wayfinder
Use to map multi-session work through issue-tracker decision tickets.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 4
- Forks
- 1
01 前言 Front Matter
- name
- wayfinder
- description
- Use to map multi-session work through issue-tracker decision tickets.
- disable-model-invocation
- true
03 SKILL.md 正文
Wayfinder
Use when the destination is larger than one context window and the route is foggy. A decision ticket resolves a question, not a build slice. Load skills per dependency loading.
Plan, don't do
Wayfinder stops when nothing remains to decide. Notes hold planning preferences and do not authorize implementation or delivery.
Invariants
- Refer to maps and tickets by name, not bare IDs.
- The map is an index, not a store. Answers live in tickets; the map keeps a gist and link.
- If
CLAUDE.mdexists, readCLAUDE.mdfirst; otherwise readAGENTS.md. Follow its Issue tracker pointer and Wayfinding operations. If absent, use the local-markdown fallback. - Claim a ticket by assigning it before work; this must be the first write. Open and unassigned means unclaimed.
- Use only
wayfinder:map/wayfinder:<type>labels, neverready-for-agent; dispatch by the type label. - Prefer native blocking. Use
Blocked by:only when native dependencies are absent. - Resolve at most one ticket per session in the primary context; explicit delegation or an
invoked
/swarmmay parallelize ready research. Otherwise do not resolve another ticket. - For authorized waves,
/efficient-frontierbudgets lanes and the coordinator synthesizes. Start each authorized research lane with a fresh context window and ticket/source pointers. - Audit another session's claims or resolved ticket with
/agent-watchdog.
Map shape
Create one issue or file marked wayfinder:map:
## Destination
<the spec, decision, or change this map must reach>
## Notes
<domain, required skills, standing planning preferences>
## Decisions so far
- [<closed ticket name>](link) -- <one-line answer gist>
## Not yet specified
<in-scope fog not sharp enough to assign>
## Out of scope
<work beyond the destination>
Query the tracker for open children; do not copy them into the map body.
Tickets
Each child is one focused question sized to one 100K-token agent session. Mark it HITL (live human judgment) or AFK (agent-driven); HITL and AFK must not be conflated:
- Research (AFK): Call the Skill tool with "research" against primary sources. Follow its artifact location; parallel lanes require consent and each lane does not invent a root file or branch.
- Prototype (HITL): Call the Skill tool with "prototype" and link the artifact.
- Grilling (HITL): Call the Skill tool twice, for "grilling" and "domain-modeling".
- Task: manual work that unblocks a decision; never delivery for its own sake.
The answer is not part of the body; record it on resolution. Assets are linked, not pasted.
Fog and scope
Do not chart what is invisible. Not yet specified excludes anything decided, already a ticket, or out of scope. A sharp but blocked question is a ticket. Fog points toward the Destination. Close a ticket discovered beyond it, explain one Out of scope line, and do not record it as a route decision.
Chart the map
- Name the Destination with separate Skill tool calls for "grilling" and "domain-modeling".
- Grill breadth-first for decisions and first steps. If this surfaces no fog, stop and ask whether to proceed without a map.
- Create the map, then only currently specifiable tickets.
- Attach every ticket through the native child/sub-issue relationship when available; otherwise link it. Re-read and verify every ticket appears as a child, then add blockers and cross-references using real IDs; placeholders can link unrelated issues.
- Resolve one ready Research ticket inline. Authorized lanes claim first, follow
/research, and return cited artifacts. - Stop after that ticket.
Work through a map
- Read the map at low resolution; select the named or first open, unblocked, unclaimed ticket.
- Claim first, then resolve it using only relevant related tickets and skills from Notes.
- Record the answer, resolve the ticket, and append its gist and link to Decisions so far.
- Add new tickets and blockers; remove graduated fog. Rule beyond-destination work Out of scope.
- Re-read tracker state before writes because other sessions may edit concurrently.
Handoff
When clear, tell the human to run /to-spec for one buildable plan, then /to-tickets.
Recheck your claims first: reread answers, gists, assets, claims, and current tracker state.
End with one command for the recommended ticket and one per safe parallel frontier ticket.
04 同仓库更多 Skills
hill-climb
Optimize a defined outcome through bounded evaluation loops. Use for hill-climb requests or repeated improvement against an executable eval, benchmark, or visual comparison; not a one-pass review.
gestures
Build or fix web drag, swipe, sheet, reorder, and hold interactions: direct tracking, rubber banding, release velocity, momentum, snap points, and interruption. Not native React Native or Swift.
hill-climb
Optimize a defined outcome through bounded evaluation loops. Use for hill-climb requests or repeated improvement against an executable eval, benchmark, or visual comparison; not a one-pass review.
gestures
Build or fix web drag, swipe, sheet, reorder, and hold interactions: direct tracking, rubber banding, release velocity, momentum, snap points, and interruption. Not native React Native or Swift.
improve-animations
Audit existing web UI motion and prioritize concrete improvements or self-contained implementation plans. Use for a motion roadmap; animate owns implementation and diff review.
effect-ts
Use when setting up a repository that uses the Effect TypeScript library.