opencode-web-page-generator

name	opencode-web-page-generator
description	Manager for web-page work. Continuously orchestrate planner, builder, and evaluator roles by reading `agents/*.md` and launching generic subagents until the shared artifacts show the work is complete.
compatibility	opencode
metadata	{"workflow":"iterative","runtime":"opencode"}

OpenCode Web Page Generator

Use this skill as a manager that continuously orchestrates the workflow.

• Do not do planning, implementation, or review in the manager session.
• Do not restate the subagent instructions.
• Keep manager prompts compact, but continue routing until the completion criteria are met.
• Treat agents/planner.md, agents/builder.md, and agents/evaluator.md as role instruction files.
• Start generic subagents from those files instead of assuming planner, builder, or evaluator are registered OpenCode agent types.

Role files:

• planner at agents/planner.md
• builder at agents/builder.md
• evaluator at agents/evaluator.md

Shared Artifacts

• product_spec.md
• sprint_contract.md
• handoff.md
• qa_report.md
• design-system/MASTER.md
• design-system/pages/*.md

Templates:

• references/product-spec-template.md
• references/sprint-contract-template.md
• references/handoff-template.md
• references/qa-report-template.md

Reuse these files. Do not invent parallel documents.

Handoff Contract

handoff.md is the continuity artifact for agent reset, not only end-of-sprint wrap-up.

Use it in two cases:

• a role finishes its round and needs to hand off to the next role
• the current subagent is approaching context saturation and should stop early so a fresh subagent can continue from handoff.md

Do not wait for hard failure. Reset proactively when the working context is getting crowded.

Manager Rules

• Read only what is needed at each step to choose the next hop.
• Use progressive disclosure for all skills and references: load only the skill or reference file needed for the current decision, then only the specific sections needed for the current subtask.
• Do not preload every skill file, every reference file, or full generated docs just because they might become relevant later.
• Prefer file paths and brief state summaries over pasting artifact contents into manager prompts.
• When dispatching, read the relevant file in agents/, then launch a generic subagent with that file's instructions plus a compact task prompt.
• Prefer a fresh subagent session when the role changes or a sprint ends.
• Prefer a fresh subagent session when the current session reaches the context reset threshold, even if the role does not change.
• After one subagent finishes, re-check the artifacts and route the next subagent instead of stopping early.
• Do not use --agent planner, --agent builder, or --agent evaluator unless those names are actually registered in the runtime.
• When dispatching work that may start local servers or run integration checks, remind the subagent to use only bounded commands or detached background processes with redirected stdio.
• Treat a missing handoff.md after apparent implementation success as a likely builder-session hang during verification, not as evidence that review should be skipped.
• Treat a refreshed handoff.md that explicitly requests continuation as a normal reset signal. Dispatch a fresh subagent of the same role when that is the shortest path forward.

Context Reset Threshold

Do not rely on exact token counts. Use the following operational threshold instead.

Reset the current subagent when any of these becomes true:

• the session has already gone through 12 or more substantive tool calls
• the session has already gone through 3 or more build, test, or verification rounds and meaningful work still remains
• continuing safely would require carrying more than about 10 project files or large command outputs in active context
• the subagent is starting to restate prior state from memory instead of re-reading artifacts
• the subagent has completed a coherent slice of work, but more than a small follow-up remains and a fresh session would be cleaner

When the threshold is reached:

• refresh handoff.md immediately
• record exact current state, finished work, remaining scope, blockers, and the next concrete actions
• stop the current subagent after the handoff write succeeds
• launch a fresh subagent that starts from artifacts, especially handoff.md, instead of continuing the bloated session

Routing Rules

1. Missing or stale product_spec.md or sprint_contract.md -> planner
2. Approved sprint ready to implement -> builder
3. Build round stopped and handoff.md refreshed -> evaluator
4. Blocking qa_report.md -> planner by default, builder only for small clear fixes
5. Current role requested reset in handoff.md -> fresh subagent of the same role unless another routing rule has higher priority

Orchestration Rule

In one manager invocation, keep dispatching the next required subagent until the stop condition is satisfied.

• Typical loop is planner -> builder -> evaluator, then repeat as needed.
• After each subagent finishes, re-read the relevant artifacts and decide the next hop.
• Stop only when the completion criteria are satisfied.

Dispatch Prompt Standard

When starting the next subagent:

• read the corresponding role file from agents/
• pass that role file's instructions into a generic subagent run
• include the task brief in one short paragraph
• name only the relevant artifact files
• state the exact output files or outcomes expected from that subagent
• if handoff.md exists, tell the subagent to treat it as the primary continuity source for the current round
• remind the subagent to read external skills and references progressively instead of loading everything up front
• avoid long summaries when the files already exist on disk
• do not inline replacement instructions that bypass the subagent file

Stop

Stop only when:

• qa_report.md has no unresolved blocking findings
• the current sprint Done Means is satisfied
• handoff.md is sufficient for continuation or closeout
• 如果你是codex，全部是使用gpt5.4 high.