Skip to content

Launch Flag Injection — Reference

When amplihack starts claude, copilot, codex, or amplifier, it builds the subprocess command line by combining flags it injects automatically with any extra arguments the user supplied. This document describes every flag that amplihack injects, the conditions under which each is injected, and how users can override the defaults.

Contents

Injected flags overview

Flag Injected when? Applicable tools Override mechanism
--dangerously-skip-permissions --skip-permissions passed AND tool is Claude-compatible claude, rusty, rustyclawd, amplifier omit --skip-permissions
--model <value> tool is Claude-compatible AND user did not pass --model, unless AMPLIHACK_DEFAULT_MODEL is set to an empty value outside the LiteLLM gateway path. Unset, the value is claude-opus-5[1m]; see --model claude, rusty, rustyclawd, amplifier pass --model, or set AMPLIHACK_DEFAULT_MODEL= (empty) to pass none
--resume only when amplihack launch --resume launch only (not claude) pass --resume to the launch subcommand
--continue only when amplihack launch --continue launch only (not claude) pass --continue to the launch subcommand

--dangerously-skip-permissions

amplihack passes --dangerously-skip-permissions to the tool subprocess only when both conditions are met:

  1. The user passed --skip-permissions on the amplihack command line.
  2. The target tool is Claude-compatible (claude, rusty, rustyclawd, or amplifier).

Tools that are not Claude-compatible (copilot, codex) never receive this flag because they do not support it.

# User explicitly opts in:
amplihack launch --skip-permissions

# amplihack spawns (the --model is amplihack's default; see --model below):
claude --dangerously-skip-permissions --model claude-opus-5[1m]

# Without --skip-permissions, the flag is NOT injected:
amplihack claude

# amplihack spawns:
claude --model claude-opus-5[1m]

# Non-Claude tools never receive it, even with --skip-permissions:
amplihack copilot --skip-permissions

# amplihack spawns:
copilot <extra_args...>

Rationale: The --dangerously-skip-permissions flag bypasses Claude's interactive confirmation prompts. It is gated behind an explicit opt-in (--skip-permissions) so that users in trusted automated environments can suppress the prompt, while interactive sessions retain the safety check.

Python launcher note: The Python launcher in amplihack/launcher/core.py may behave differently. Verify Python launcher behavior independently.

--model

amplihack passes a --model to Claude-compatible tools (claude, rusty, rustyclawd, amplifier) unless you pass one yourself or turn it off. Non-Claude tools (copilot, codex) never receive an injected --model and use their own default model selection.

This section summarises the rules. The full reference for the value, the dotted-id rewrite and every stderr line amplihack prints about the model is AMPLIHACK_DEFAULT_MODEL.

Which value is passed

Highest precedence first:

Situation What the tool receives
--model <id> or --model=<id> on the amplihack command line that argument, exactly as typed; amplihack adds no --model of its own
a LiteLLM gateway variable is set --model with AMPLIHACK_LITELLM_MODEL, unchanged
AMPLIHACK_DEFAULT_MODEL is empty or whitespace-only no --model; the tool picks, and the "model" in ~/.claude/settings.json takes effect
AMPLIHACK_DEFAULT_MODEL is set to an id --model with that id, trimmed; a dotted Claude id is rewritten to hyphens
AMPLIHACK_DEFAULT_MODEL is unset --model claude-opus-5[1m], amplihack's built-in default

Default: a concrete model id

With neither --model on the command line nor AMPLIHACK_DEFAULT_MODEL in the environment, amplihack passes its built-in default:

# User runs:
amplihack claude

# amplihack spawns:
claude --model claude-opus-5[1m]

Why a concrete id and not an alias (issue #1421): amplihack used to force --model opus[1m]. That string is an alias resolved by the tool, not by amplihack, and what it resolves to depends on the tool's version. On one install it resolved to the retired claude-opus-4-1-20250805, so every agent step failed with API Error: 404 ... model: claude-opus-4-1-20250805, an id the user had never chosen and could not find in any config file or binary. A concrete id cannot drift that way: a given Claude Code version either accepts claude-opus-5[1m] or fails naming that exact string.

Because amplihack passes --model by default, its choice outranks the "model" in ~/.claude/settings.json. To let the tool and that file decide, set the variable to an empty value:

AMPLIHACK_DEFAULT_MODEL= amplihack claude

# amplihack spawns:
claude

Pinning a model via environment variable

Set AMPLIHACK_DEFAULT_MODEL to pin a model for every session without changing the command line:

export AMPLIHACK_DEFAULT_MODEL=claude-sonnet-4-5
amplihack claude

# amplihack spawns:
claude --model claude-sonnet-4-5

A dotted Claude id, the spelling GitHub Copilot CLI uses, is rewritten to the hyphenated id Claude Code accepts (issue #1527):

export AMPLIHACK_DEFAULT_MODEL='claude-opus-5.5[1m]'
amplihack claude

# amplihack spawns:
claude --model claude-opus-5-5[1m]

The exact form that is rewritten, and the values left alone, are listed under AMPLIHACK_DEFAULT_MODEL.

