `production` — full multi-subagent analysis with cross-validation\n\n---\n\n## Step 3 — Announce the plan\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nOutput a structured plan block before executing. Format:\n\n```\nDetected: \u003CCREATE | EVOLVE | POLISH | REMOVE | FIX | AUDIT | REALIGN | DESIGN | INFRA | RESEARCH | GENERAL>\nRisk:     \u003Clow | medium | high — one-line reason>\nMode:     \u003Cfast | balanced | production | plan-only — one-line reason or \"user-specified\">\nPipeline:\n  1. \u003Cphase>  → \u003Ccore skill, sub-skill, or subagent>\n  2. \u003Cphase>  → \u003Ccore skill, sub-skill, or subagent>\n  ...\n```\n\nPipeline numbering must match what the routed core skill or ad-hoc plan will actually run at the chosen mode. Do not invent phases the routed skill won't execute.\n\n**Confirmation gating depends on mode:**\n\n| Mode | Gating |\n|---|---|\n| `production` | Use `question` \"Proceed with this pipeline?\" before Step 4. Decline → stop. |\n| `balanced` | Print the plan inline, then proceed to Step 4 in the same turn. User can abort with ESC or a new prompt before execution begins. |\n| `fast` | Print the plan inline as a one-line preamble, then execute immediately. No confirm prompt. |\n| `* + plan-only` | Print the plan. Skip Step 4. Proceed to Step 5 with save-to-memory. |\n\n---\n\n## Step 4 — Execute\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\n**Plan-only gate:** If mode contains `plan-only`, skip execution entirely. Save the plan to agentmemory via `agentmemory_memory_save` with type=\"plan\" and concepts including the detected intent and file paths. Then proceed to Step 5 without calling any core skill.\n\n**If routing to a core skill:** Use `skill({ name: \"\u003Ccore-skill>\" })`. The skill reads the user's goal and mode\u002Finclude\u002Fskip from the conversation context — no explicit args needed.\n\n**If routing to DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL:** Orchestrate directly:\n1. Launch appropriate `task` subagents for analysis\n2. Use `codegraph_explore` \u002F `codegraph_search` for understanding\n3. Invoke any matching sub-skills (add-migration, add-observability, etc.) when their gates trigger\n4. Synthesize and present results\n\n**Host-portability fallback.** If `skill` is not available:\n1. Read the routed skill's SKILL.md directly.\n2. Execute its instructions inline.\n3. Surface the degradation in the Step 3 announcement: add a line like `Routing: inline (skill tool unavailable)`.\n\n**Respect downstream gates.** Core skills have their own approval gates (e.g., `add-feature mode=production`'s Plan-approval). Let them fire. Don't bypass them from \u002Fstart.\n\n**Adjunct skills live downstream.** \u002Fstart only chooses the top-level workflow. The routed skill owns stack\u002Fplugin-specific handoffs.\n\n---\n\n## Step 5 — Report and hand off to git\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nOutput a visible pipeline summary:\n\n```\n✔ \u003Cphase 1>  — \u003Cone-line outcome>\n✔ \u003Cphase 2>  — \u003Cone-line outcome>\n✔ \u003Cphase 3>  — \u003Cone-line outcome>\n...\n\nFindings:\n  - \u003Ceach finding surfaced by the routed skill or audits>\n```\n\nIf execution was skipped (plan-only mode), append:\n```\nPlan saved to agentmemory. Switch to Build mode and run \u002Fcontinue to execute.\n```\n\nIf execution completed:\n```\nCode is production-ready. To publish:\n  • \u002Fcommit           — group the working tree into staged commits without pushing\n  • \u002Fcommit-and-push  — commit then push the current branch\n  • \u002Fopen-pr          — open a GitHub PR\n  • \u002Fcontinue         — continue this work if something still needs fixing\n```\n\nSurface findings, not just \"done.\" If a sub-skill audit returned issues that weren't auto-fixed, name them here.\n\n**Do not commit. Do not push.** The user picks the publish path.\n\n---\n\n## NEVER\n\n- **NEVER commit, push, or open PRs from inside this skill**\n  **Instead:** Stop at Step 5 and direct the user to `\u002Fcommit`, `\u002Fcommit-and-push`, `\u002Fopen-pr`, or `\u002Fcontinue`.