[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7ADPDCMHF7VPVTDHCS4M4WP":3,"skill-same:01M7ADPDCMHF7VPVTDHCS4M4WP":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":6,"updatedAt":34,"install":35,"resources":48},"01M7ADPDCMHF7VPVTDHCS4M4WP","leo",null,"Use when routing and orchestrating any work across this Devin bundle through one universal entry point, including direct specialist skills, multi-skill flows, and local AFK issues.","Leostruka\u002Fdevin-bundle","Leostruka","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F81370685?v=4","01M7AD8MAK63VDNAWRDRRRW04N","leandro ostruka","01M7AD8MBS65YDS1KMMV7YPM07",0,"MIT","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fleo\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: leo\ndescription: Use when routing and orchestrating any work across this Devin bundle through one universal entry point, including direct specialist skills, multi-skill flows, and local AFK issues.\ntriggers: [user, model]\n---\n\n# \u002Fleo - Universal bundle orchestrator\n\n## TL;DR\n\n1. Self-check (scope, skills, verify, no opinion).\n2. Classify objective; route directly to matching skill(s).\n3. Start with `using-skills`; use `ask-bundle` only if unclear.\n4. For 3+ steps, write `todo_write`; mark `in_progress` then `completed`.\n5. Define Verification Function (gate\u002Fexpect\u002Fevidence) per step.\n6. Dispatch `qa-ci` to re-run every gate on clean checkout.\n7. Run the project's verification commands (e.g., `python audit.py` and `python -m pytest` when present) before claiming done.\n\n## Goal\n\nProvide one entry point for the bundle's universal orchestration: classify the objective, invoke specialist skills directly, compose multi-skill flows, and retain control through verified completion. `ask-bundle` is the full routing reference, not a mandatory hop; `wayfinder` and every other clear specialist route are first-class entry paths.\n\n## When to use\n\n- The user starts a session with \"leo\", a vague request, or no clear objective.\n- The user needs help choosing the first skill or flow.\n- The user wants to prepare local AFK issues for unattended implementation.\n- At any session start when the next action is non-trivial.\n- The user wants Leo to coordinate an objective across one or more specialist skills.\n\n## Procedure\n\n1. **Self-check before responding**\n   - Scope: do EXACTLY what was asked, no more, no less.\n   - Telegraphic output: no preamble, filler, opinion, or unsolicited explanation.\n   - Skills: for non-trivial tasks, invoke matching skills before acting.\n   - Verify: use `read`\u002F`exec`\u002F`grep`\u002F`glob`\u002F`run_subagent` before asserting.\n   - Opinion-silent: don't critique, reframe, or suggest unless asked.\n   - If the objective is clear, route to the matching flow in the situation router.\n   - If the objective is unclear or the user just says \"leo\" \u002F \"start\", ask the quick-start menu (see **Quick-start menu** below) with `ask_user_question` before routing.\n\n2. **Direct routing and composition**\n   - Classify the objective from the user's words and observed repository state.\n   - Route directly to every clearly matching specialist skill; invoke multiple skills when their responsibilities compose.\n   - Start with `using-skills` to reinforce the skill-first rule.\n   - Use `ask-bundle` only when the idea-to-ship path or phase boundary is unclear; it is a routing reference, not Leo's sole downstream router.\n   - If no bundled skill matches, use `tool-and-skill-discovery` or `skill search`\u002F`skill list`.\n   - For fast domain lookup, read `docs\u002FSKILL-TIERS.md`.\n\n3. **Planning**\n   - For 3+ step tasks, create a `todo_write` immediately.\n   - Every plan must be detailed, sequential, verifiable, with no loose ends.\n   - Mark `in_progress` when starting, `completed` when done - no batching.\n   - For explicit acceptance criteria, use `unlazy` or `autonomous-gates`.\n\n4. **Execution and verification (modular atomic action mode)**\n   - Never deduce. Use tools to observe reality first.\n   - Each step is atomic: one verifiable action, completed and verified before the next starts.\n   - For each step, define a **Verification Function (VF)** before executing it:\n     - `gate:` the exact command that proves the step done (build\u002Ftest\u002Flint\u002Ftypecheck\u002Fdry-run).\n     - `expect:` the output or exit code that constitutes pass.\n     - `evidence:` where the raw output will be recorded (ledger line or file path).\n   - Mark `in_progress` when starting a step, `completed` only after verification passes - no batching.\n   - **Independent QA\u002FCI verification (anti-gaming, mandatory for non-trivial steps):**\n     - After a step claims done, dispatch the `qa-ci` subagent (Medium effort via `BUNDLE_MEDIUM_MODEL` \u002F `data\u002Fbundle-models.json`, no write tools) to re-run every gate independently.\n     - The QA\u002FCI subagent sees only the diff and the spec - never the implementer's report.