Agent guide, a gibhub skill, and a BRC-73 permission manifest

[shruggr]

Three things, each verified against the code rather than against memory.

1. AGENTS.md, with CLAUDE.md as a symlink to it

CLAUDE.md was 46 lines and had drifted. It is replaced by AGENTS.md
(git mode 120000 symlink at CLAUDE.md), the arrangement the gib CLI
uses.

It covers every route and what it renders, the lib modules that matter,
the head token's exact fields and vocabulary, the overlay and ORDFS
routes the site calls and what each returns, the fact that
lib/gib-head.ts decodes heads locally so /me works with no overlay,
and the conventions already present in the code — server components by
default, server-side handle resolution behind one shared cache, the
react-query keys in use, and the standing rule not to reimplement
anything the @1sat SDK ships.

The "things that will bite you" section is drawn from real facts: the
repository origin resolving to the genesis tree forever, handle
verification being per head rather than per commit, the hard-coded gib
basket, and CI's bun pin.

2. skills/gibhub/SKILL.md

A short skill for agents working in this codebase, listed in a catalog
table in AGENTS.md. This is not a monorepo, so skills/ at the root
is the only copy; there is nothing to mirror it to.

3. A BRC-73 grouped permission manifest

The site made the wallet prompt separately for each permission as the
user reached it. /manifest.json now carries a
metanet.groupPermissions block so a BRC-100 wallet asks once.

The permissions block went into the existing app/manifest.json/route.ts
rather than a new public/manifest.json, because the babbage.trust
block there is conditional on an environment variable and so the
document has to be built per request. Nothing was clobbered.

Declared, and nothing more:

Declared	Reached by
[1, "identity key retrieval"]	getPublicKey({ identityKey: true }) on connect
[1, "gib branch"]	pushDropLock / unlockByScript
[1, "action label gib push"]	mintHead
[1, "action label gib delete"]	burnHead
basket gib	listOutputs({ basket: "gib" }), and where minted heads are filed
spendingAuthorization, 10 000 sat/month	createAction's net spend for the 1-satoshi head plus fees

Spending authorization is declared because the code needs one, not on
assumption: createAction computes netSpent > 0 for the 1-satoshi
head and calls ensureSpendingAuthorization.

Three details are not guessable from the spec name and are recorded in
AGENTS.md:

• Action labels are not a category. GroupedPermissions has only
  description, spendingAuthorization, protocolPermissions,
  basketAccess, certificateAccess. A label is gated as the level-1
  protocol action label <name>.
• Level-1 entries carry no counterparty. The manager forces it to
  "" at level 1, and a literal "" is a reserved slot that drops the
  entry — so no entry here declares one, even though gib-wallet.ts
  signs with counterparty anyone.
• No extra keys. The grant path deep-equals granted entries against
  the requested ones, which are these literal objects.

The response also sets access-control-allow-origin: *, matching BSV
Desktop's own manifest route: a wallet running in a page fetches this
cross-origin, and a failed fetch is cached as "no manifest" for five
minutes.

Verification

bun run lint, bun run typecheck, bun test and bun run build all
pass. The manifest was fetched from a running dev server and checked two
ways: that it parses and matches the GroupedPermissions shape with no
dead (level 0), dropped (counterparty: "", level 2 without a
counterparty) or non-spec entries; and that the manager's
isRequestIncludedInGroupPermissions matcher, transcribed from source,
covers all six requests this app actually makes and nothing else.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EM5trZ4BfZpW7pBjq2onqX

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
gibhub	Ready	Preview	Sep 20, 2026 5:54am UTC