\n\n- **NEVER bypass a routed core skill's own approval gate**\n  **Instead:** Let the gate fire. The user interacts with the core skill's gate, not with \u002Fstart.\n\n- **NEVER replicate the core skill's pipeline inline from memory**\n  **Instead:** Always delegate via `skill(...)`. When unavailable, read the SKILL.md and follow it verbatim.\n\n- **NEVER hide which mode and pipeline you picked**\n  **Instead:** Announce in every mode. Even `fast` prints the one-line preamble.\n\n- **NEVER guess between two plausible intents without asking**\n  **Instead:** Use Step 0's Q&A to resolve ambiguities before routing.\n\n- **NEVER honor a `mode=fast` override on a high-risk change without surfacing the conflict**\n  **Instead:** Pause and ask: \"Detected high-risk signals (e.g., payments). You requested fast mode. Confirm fast anyway, or upgrade to production?\"\n\n- **NEVER reject a request that doesn't fit the first 9 categories (INIT, CREATE, EVOLVE, POLISH, REMOVE, INTEGRATE, FIX, AUDIT, REALIGN)**\n  **Instead:** Route to DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL or orchestrate ad-hoc. There is always a path forward.\n\n- **NEVER execute in plan-only mode**\n  **Instead:** Stop at Step 3, save the plan, report. The user will return in Build mode via \u002Fcontinue.\n  **Why:** Plan mode is read-only. Executing code changes violates the user's mode contract.\n\n---\n\n## Appendix — Adjunct skills and sub-skills\n\n\u002Fstart routes to core skills. Those core skills own the following adjunct invocations (gate-driven, not automatic):\n\n### CREATE → `add-feature` may invoke\n\n- **UI scaffolding:** `add-empty-error-states`, `polish-ui`, `propagate-ui-pattern`\n- **Backend scaffolding:** `add-migration`, `add-observability`, `reviewer-authz`\n- **Tests:** `write-tests`, `add-e2e-test`\n- **Audits (Phase 7):** `reviewer-contracts`, `reviewer-concurrency`, `reviewer-data-integrity`, `reviewer-security-regression`, `reviewer-error-boundaries`, `reviewer-loading-states`, `reviewer-accessibility-regression`, `reviewer-client-bundle`, `reviewer-observability-coverage`, `reviewer-perf`, `audit-responsive`\n- **Cleanup:** `simplify`, `polish-ui`\n\n### EVOLVE → `modify-feature` may invoke\n\n- **UI extensions:** `add-empty-error-states`, `polish-ui`\n- **Backend extensions:** `add-migration`, `add-observability`, `reviewer-authz`\n- **Tests:** `write-tests`, `add-e2e-test`\n- **Audits:** `reviewer-contracts`, `reviewer-concurrency`, `reviewer-observability-coverage`, `reviewer-data-integrity`, `reviewer-security-regression`, `reviewer-error-boundaries`, `reviewer-loading-states`, `reviewer-accessibility-regression`, `reviewer-client-bundle`\n- **Cleanup:** `simplify`, `polish-ui`\n\n### POLISH → `polish-ui`\n\nRuns UX checklist; no fan-out.\n\n### FIX → `fix-bug` may invoke\n\n- `add-regression-test`\n- `polish-ui` (if UI changed)\n\n### REMOVE → `remove-feature` may invoke\n\n- `add-migration` (if schema change)\n- `reviewer-data-integrity`, `reviewer-contracts`\n\n### AUDIT → `audit` may invoke\n\nFull reviewer-* family across the repo, plus `audit-responsive`, `audit-seo-meta`, `audit-analytics`, `simplify`, `harden-types`.\n\n### REALIGN → `realign` may invoke\n\n- `add-migration` (if persisted data needs migration)\n- `reviewer-data-integrity` (post-step audit)\n\n### DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL\n\nThese are ad-hoc orchestration paths. They may invoke any of the above sub-skills on demand, plus direct subagent work, codegraph exploration, web research, and CLI tooling. No fixed adjunct list.\n","> **Tool mapping (OpenCode):**\n> - `question` tool for user prompts (not `AskUserQuestion`)\n> - `task` tool for subagents (not `Agent` or `Task` subagent syntax)\n> - `skill({ name: \"...\" })` to load other skills (not `Skill(skill=\"...\", args=\"...\")`)\n> - `codegraph_explore`, `codegraph_search`, `codegraph_context` for codebase exploration (preferred)\n> - `todowrite` for phase\u002Fprogress tracking\n> - `grep`, `glob`, `read`, `bash` as fallback when `.codegraph\u002F` not initialized\n>\n> **Context rule:** This skill supersedes any prior skill instructions. Follow ONLY these instructions now. Read the user's goal and any mode\u002Finclude\u002Fskip from the conversation context.