ToRvaLDz/trailhead

Start and drive large projects as a map of tickets on GitHub Issues, resolving one at a time until the way to the destination is clear.

trailhead is an orchestrator skill for coding agents, running on Claude Code and Codex CLI from a single source. It gives a big, foggy idea a place to begin (a trailhead) and a disciplined way to walk it to a working result, with the whole plan living on your issue tracker instead of in scattered local files.

Two approaches to agent-driven project work each nail one half of the problem:

• Wayfinder (by Matt Pocock) is brilliant at starting. It turns a loose idea into a shared map of decision tickets on your issue tracker, named through a proper grilling conversation, with a visible frontier and a "fog of war" for what isn't sharp enough to plan yet. You resolve one decision per session and the map grows outward. What it deliberately doesn't do is build: it plans, then hands off.

• GSD (the open-gsd workflow system) is brilliant at doing. Its discuss → plan → execute → verify cycle, atomic commits, and zero-friction capture make the actual construction reliable and reviewable. But its state lives in a local .planning/ tree, separate from where the team can see it.

Working across both, one pattern kept recurring: use a Wayfinder-style map and onboarding to decide the shape of the work, then a GSD-style engine to build each piece, without ever leaving the issue tracker. trailhead is that pattern made into a set of cohesion-clustered skills over a shared core, its own self-contained take on both ideas rather than a wrapper around either.

trailhead is self-contained in the sense that it has zero external-plugin dependencies: trailhead-owned agents and the host's built-ins are fine, and no other skill or plugin is required. Wayfinder and GSD are inspiration, not dependencies. Every technique those systems package as a separate skill (grilling, TDD, systematic debugging, codebase mapping, code review) is built in here as an inline protocol. trailhead's core invokes no other skill; each engine technique dispatches to trailhead's own trailhead-* agent (trailhead-executor, trailhead-plan, trailhead-research, trailhead-code-review, trailhead-debug, plus trailhead-fix and trailhead-codebase-map), authored once as a committed agents/trailhead-*.md file and projected to each host (Codex gets ~/.codex/agents/trailhead-*.toml generated from that same source). These agents are thin (identity + tool policy + a config-driven model, with the technique protocol injected at spawn) and run on the host's native subagent toolkit (Claude Code's built-in Agent tool, or Codex's multi_agent toolkit). The only thing to install is trailhead itself. There is one opt-in exception: hosted mockups. design: artifacts uses Claude Code's built-in Artifact tool (nothing to install), and design: stitch uses the Google Stitch MCP (offered as a one-time install). See Mockups before code.

The design choices that fall out of it:

• Everything lives on GitHub Issues. The map is the parent issue; each ticket is a child issue; discussion, plan, and verification are comments. There is no .planning/ directory: the repo holds code only, and the plan is always visible to anyone with the repo.

• Execution happens inside the map. Unlike pure Wayfinder ("plan, don't do"), build tickets graduate from the fog and are executed as children of the map. The destination is the working artifact (a deployed app), not a spec document, though a project can override that and stop at the spec.

• Lean by default. The heavy machinery (codebase mapping, systematic debugging) runs once where it earns its keep; the per-ticket cycle stays light.

• Built for teams, because the plan is shared. The tracker is the single source of truth, so the map is a shared workspace, not one person's local file. Many people (and their agents) work it at once, and the mechanics keep them from colliding:

  • Claiming. You take a ticket by assigning it to yourself; the assignee is the lock. A collision stops you and asks, rather than clobbering someone else's work. Whoever claims a ticket owns it end to end and closes it, no separate approver.
  • A shared frontier, one ticket per session. The frontier (open, unassigned, unblocked, not paused tickets) is the common queue of what's takeable now; each person picks a different one. Resolving a ticket ends the session, so work stays in reviewable, single-ticket chunks.
  • Isolation for concurrent sessions on one machine. Two agents editing one checkout corrupt each other, so the isolation: convention (default auto, which asks at each ticket's work-start whether you'll work more than one at once and then picks the right mode by repo shape) gives each in-flight ticket its own git worktree (or a full clone for path-bound apps, or repos with submodules, a worktree can't safely isolate) when you run more than one at once; disjoint Scope: lines let non-overlapping work run in parallel and serialise only where it truly overlaps.
  • Pause, resume, split. Any ticket can be paused with a checkpoint comment and resumed later by anyone (once released), or split into children when it outgrows a session, with blockers re-pointed so the frontier stays honest.
  • Several maps at once. A repo can carry more than one live map (parallel milestones or features), each with its own scoped frontier, so two people can drive two efforts on the same repo without stepping on each other, and you can park one map to work another.
  • Outsiders and trust. Issues opened by other people flow through an inbox that reframes the worthwhile ones into tickets in place (keeping the reporter's authorship), while a trust and provenance guard quarantines trailhead:*-labelled issues from untrusted sources until a maintainer adopts them, backed by a repo-side label-guard workflow.

  The full protocol lives in the Working as a team section below.

Runs on Claude Code and Codex CLI. trailhead is authored once; the installer projects host-native artifacts. On Claude Code: a skill + /trailhead:* commands + hooks. On Codex CLI: a native skill invoked $trailhead, one $trailhead-<verb> skill per verb for discoverability, and the guardrails as native Codex hooks. Choose the host with --claude / --codex, or let the installer auto-detect. More hosts can follow (the installer uses a per-host descriptor).

Prerequisites: an authenticated gh CLI (the tracker is GitHub Issues) and a GitHub repo to work in. The npm path also needs Node 18+.