\n     - It re-executes each gate on a clean checkout (fresh worktree when feasible), runs `tests\u002Fheld-out\u002F` if present, and audits the diff for overfitting (hard-coded constants, mocked gates, skipped tests, phantom guardrails).\n     - A step is `completed` only when the QA\u002FCI subagent returns `Verdict: PASS` with fresh command output + exit code as evidence.\n     - If QA\u002FCI returns `FAIL`, the step stays `in_progress`; feed the failure back and re-execute the fix loop. Never override a QA\u002FCI FAIL with self-report.\n   - Techniques that make verification gaming-resistant:\n     - **Held-out tests** (`tests\u002Fheld-out\u002F`): the implementer cannot see or edit them; QA\u002FCI runs them. Passing visible tests while held-out fails = FAIL.\n     - **Independent verifier**: the QA\u002FCI subagent has no `write`\u002F`edit` tools, so it cannot alter the proof.\n     - **Clean-room re-execution**: gates run on a fresh checkout, never on the implementer's polluted workspace.\n     - **Evidence-based completion**: every \"done\" requires a fresh command + output + exit code from QA\u002FCI, not self-report.\n     - **CI as external immutable gate**: GitHub Actions \u002F hooks run `audit.py`, AI-signature check, and push-green check independently of the agent.\n   - Before claiming the session done, run `python audit.py` and `python -m pytest` and attach output.\n   - After routing, verify the target skill is loaded and its first step is started.\n\n5. **Return to Leo**\n   - After each specialist finishes, compare its verified result with the original objective and current todo state.\n   - Route the next unmet responsibility to the appropriate specialist; preserve artifacts and decisions across handoffs.\n   - Stop only at verified completion, an explicit user decision boundary, or a reported blocker. Leo owns orchestration; specialists own domain procedure.\n\n## Bundle context\n\nKeep this in mind for every session:\n\n- Devin CLI validated release: `{{VALIDATED_CLI_VERSION}}` from `data\u002Fbundle-identity.json` (see `docs\u002FDEVIN-CLI-COMPATIBILITY.md`).\n- Models: SWE-2 routes by effort level - parent `BUNDLE_DEFAULT_MODEL` \u002F `data\u002Fbundle-models.json` (`default_parent_model`, High); custom subagents `BUNDLE_MAX_MODEL` (`max_role_model`, Max) or `BUNDLE_MEDIUM_MODEL` (`medium_role_model`, Medium), all free. Never use non-bundle aliases (`opus`, `sonnet`, `gpt`, etc.) when the parent is free.\n- Issue tracker: local Markdown under `.devin\u002Fscratch\u002F\u003Cfeature-slug>\u002F`, conventions in `.devin\u002Fagents\u002Fissue-tracker.md` and `.devin\u002Fagents\u002Ftriage-labels.md`.\n- Skills: discover via `tool-and-skill-discovery` or `docs\u002FSKILL-TIERS.md`.\n- Hooks: 8 lifecycle events (see `docs\u002FTOOLS-MAP.md` and `config.json`).\n- Verification baseline: run the project's own verification commands; common defaults are `python audit.py` and `python -m pytest` when they exist.\n\n## Situation router\n\nPick the entry skill from the user's situation. If the situation is not in this table, route to `ask-bundle` for the full map or `tool-and-skill-discovery` for an external skill.\n\n| Situation | Entry skill | Next |\n|---|---|---|\n| Assemble \u002F expand \u002F elevate a raw project idea or prompt with OSINT teardown, first principles, and atomic\u002Fmodular hardening | `montador-de-ideias` | → `project-orchestrator` or `grilling` → `planning-pipeline` |\n| Build \u002F change \u002F implement something | `grilling` (With-docs) to capture intent and user impact, or `review-cadence` if trivial | → `planning-pipeline` Spec → Tickets (estimate PR size; split >500 lines) → `implement` + `tdd` → `security-audit` (if API, DB, secrets, endpoints, or infra) → `code-review` → `verification-before-completion` → `finishing-a-development-branch` |\n| Run AFK \u002F unattended on a feature | `afk-loop` only if issues exist and are `ready-for-agent`; otherwise use **Quick AFK issue creation** below | `afk-loop` |\n| Hard \u002F intermittent \u002F unclear bug | `diagnosing-bugs` | → `tdd` regression → `improve-codebase-architecture` if no seam |\n| CI is failing | `debug-ci-failures` | |\n| Triage incoming issues \u002F requests | `triage` | → `implement` |\n| Large, foggy, multi-session effort | `wayfinder` | → `planning-pipeline` Spec → ... |\n| Prototype to settle a design question | `prototype` (via `handoff`) | back to `grilling` \u002F `planning-pipeline` |\n| Research \u002F deep codebase exploration | `research` or `deep-mode` | feed into `grilling` or Spec |\n| Need input from another person | `planning-pipeline` Questionnaire | → `grilling` or Spec |\n| Git merge \u002F rebase conflict | `resolving-merge-conflicts` | |\n| Improve architecture \u002F find deep modules | `improve-codebase-architecture` | → `grilling` if it generates an idea |\n| Improve \u002F evolve a skill, rule, hook, or MCP | `continuous-improvement` | → `writing-skills` if editing one skill, `self-extend` if adding a new capability |\n| Release \u002F deploy \u002F rollback | `deploy` | → `gh` for GitHub operations, then smoke tests |\n| Security assessment | `security-audit` | → `implement` for approved remediations |\n| Data query \u002F analysis \u002F charts | `data-analyst` | |\n| UI \u002F UX design or polish | `impeccable` | → `a11y-audit` \u002F `e2e-testing` when applicable |\n| Logging \u002F metrics \u002F tracing \u002F quality infrastructure | `observability-quality` | |\n| API or database design | `api-design` or `database` | → `implement` after contract\u002Fschema decisions |\n| Performance or cost optimization | `performance` or `cost-optimization` | |\n| Human-only procedure \u002F provisioning | `wizard` | |\n| Guided learning | `teach` | |\n| Set up this repo for Devin | `project-setup` or the engineering-skills setup | |\n| Not sure which skill \u002F flow fits | `ask-bundle` | full map |\n| No skill matches | `tool-and-skill-discovery` | evaluate \u002F install |\n\nFor the main flow details (idea → ship, on-ramps, phase boundaries), see `ask-bundle`.\n\n### Build\u002Fchange flow gates\n\nThe default build\u002Fchange flow is: `grilling` (intent and user impact) → `planning-pipeline` (spec + tickets) → `implement` + `tdd` → `security-audit` (when the change touches API, DB, secrets, endpoints, or infrastructure) → `code-review` → `verification-before-completion` → `finishing-a-development-branch`.\n\nGates applied inside that flow:\n\n- **Intent and impact**: `grilling` captures the final goal, where the feature lands, and how the user is affected. Implementation does not start until intent is clear.\n- **Size and boundaries**: `planning-pipeline` estimates the PR size. If the estimate exceeds ~500 lines, split the task into smaller tickets with clear input\u002Foutput boundaries. Target PR size is ~300 lines.\n- **TDD**: `tdd` is a default step, not optional. Tests verify intent; don't delete or skip them without explicit approval.\n- **Pre-commit**: `setup-pre-commit` runs lint-staged and tests before commit when the project is set up by `project-setup` or `implement` touches pre-commit hooks.\n- **Security audit**: `security-audit` is a gate, not optional, whenever the change touches API, DB, secrets, endpoints, or infrastructure.\n- **Destructive work**: `docker` or a dev container is recommended for tasks that could be destructive (file system, DB, network, credentials).\n- **Parallelism**: limit concurrent subagents to 1-3. If a task suggests more than 3 parallel agents, alert the user and ask for approval.\n\n### Tool-output validation\n\nAfter executing critical tools (API calls, DB writes, file-system changes, network calls), validate the output before allowing side effects or reporting success:\n\n- **Type\u002Fschema check**: validate tool arguments and responses with Pydantic or an equivalent schema check before using them.\n- **Ontology\u002Freasonableness check**: check the result against the domain ontology or business rules (e.g., status values must be `paid`, `shipped`, or `refunded`; entities must be the expected type). If the result is not reasonable, go back to the LLM or ask the human - do not apply side effects.\n- **No side effects until validated**: don't let a tool output drive destructive or state-changing actions before it passes the checks.\n\n## Quick-start menu\n\nWhen the user says \"leo\", \"start\", or the objective is unclear, ask a focused `ask_user_question` with 2–4 options. The tool adds an \"Other\" option automatically; route \"Other\" to `ask-bundle` or a follow-up question.\n\n- **Assemble a raw idea \u002F prompt, or build \u002F change something in the repo** → `montador-de-ideias` to expand and harden a seed idea\u002Fprompt with OSINT + first principles, `grilling` (With-docs) if design decisions remain, or `review-cadence` if trivial\n- **Improve \u002F add a skill, rule, hook, or MCP** → `continuous-improvement`\n- **Debug, fix, or research a problem** → `diagnosing-bugs` for bugs, `research` \u002F `deep-mode` for exploration\n- **Set up this repo for Devin or run AFK work** → `project-setup` \u002F the engineering-skills setup for setup, `afk-loop` if issues already exist\n\n## Quick AFK issue creation\n\nWhen the user wants unattended work but there are no local tickets yet:\n\n1. Confirm the feature slug and the high-level objective.\n2. If there is no spec, run `grilling` (With-docs mode) or `planning-pipeline` (Spec mode) to write `.devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fspec.md`.\n3. Run `planning-pipeline` (Tickets mode) to break the spec into vertical-slice tracer bullets.\n4. Publish one file per ticket to `.devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fissues\u002F\u003CNN>-\u003Cslug>.md`:\n   - Number from `01` in dependency order (blockers first).\n   - Use the local ticket template from `planning-pipeline`.