\n\nRead `mode=`, `include=`, and `skip=` from the conversation context. Default to the mode specified in this skill if not provided.\n\n# start\n\nThe user gave you an engineering goal. Any goal. Clarify it when needed, pick the right depth, route to the best toolchain, run it, report. Stop before git.\n\nThe tax on a vibe coder is picking *which* workflow to invoke and *how* thorough to be. This skill removes that tax while keeping full visibility of what was decided.\n\n---\n\n## Step 0 — Clarify (conditional Q&A)\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nAnalyze the user's request for:\n\n1. **Ambiguity** — does the goal have multiple interpretations?\n2. **Missing inputs** — is key info absent (which system, what tech, expected behavior)?\n3. **Decision-tree branches** — does the problem have unresolved forks (e.g., \"fix auth\" → fix login, fix signup, fix token refresh, fix permissions)?\n\n**If clear** → skip this step entirely. Proceed to Step 1.\n\n**If ambiguous** → ask questions to resolve. Batch related questions in a single `question` call when possible. Only drill one-at-a-time (grill-me style) for genuinely complex decision trees where answers change follow-up questions.\n\n**No question limit.** But don't over-ask — one round of batched questions resolves 90% of cases.\n\nExamples of when to ask:\n- \"Add a payment feature\" → \"Which payment provider? Stripe, Paddle, or custom?\"\n- \"Fix the dashboard\" → \"What's broken — data not loading, wrong calculations, layout issues, or something else?\"\n- \"Build infra\" → \"Self-hosted or cloud? Which provider? What's the stack?\"\n\nExamples of when to skip:\n- \"Add a dark mode toggle\" → clear, specific, self-contained\n- \"Fix the 500 error on \u002Fcheckout when clicking 'Pay Now'\" → clear + has reproduction\n\n---\n\n## Step 1 — Analyze & Route\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nRead the user's prompt and classify the request into one of these categories:\n\n| Task type | Intent | Routes to |\n|---|---|---|---|\n| New project or service from scratch | INIT | `init-project` |\n| New feature, screen, component, endpoint, module in an existing codebase | CREATE | `add-feature` |\n| Update, extend, change behavior, small cosmetic tweak on an existing feature | EVOLVE | `modify-feature` |\n| Apply UX polish checklist without behavior change | POLISH | `polish-ui` |\n| Remove, delete, deprecate, rip out a feature | REMOVE | `remove-feature` |\n| Integrate with an external API — typed client, retry, observability | INTEGRATE | `integrate-api` |\n| Bug, broken, not working, silent failure, regression | FIX | `fix-bug` |\n| Codebase audit, tech-debt sweep, deep clean, production-readiness | AUDIT | `audit` |\n| Domain model rename, enum migration, status\u002Fvocabulary change with persisted data | REALIGN | `realign` |\n| Architecture design, system design, thinking through structure | DESIGN | `general` subagent(s) + codegraph exploration |\n| Infrastructure setup, debugging, or review | INFRA | `general` subagent(s) + relevant tooling |\n| Codebase research, explain a system, understand a flow | RESEARCH | `explore` subagent(s) + codegraph |\n| Anything else \u002F multi-faceted \u002F doesn't fit above | GENERAL | Orchestrate ad-hoc with subagents + available tooling |\n\n**Routing rules:**\n\n- If the request maps to a core skill (INIT \u002F CREATE \u002F EVOLVE \u002F POLISH \u002F REMOVE \u002F INTEGRATE \u002F FIX \u002F AUDIT \u002F REALIGN), route there with mode= and any include=\u002Fskip= overrides.\n- If the request maps to DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL, do NOT try to force-fit into a core skill. Instead:\n  1. Use `task` with appropriate subagent types to do the work\n  2. Use codegraph_explore for understanding\n  3. Use the full MCP + tool arsenal\n  4. If a sub-skill exists that partially matches (add-migration, add-observability, etc.), invoke it as part of the orchestration\n- **Multi-intent** — if the user lists clearly separable goals (\"add X and remove Y\"), execute sequentially as separate \u002Fstart routings. State the order before starting.