Duntak1 Agentworkflow Agent Workflow

agent-workflow（运行时精简版）

Truth: package/INVOCATION.md after install. Reference: reference.md. Manual: package/AGENTWORKFLOW_MANUAL.html for humans; do not load it into model context unless the user asks.

Runtime Rule

This skill is a workflow gate, not a documentation bundle. Keep context small:

• Read only this SKILL.md at skill activation.
• Read package/INVOCATION.md only when a step is ambiguous or a command is needed.
• Read reference.md only for command lookup.
• Never read the HTML manual as AI context by default.
• Prefer CLI summaries (aw status, aw next, aw task brief, aw context plan, aw sync inbox) over opening full ledgers.

Updating aw

When the user says 更新 aw, update aw, 升级 aw, or asks to refresh an already installed skill, run the project-local updater:

./scripts/aw upgrade --from-github --adapters

./scripts/aw update --from-github --adapters is the same path. This fetches https://github.com/duntak1/agentworkflow.git, removes the old local skill install, installs the fresh skill, and replaces the current project's agent-workflow/ package and scripts/ while preserving business docs/, reference/, and workflow state. Use --repo <url> or --ref <branch-or-tag> only when the engineer explicitly asks for a fork, branch, or release tag.

Startup

When the user says 启动 aw, @aw, aw start, or equivalent:

1. Declare AgentWorkflow active.
2. Ask the role before any scan or coding:
   • 1=产品
   • 2=前端
   • 3=后端
   • 4=全栈
3. After role confirmation, run/present aw project scan, summarize docs/PROJECT_SCAN.md, and ask the engineer to confirm new vs existing.
4. Ask sync-center decision and record it with aw config init --sync-center 1|2|3.
5. Ask code hosting provider and build target.

全栈 means frontend/backend code is in one repository by default, build target 3=fullstack; sync center is optional unless the engineer confirms split repos, different computers, or PM-managed collaboration.

Token Budget

Default budgets per AT-T:

Do not exceed a budget without saying why and asking the engineer to confirm the expansion. Never do an aimless full-repo scan. Avoid generated/cache/dependency/build/log directories: .git, node_modules, dist, build, coverage, .next, .nuxt, target, vendor, tmp, logs.

Use this lean order:

./scripts/aw status --json
./scripts/aw next
./scripts/aw task brief <AT-T>
./scripts/aw context plan --task <AT-T>
./scripts/aw context gate --task <AT-T>

Read only files listed in docs/context/tasks/CTX-<AT-T>.md. Prefer CodeGraph / aw code-map / aw context symbol, caller/callee, impact, and affected-test queries. If CodeGraph is unavailable, use CODE_MAP, CODE_CONTEXT_INDEX, FILE_INDEX, and precise rg.

Code map is automatic by default. aw context plan, aw task start, aw task complete, aw watch index, and aw gate pre-commit refresh/check docs/context/CODE_MAP.md unless AW_CODE_MAP_AUTO=0 is set for an explicit exception. Before coding in existing or large projects, query the map instead of scanning the repo:

./scripts/aw code-map build
./scripts/aw code-map query "<feature|symbol|route>"
./scripts/aw code-map impact "<symbol|route>"

CODE_MAP.md is a locator index, not permission to read full files. Convert findings into the task Context Plan and wait for confirmation before opening file contents.

Required Gates

When the user says “执行研发任务”, “开始开发”, “做下一个任务”, or equivalent, do not start coding. Start with aw next and aw task brief.

Context Continuity

Use aw compact "focus" --write --snapshot before Codex context compaction, new chats, model switches, long pauses, or after a large requirement/AT-T batch. Add --memory-summary only for reusable decisions or follow-ups. Do not store secrets.

New session lean resume:

./scripts/aw handoff --check
./scripts/aw memory inject
./scripts/aw status --json
./scripts/aw next

Read handoff/memory selectively. Do not read ENGINEERING_INDEX.md into AI context; it is for human engineers.

DSL / Plan / PM

• DSL suite: aw dsl suite <slug> "title" creates multi-file DSL coverage for requirements, pages, interactions, events, boundaries, and acceptance.
• aw dsl review <dsl> --write creates the engineer review package before aw approve dsl.
• aw approve dsl <dsl> --plan may guide Plan generation; PM/three-side projects should use aw pm plan --write before dispatch.
• aw plan change, aw plan task-add, and aw task split handle development-time changes without bloating the active task.
• Product-led work uses aw pm start and aw pm init <project-harness>; PM owns shared references, Pencil intake, shared DSL/Plan, dispatch, dashboard, and lifecycle gates.

Cross-Project Sync

For split frontend/backend repositories:

• Configure sync center with aw sync init <project-harness> --project ... --agent ... --role ....
• Start dependent work with aw sync pull + aw sync gate --task <id>.
• Read only aw sync inbox --from <peer> summaries and relevant event/contract files.
• aw sync pull imports read-only inbox files; never treat inbox as local truth until adopted through REQ / Bug / Plan / Handoff.
• API changes use aw contract change, aw contract test, and aw contract gate.
• Long-lived worker identities use aw agents register; common roles can be bootstrapped with aw agents register --defaults.
• Multi-agent implementation uses aw agents claim, aw agents heartbeat, and aw agents gate --strict.

Engineering Principles

Also maintain completeness, traceability, maintainability, and handoffability. Use aw audit, aw policy, aw security, aw service-catalog, aw release, aw trace check, aw metrics summary, aw ops gate, aw report handoff, aw report release, aw score record, and aw recover ... when the task touches risk, production, dependencies, services, release, multi-agent work, or handoff.

File / Bug / Git Rules

• Record spoken new requirements with aw req new; record development-time changes with aw req change, then update DSL/Plan/ATOMIC before continuing.
• Every bug or suspected bug goes to aw bug add.
• Update docs/FILE_INDEX.md when adding/deleting/renaming business files.
• Ask before Git commit/push. Use aw commit --task <id> --changelog "..."; do not commit automatically.
• Comment AI-written code for non-obvious business rules, boundaries, tradeoffs, side effects, and temporary decisions; avoid obvious comments.

Pencil

Pencil design files are formal inputs. Store .pen sources under global/references/design/pencil/source/, exports under exports/, screenshots under screenshots/, and changes in DESIGN_CHANGELOG.md. Do not parse encrypted .pen files as plain text; use Pencil tooling or exported artifacts.

Source: duntak1/agentworkflow — distributed by TomeVault.