\n   - Include `Status: ready-for-agent`.\n   - Include `Blocked by: \u003Cnumbers>` or `None - can start immediately`.\n   - List acceptance criteria as checkboxes.\n   - Focus on end-to-end \"what to build\" from the user's perspective, not layers.\n5. Verify the files exist and the DAG is consistent:\n   - `glob .devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fissues\u002F*.md`\n   - `read .devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fspec.md`\n   - Parse `Status:` and `Blocked by:` from each issue.\n6. Only then run `afk-loop` if the user explicitly authorizes unattended work.\n\n## Specifications\n\n- `AGENTS.md` rules are respected.\n- `.devin\u002Fglobal_rules.md` is read when working inside this repo.\n- `docs\u002FSKILL-TIERS.md` is the fast path for skill discovery.\n- Matching skills are invoked before non-trivial actions.\n- 3+ step tasks use `todo_write` with updated states.\n- Every claim is verified by a tool, not by reasoning.\n- Output is terse, structured, and cited when making factual claims.\n\n## Advice\n\n- Subagents: use custom profiles pinned to bundle effort variants (`swe-2-max` \u002F `swe-2-flash-high` \u002F `swe-2-flash-medium`, free). Never use `subagent_explore` or other paid aliases when the parent is free.\n- Deep search: use the `researcher` subagent profile (`swe-2-max`).\n- Models: parent is `BUNDLE_DEFAULT_MODEL` \u002F `data\u002Fbundle-models.json` (`default_parent_model`); subagents are `BUNDLE_MAX_MODEL` (`max_role_model`), `BUNDLE_MEDIUM_MODEL` (`medium_role_model`) or `BUNDLE_FAST_MODEL` (`fast_role_model`). Don't use paid aliases unless the parent is paid.\n- Context: prefer `clear` between unrelated tasks; don't paste large documents into chat.\n- Secrets: never display `.env`\u002F`credentials.toml` values; name the variable and symptom only.\n- Fact doubt → research (`web_search`, `webfetch`, `grep`, `exec`).\n- Intent doubt → ask (`ask_user_question`).\n- Cross-skill: `leo` is the session-start wrapper. Use `using-skills` to reinforce skill-first behavior. Use `ask-bundle` for the full flow map when the quick router is not enough. Use `tool-and-skill-discovery` for external or missing skills.\n\n## Forbidden Actions\n\n- Deduce state, file content, or command output without using tools.\n- Start non-trivial work without skill discovery.\n- Mark a step `completed` without an independent `qa-ci` PASS for non-trivial steps.\n- Override a `qa-ci` FAIL with self-report, \"should work\", or confidence.\n- Push or commit with failing local checks (Rule 5).\n- Sign commits, files, PRs, or docs as AI (Rule 2).\n- Display secret values.\n- Run destructive\u002Firreversible actions without explicit user confirmation.\n- Use `subagent_explore` or paid models when the parent is free.\n- Compact when `clear` is sufficient; let context grow unchecked.\n- Start `afk-loop` without local issues in `ready-for-agent` state.\n- Edit or read `tests\u002Fheld-out\u002F` from the implementer context; only `qa-ci` may run them.\n- Reuse the implementer's shell, caches, or installed deps for QA\u002FCI verification without re-installing from pinned manifests.\n\n## Required from User\n\n- A clear objective at the start of the session, or a choice from the quick-start menu.\n- Clarification when the request is ambiguous and the deliverable would change.\n\n## Priority Hierarchy\n\nHard constraints (never violated):\n1. Safety and pinned `AGENTS.md` rules (Rules 2, 5, 7, 12-19, 21).\n2. Verify with tools before asserting (Rule 17).\n3. Execute exactly what was asked, without opinion (Rule 7).\n\nDesign preferences (when constraints allow):\n4. User usefulness and real benefit.\n5. Ease and pleasantness of interaction.\n6. Quality of experience.\n7. Technical coherence.\n8. Performance and responsiveness.\n","# \u002Fleo - Universal bundle orchestrator\n\n## TL;DR\n\n1. Self-check (scope, skills, verify, no opinion).\n2. Classify objective; route directly to matching skill(s).\n3. Start with `using-skills`; use `ask-bundle` only if unclear.\n4. For 3+ steps, write `todo_write`; mark `in_progress` then `completed`.\n5. Define Verification Function (gate\u002Fexpect\u002Fevidence) per step.\n6. Dispatch `qa-ci` to re-run every gate on clean checkout.\n7. Run the project's verification commands (e.g., `python audit.py` and `python -m pytest` when present) before claiming done.\n\n## Goal\n\nProvide one entry point for the bundle's universal orchestration: classify the objective, invoke specialist skills directly, compose multi-skill flows, and retain control through verified completion. `ask-bundle` is the full routing reference, not a mandatory hop; `wayfinder` and every other clear specialist route are first-class entry paths.\n\n## When to use\n\n- The user starts a session with \"leo\", a vague request, or no clear objective.\n- The user needs help choosing the first skill or flow.\n- The user wants to prepare local AFK issues for unattended implementation.