\n\n**Never reject a request that doesn't fit the first 9 categories (INIT, CREATE, EVOLVE, POLISH, REMOVE, INTEGRATE, FIX, AUDIT, REALIGN).** DESIGN, INFRA, RESEARCH, and GENERAL cover everything else. If none of those fit either, the problem is so unusual that ad-hoc orchestration is the right answer — still don't reject.\n\n---\n\n## Step 2 — Infer the depth mode\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nThree base modes — `fast`, `balanced`, `production`. Pick one before announcing.\n\n**Plan-only modifier:** `plan-only` can be appended to any base mode (e.g. `mode=production+plan-only` or just `mode=plan-only`). When present, the skill stops after Step 3 (announce plan) — no execution, no changes. The plan is saved to agentmemory for later `\u002Fcontinue` in Build mode. Use this when the user is in Plan mode (read-only) and wants to review before executing.\n\n**Risk signals (any one → `production`)**:\n- Touches auth, permissions, payments, billing, secrets, or external webhooks\n- Schema migration or persisted-data rewrite\n- Destructive deletion or deprecation of an external\u002Fpublic contract\n- Background jobs, queues, cron, retries, email\u002FSMS\u002Fpush, imports\u002Fexports, file writes, spawned processes, IPC, or external APIs\n- Caching, query invalidation, feature flags, analytics\u002Fbusiness reporting, or concurrency-sensitive mutations\n- Multi-subsystem in the same change (frontend + backend + DB)\n\n**Tiny-scope signals (all four → `fast`)**:\n- Single file\n- Cosmetic \u002F copy \u002F styling only\n- No data layer touched\n- No new public API surface\n\n**Default → `balanced`**.\n\n**Override:** explicit `mode=fast`, `mode=balanced`, or `mode=production` in the user's prompt always wins — except when it conflicts with a high-risk signal (see NEVER below).\n\nFor DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL tasks, mode means:\n- `fast` — minimal analysis, direct answer\n- `balanced` — moderate depth with subagent exploration\n- `production` — full multi-subagent analysis with cross-validation\n\n---\n\n## Step 3 — Announce the plan\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nOutput a structured plan block before executing. Format:\n\n```\nDetected: \u003CCREATE | EVOLVE | POLISH | REMOVE | FIX | AUDIT | REALIGN | DESIGN | INFRA | RESEARCH | GENERAL>\nRisk:     \u003Clow | medium | high — one-line reason>\nMode:     \u003Cfast | balanced | production | plan-only — one-line reason or \"user-specified\">\nPipeline:\n  1. \u003Cphase>  → \u003Ccore skill, sub-skill, or subagent>\n  2. \u003Cphase>  → \u003Ccore skill, sub-skill, or subagent>\n  ...\n```\n\nPipeline numbering must match what the routed core skill or ad-hoc plan will actually run at the chosen mode. Do not invent phases the routed skill won't execute.\n\n**Confirmation gating depends on mode:**\n\n| Mode | Gating |\n|---|---|\n| `production` | Use `question` \"Proceed with this pipeline?\" before Step 4. Decline → stop. |\n| `balanced` | Print the plan inline, then proceed to Step 4 in the same turn. User can abort with ESC or a new prompt before execution begins. |\n| `fast` | Print the plan inline as a one-line preamble, then execute immediately. No confirm prompt. |\n| `* + plan-only` | Print the plan. Skip Step 4. Proceed to Step 5 with save-to-memory. |\n\n---\n\n## Step 4 — Execute\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\n**Plan-only gate:** If mode contains `plan-only`, skip execution entirely. Save the plan to agentmemory via `agentmemory_memory_save` with type=\"plan\" and concepts including the detected intent and file paths. Then proceed to Step 5 without calling any core skill.\n\n**If routing to a core skill:** Use `skill({ name: \"\u003Ccore-skill>\" })`. The skill reads the user's goal and mode\u002Finclude\u002Fskip from the conversation context — no explicit args needed.