[omartelo]

The lich half of lich-plugin#12
(released as v0.9.0),
which gives opencode sessions the seven operations as tools of its own. Two
commits, both about the same thing: which sessions can drive the others, and how
lich talks to them about it.

1. Crush reads the server out of its plugin install

Claude Code and Codex are told about lich's MCP server on their own command
line, at spawn: per session, no file, no secret in argv. Crush takes no such
flag — its only route is its config, and the plugin install is already writing
there, inside markers an update replaces and an uninstall removes (#220).

So the registration joins that block:

# >>> lich-plugin v0.9.0 >>>
# Managed by lich. Edit through Settings, not by hand.
hook add PreToolUse --name lich-session-id --command '…/report-session-start.sh crush' --timeout 5
hook add PreToolUse --name lich-touched --matcher '^(edit|write|multiedit|bash)$' --command '…' --timeout 5
mcp add lich --type stdio --command '/usr/bin/lich' --args mcp --timeout 10
# <<< lich-plugin <<<

The line names the absolute path of the binary that installed it: a line in a
file cannot expand $LICH_BIN per session, and it does not need to — the binary
is the transport, and which lich a session reaches is decided by the
coordinates in its PTY. An executable that will not resolve leaves the
registration out rather than writing a command that cannot run.

2. The relayed message asks who has the tools

The message lich types names the reply_to_session tool where there is one and
the shell command everywhere. Which of the two it offered came from
providers.AcceptsMCPServer — a fact about the provider, from when only Claude
Code and Codex could be handed tools at all.

That went stale with the plugin carrying them. Now the relay asks
agentplugin.HasTools: true for the two told at spawn, and for opencode and
Crush once the installed plugin is ≥ 0.9.0. The version matters — an agent
pointed at a tool its plugin predates loses the turn to an error, while the
command costs nothing and works everywhere. Nothing wired, an old plugin and an
absent one all fall back to it.

It rides the seam that already tells the relay which providers report their
state, now one small Plugins interface instead of a lone callback.

Test plan

• internal/agentplugin: the registration line pinned as the literal Crush reads; the path quoted for the shell reading crushrc; no binary means no registration but the hooks stay; the line lives inside lich's markers, so an uninstall takes it and leaves what the user wrote.
• internal/agentplugin: HasTools is true at spawn for Claude Code and Codex, and elsewhere follows the installed version — 0.9.0 and 0.10.0 yes, 0.8.0 and nothing installed no. Versions pinned as literals, not read off the constant.
• internal/relay: the message names the tool for a target that has it, names only the command for one that does not, and falls back to the command with no plugin check wired at all.
• go test ./..., go vet ./..., gofmt; pnpm test, pnpm check, tsc --noEmit; cross-compile build + vet for linux/darwin/windows. Coverage: relay 94.9%, agentplugin 85.5%.
• mcp add syntax checked against Crush's own docs — and against the fact that crush mcp is not a CLI subcommand, which is why this is a line in a file rather than a command lich runs.