mvp

Build the thing, show it working, then turn what it taught into real work. The MVP branch is a reference, never a delivery: production code reaches main only through the Tasks the harvest derives.

Before anything else, run the tracker preflight (${CLAUDE_PLUGIN_ROOT}/skills/shared/references/preflight.md); a failure stops here with its single ## Action needed block.

$ARGUMENTS

Rebuild first

On every invocation, and after any compaction or resume, before other work: get-issue the ticket and take the newest # MVP ledger comment's agent_detail. It is the full live state — every entry is a snapshot, so the newest alone is enough. If no ledger entry is on the page, or the ledger records a harvest SHA and no manifest is on the page, and comments_truncated is set, stop loudly rather than rebuild from less. Also note whether a # MVP coverage manifest is on the page. Read the shortcuts taken so far in the MVP worktree:

git log --format='%h %(trailers:key=Shortcut,valueonly)' origin/main..HEAD

A reference read earlier in this session is only a summary after compaction; re-read the current state's reference before acting on it.

Invariants

• No /sy:spec, no GATE, no merge standards on the MVP branch.
• The branch never merges into main. It may merge main in; it is never rebased or force-pushed.
• Localhost first. Deploy only on the user's word.
• A deploy is verified by the brief's version check reporting the pushed SHA, never by readiness alone.
• A ledger entry is posted on each deploy, each new user rule, and each hold change (references/loop.md).
• Every shortcut gets a Shortcut: <what was faked> -> <what production needs> trailer on the commit that takes it.
• Ask only through AskUserQuestion (${CLAUDE_PLUGIN_ROOT}/skills/shared/references/user-interaction.md).
• Tracker mechanics belong to the tracker skill (/sy:tracker); this skill names canonical verbs only.

Agents on their own

The MVP session builds directly; these run beside it, never in place of the user's word on a deploy:

• sy:sweep for broad reconnaissance of a large surface;
• sy:trace for one load-bearing path, at most the resolved limits.max_depth_agents cap in flight — resolve per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/config-values.md;
• sy:slice for a bounded build step in its own worktree; its brief carries the trailer rule above;
• sy:img-inspector for any screenshot or rendered visual (${CLAUDE_PLUGIN_ROOT}/skills/shared/references/image-inspection.md);
• /sy:ci for a pushed branch's CI.

Every subagent dispatch resolves its model and effort from config and passes them as the Agent invocation's actual model and effort parameters, per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/model-dispatch.md; a nested dispatch inherits nothing and must resolve again.

State router

Classify before loading detailed procedure:

KICKOFF   no ticket, brief incomplete, or no `# MVP ledger` entry → references/kickoff.md
HARVEST   harvest SHA in ledger, no manifest → references/harvest.md from step 2, on that SHA
HELD      manifest posted, hold open       → build nothing; report the hold; the user releases it with /sy:shutdown <ticket>
LOOP      ledger posted, building          → references/loop.md
HARVEST   the user says the MVP is done    → references/harvest.md
RELEASE   harvest done, hold released      → suggest /sy:shutdown <ticket>

Load only the reference for the current state. Do not preload mutually exclusive procedures.