\n\n**If routing to DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL:** Orchestrate directly:\n1. Launch appropriate `task` subagents for analysis\n2. Use `codegraph_explore` \u002F `codegraph_search` for understanding\n3. Invoke any matching sub-skills (add-migration, add-observability, etc.) when their gates trigger\n4. Synthesize and present results\n\n**Host-portability fallback.** If `skill` is not available:\n1. Read the routed skill's SKILL.md directly.\n2. Execute its instructions inline.\n3. Surface the degradation in the Step 3 announcement: add a line like `Routing: inline (skill tool unavailable)`.\n\n**Respect downstream gates.** Core skills have their own approval gates (e.g., `add-feature mode=production`'s Plan-approval). Let them fire. Don't bypass them from \u002Fstart.\n\n**Adjunct skills live downstream.** \u002Fstart only chooses the top-level workflow. The routed skill owns stack\u002Fplugin-specific handoffs.\n\n---\n\n## Step 5 — Report and hand off to git\n\nUse `todowrite` to mark this step as in_progress. Mark it completed when done.\n\nOutput a visible pipeline summary:\n\n```\n✔ \u003Cphase 1>  — \u003Cone-line outcome>\n✔ \u003Cphase 2>  — \u003Cone-line outcome>\n✔ \u003Cphase 3>  — \u003Cone-line outcome>\n...\n\nFindings:\n  - \u003Ceach finding surfaced by the routed skill or audits>\n```\n\nIf execution was skipped (plan-only mode), append:\n```\nPlan saved to agentmemory. Switch to Build mode and run \u002Fcontinue to execute.\n```\n\nIf execution completed:\n```\nCode is production-ready. To publish:\n  • \u002Fcommit           — group the working tree into staged commits without pushing\n  • \u002Fcommit-and-push  — commit then push the current branch\n  • \u002Fopen-pr          — open a GitHub PR\n  • \u002Fcontinue         — continue this work if something still needs fixing\n```\n\nSurface findings, not just \"done.\" If a sub-skill audit returned issues that weren't auto-fixed, name them here.\n\n**Do not commit. Do not push.** The user picks the publish path.\n\n---\n\n## NEVER\n\n- **NEVER commit, push, or open PRs from inside this skill**\n  **Instead:** Stop at Step 5 and direct the user to `\u002Fcommit`, `\u002Fcommit-and-push`, `\u002Fopen-pr`, or `\u002Fcontinue`.\n\n- **NEVER bypass a routed core skill's own approval gate**\n  **Instead:** Let the gate fire. The user interacts with the core skill's gate, not with \u002Fstart.\n\n- **NEVER replicate the core skill's pipeline inline from memory**\n  **Instead:** Always delegate via `skill(...)`. When unavailable, read the SKILL.md and follow it verbatim.\n\n- **NEVER hide which mode and pipeline you picked**\n  **Instead:** Announce in every mode. Even `fast` prints the one-line preamble.\n\n- **NEVER guess between two plausible intents without asking**\n  **Instead:** Use Step 0's Q&A to resolve ambiguities before routing.\n\n- **NEVER honor a `mode=fast` override on a high-risk change without surfacing the conflict**\n  **Instead:** Pause and ask: \"Detected high-risk signals (e.g., payments). You requested fast mode. Confirm fast anyway, or upgrade to production?\"\n\n- **NEVER reject a request that doesn't fit the first 9 categories (INIT, CREATE, EVOLVE, POLISH, REMOVE, INTEGRATE, FIX, AUDIT, REALIGN)**\n  **Instead:** Route to DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL or orchestrate ad-hoc. There is always a path forward.\n\n- **NEVER execute in plan-only mode**\n  **Instead:** Stop at Step 3, save the plan, report. The user will return in Build mode via \u002Fcontinue.\n  **Why:** Plan mode is read-only. Executing code changes violates the user's mode contract.