Installing a SEED

Reference implementation of SEED installation. The install procedure lives in SEED.md; this skill is the call site.

Two contracts govern every install run. Read both before starting:

1. Display gate — every repo-supplied shell block is shown in full and announced in one line before executing. No per-block confirmation. See §Display gate.
2. Output: install-report.json — the skill continuously writes a structured JSON ledger at <repo-root>/install-report.json, one item per step, flushed via atomic rename after every step. This artifact is the single source of truth: the user-facing post-terminal summary renders from it, and the feedback dispatch uploads it verbatim on user consent. See §Install report.

Do not start the install procedure without committing to both. The display gate without the report is a black-box install; the report without the gate violates trust. They run together.

Display gate

Every shell block under ## Dependencies and every shell prompted by ## Verify MUST be displayed in full before execution. After displaying, the agent prints a one-line summary (10–20 words: what runs and why) and executes immediately. No yes/no prompt, no "approve all", no batching question. The user gave consent once at the install entry point; per-block prompts are redundant friction.

The skill's own pre-SEED commands (clone, cd) follow the same pattern: display, announce in one line, run.

The only question the skill MAY ask is "where do you want to clone?" — and only when the target is a remote git URL. Local-path and CWD installs ask nothing.