/plugin marketplace add ToRvaLDz/trailhead
/plugin install trailhead@trailhead

Update with /plugin update trailhead; remove with /plugin uninstall trailhead.

npx @marcomigozzi/trailhead              # auto-detect the host (defaults to ~/.claude)
npx @marcomigozzi/trailhead --claude     # Claude Code (~/.claude or $CLAUDE_CONFIG_DIR)
npx @marcomigozzi/trailhead --codex      # Codex CLI ($CODEX_HOME or ~/.codex)
npx @marcomigozzi/trailhead --symlink    # dev install (Claude; symlink to the checkout, edits go live)
npx @marcomigozzi/trailhead --uninstall  # remove everything it added (pair with --claude / --codex)
npx @marcomigozzi/trailhead --dir=<path> # target a specific config dir

On Claude Code it copies the skill (+ its references/), the /trailhead:* commands, the hooks (into hooks/, registered in settings.json) including a SessionStart update check, and the templates (label-guard workflow, statusline, the git commit-msg hook and the helper scripts the engine runs), idempotently.

On Codex CLI (--codex) it projects a native Codex skill invoked as $trailhead ($trailhead work, $trailhead new "idea", …), one thin $trailhead-<verb> skill per verb for discoverability, the seven guardrails as native Codex hooks (~/.codex/hooks.json, enabling features.hooks; the four Bash guards share one process behind a single registration; the search guard is Claude-Code-only and isn't projected to Codex), and per-technique subagent model pins from models.codex.* into ~/.codex/agents/. Codex 0.145.0+ is required (the installer gates on codex --version). The artifacts mockup mode (Claude Code's Artifact tool) isn't available on Codex; the cross-host design: stitch mode (Google Stitch MCP) is the hosted option there, with local disk as the fallback.

Re-run the install to update, or run /trailhead:update (Claude) / $trailhead update (Codex) from inside the agent: it detects how trailhead was installed and installs the newer version where that is safe (a git pull for a dev-symlink, an npx re-run for npm, or /plugin update for the plugin). When a newer version exists, the statusline shows a ⬆ trailhead <version> flag (Claude).

Restart or reload your agent so the commands register, then run /trailhead to start (smart entry), or /trailhead:new "<idea>" to chart a map. On Codex the surface is $trailhead instead: run $trailhead for smart entry, $trailhead new "<idea>" to chart, or a per-verb $trailhead-<verb> skill. Note: once installed, the commit guard hook runs on every git commit (enforcing Conventional Commits and blocking Co-Authored-By); disable the plugin's hooks in settings if you don't want that. Either way trailhead is self-contained: no external skill or plugin is required, no external-plugin preflight, no version drift.

A map is one GitHub issue labelled trailhead:map. It's an index, not a store: a Destination, standing Notes, the Decisions so far, the Not yet specified fog, and what's Out of scope.

What earns a map? A map is the unit for an effort that is too big to hold in one session and carries open decisions, not just execution. In practice that's a milestone (a release, a version, "v2 of the dashboard"), or a single feature big enough that you must decide before you can build it: where the data lives, which library, what the screen actually does. Charting a map names its destination and fans that effort out into many tickets, wired by their dependencies. As the decisions resolve, the fog clears and the build tickets graduate, until the destination is reached.

What does not need its own map is a single, defined piece of work with no open questions, small enough to finish in one session. That is just one ticket, a direct build (or a todo): a work you take with /trailhead:work and resolve in a sitting, no charting required. Rule of thumb: decisions to make, or too big for one sitting → a map; "just do X" → a single work. A feature that starts as a lone work but turns out to need decisions gets promoted to its own map (or split); a map whose questions all evaporate collapses back to a handful of works. And a repo can hold several maps at once (parallel milestones or features), each with its own frontier, so you can park one and work another.

Each ticket is a child issue with a type label and a one-question body. The frontier is the set of open, unassigned tickets whose blockers are all closed and that are not paused: what's takeable right now. You claim a ticket by assigning it to yourself, resolve it with the engine for its type, then record the answer as a comment, close it, and gist it back onto the map.

As tickets resolve, the fog clears: questions that were too vague to phrase become sharp enough to ticket, one at a time, until nothing is left to decide or build and the destination is reached.

A map is scoped to one effort, so knowledge that belongs to the repo (not to any single map) lives in repo-scoped anchor issues, created once and shared by every map (each map's Notes just links them, so nothing is stranded when a map finishes):

• trailhead:codebase: the distilled codebase map (architecture, stack, conventions, risks, test/build), written once at adopt and refreshed only on major drift.
• trailhead:conventions: the project's way of working, readable by everyone: a small machine-read header the engine obeys (git: main|pr, release: command|auto, isolation: auto|none|worktree|clone) over human prose. /trailhead:adopt and :new ask for it up front. isolation: auto (the default) decides per ticket at work-start: it asks whether you'll work more than one ticket at once and, if so, uses a git worktree (plain repo) or a clone (submodule superrepo); otherwise it works in the current checkout. isolation: worktree gives each executing ticket its own git worktree + branch (concurrent sessions on one clone never share a working tree); isolation: clone gives it a dedicated clone instead, for path-bound apps a worktree can't build and for submodule-bearing superrepos (where worktree is escalated to clone).
• trailhead:dashboard: the pinned index of the whole surface, a link to every open map (with GitHub's native progress bars), the whiteboard, and live counts (inbox, whiteboard frontier). Refreshed when a map is charted or exhausted, when a whiteboard ticket is born or resolved (it has no native progress bar, so the dashboard is the only place it shows; a quick-born ticket, worked and resolved in one sitting, refreshes at its handoff or pause, not at creation), and on demand via /trailhead:dashboard.

