Amp Runner

Below is the implementation spec for the AMP CLI (Sourcegraph) runner shipped in Untether.

---

Scope

Goal

Provide the amp engine backend so Untether can:

• Run AMP non-interactively via the AMP CLI (amp).
• Stream progress by parsing --stream-json output. AMP uses a Claude Code-compatible JSONL protocol.
• Support resumable sessions via amp threads continue <thread-id>.

Non-goals (v1)

• Thread management commands — amp threads list/search/share etc. are not exposed via Telegram.

---

UX and behavior

Engine selection

• Default: use default_engine from config
• Override: /amp <prompt> in Telegram

Resume UX (canonical line)

Untether appends a single backticked resume line at the end of the message:

`amp threads continue T-2775dc92-90ed-4f85-8b73-8f9766029e83`

Notes:

• The resume token is the thread ID (format: T-<uuid>), captured from the system(init) event’s session_id field.
• AMP calls sessions “threads” — amp threads continue resumes them.

Non-interactive runs

The runner invokes:

amp --dangerously-allow-all --mode <mode> --model <model> -x --stream-json <prompt>

Flags:

• --dangerously-allow-all — auto-approve all tool calls (default, configurable)
• --mode <mode> — optional (deep|free|rush|smart)
• --model <model> — optional, from config or /config override
• -x — execute mode (non-interactive)
• --stream-json — JSONL output
• --stream-json-input — optional; enables stdin streaming (preliminary support, configurable)

For resumed sessions:

amp threads continue <thread-id> --dangerously-allow-all -x --stream-json <prompt>

---

Config additions

=== “untether config”

```sh
untether config set default_engine "amp"
untether config set amp.model "claude-sonnet-4-6"
untether config set amp.mode "smart"
untether config set amp.dangerously_allow_all true
```

=== “toml”

```toml
# ~/.untether/untether.toml

default_engine = "amp"

[amp]
model = "claude-sonnet-4-6"       # optional; passed as --model
mode = "smart"                     # optional; deep|free|rush|smart
dangerously_allow_all = true       # default: true
stream_json_input = false          # default: false; passes --stream-json-input
```

Notes:

• mode controls model selection, system prompt, and tool availability within AMP.
• dangerously_allow_all defaults to true since Untether runs headless.
• stream_json_input enables --stream-json-input for stdin streaming. This is preliminary plumbing — the interactive control flow (approve/deny via Telegram) is not yet wired.

---

Code changes (by file)

src/untether/runners/amp.py

Exposes BACKEND = EngineBackend(id="amp", build_runner=build_runner, install_cmd="npm install -g @sourcegraph/amp").

Runner invocation

amp [threads continue <thread-id>] --dangerously-allow-all [--mode <mode>] [--model <model>] -x --stream-json [--stream-json-input] <prompt>

Event translation

AMP uses a Claude Code-compatible JSONL protocol with a type discriminator. The runner translates:

• system(subtype="init") -> StartedEvent (captures session_id)
• assistant (tool_use blocks) -> ActionEvent (phase: started)
• user (tool_result blocks) -> ActionEvent (phase: completed)
• assistant (text blocks) -> text accumulation for final answer
• result -> CompletedEvent (with accumulated usage)

Subagent tracking

parent_tool_use_id from assistant/user messages is stored in action.detail["parent_tool_use_id"] when present. This tracks which tool calls belong to subagent invocations.

Usage accumulation

Unlike Gemini (which reports usage once in result.stats), AMP reports per-message usage in assistant messages. The runner accumulates input_tokens and output_tokens across all assistant messages and builds the final usage dict at completion.

---

Installation and auth

Install the CLI globally:

npm install -g @sourcegraph/amp

Run amp login to authenticate with Sourcegraph.

---

Known pitfalls

• AMP uses amp threads continue <thread-id> for resume, not --resume.
• Thread IDs use the format T-<uuid> (e.g., T-2775dc92-90ed-4f85-8b73-8f9766029e83).
• --stream-json-input is passed when stream_json_input = true in config. The interactive control flow (approve/deny buttons in Telegram) is not yet wired — this is preliminary plumbing.
• AMP’s --model flag may have no effect when using hosted models (model is controlled server-side by --mode).