\n- At any session start when the next action is non-trivial.\n- The user wants Leo to coordinate an objective across one or more specialist skills.\n\n## Procedure\n\n1. **Self-check before responding**\n   - Scope: do EXACTLY what was asked, no more, no less.\n   - Telegraphic output: no preamble, filler, opinion, or unsolicited explanation.\n   - Skills: for non-trivial tasks, invoke matching skills before acting.\n   - Verify: use `read`\u002F`exec`\u002F`grep`\u002F`glob`\u002F`run_subagent` before asserting.\n   - Opinion-silent: don't critique, reframe, or suggest unless asked.\n   - If the objective is clear, route to the matching flow in the situation router.\n   - If the objective is unclear or the user just says \"leo\" \u002F \"start\", ask the quick-start menu (see **Quick-start menu** below) with `ask_user_question` before routing.\n\n2. **Direct routing and composition**\n   - Classify the objective from the user's words and observed repository state.\n   - Route directly to every clearly matching specialist skill; invoke multiple skills when their responsibilities compose.\n   - Start with `using-skills` to reinforce the skill-first rule.\n   - Use `ask-bundle` only when the idea-to-ship path or phase boundary is unclear; it is a routing reference, not Leo's sole downstream router.\n   - If no bundled skill matches, use `tool-and-skill-discovery` or `skill search`\u002F`skill list`.\n   - For fast domain lookup, read `docs\u002FSKILL-TIERS.md`.\n\n3. **Planning**\n   - For 3+ step tasks, create a `todo_write` immediately.\n   - Every plan must be detailed, sequential, verifiable, with no loose ends.\n   - Mark `in_progress` when starting, `completed` when done - no batching.\n   - For explicit acceptance criteria, use `unlazy` or `autonomous-gates`.\n\n4. **Execution and verification (modular atomic action mode)**\n   - Never deduce. Use tools to observe reality first.\n   - Each step is atomic: one verifiable action, completed and verified before the next starts.\n   - For each step, define a **Verification Function (VF)** before executing it:\n     - `gate:` the exact command that proves the step done (build\u002Ftest\u002Flint\u002Ftypecheck\u002Fdry-run).\n     - `expect:` the output or exit code that constitutes pass.\n     - `evidence:` where the raw output will be recorded (ledger line or file path).\n   - Mark `in_progress` when starting a step, `completed` only after verification passes - no batching.\n   - **Independent QA\u002FCI verification (anti-gaming, mandatory for non-trivial steps):**\n     - After a step claims done, dispatch the `qa-ci` subagent (Medium effort via `BUNDLE_MEDIUM_MODEL` \u002F `data\u002Fbundle-models.json`, no write tools) to re-run every gate independently.\n     - The QA\u002FCI subagent sees only the diff and the spec - never the implementer's report.\n     - It re-executes each gate on a clean checkout (fresh worktree when feasible), runs `tests\u002Fheld-out\u002F` if present, and audits the diff for overfitting (hard-coded constants, mocked gates, skipped tests, phantom guardrails).\n     - A step is `completed` only when the QA\u002FCI subagent returns `Verdict: PASS` with fresh command output + exit code as evidence.\n     - If QA\u002FCI returns `FAIL`, the step stays `in_progress`; feed the failure back and re-execute the fix loop. Never override a QA\u002FCI FAIL with self-report.\n   - Techniques that make verification gaming-resistant:\n     - **Held-out tests** (`tests\u002Fheld-out\u002F`): the implementer cannot see or edit them; QA\u002FCI runs them. Passing visible tests while held-out fails = FAIL.\n     - **Independent verifier**: the QA\u002FCI subagent has no `write`\u002F`edit` tools, so it cannot alter the proof.\n     - **Clean-room re-execution**: gates run on a fresh checkout, never on the implementer's polluted workspace.\n     - **Evidence-based completion**: every \"done\" requires a fresh command + output + exit code from QA\u002FCI, not self-report.\n     - **CI as external immutable gate**: GitHub Actions \u002F hooks run `audit.py`, AI-signature check, and push-green check independently of the agent.\n   - Before claiming the session done, run `python audit.py` and `python -m pytest` and attach output.\n   - After routing, verify the target skill is loaded and its first step is started.\n\n5. **Return to Leo**\n   - After each specialist finishes, compare its verified result with the original objective and current todo state.\n   - Route the next unmet responsibility to the appropriate specialist; preserve artifacts and decisions across handoffs.\n   - Stop only at verified completion, an explicit user decision boundary, or a reported blocker. Leo owns orchestration; specialists own domain procedure.\n\n## Bundle context\n\nKeep this in mind for every session:\n\n- Devin CLI validated release: `{{VALIDATED_CLI_VERSION}}` from `data\u002Fbundle-identity.json` (see `docs\u002FDEVIN-CLI-COMPATIBILITY.md`).