These three fill GitHub's 3 pinned-issue slots, so a repo's trailhead anchors stay one click away; maps themselves are never pinned (they're indexed by the dashboard instead, since the "active" map is per-checkout local state, not a repo-global fact). Project config (models, TDD, design…) is separate again, a plain .trailhead/config.json file at the repo root, never in an issue.

The lifecycle. Every project walks the same loop:

1. Start: /trailhead:new "<idea>" (greenfield) or /trailhead:adopt (existing code: map the codebase once, then go lean). This charts the map: name the destination, then map the frontier breadth-first into the first tickets and the fog.
2. Work the frontier, one ticket per session: /trailhead:work takes the next takeable ticket (or one you name) and runs the engine for its type: research gathers a fact · decision diverges options then grills to choose · prototype makes a rough artifact to react to · build runs discuss → plan → execute → verify · bug runs repro → diagnose → fix → verify · task is manual plumbing that unblocks a decision. /trailhead:auto chains the frontier unattended, one ticket after another, stopping only at the safety rail, fog, or a human-necessary decision. At a ticket's start, work and quick show a short brief with a Proceed / Change something choice; it is skipped (a one-line note instead) for a ticket you created or just settled in a grilling this session, and auto never asks.
3. Close & unblock: resolve the ticket (comment + close), gist it into Decisions so far, and drop trailhead:blocked from any dependent whose last blocker just closed, graduating it onto the frontier.
4. Repeat until the frontier is empty and the fog has cleared: the destination (a working artifact) is reached.

Along the way: capture ideas mid-work without derailing, split a ticket that grew too big, pause/resume across sessions, and let several people work unblocked tickets in parallel.

Destination: users sign in with Google and GitHub, alongside email/password. Charting fans it into six tickets (one of each type) wired by their dependencies:

① research   "OAuth providers + library: Supabase Auth native or Auth.js?"   ← frontier (AFK)
② decision   "approach: Supabase Auth native vs Auth.js custom"              ← blocked by ①
③ task       "register OAuth apps on Google + GitHub, get client id/secret"  ← frontier (HITL)
④ prototype  "how the login screen with social buttons looks"                ← frontier (HITL)
⑤ build      "implement Google + GitHub login"                               ← blocked by ②③④
⑥ bug        (appears after ⑤ ships)

Initial frontier = ①③④, three people can start in parallel. Then:

• ① research → a subagent reads the docs, returns a decision-ready finding, closes → unblocks ②.
• ② decision → diverge the options, grill to choose native vs custom, record the why.
• ③ task → the agent hands you a checklist (it needs your Google/GitHub accounts); you register the apps, it records where the credentials live.
• ④ prototype → a rough login screen (on the configured Artifacts canvas), approved before UI code.
• With ②③④ closed, ⑤ build graduates → discuss → plan → execute (atomic commits, TDD at the auth seams) → verify (tests + code review + acceptance: asked who drives, the agent goes AI-driven and drives the browser through the real login flow).
• After it ships you spot a wrong redirect in prod → /trailhead:bug --of ⑤ "GitHub redirect goes to localhost" → a new ticket carrying Regression of: ⑤, worked repro → diagnose → fix → verify. ⑤ stays closed.

① research ─┐
③ task ─────┼─► ② decision ─┐
④ prototype ┘                ├─► ⑤ build ──(ship)──► ⑥ bug (Regression of: ⑤)
③ task ──────────────────────┘

The preparatory types (research/decision/prototype/task) unblock the constructive ones (build/bug); parallelism is real but across sessions on the frontier, not inside one.

The app is live, and a user not on the team files a plain issue: "can we add Apple sign-in too?" No trailhead:* label, so it isn't a ticket yet. Here's how an outside voice becomes a first-class ticket on the map, without losing the reporter's authorship:

1. Inbox: /trailhead:inbox lists it under New: to triage. It's in scope but not sharp: web only or native? does it need the paid Apple Developer account? You can't phrase one answerable Question yet, so you don't force a ticket.
2. Fog: label it trailhead:fog and keep the issue open as a clarification space. A comment posts the sharpening questions (async grilling); the reporter and anyone interested refine it in the thread. It's off the frontier but tracked: gh issue list --label trailhead:fog is the durable list, and the reporter follows their own issue.
3. Graduation: the discussion converges. Next time you run /trailhead:inbox, its Parked fog section surfaces the thread as recently-active: that's how you notice the fog cleared. Now the Question is phrasable, so you adopt it in place: swap trailhead:fog for trailhead:ticket + a type label, add the Parent: line and the ## Question. Same issue number, same reporter credited, it lands on the frontier, ready for /trailhead:work.

suggestion issue ──inbox──► 🌫️ fog (kept open, discussed) ──sharpens──► 🎫 ticket on the frontier
                                                                       (same #, reporter still credited)

No separate backlog, no lost credit: the map absorbs external suggestions the same way it grows its own, through fog that graduates when it's sharp.

trailhead is user-invoked (it won't fire on its own). The first word is the verb; the rest is text or a ticket number. With no verb, /trailhead does smart entry: it inspects the repo and proposes the right next move.

Every verb is also a namespaced command (/trailhead:new, /trailhead:work, /trailhead:bug, …) so typing /trailhead lists them all in the command picker. /trailhead <verb> and /trailhead:<verb> are equivalent.

The trailhead mod (Claude Code 2.1.287+) ships inside the plugin and draws the work beside the conversation. Codex keeps the command hooks and has no panel.

The map panel lists the open maps to view, the frontier with a work button per ticket that clears the context and starts the ticket, who holds what (your own ticket is marked this session), the pending UAT follow-ups in a UAT group (each naming the ticket it tests), the paused ones in a Paused group, and what is blocked. The footer has refresh and close. /trailhead:panel toggles it, and it opens by itself once in fullscreen.

