WCAG 2.2 AA compliance, ARIA patterns, keyboard navigation, screen reader optimization
中文简介正在准备中。当前显示原文说明。
Implement technical plans from thoughts/shared/plans with verification
中文简介正在准备中。当前显示原文说明。
查看安装方法安装前可以查看交给 AI 代理的指令内容。
Five modal logics via fusion with bridge principles:
# Justification terms
[h]:context(task_n) # Handoff h justifies task context
[v]:verified(phase_n) # Verification v justifies completion
[p]:plan(tasks) # Plan p justifies task list
# Evidence production
[read(f)]exists e. [e]:content(f)
[verify(c)]exists v. [v]:pass(c) | [v]:fail(c)
# Handoff chain: evidence propagates
[h_n]:complete(task_n) -> [h_{n+1}]:context(task_{n+1})
proceed(task) <-> exists h. [h]:validated
# Mode and blocker questions
?{direct, orchestration} # Mode selection
?{continue, retry, ask_user} # Blocker resolution
mismatch(plan, reality) -> ?{how_proceed}
no_validation -> ?{run_validation_first}
# File reading constraints
[](mentioned(f) -> <>read_fully(f)) # Eventually read
[](mentioned(f) -> not spawn U read_fully(f)) # No spawn until read
[](partial_read(f) -> false) # Partial reads forbidden
# Phase sequencing
[](phase(n) -> P(phase(n-1) & verified(n-1))) # Verified before next
[](automated_pass -> <>manual_verify) # Automated gates manual
[](manual_pass(n) -> <>phase(n+1)) # Manual gates next phase
# Handoff persistence
[](handoff_created(h) -> []exists_on_disk(h)) # Survives compaction
# Termination
<>(all_complete | abandoned)
# Reading obligations
O(read_fully(plan))
O(read_fully(f)) <- mentioned_in_plan(f)
O(check_existing_checkmarks)
F(partial_read)
# Verification obligations
O(run_automated) <- impl_complete
O(pause_for_manual) <- automated_pass
O(present_manual_checklist)
F(checkoff_manual) <- not user_confirmed
# Mode selection
O(orchestration) <- tasks >= 4
P(direct) <- tasks <= 3
O(respect_user_preference)
# Orchestration obligations
O(read_previous_handoff) <- exists_handoff(task_{n-1})
O(create_handoff) <- agent_completes
O(update_ledger) <- task_complete
F(batch_tasks) # One agent per task
F(proceed_on_mismatch) <- not user_guidance
# Implementation actions
[read(plan)]K(tasks) & K(phases) & K(criteria)
[read(handoff_n)]K(context_{n+1})
[spawn(agent, task)]<>result(agent)
[verify(c)](K(pass) | K(fail))
# Composed workflows
[select_direct][implement ; verify_auto ; present_manual ; wait]*
[select_orchestration][prepare ; spawn ; wait ; read_handoff ; update]*
# Recovery
[compaction ; read_ledger ; list_handoffs ; read_last]resume
# Mismatch
[detect_mismatch ; stop ; present ; wait]proceed_or_abort
# Evidence persistence (JL-TEL)
[h]:context(n) -> [][h]:context(n)
# Evidence obligations (JL-SDL)
O(exists h. [h]:validated) <- pre_implement
O(exists v. [v]:pass(auto)) <- pre_manual
# Handoff chain (full integration)
[h_n]:complete(n) -> O([spawn]<>[h_{n+1}]:context(n+1))
compaction -> (forall h. persists(h))
INIT --> READ_PLAN --> MODE_SELECT --+--> DIRECT: [IMPL -> AUTO -> MANUAL -> WAIT]*
|
+--> ORCHESTRATION: [PREP -> SPAWN -> WAIT -> HANDOFF]*
|
v
COMPLETE
handoff_path: "thoughts/handoffs/<session>/task-[NN]-[desc].md"
schema:
required: [status, task_desc, files_modified[], verification_results, context_for_next]
optional: [blocker, decisions[], open_questions[]]
tracking:
plan: "- [x] Task N: description"
ledger: "[x] Task N"
| Tasks | Context Critical | Mode |
|---|---|---|
| 1-3 | No | Direct |
| 1-3 | Yes | Orchestration |
| 4+ | Any | Orchestration |
User preference overrides.
Mismatch:
Issue in Phase [N]:
Expected: [plan says]
Found: [actual]
How should I proceed?
Manual Verification Pause:
Phase [N] Complete - Ready for Manual Verification
Automated passed: [list]
Please verify: [manual items from plan]
Let me know when done.
Agent Spawn:
Task(subagent_type="general-purpose", model="claude-opus-4-5-20251101", prompt="""
[implement_task SKILL.md]
## Context
- Ledger: [content]
- Plan: [section]
- Task: [N]/[Total]: [desc]
- Previous Handoff: [content or "first task"]
- Handoff Dir: thoughts/handoffs/<session>/
""")
Recovery (post-compaction):
ls thoughts/handoffs/<session>/forall phase. has_auto_criteria(phase) & has_manual_criteria(phase)
forall task. one_agent_per_task(task)
forall h. on_disk(h) -> recoverable(h)
compaction -> (forall h. persists(h))
forall i < j. completed(task_i) before started(task_j)
还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
WCAG 2.2 AA compliance, ARIA patterns, keyboard navigation, screen reader optimization
中文简介正在准备中。当前显示原文说明。
axe-core integration, WCAG 2.2 AA checklist, keyboard navigation testing, screen reader testing, and ARIA pattern validation.
中文简介正在准备中。当前显示原文说明。
Steam-style achievement system with XP, levels, streaks, and skill trees. Gamifies the development workflow. 25 achievements across 5 categories.
中文简介正在准备中。当前显示原文说明。
Framework for measuring and tracking agent response quality over time. Detects regressions before they reach production. Use when evaluating agent changes, auditing quality, or establishing performance baselines.
中文简介正在准备中。当前显示原文说明。
Agent Context Isolation
中文简介正在准备中。当前显示原文说明。
Agent ve skill dosyalarinin yapisal dogrulamasi. Frontmatter kontrol, naming convention, zorunlu bolum kontrolu, tutarlilik denetimi. Yeni agent/skill eklendiginde veya mevcut dosyalar duzenlediginde otomatik calistirilir.
中文简介正在准备中。当前显示原文说明。