\n- Models: SWE-2 routes by effort level - parent `BUNDLE_DEFAULT_MODEL` \u002F `data\u002Fbundle-models.json` (`default_parent_model`, High); custom subagents `BUNDLE_MAX_MODEL` (`max_role_model`, Max) or `BUNDLE_MEDIUM_MODEL` (`medium_role_model`, Medium), all free. Never use non-bundle aliases (`opus`, `sonnet`, `gpt`, etc.) when the parent is free.\n- Issue tracker: local Markdown under `.devin\u002Fscratch\u002F\u003Cfeature-slug>\u002F`, conventions in `.devin\u002Fagents\u002Fissue-tracker.md` and `.devin\u002Fagents\u002Ftriage-labels.md`.\n- Skills: discover via `tool-and-skill-discovery` or `docs\u002FSKILL-TIERS.md`.\n- Hooks: 8 lifecycle events (see `docs\u002FTOOLS-MAP.md` and `config.json`).\n- Verification baseline: run the project's own verification commands; common defaults are `python audit.py` and `python -m pytest` when they exist.\n\n## Situation router\n\nPick the entry skill from the user's situation. If the situation is not in this table, route to `ask-bundle` for the full map or `tool-and-skill-discovery` for an external skill.\n\n| Situation | Entry skill | Next |\n|---|---|---|\n| Assemble \u002F expand \u002F elevate a raw project idea or prompt with OSINT teardown, first principles, and atomic\u002Fmodular hardening | `montador-de-ideias` | → `project-orchestrator` or `grilling` → `planning-pipeline` |\n| Build \u002F change \u002F implement something | `grilling` (With-docs) to capture intent and user impact, or `review-cadence` if trivial | → `planning-pipeline` Spec → Tickets (estimate PR size; split >500 lines) → `implement` + `tdd` → `security-audit` (if API, DB, secrets, endpoints, or infra) → `code-review` → `verification-before-completion` → `finishing-a-development-branch` |\n| Run AFK \u002F unattended on a feature | `afk-loop` only if issues exist and are `ready-for-agent`; otherwise use **Quick AFK issue creation** below | `afk-loop` |\n| Hard \u002F intermittent \u002F unclear bug | `diagnosing-bugs` | → `tdd` regression → `improve-codebase-architecture` if no seam |\n| CI is failing | `debug-ci-failures` | |\n| Triage incoming issues \u002F requests | `triage` | → `implement` |\n| Large, foggy, multi-session effort | `wayfinder` | → `planning-pipeline` Spec → ... |\n| Prototype to settle a design question | `prototype` (via `handoff`) | back to `grilling` \u002F `planning-pipeline` |\n| Research \u002F deep codebase exploration | `research` or `deep-mode` | feed into `grilling` or Spec |\n| Need input from another person | `planning-pipeline` Questionnaire | → `grilling` or Spec |\n| Git merge \u002F rebase conflict | `resolving-merge-conflicts` | |\n| Improve architecture \u002F find deep modules | `improve-codebase-architecture` | → `grilling` if it generates an idea |\n| Improve \u002F evolve a skill, rule, hook, or MCP | `continuous-improvement` | → `writing-skills` if editing one skill, `self-extend` if adding a new capability |\n| Release \u002F deploy \u002F rollback | `deploy` | → `gh` for GitHub operations, then smoke tests |\n| Security assessment | `security-audit` | → `implement` for approved remediations |\n| Data query \u002F analysis \u002F charts | `data-analyst` | |\n| UI \u002F UX design or polish | `impeccable` | → `a11y-audit` \u002F `e2e-testing` when applicable |\n| Logging \u002F metrics \u002F tracing \u002F quality infrastructure | `observability-quality` | |\n| API or database design | `api-design` or `database` | → `implement` after contract\u002Fschema decisions |\n| Performance or cost optimization | `performance` or `cost-optimization` | |\n| Human-only procedure \u002F provisioning | `wizard` | |\n| Guided learning | `teach` | |\n| Set up this repo for Devin | `project-setup` or the engineering-skills setup | |\n| Not sure which skill \u002F flow fits | `ask-bundle` | full map |\n| No skill matches | `tool-and-skill-discovery` | evaluate \u002F install |\n\nFor the main flow details (idea → ship, on-ramps, phase boundaries), see `ask-bundle`.\n\n### Build\u002Fchange flow gates\n\nThe default build\u002Fchange flow is: `grilling` (intent and user impact) → `planning-pipeline` (spec + tickets) → `implement` + `tdd` → `security-audit` (when the change touches API, DB, secrets, endpoints, or infrastructure) → `code-review` → `verification-before-completion` → `finishing-a-development-branch`.\n\nGates applied inside that flow:\n\n- **Intent and impact**: `grilling` captures the final goal, where the feature lands, and how the user is affected. Implementation does not start until intent is clear.\n- **Size and boundaries**: `planning-pipeline` estimates the PR size. If the estimate exceeds ~500 lines, split the task into smaller tickets with clear input\u002Foutput boundaries. Target PR size is ~300 lines.\n- **TDD**: `tdd` is a default step, not optional. Tests verify intent; don't delete or skip them without explicit approval.\n- **Pre-commit**: `setup-pre-commit` runs lint-staged and tests before commit when the project is set up by `project-setup` or `implement` touches pre-commit hooks.