gan-xing/codexbridge/agent-draft-router

Agent Draft Router

Use this skill only for first-host /agent draft intake. Do not use it for:

• explicit /agent confirm|edit|cancel|list|show|result|stop|retry|delete|rename|send
• Mission runtime execution
• package-owned lifecycle control

Primary reference:

• ../../docs/architecture/agent-draft-templates.md

Read that document when you need the full code or generic template contract.

Workflow

1. Determine whether the request is draft intake at all.
2. Keep explicit /agent subcommands deterministic and out of model routing.
3. For bare /agent <text> or /agent add <text>, emit only a bounded action.
4. If the action is add/create, enter create-flow:

• determine task type inside the skill
• clarify only when the missing information is truly blocking
• generate the formal checklist
• generate the immutable prompt
• generate loop policy

1. Produce a checklist-first draft rather than a generic lifecycle plan.

Routing Rules

Allowed bounded actions for natural-language intake:

• create_draft
• update_pending_draft
• clarify
• query_jobs
• show_job
• show_result
• propose_stop
• propose_retry
• reject

Prefer clarification over forced routing when confidence is low.

Do not:

• invent jobs or drafts
• mutate authoritative mission state
• bypass deterministic bridge handlers
• bypass package-owned approval or change gates

Create-Flow Rules

Only continue into create-flow when the action clearly resolves to add/create.

Inside create-flow:

1. Determine task type from the current request and repo context.
2. If the task is genuinely ambiguous or underspecified, ask one narrowing question.
3. Generate a formal checklist, not a generic software lifecycle.
4. Generate an immutable prompt scaffold that can survive autonomous loops.
5. Keep internal substeps separate from formal checklist mutation.

code Missions

For code missions, prefer:

• repo-aware checklist items
• fixed immutable prompt scaffolding
• explicit verification commands
• explicit execution boundaries
• bilingual Conventional Commit requirements when repository changes are in scope
• repo context from the invocation payload over fixed path assumptions
• distill the user input into one direct sentence with at most one comma; do not expand it into rationale, checklist, or scope narration
• generate plan[] from the user goal plus repo context, and never leave it empty
• if at least 3 concrete checklist items cannot be derived from the goal and context, return clarify instead of forcing create_draft

Do not fall back to:

• analyze
• design
• code
• test
• deploy

unless those are truly the correct bounded checklist items.

Generic Non-Code Missions

For generic non-code missions:

• keep the prompt lighter than code
• still require immutable goal, checklist, acceptance criteria, and loop policy
• still require checklist/progress updates each cycle

State-Aware Loop Expectations

When shaping immutable prompts, make sure the loop can use Mission Control status semantics:

• running / verifying / repairing: take the smallest next executable step
• waiting_user: ask one concrete blocking question
• needs_human: summarize why autonomy is no longer reasonable
• handoff: preserve current state and recommended next owner action
• blocked: describe the blocker precisely and offer options
• max_loops_reached: explain which budget was exhausted and the next recovery path

To avoid "do two steps then stall" behavior:

• prefer one bounded next step over broad re-planning
• prefer checklist refinement suggestions over vague failure language
• only clarify when the missing information is truly blocking

Output Requirements

When generating or updating drafts:

• treat plan[] as the formal confirmed checklist / TODO
• preserve user-confirmed scope unless a formal refinement is required
• ensure the immutable prompt explicitly requires checklist-status updates,

overall progress updates, blockers, and next-step updates each cycle

• for code missions, include bilingual Conventional Commit rules whenever

repository changes are allowed

When editing a draft:

• return the full updated draft, not a patch
• preserve fields the user did not request to change