A click only views a map, and back to active returns to the active one. The active map changes when you work one of its tickets. A Whiteboard row above the maps list, with its N open / M counts, shows the repo index instead: every open map (a click opens it) and the whiteboard's loose tickets by state, each frontier one with its work button; map goes back.

The burn-down shows the open tickets over time, from the map's first ticket to now, with done and removed counted apart.

The pipeline band above the prompt and the status entry under it show the ticket in play and its step.

The four fog/ticket captures form a spectrum of commitment and timing: note < idea < seed < todo.

When a map is open, a capture that produces a ticket (todo/bug/seed/a sharp idea) asks whether to file it on the active map or the whiteboard, the home for loose, map-less work that doesn't belong to any map (or isn't worth charting one). With no map open it lands on the whiteboard. Work a whiteboard ticket with /trailhead:quick (which also opens one from "text" and works it in the same sitting), and see them all with /trailhead:whiteboard.

Deep in a map, something unrelated surfaces: a bug in another area, a chore, a quick idea you want to act on now. Forcing it onto the map's frontier pollutes the map; charting a whole new map for it is overkill. That is what the whiteboard is for, loose map-less work, and two moves keep you from getting stuck:

• Capture it aside. A todo/bug/seed/sharp idea fired while a map is open asks map or whiteboard?. Send it to the whiteboard and it stays off the map: tracked, but out of the way, so the map's frontier keeps meaning "the way to this destination".
• Do it on the fly. /trailhead:quick "<text>" opens a whiteboard ticket and works it end to end in the same sitting, the full discuss → plan → execute → verify engine (atomic commits, code review, the lot), except it grills only if needed and never splits, and skips every map book-keeping step. On a small piece it size-triages and offers to skip the heavy steps (on an explicit yes only; the correctness spine always stays), so quick stays fast without setting effort: lean. /trailhead:quick <n> does the same for a ticket that already exists.

See the whole whiteboard with /trailhead:whiteboard. Nothing about the map changes: you just stepped off it, did the thing, and step back on when you're ready.

The three that trip people up are idea, seed, todo, so here they are spelled out:

• idea = maybe, and not yet clear. It lands in the fog (Not yet specified), not as a ticket, because you can't even phrase the question sharply yet, so there is nothing to work. It graduates into a ticket later, when the frontier reaches it or it simply gets clearer.
• seed = yes, but not yet. It is a ticket, but parked (trailhead:blocked) on a trigger you name ("when the public API ships", "when we pass 1k users"). The question is already sharp; what's missing is a condition, not clarity. When the trigger fires, it graduates onto the frontier.
• todo = yes, now. Defined work you will just do: it's born a build ticket on the frontier, takeable immediately.

The two cuts that matter: idea vs seed is waiting on clarity vs waiting on a condition (both land "later", for different reasons); seed vs todo is committed later vs committed now. Below all three sits note (raw text to remember, maybe never work), and off to the side is bug (a defect, not a commitment tier). The test for idea vs ticket is always: can you phrase the question precisely now? Yes → a ticket (todo if takeable, seed if gated); no → the fog (idea).

Each ticket carries a type label; each type has its own way of being resolved.

Each type has its own inline engine: no external skill is invoked.

Two rules of thumb: build tickets never auto-grill: on blocking ambiguity the skill stops and asks; and brainstorming (divergence) lives in charting, in ticket's micro-charting, and in a decision's option phase, never in the grilling itself, which only converges.

When trailhead puts a forced choice to you, an advisory one (not a decision that shapes the map) may also offer "let the agent pick (with rationale, then confirm)", and any choice may offer "defer this decision", which parks it as a blocked decision ticket or a seed until its prerequisite resolves.

Every ticket that changes user-facing UI (a new screen, a redesign, a changed layout) is mocked up before any UI code is written. You react to the mockup, approve it, and the build reproduces it faithfully; acceptance testing later checks the result against it. Where the mockup lives is the design key:

A UI question that's still open gets 3 structurally different variants (different layouts, not colours or copy) to pick from, not one take.

you       › /trailhead:work                       # ⑤ build: "login screen"
trailhead › This ticket changes UI. Mock it up first?            → yes
trailhead › No canvas for this map yet: pick one, paste a URL, or create one?
you       › create one, "Acme app"
trailhead › Installed your design system "Acme DS" on the canvas.
            Canvas: https://claude.ai/artifact/…  ·  page "#42 Login"
            Artboards: Login-A (split hero) · Login-B (centered card) · Login-C (stepper)
            Which direction?
you       › B, but with the social buttons above the form   # or edit it on the canvas yourself
trailhead › Updated Login-B. Approve?                         → approve
trailhead › Re-read the approved artboard (your live edits included) → planning the build

• One canvas per map, cached in design.project; each ticket gets its own canvas page, each screen (and variant) its own artboard. App screens get a fixed frame (e.g. a 390×844 phone), web pages a fluid one.
• Your design system on the canvas. design.system (or your account's default) is installed, so the canvas editor offers its colours, type and components instead of look-alikes.
• Approve with the link in front of you. Every approval ask carries the canvas URL and the artboard names; the mockup-link guard blocks one that doesn't.
• Edit freely. Change the canvas directly or leave comments there; on approval trailhead re-reads the artboards and builds from that version, not from what it last published.
• design.approval: explicit (default) waits for your go-ahead before UI code; auto surfaces the mockup and proceeds.

design: stitch uses the Google Stitch MCP, Google's Gemini-powered UI generator, and is the hosted option that runs on both Claude Code and Codex. design.surface picks a Stitch project of screens (canvas, default) or a shared Stitch design system (design-system). If the MCP isn't connected, trailhead offers to install it once (decline and it falls back to disk):

claude mcp add --scope user --transport http stitch https://stitch.googleapis.com/mcp

Where artifacts can't run (Codex, or an API-key session), trailhead offers to switch to stitch before falling back to disk, and says so: it never silently swaps a hosted mode for a local file.

