to-goal
Turn an approved spec, agent-ready ticket, tracker frontier, or partially implemented ticket into a verifiable execution goal without re-interviewing the user. Use after to-tickets or triage, before starting a fresh implementation session; use --all only for an explicitly requested cross-ticket goal.
tt-a1i · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add tt-a1i/matt-skills-with-to-goal --skill to-goal技能说明
name: to-goal description: Turn an approved spec, agent-ready ticket, tracker frontier, or partially implemented ticket into a verifiable execution goal without re-interviewing the user. Use after to-tickets or triage, before starting a fresh implementation session; use --all only for an explicitly requested cross-ticket goal. disable-model-invocation: true
To Goal
Compile existing planning and repository evidence into an execution goal. Do not implement, mutate the tracker, create a branch, or modify files.
Source of truth
Read ../goal-crafter/SKILL.md only for compiled-handoff mode, Phase 2 harness formats, Phase 3's four self-checks, and Special Rules (never re-interview; one verifiable condition per checkbox). Skip Phase 1 and Examples. This skill owns context gathering, frontier selection, readiness checks, and execution handoff. Do not reopen decisions already made by to-spec, to-tickets, or triage.
Accepted inputs
Resolve one of:
- no argument: inspect the configured tracker and select the current unblocked, agent-ready frontier ticket;
- a ticket number or URL: read that ticket in full;
- a parent spec issue: read the spec, sub-issues, and blocking graph, then select its current frontier;
- a local spec or ticket path: read the complete file and any directly referenced local planning document (local tracker:
.scratch/<feature>/spec.mdand one file per ticket under.scratch/<feature>/issues/<NN>-<slug>.md); --all <parent>: generate one dependency-ordered cross-ticket goal.
Always read ticket comments. For a tracker parent, use native sub-issue and dependency relationships when available; otherwise use explicit blocker text. Do not infer that a ticket is ready merely from its label.
If no argument yields several frontier tickets, list them and ask the user to choose one. Do not silently combine them. If a requested ticket is blocked, report its blockers and do not generate an implementation goal.
For a spec with no tickets, generate a goal directly only when the entire work fits one fresh context window. If it is multi-session work, route it through to-tickets instead of bypassing the context boundary.
Gather current evidence
Before drafting:
- Read the complete source spec and selected ticket, including acceptance criteria and comments.
- Inspect the repository instructions and relevant design vocabulary.
- Inspect the current branch, HEAD, worktree status, recent commits, and diff. Record the pre-implementation HEAD as the code-review fixed point.
- Compare current behavior and tests with every acceptance criterion.
- Classify criteria as evidenced complete, demonstrably incomplete, or unverified. A commit message is not evidence.
- Discover validation commands from the repository's own scripts, CI, documentation, and existing tests.
- Preserve user-established permissions and workspace boundaries from the source context.
Keep this work read-only. Do not create status artifacts merely to build the goal.
Select scope
Default to exactly one frontier ticket. The generated goal must fit one fresh context window and must not attract work from downstream blocked tickets.
For --all:
- preserve the complete dependency order;
- distinguish the current frontier from future work;
- carry forward partially completed work without treating it as done;
- warn that the goal requires a persistent harness capable of context renewal;
- never present
--allas the normal Matt workflow.
Recommend execution capacity
Classify the implementation session by required capability, not by a hard-coded model name. The recommendation must remain portable across Claude Code, Codex, Pi, and other coding agents.
Choose exactly one capability tier:
- Lightweight: bounded search, inventory, formatting, mechanical edits, or a small change following an established pattern with low failure cost.
- Standard: normal feature work, focused bug fixes, tests, or moderate multi-file changes with clear repository patterns. This is the default.
- Advanced: difficult root-cause analysis, cross-module design, security or authorization changes, schema/data migrations, concurrency, long-context synthesis, or work where a plausible mistake has high cost.
Choose exactly one reasoning intensity:
- Low: deterministic work with little ambiguity and cheap verification.
- Medium: some design judgment, multiple affected files, or non-trivial tests. This is the default.
- High: ambiguous behavior, interacting invariants, risky migrations, concurrency, security boundaries, or expensive failure modes.
Recommend the lowest tier and intensity that can reliably complete the ticket. Include one short evidence-based reason. Do not recommend a stronger tier merely because the ticket is large; prefer splitting work when it cannot fit one fresh context window.
Only name a concrete model when the target harness and its available model choices are known from current context. When naming one, present it as an optional mapping after the portable recommendation, not as the recommendation itself. Never assume a fixed set such as Luna, Terra, or Sol.
Readiness checklist
This list is for this compiler; do not put it in the paste block.
Required propositions — tick every item before drafting. If any of these is unchecked, stop and do not invent a goal.
- Source is agent-ready: every required product decision and completion condition is in the evidence. If not, stop, name what is missing, and do not reopen a planning interview.
- Ticket is unblocked. If blocked, report the blockers only; do not generate an implementation goal.
- Exactly one frontier. If several, list them and ask the user to choose; do not silently combine them.
- Pre-implementation HEAD recorded as the code-review fixed point.
- Every acceptance criterion classified: evidenced complete / demonstrably incomplete / unverified.
- Valida