feat: add peer completion for bash, zsh, and fish

[ayushsingh82]

Running peer at a terminal today gives you nothing on tab — no command names, no flags. This adds peer completion <bash|zsh|fish> so you can wire up real completion:

source <(peer completion bash)
source <(peer completion zsh)
peer completion fish | source

The script isn't hand-written. It's generated from commandDefinitions and the program-level option list, so it can't drift from the actual command tree — add a command, rebuild, the completion follows.

How it's put together:

• src/completion/index.ts walks the registry into a small tree (each node = its sub-commands + its long flags, with parent groups filled in from child paths), then renders per shell: a case "$path" lookup for bash, #compdef + _describe for zsh (keeps the command descriptions), and declarative complete -c peer lines behind a path-precise guard for fish.
• src/commands/global-options.ts pulls the 15 global options out into a list. src/cli.ts now loops over that list instead of 15 literal program.option(...) calls, so the CLI and the completion generator are reading the same source.

It's a read-only command — passthrough: true so the raw script lands on stdout and stays source-able, errors still go out as the usual JSON envelope, and an unknown shell gets a VALIDATION_ERROR with exit 1. It's exposeInMcp: false for the same reason peer mcp is: a shell-script emitter doesn't belong in the MCP tool surface. MCP profile counts don't change.

Regenerated agents/tool-catalog.json and agents/runtime-manifest.json (CLI count 147 → 148). README gets a "Shell completion" section, plus updates to CHANGELOG.md, llms.txt, and AGENTS.md.

npm run check (247 tests, lint, typecheck, build) and npm run test:e2e (4/4) both pass; npm run audit is clean. tests/completion.test.ts covers the model builder, all three renderers, and the command including the error paths — the new files land at 100% coverage.

[ayushsingh82]

Hey @ADWilkinson — heads up that I put this together on my own, it wasn't asked for, so genuinely no hard feelings if it's not something you want in the CLI.

The reason I did it: I kept reaching for tab completion while poking around with peer and there just isn't any yet. Rather than write completion scripts by hand and have them rot, I let them generate from the command registry so they stay honest as commands come and go. It's read-only and stays well away from anything protocol- or signing-related. I also tried to match how the repo already does things and moved the docs and generated catalogs along with it. Tests and the e2e run are passing.

If the approach isn't right, or you'd just rather not own this surface, tell me and I'll close it — no problem either way.