Moving from claude.ai/design. The standalone Claude Design site closes on 2026-12-14 and its designs now live in Claude as Artifacts. A config that still says design: claude.ai/design keeps working as artifacts (trailhead tells you once and offers to rewrite the key), but a cached claude.ai/design project can't be reused: migrate your design systems from the Artifacts page banner and let trailhead attach a new canvas.

Passing tests prove the code does what the code intended; acceptance testing proves it does what you wanted. For anything you can see (a new screen, data that now appears, a changed flow), a green suite does not close the ticket: Verify runs a UAT too, after tests and code review.

Before it starts, trailhead checks the base: it measures how far your working branch is behind the trunk, syncs it (fast-forward, rebase or merge; a conflict aborts and restores), re-runs the tests and runs your preflight.checks, and the same check precedes integration; anything it cannot verify stops the ticket. Then it prepares the environment: the uat.setup steps (start the dev server, boot the emulator; a step whose check already passes is skipped) run before the ask below, a failed step stops with a clear reason, and uat.accounts names the test logins (only $ENV references are stored). With no setup configured it can probe the project and propose one to save. /trailhead:auto runs the configured setup only.

1. It asks who drives. Before any step runs: AI-driven, human-driven or skip. It waits for a clear answer (silence re-asks). If no step is something the agent can do this session (a device it can't reach, credentials it doesn't hold, a judgement of feel or sound), it drops the AI-driven option and asks human-driven or skip, telling you why in one line. Skip is a waiver, not a deferral: you forgo acceptance, the VERIFY comment records it as waived by the user (unconfirmed), never as a pass, and no follow-up is opened. The agent never picks it for you, not even under /trailhead:auto.
2. AI-driven: it drives the app itself. For a web app it finds the target URL on its own (a running dev server, the dev script, the framework's configured port), proposes it for you to confirm, then navigates, fills, clicks and asserts what's on screen, with a screenshot as evidence in the ticket's VERIFY comment. Beyond the browser, it drives whatever device the session can reach: a phone or emulator over adb, an iOS simulator, a smart TV, a desktop app, through the commands and tools you have connected (a device MCP, computer use, the platform's CLI). Only the steps it truly can't do come back to you, one at a time.
3. Human-driven, or what the agent can't reach: a guided walkthrough. A device that isn't connected, an account it doesn't hold, a call on feel or sound: trailhead walks you through it one step at a time in the chat (what to do, where to tap, what you should see), waits for what you actually saw, then moves on. Never a checklist dumped on you to self-serve.
4. Against the approved mockup. When the ticket had one, the built screen is checked against it element by element and state by state (layout, spacing, colour, type, copy; empty, loading, error, disabled…), from the re-read mockup, never from memory. An unforced deviation is a finding, not "close enough".
5. Failures become tickets on the spot. A failed step records expected vs. seen and opens a bug follow-up (or a fix on this ticket) right away, then asks whether to carry on.
6. Nothing slips through as "done". If the UAT can't run now (the device isn't at hand, the backend isn't deployed), trailhead asks: run it now, or open a follow-up ticket, linked back with a UAT of: line and tagged with the trailhead:uat label so every pending UAT is findable by one label, that carries the full checklist. A user-facing ticket never closes on a UAT nobody ran unless you explicitly waive it, recorded as unconfirmed.

trailhead › Tests and review are green. Who drives the UAT: AI or you?   → AI
trailhead › Acceptance will run against http://localhost:5173, correct?   → yes
trailhead › ✓ Sign in with GitHub → lands on /dashboard        (screenshot)
            ✓ Login-B vs. the build: 14/14 elements, 4/4 states match
            ✓ Android app on the connected phone (adb): GitHub sign-in → home  (screenshot)
            ? Step 6 is the haptic on "Signed in": does it buzz once, softly?

Config: acceptance.browser (auto default, on, off) governs browser driving; off keeps AI-driven checks via commands and the suite, and still asks who drives. testing.webapp / testing.url say whether the app is browser-drivable and where. Under /trailhead:auto, it goes AI-driven; steps only you can do are never faked: the ticket is set aside and listed in the run's summary; /trailhead:auto never picks skip.

Everything the map needs is expressed as GitHub labels, so state is queryable in the tracker UI:

• Structural: trailhead:map, trailhead:ticket
• Repo-scoped anchors (one each per repo, pinned): trailhead:codebase 🧱 (the distilled codebase map), trailhead:conventions 📜 (the way of working), trailhead:dashboard 📊 (the pinned index of maps + whiteboard + counts); each title carries its icon (like the map's 🗺) so the pinned anchors stand apart at a glance
• Type (one per ticket): 🧭 trailhead:decision · 🔬 research · 🎨 prototype · 🔨 build · 🐛 bug · 🔧 task
• State: trailhead:blocked (has an open blocker) · seed (parked on a trigger) · out-of-scope (closed, beyond the destination) · superseded (closed, split into children) · paused (checkpointed, off the frontier) · unverified (quarantined until a maintainer adopts it) · fog (kept open for clarification)
• Container: trailhead:whiteboard (a loose, map-less ticket, off every map's frontier, on the whiteboard's own)
• Marker: trailhead:uat (a pending UAT follow-up; never a frontier filter) · trailhead:map-<n> (scopes a ticket to its map)

The frontier is then a single query (open, unassigned, not trailhead:blocked, not paused), no body-parsing needed.