\n- **Security audit**: `security-audit` is a gate, not optional, whenever the change touches API, DB, secrets, endpoints, or infrastructure.\n- **Destructive work**: `docker` or a dev container is recommended for tasks that could be destructive (file system, DB, network, credentials).\n- **Parallelism**: limit concurrent subagents to 1-3. If a task suggests more than 3 parallel agents, alert the user and ask for approval.\n\n### Tool-output validation\n\nAfter executing critical tools (API calls, DB writes, file-system changes, network calls), validate the output before allowing side effects or reporting success:\n\n- **Type\u002Fschema check**: validate tool arguments and responses with Pydantic or an equivalent schema check before using them.\n- **Ontology\u002Freasonableness check**: check the result against the domain ontology or business rules (e.g., status values must be `paid`, `shipped`, or `refunded`; entities must be the expected type). If the result is not reasonable, go back to the LLM or ask the human - do not apply side effects.\n- **No side effects until validated**: don't let a tool output drive destructive or state-changing actions before it passes the checks.\n\n## Quick-start menu\n\nWhen the user says \"leo\", \"start\", or the objective is unclear, ask a focused `ask_user_question` with 2–4 options. The tool adds an \"Other\" option automatically; route \"Other\" to `ask-bundle` or a follow-up question.\n\n- **Assemble a raw idea \u002F prompt, or build \u002F change something in the repo** → `montador-de-ideias` to expand and harden a seed idea\u002Fprompt with OSINT + first principles, `grilling` (With-docs) if design decisions remain, or `review-cadence` if trivial\n- **Improve \u002F add a skill, rule, hook, or MCP** → `continuous-improvement`\n- **Debug, fix, or research a problem** → `diagnosing-bugs` for bugs, `research` \u002F `deep-mode` for exploration\n- **Set up this repo for Devin or run AFK work** → `project-setup` \u002F the engineering-skills setup for setup, `afk-loop` if issues already exist\n\n## Quick AFK issue creation\n\nWhen the user wants unattended work but there are no local tickets yet:\n\n1. Confirm the feature slug and the high-level objective.\n2. If there is no spec, run `grilling` (With-docs mode) or `planning-pipeline` (Spec mode) to write `.devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fspec.md`.\n3. Run `planning-pipeline` (Tickets mode) to break the spec into vertical-slice tracer bullets.\n4. Publish one file per ticket to `.devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fissues\u002F\u003CNN>-\u003Cslug>.md`:\n   - Number from `01` in dependency order (blockers first).\n   - Use the local ticket template from `planning-pipeline`.\n   - Include `Status: ready-for-agent`.\n   - Include `Blocked by: \u003Cnumbers>` or `None - can start immediately`.\n   - List acceptance criteria as checkboxes.\n   - Focus on end-to-end \"what to build\" from the user's perspective, not layers.\n5. Verify the files exist and the DAG is consistent:\n   - `glob .devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fissues\u002F*.md`\n   - `read .devin\u002Fscratch\u002F\u003Cfeature-slug>\u002Fspec.md`\n   - Parse `Status:` and `Blocked by:` from each issue.\n6. Only then run `afk-loop` if the user explicitly authorizes unattended work.\n\n## Specifications\n\n- `AGENTS.md` rules are respected.\n- `.devin\u002Fglobal_rules.md` is read when working inside this repo.\n- `docs\u002FSKILL-TIERS.md` is the fast path for skill discovery.\n- Matching skills are invoked before non-trivial actions.\n- 3+ step tasks use `todo_write` with updated states.\n- Every claim is verified by a tool, not by reasoning.\n- Output is terse, structured, and cited when making factual claims.\n\n## Advice\n\n- Subagents: use custom profiles pinned to bundle effort variants (`swe-2-max` \u002F `swe-2-flash-high` \u002F `swe-2-flash-medium`, free). Never use `subagent_explore` or other paid aliases when the parent is free.\n- Deep search: use the `researcher` subagent profile (`swe-2-max`).\n- Models: parent is `BUNDLE_DEFAULT_MODEL` \u002F `data\u002Fbundle-models.json` (`default_parent_model`); subagents are `BUNDLE_MAX_MODEL` (`max_role_model`), `BUNDLE_MEDIUM_MODEL` (`medium_role_model`) or `BUNDLE_FAST_MODEL` (`fast_role_model`). Don't use paid aliases unless the parent is paid.\n- Context: prefer `clear` between unrelated tasks; don't paste large documents into chat.\n- Secrets: never display `.env`\u002F`credentials.toml` values; name the variable and symptom only.\n- Fact doubt → research (`web_search`, `webfetch`, `grep`, `exec`).\n- Intent doubt → ask (`ask_user_question`).\n- Cross-skill: `leo` is the session-start wrapper. Use `using-skills` to reinforce skill-first behavior. Use `ask-bundle` for the full flow map when the quick router is not enough. Use `tool-and-skill-discovery` for external or missing skills.