slash-chain

A Claude Code Mod that runs slash commands joined with && one after another, as a shell does: /tiny && /context runs /context once /tiny ended well. Without the mod the engine runs the first command alone and drops the rest.

What it does

1. The engine hands everything after the first command's name to that command as its arguments. The mod splits them at each && /<name>: the first command runs with its own arguments alone, and every /<name> after a && becomes a step with its own arguments. A && that no /<name> follows stays in the arguments, so /commit build && test keeps its text.

2. Each step waits for what it started before the next one runs:

   • A command that hands the model a prompt (a skill or a prompt command) waits for its turn to end. A turn that ends with answer runs the next step. A turn that ends with error, refusal or aborted (Esc) stops the chain. A question the model asks with AskUserQuestion stays inside that turn, so the chain waits for the answer too.
   • A command that opens a focused pane (/disk-janitor) waits until that pane closes.
   • A built-in dialog (/cost, /model) holds its command until it closes.
   • Any other command is done when it returns.

3. A command that throws stops the chain.

4. Each step gets one transcript line, and the chain gets one last line:

   1/2: /tiny
   2/2: /context
   all 2 command(s) ran
   stopped after /tiny: its turn ended with aborted; not run: /context
   

5. A prompt or a new chain you type while a chain waits cancels it: cancelled; not run: /context. The new chain starts from there. A single command you type (/cost) runs beside the waiting chain and does not cancel it, because the mod hooks only the commands whose arguments hold && /<name>. The engine still writes this mod's name in front of every plugin command's output (task-poke+slash-chain: ...): it picks the names it writes by a hook's command matcher alone, and a chain can start with any command, so this hook cannot name one (measured on 2.1.281 with two probe plugins: a hook with only an args matcher that never ran was named, a hook with a command matcher was not).

When a step fails in words

The engine gives a turn no exit code: a model that could not do a step, says so and ends its turn normally ends it with answer, and the next step would run. So a chained skill or prompt step that has steps after it gets one line after its text:

[slash-chain] This is step 1/2 of a chain; after it: /exit. If you could not do what this step asks, call the mcp__slash-chain__fail tool with the reason before you end your turn, and the steps after it do not run.

The mod declares the tool mcp__slash-chain__fail (reason) at the session's start and keeps it in the model's tool list, not behind ToolSearch, so the model can call it at once. A call stops the chain:

stopped after /fail: the model reported it failed: Writing /nonexistent-dir/CLAUDE.md failed: EROFS read-only file system, could not create /nonexistent-dir.; not run: /exit

Measured on 2.1.281: /fail && /exit, where /fail asked for a write that failed, called the tool, stopped before /exit, and the session stayed open; /ok && /context did not call it and ran both. A call while no chained step waits on the turn, from a subagent, or without a reason is refused. The last step of a chain and a command outside a chain get no line.

Command

/slash-chain          the setting and the chain that runs, with what it waits for
/slash-chain stop     cancels the chain that runs
/slash-chain on | off on by default; off leaves the command as the engine hands it

Install

claude plugin marketplace add KilimcininKorOglu/claude-code-mods
claude plugin install slash-chain@kilimcininkoroglu-mods

Function hooks are early access. Nothing loads without the flag. To keep it on, add this to ~/.claude/settings.json:

{ "env": { "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1" } }

After installing

1. Restart Claude Code.

What it can reach

Validated with claude plugin validate on Claude Code 2.1.281:

❯ ./register.ts hooks: session.start, tool.describe{tool=mcp__slash-chain__fail}, tool.call{tool=mcp__slash-chain__fail}, command.run{command=slash-chain}, command.run{args=/"(?:^|\\s)&&\\s*\\/[A-Za-z0-9_:.-]+(?=\\s|$)"/}, skill.prompt, ui.open, ui.close, turn.complete, prompt.submit
❯ ./register.ts calls: $.clock.after (via advance), $.command.register, $.command.run (via runStep), $.store.get, $.store.set (via setEnabled), $.tool.register, $.ui.log (via advance, cancel, runFirst, stop)

Reach L2, it drives Claude: it runs the slash commands you chained.

1. Reads:    the arguments of every slash command, the name of each skill prompt, each pane's id and focus, and each main-loop turn's end reason. It reads no file, no prompt text and no answer; it reads the reason of a `fail` call.
2. Runs:     each command after the first of a chain, once, with the arguments you typed for it; declares one tool, `mcp__slash-chain__fail`, at the session's start
3. Sends:    to the model, the `fail` tool in its tool list and one line after the text of a chained skill or prompt step that has steps after it; nothing to the network
4. Persists: in $.store, the on/off setting
5. Hostile input: the chain comes from the text you typed; a step runs only a command the engine knows, with its own arguments, and a pasted text that holds `&& /<name>` in a command's arguments runs that command too

Limits

• The fail tool stands in the tool list of every session, also with no chain, because a tool cannot be taken back once declared.
• The steps after the first run as a plugin (origin.kind: 'plugin'), so a command that answers only the person refuses there. /disk-janitor delete <path> is one.
• A local command succeeds when it does not throw. The engine gives no other error signal, so a command that prints an error and returns counts as done.
• A skill or prompt command succeeds when its turn ends with answer and the model did not call fail. A model that says it failed without calling the tool still counts as done: measured on 2.1.281 before the tool, /fail && /exit, where /fail answered I could not write CLAUDE.md., ran /exit, and the session closed. A turn you stop with Esc does stop the chain: /slow && /exit stopped with stopped after /slow: its turn ended with aborted; not run: /exit, and the session stayed open.
• /exit runs as a later step: the engine takes it from a plugin, and the session closes.
• An unknown first command never reaches the mod: the engine answers Unknown skill, and nothing after it runs. An unknown later command stops the chain with the engine's error: stopped after /x: it failed: $.command.run: no command named /x in this session.
• Only && is read. ||, ; and | stay in the arguments.
• A focused pane is recognized when it opens during its step. A pane a command opens later, from a timer, is not waited for.
• The chain was checked in an interactive session alone. A headless session (claude -p) was not measured.

Development

make install     # eslint, typescript-eslint, typescript
make lint        # complexity limit 10, fails the build above it
make typecheck   # needs .claude/types/ from /plugin-types
make validate
make test        # claude plugin test