Many people (and their agent sessions) share one map and work it concurrently:

• Claim = assign to yourself. The claimer owns the ticket end to end and closes it: there is no approver and no one to wait on. On a claim collision the session stops and asks you, rather than resolving it silently.
• Split an oversized ticket into children and supersede the original (/trailhead:split).
• Pause/resume via a PAUSED checkpoint comment and a trailhead:paused label (added by pause, removed by resume), so any session can pick a ticket back up. Paused tickets are listed under a Paused group in the map, whiteboard and dashboard views.
• Isolate concurrent sessions with worktrees. By default (isolation: auto) trailhead does this for you: at each ticket's work-start it asks whether you'll run more than one at once and picks the mode by repo shape. To force it eagerly, set isolation: worktree in the conventions header, and each executing ticket runs in its own git worktree on a trailhead/t<n> branch, integrated back to the trunk at Resolve. It implies a branch per ticket even under git: main (git won't check the trunk out twice), and it is the right posture for a monorepo (a worktree shares the object store, far cheaper than a second clone). One caveat: a worktree isolates source but is a fresh path with no installed deps, so a path-bound package (React Native/Expo, native toolchains, anything tied to a local node_modules) won't build from it. For those there's isolation: clone: a dedicated per-ticket clone that installs its own deps and builds/UATs like the real app (heavier than a worktree, and the right tool for an Expo/RN submodule). Or keep isolation: none and serialize. Otherwise run concurrent sessions in separate clones. A fresh worktree or clone has no .env: list the gitignored files it needs under isolation.copy in .trailhead/config.json and trailhead copies them in right after creating the workspace, never overwriting (see Configuration).
• A session-ticket marker for your statusline. When it starts working a ticket, trailhead drops a one-line, gitignored .trailhead/session-ticket (#<n> <title>) at the working root and clears it at handoff. It's a cheap, offline hint so a statusline (or any tool) can show which ticket this session is on without hitting the tracker; trailhead itself still treats the Issues as the source of truth. It also does more than that: while it's present, the git commit-msg hook requires every commit to carry a Refs: #<n> trailer matching it.
• It picks isolation for you by default (auto). With no explicit isolation: set (the default is auto), at a ticket's work-start trailhead asks whether you'll work more than one at once on this machine this session; if so it uses a git worktree (plain repo) or a clone (submodule superrepo), otherwise it just works in the current checkout. Nothing is written into the conventions: the answer is remembered on this machine in the gitignored .trailhead/isolation-answer.json and reused, with a one-line note, while your other assigned tickets stay the same and the last work-start was under 8 hours ago; a change or an expired record re-asks. You can still force none/worktree/clone explicitly, and /trailhead:auto and explicit modes never use the remembered answer.
• Scope tickets by path on a monorepo. A ticket can carry a Scope: line naming the package/dir(s) it touches: disjoint scopes parallelize safely, and it scopes build/test and the worktree's commits to the affected package. /trailhead:adopt detects a monorepo and proposes both isolation: worktree and the Scope: convention. Under isolation: none, Scope: also serializes a shared area: at claim time a session holds off if another in-progress ticket's scope overlaps, so two tickets scoped to the same submodule/package can't run at once (how you keep two sessions off one path-bound app).
• Git submodules are handled as separate repos. Work that lands inside a submodule is committed inside the submodule (on a branch) and published before a gitlink bump in the parent points at it. Because superproject worktrees share one submodule git dir, a submodule-bearing repo resolves to clone (under the default auto, or by escalating an explicit worktree), so each ticket gets independent submodule state; the clone provisions its submodules at setup (--recurse-submodules) and Verify checks git status inside each touched submodule. A ticket spanning two submodules stays atomic via a single parent commit that bumps both gitlinks together; its Scope: names both, and the same submodule counts as overlapping scope, so two tickets on it serialise. /trailhead:adopt detects .gitmodules and writes the rule into the conventions.
• State lives on the tickets (claim, labels, comments; all conflict-free); the map body is a re-readable index, not the source of truth. The durable record of a resolution is the ticket's own comment.
• Labels are protected from outsiders. trailhead:* labels drive the map, so trailhead trusts a labelled issue only when a write-access collaborator applied the label and (for a ticket) it carries a valid Parent:. Anything else is quarantined as trailhead:unverified, off the frontier. GitHub already blocks non-write users from labelling; against triage users and automations, trailhead can install the label guard GitHub Action for you (from templates/trailhead-label-guard.yml) into your repo's .github/workflows/, where it strips unauthorized trailhead:* labels at the source. Installing a workflow needs a token with the workflow scope, so trailhead runs gh auth refresh -s workflow if the push is rejected.
• Issues opened by others are triaged, not trusted blindly. /trailhead:inbox lists inbound issues (bug reports, requests, questions that aren't yet trailhead tickets) and integrates the worthwhile ones in place: reframing them as tickets, keeping the reporter's authorship, and applying the labels on adoption. The rest routes to fog, out-of-scope, duplicate, or needs-info.

A ticket is one answerable Question, sized to one session, so you never grow it in flight. When new scope surfaces while you're working (testing especially sparks ideas), you decide per idea: small and part of the same Question → just do it; balloons the ticket past one session → split it; separate work or a follow-up → capture it (idea/todo/ticket/bug) and keep going. Extras never become silent lines inside the ticket you're on.