\n\n## Forbidden Actions\n\n- Deduce state, file content, or command output without using tools.\n- Start non-trivial work without skill discovery.\n- Mark a step `completed` without an independent `qa-ci` PASS for non-trivial steps.\n- Override a `qa-ci` FAIL with self-report, \"should work\", or confidence.\n- Push or commit with failing local checks (Rule 5).\n- Sign commits, files, PRs, or docs as AI (Rule 2).\n- Display secret values.\n- Run destructive\u002Firreversible actions without explicit user confirmation.\n- Use `subagent_explore` or paid models when the parent is free.\n- Compact when `clear` is sufficient; let context grow unchecked.\n- Start `afk-loop` without local issues in `ready-for-agent` state.\n- Edit or read `tests\u002Fheld-out\u002F` from the implementer context; only `qa-ci` may run them.\n- Reuse the implementer's shell, caches, or installed deps for QA\u002FCI verification without re-installing from pinned manifests.\n\n## Required from User\n\n- A clear objective at the start of the session, or a choice from the quick-start menu.\n- Clarification when the request is ambiguous and the deliverable would change.\n\n## Priority Hierarchy\n\nHard constraints (never violated):\n1. Safety and pinned `AGENTS.md` rules (Rules 2, 5, 7, 12-19, 21).\n2. Verify with tools before asserting (Rule 17).\n3. Execute exactly what was asked, without opinion (Rule 7).\n\nDesign preferences (when constraints allow):\n4. User usefulness and real benefit.\n5. Ease and pleasantness of interaction.\n6. Quality of experience.\n7. Technical coherence.\n8. Performance and responsiveness.\n",{"name":5,"triggers":24,"description":7},[25,26],"user","model","main","0032dd465157e15d3ddd16afb75f4fe2dc1a356c","skills\u002Fleo",1,"https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle","https:\u002F\u002Fleostruka.github.io\u002Fdevin-bundle\u002F","https:\u002F\u002Fgithub.com\u002FLeostruka","2026-11-15 05:19:58",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add Leostruka\u002Fdevin-bundle --skill leo",{"tool":41,"command":42},"bunx","bunx skills add Leostruka\u002Fdevin-bundle --skill leo",{"tool":44,"command":45},"pnpm","pnpm dlx skills add Leostruka\u002Fdevin-bundle --skill leo","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fleo\u002FSKILL.md",true,[],[50,57,64,71,78,85],{"skillNo":51,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":19},"01M7E0K3MG2XC4FSRV0MK6SFR1","montador-de-ideias","Use when assembling, expanding, or elevating a raw project idea or initial prompt into an ultra-detailed, first-principles dossier and compiled master prompt with journalistic OSINT market research, Tech Lead and Staff Architect justifications, atomic and modular hardening\u002Foptimization, and high-impact contact surfaces.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fmontador-de-ideias\u002FSKILL.md",[],[],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M7E0J7JJWZ8QMK01RZ3B3C9Z","git-github-adaptive-engineering","Use when any task touches git history, GitHub issues\u002FPRs\u002Freleases\u002FCI, repo governance, or branch and merge decisions. Runs the Fase Zero situational-awareness battery, picks one of five delivery flows via the adaptive decision engine, and enforces commit\u002FPR information-density standards. Orchestrates the gh, git-workflows, github-actions, pr, and debug\u002Ffinish skills.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fgit-github-adaptive-engineering\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M7E0GJFRPFEVQQMXEM58EWDM","adaptive-agenda-calendar-routine-engine","Use when planning or executing work across hours, days, weeks, or sprints. Turns the calendar into a strict time-blocked agenda (HH:MM-HH:MM blocks carrying a routine profile, expected inputs, and a handoff output artifact), runs the 5-stage cognitive flow (Radar and Hunt -> Triage plus Steve Jobs conception -> workbench manifest -> >=32-cycle loop -> Close plus Open-Minded Scrum), and unfolds macro calendar milestones into micro agenda blocks through extensions\u002Fgcal\u002Fwrapper.py.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fadaptive-agenda-calendar-routine-engine\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M7AE2K9FTY0BVS48N0A9JFKF","demo","Use when demonstrating a skill in a fixture project.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fdevin-config\u002Fproject\u002F.devin\u002Fskills\u002Fdemo\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M7AE2G3DP38PTV3B1N7MW5TS","duplicate-skill","Use when testing duplicate skill name detection.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fdevin-config\u002Fproject\u002F.devin\u002Fskills\u002Fbeta\u002FSKILL.md",[],[],{"skillNo":86,"name":80,"nameZh":6,"description":81,"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},"01M7AE2DBR8ZHT6T875MJ392GK","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fdevin-config\u002Fproject\u002F.devin\u002Fskills\u002Falpha\u002FSKILL.md",[],[]]