\n\n---\n\n## Appendix — Adjunct skills and sub-skills\n\n\u002Fstart routes to core skills. Those core skills own the following adjunct invocations (gate-driven, not automatic):\n\n### CREATE → `add-feature` may invoke\n\n- **UI scaffolding:** `add-empty-error-states`, `polish-ui`, `propagate-ui-pattern`\n- **Backend scaffolding:** `add-migration`, `add-observability`, `reviewer-authz`\n- **Tests:** `write-tests`, `add-e2e-test`\n- **Audits (Phase 7):** `reviewer-contracts`, `reviewer-concurrency`, `reviewer-data-integrity`, `reviewer-security-regression`, `reviewer-error-boundaries`, `reviewer-loading-states`, `reviewer-accessibility-regression`, `reviewer-client-bundle`, `reviewer-observability-coverage`, `reviewer-perf`, `audit-responsive`\n- **Cleanup:** `simplify`, `polish-ui`\n\n### EVOLVE → `modify-feature` may invoke\n\n- **UI extensions:** `add-empty-error-states`, `polish-ui`\n- **Backend extensions:** `add-migration`, `add-observability`, `reviewer-authz`\n- **Tests:** `write-tests`, `add-e2e-test`\n- **Audits:** `reviewer-contracts`, `reviewer-concurrency`, `reviewer-observability-coverage`, `reviewer-data-integrity`, `reviewer-security-regression`, `reviewer-error-boundaries`, `reviewer-loading-states`, `reviewer-accessibility-regression`, `reviewer-client-bundle`\n- **Cleanup:** `simplify`, `polish-ui`\n\n### POLISH → `polish-ui`\n\nRuns UX checklist; no fan-out.\n\n### FIX → `fix-bug` may invoke\n\n- `add-regression-test`\n- `polish-ui` (if UI changed)\n\n### REMOVE → `remove-feature` may invoke\n\n- `add-migration` (if schema change)\n- `reviewer-data-integrity`, `reviewer-contracts`\n\n### AUDIT → `audit` may invoke\n\nFull reviewer-* family across the repo, plus `audit-responsive`, `audit-seo-meta`, `audit-analytics`, `simplify`, `harden-types`.\n\n### REALIGN → `realign` may invoke\n\n- `add-migration` (if persisted data needs migration)\n- `reviewer-data-integrity` (post-step audit)\n\n### DESIGN \u002F INFRA \u002F RESEARCH \u002F GENERAL\n\nThese are ad-hoc orchestration paths. They may invoke any of the above sub-skills on demand, plus direct subagent work, codegraph exploration, web research, and CLI tooling. No fixed adjunct list.\n",{"name":5,"license":15,"metadata":24,"description":7,"compatibility":26},{"requires_codegraph":25},true,"opencode","main","a40966cd3b7323e693a9fd665ab9c720d8bf9634","skills\u002Fstart",0,"https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system","https:\u002F\u002Fwww.npmjs.com\u002Fpackage\u002F@acauhi\u002Fopencode-agent-system","https:\u002F\u002Fgithub.com\u002FAcauhi99","@github Backend Engineer","2026-07-07 22:24:45",{"source":8,"skillName":5,"commands":37,"prompt":47,"zipAvailable":25},[38,41,44],{"tool":39,"command":40},"npx","npx skills add Acauhi99\u002Fopencode-agent-system --skill start",{"tool":42,"command":43},"bunx","bunx skills add Acauhi99\u002Fopencode-agent-system --skill start",{"tool":45,"command":46},"pnpm","pnpm dlx skills add Acauhi99\u002Fopencode-agent-system --skill start","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Fstart\u002FSKILL.md",[],[50,56,62,69,75,82],{"skillNo":51,"name":52,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M74KEHKFCSF2H23R44MR5DN8","write-tests","https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Fwrite-tests\u002FSKILL.md",[],[],{"skillNo":57,"name":58,"nameZh":6,"description":6,"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},"01M74KEEXZW2185BZ3VJCM7AF6","update-changelog","https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Fupdate-changelog\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},"01M74KDVH470SS6WPS8WK3VSW3","triage","Move issues and external PRs through a state machine of triage roles — categorise, verify, grill if needed, and write agent-ready briefs.","https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Ftriage\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M74KDKGNY0MQXQ0ZHAGYSJJ4","to-prd","https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Fto-prd\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M74KD9YN0K59GWMWDT5YHBKY","to-issues","Break a plan, spec, or PRD into independently-grabbable issues on the project issue tracker using tracer-bullet vertical slices.","https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Fto-issues\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M74KD14VJAJZ8E4ENN0KP9GV","testing-plan","https:\u002F\u002Fgithub.com\u002FAcauhi99\u002Fopencode-agent-system\u002Fblob\u002Fmain\u002Fskills\u002Ftesting-plan\u002FSKILL.md",[],[]]