Tickets that spin off from other work carry a lineage pointer in their body, so the map stays traceable:

Three layers: nearest wins, key by key (a key unset at one layer inherits the next):

• Project: a .trailhead/config.json file at the repo root: overrides for this project, committed so a team shares one config. Config lives in a plain file, never in the map issue; it's yours to change at any time.
• Global: ~/.claude/trailhead/config.json: your standing defaults across every project.
• Defaults: the built-in values.

/trailhead:config runs a guided, menu-driven setup: pick the scope, then walk each setting (⚙️ way of working (git · release · isolation) · 🌐 ticket language · 🧠 models (a High / Balanced / Low profile, or Manual) · 🎨 design + approval · 🧪 TDD · ⚖️ effort · 🖥️ acceptance testing (browser, UAT setup, a test account) · 🧑‍⚖️ plan review · 📊 statusline) as an icon-labelled menu; no hand-editing JSON. Every step is asked (none skipped), and plan and execute models are always two separate, version-pinned choices. config get prints the effective merged config; config set <key> <value> writes one key.

/trailhead:prune [map] reclaims a map's sub-issue slots under GitHub's 100-cap (drops closed/orphan edges, keeps every reference). Offered automatically when a map is near or at the cap, and runnable on demand.

/trailhead:audit [map] [--all] hunts regressions: it checks that the tickets closed since the last audit are still true against the current code. Code a ticket's commits left untouched is classed unchanged with no agent (a fast, deterministic git check that relies on the Refs: #<n> commit trailers), decisions, research, tasks and prototypes get a light inline check, and a full verification agent runs only where the code changed or the fast path cannot decide (no Refs: commits, for example). Each ticket is classed implemented, unchanged, evolved (a later ticket changed it on purpose, listed with no proposal), implemented wrong, or not implemented. It then compares the open ones for duplicates and groupable clusters. It is read-only until you confirm: every fix (a regression bug, a reopen, a duplicate close, a supersede-and-merge) is a proposal applied only on an explicit yes. --all re-audits every closed ticket.

The 📊 statusline step offers to install trailhead's Claude Code status bar: one line with model · project · branch · plan usage (5h % · reset · 7d %) · a context-window bar, plus a second line with the active ticket (▸ #N Title) whenever you're working one, and a ⬆ trailhead <version> flag when a newer trailhead is available (run /trailhead:update). The project is always the main repo's name even from an isolated checkout, and the branch carries a (WT) tag in a worktree or (C) in a per-ticket clone (nothing on the original checkout). It's a global Claude Code setting; if you already run a statusline (e.g. ccstatusline) the setup asks before replacing it, and the script also exposes --ticket-only / --context-only / --usage-only segments to slot into an existing tool.

First-use offer (once per project). You don't have to seek this out: the first time you set a project up (at the end of /trailhead:new/:adopt, or on a bare /trailhead when a map exists but no config does) trailhead offers to run the guided setup, or to continue on defaults. It asks only once: configuring writes .trailhead/config.json, and "continue on defaults" writes an empty {}. Once that file exists it's never offered again, and it never interrupts mid-ticket.

Models. Each key runs its activity as a subagent on the model you name, so the whole per-activity split applies within a single work session, whatever model that session runs on:

• plan and execute are the build/bug engine's steps. Execute is a subagent too, and it commits to main.
• research, review, debug, verify (the goal-backward Verify check), and the codebase-map fan-out each run on their own key.

Choose plan and execute separately, each by full versioned id (never a bare opus or sonnet). A key left unset uses its Balanced-profile default, and one equal to the session model simply runs inline; set inherit session to use the session model.

The main session stays the orchestrator. It holds the interactive moments (a build's Discuss, and Verify's acceptance/UAT) and it does charting, grilling, and ticket-writing on its own model. No key governs ticket quality, so run charting sessions on your strong model.

Example config file (the same shape works for the project .trailhead/config.json and the global ~/.claude/trailhead/config.json):

{ "ticket": { "language": "en" }, "models": { "plan": "claude-opus-5-5", "execute": "claude-sonnet-5-5" }, "tdd": "seams", "acceptance": { "browser": "auto" } }

Design mockups and acceptance testing. How the design.* keys play out (canvas, design system, approval, moving from claude.ai/design) is in Mockups before code; how acceptance.browser and testing.* drive Verify is in Acceptance testing.

Nested or flat. The list keys above (isolation.copy, preflight.checks, uat.setup, uat.accounts) can be written nested ({"uat": {"setup": [...]}}) or flat ("uat.setup": [...]); if a file has both, the nested form wins with a warning, and a project value replaces the global one as a whole (lists are not merged). git, release and the isolation mode stay in the conventions header; only isolation.copy lives in config.json.

Beyond the skill's instructions, trailhead ships eight of its own hooks (in hooks/, self-contained via ${CLAUDE_PLUGIN_ROOT}) that enforce the parts of the discipline the model shouldn't be trusted to remember:

• Commit guard (PreToolUse on git commit): hard-blocks a commit whose subject isn't Conventional Commits, and hard-blocks any Co-Authored-By trailer. This mirrors GSD's commit validation.
• Secret guard (PreToolUse on gh issue/PR writes): trailhead posts a lot to the tracker (ticket bodies, engine comments, codebase/conventions issues, resolutions). This scans what a gh issue/gh pr/gh api write is about to post (the inline --body, a heredoc, or a --body-file) and hard-blocks it if it matches a credential pattern (private keys, GitHub/AWS/OpenAI/Slack/Google/Stripe tokens, JWTs, or a hardcoded password/secret/token=…). The block is a clean-and-retry cue, not a dead end: trailhead redacts the flagged value in place (<REDACTED> or an env-var reference) and re-posts automatically, so the cleaned content still lands: the secret just never reaches the tracker. It stays a fail-safe block rather than a silent auto-rewrite on purpose: a block never leaks, whereas a redaction that quietly failed to apply would. It never scans reads, only outbound writes.
• Install guard (PreToolUse on package installs): a slopsquatting barrier. It hard-blocks a named public-registry install (npm / PyPI / crates.io / Go / RubyGems) until the package is vetted (typosquat similarity to popular names, age / first-seen date, download counts, repository metadata present and matching); once vetted, re-run the exact command prefixed with TRAILHEAD_VERIFIED_INSTALL=1 to proceed. The Execute / Fix engine carries the same block-and-ask. It reads commands the way a shell does: text inside quotes or a heredoc (an echo, a commit message) is not an install, lockfile installs pass, and wrappers such as sudo, time, env or xargs are seen through.
• Search guard (PreToolUse on Bash): structural backstop for the search-command-hygiene rule (never cd then read/search a relative path). It hard-blocks a single Bash command that both changes directory (cd/pushd/env -C) and, in a joined or later statement, reads/searches a relative path (grep/rg/ag/cat/head/tail/wc/nl/sed/awk/less/more, or git grep/log/show/diff without a -C), so the shape that trips a Read() deny rule under bypass permissions (falling back to a manual approval prompt instead of running headless) never runs at all, rather than merely being discouraged in the agent's own prose. Redirection targets (2>/dev/null) are not treated as reads, and a cd to the session's own directory does not arm it. Claude Code only; not projected to Codex, which has no such Read() deny rule.
• Body guard (PreToolUse on gh issue/PR body writes): hard-blocks a gh issue edit / gh pr edit / gh api write that would set an issue body to empty or whitespace-only, so a read-modify-write whose read returned an empty/degraded body (e.g. a transient HTTP 5xx on gh issue view --json body) can never clobber the existing body.
• Secret-read guard (PreToolUse on Read|Bash): denies reads of .env/.env.*/.secrets, whether via the Read tool or a Bash command (a file operand, an option value, or a < redirect target). This is a hook deny decision rather than a Read() permission rule, so secrets stay unreadable without arming Claude Code 2.1.259's cd-compound-read approval prompt.
• Mockup-link guard (PreToolUse on AskUserQuestion): structural backstop for the Prototype rule that a mockup approval ask carries the mockup's link. It hard-blocks an AskUserQuestion that asks to approve a mockup (or pick between mockup variants) when neither the reply text right before it nor the approval question itself holds a URL or file path (when the current message is not yet in the transcript, the guard cannot judge and allows), so the agent puts the clickable link in the message before asking again. On Codex, which has no AskUserQuestion tool and asks as a plain-text numbered list, its twin mockup-link-stop runs on Stop: when the final reply asks for that approval without a link, Codex continues the turn so the agent rewrites the ask with it.
• Issue injection scanner (PostToolUse on gh reads): trailhead reads issue/PR/comment text written by anyone with repo access, i.e. untrusted input. When that text contains prompt-injection phrases, the hook injects an advisory reminding the agent to treat it as data, never as commands. Advisory only: it never blocks.

The commit, secret, install, search, and body guards run in one process: a single trailhead-bash-guards.js registration evaluates them in that order, so a Bash call pays one node start-up instead of five (three hook processes per Bash call in all: the merged hook, the secret-read guard, and the injection scanner). They remain five distinct guards with unchanged verdicts; the first block wins, and a guard that errors allows on its own without stopping the others. A guard whose module is missing or broken is skipped too, but never silently: the hook tells the session which guard did not check the command, so a partial install shows up. On Codex the same hook is registered with --guards=commit,secret,install,body, since there is no search guard there.

The secret, body, and secret-read guards also fail closed on shell constructs they cannot interpret with certainty (eval, source / ., a bash 5.3 ${ cmd; }, printf -v, a function definition, an alias, a quoted gh api body key, a backslash-newline inside a word): when one can hide or precede a write (or sits next to a secret path), the command is blocked with a "rewrite it in a simple form" message instead of being emulated; the same constructs with no write or secret nearby pass.

Alongside them trailhead installs a git commit-msg hook (a POSIX sh wrapper, so it also works in "type": "module" repos) that enforces the same commit rules on every host, plus the Refs: #<n> trailer while a session ticket is active. If it cannot run (script missing, no node) it prints trailhead: commit-msg hook inactive (<reason>); commit rules not enforced. to stderr and lets the commit through; an installed hook is upgraded at the next ticket start.

All eight are crash-safe (any error → allow) and active whenever the plugin is installed. The commit, secret, install, search, body, and secret-read guards therefore apply in every repo, not only trailhead projects: intentional, since conventional commits, no Co-Authored-By, no leaked secrets, no unvetted installs, no cd-then-relative-read, no clobbered issue bodies, and no readable .env/.secrets are trailhead's standing rules. To opt out, disable the plugin's hooks in your Claude Code settings.

trailhead stands on the shoulders of two bodies of work. It is an original, self-contained methodology inspired by them. It does not copy or redistribute their code, and it invokes none of their skills; it reimplements the ideas as its own inline protocols:

• Matt Pocock: for Wayfinder, whose map / frontier / fog-of-war model is the backbone of trailhead's planning half, and whose grilling, domain-modeling, prototype, research, and tdd skills shaped trailhead's inline techniques. See mattpocock/skills.
• The open-gsd team: for GSD, whose discuss → plan → execute → verify cycle, capture model, codebase mapping, and systematic debugging are the engine of trailhead's building half.

Thank you. If you like the ideas here, go star and use the originals: they go far deeper than trailhead's compact protocols.

MIT © Marco Migozzi (@ToRvaLDz)