For teams or CI environments that standardise on one model:

# .github/workflows/ai-tasks.yml
env:
  AMPLIHACK_DEFAULT_MODEL: "claude-sonnet-4-5"
  AMPLIHACK_NONINTERACTIVE: "1"

steps:
  - run: amplihack claude --print 'Run the lint checks'
    # spawns: claude --model claude-sonnet-4-5 --print 'Run the lint checks'

The stderr line

Whenever amplihack adds --model, it prints one line to stderr naming the model and where it came from, so a later "model not found" error can be traced back to it:

amplihack: passing `--model claude-sonnet-4-5` to `claude` (from AMPLIHACK_DEFAULT_MODEL). Set AMPLIHACK_DEFAULT_MODEL to override it, or to an empty value to let claude choose its own default model.

The source reads amplihack's built-in default when the variable is unset, and the line also names the original spelling when a dotted id was rewritten. On the LiteLLM gateway path it names AMPLIHACK_LITELLM_MODEL as both the source and the variable to change. The environment variable reference quotes each form.

Override via command-line flag

Pass --model directly and amplihack adds none of its own; AMPLIHACK_DEFAULT_MODEL is ignored. The value is forwarded exactly as typed and is never rewritten:

amplihack claude --model claude-haiku-4-5

# amplihack spawns:
claude --model claude-haiku-4-5

A dotted Claude id given to --model is also forwarded unchanged. Claude model ids use hyphens, and the launched tool may not report the problem itself (issue #1527), so amplihack prints a warning to stderr naming the hyphenated spelling:

amplihack claude --model claude-opus-5.5

# amplihack spawns:
claude --model claude-opus-5.5
# and warns: ... Use `--model claude-opus-5-5`.

There is no warning for amplihack copilot or amplihack codex, since the dotted spelling is GitHub Copilot CLI's own, and none on the LiteLLM gateway path, where the model is a gateway route name and a dot in it may be correct.

Detection is exact: an argument equal to --model, or one starting with --model=, counts as an explicit model. Arguments that only begin with --model, such as --model-config, do not.

Supported model identifiers

Apart from the dotted-id rewrite of AMPLIHACK_DEFAULT_MODEL, amplihack does not validate the model string; it forwards it to the tool. Aliases such as opus[1m], sonnet, and haiku are resolved by the tool, and which concrete model each maps to changes with the tool's version. This document deliberately does not tabulate those mappings: a table here would go stale exactly the way the old hardcoded alias did. Ask the tool (claude --help, or its release notes) for the current list.

--resume and --continue

These flags are passed through only when the user explicitly requests them on the launch subcommand. They are never injected automatically.

Important: The claude subcommand does not support --resume or --continue. Only launch exposes these flags.

amplihack launch --resume --skip-permissions
# spawns: claude --dangerously-skip-permissions --model claude-opus-5[1m] --resume

amplihack launch --continue --skip-permissions
# spawns: claude --dangerously-skip-permissions --model claude-opus-5[1m] --continue

The claude, copilot, codex, and amplifier subcommands do not support --resume or --continue.

Extra args passthrough

All positional arguments and flags after the subcommand name are forwarded verbatim to the tool subprocess after the injected flags. Order is:

<binary> [--dangerously-skip-permissions] [--model <value>] [--resume|--continue] <extra_args...>

amplihack claude --print 'Fix the failing tests' --output-format json

# amplihack spawns:
claude --model claude-opus-5[1m] --print 'Fix the failing tests' --output-format json

There is no processing or escaping of extra_args. What the user types is what the subprocess receives.

Complete command-line assembly

build_command_for_dir() in crates/amplihack-cli/src/commands/launch/command.rs assembles the final command line. The assembly order is:

  1. Binary path (resolved by bootstrap::ensure_tool_available())
  2. --dangerously-skip-permissions — only if skip_permissions == true and the tool is Claude-compatible (claude, rusty, rustyclawd, amplifier)
  3. --model <value>: only if the tool is Claude-compatible and --model is not already present in extra_args. The value is AMPLIHACK_LITELLM_MODEL on the LiteLLM gateway path, otherwise AMPLIHACK_DEFAULT_MODEL (a dotted Claude id rewritten to hyphens), otherwise claude-opus-5[1m]. An empty AMPLIHACK_DEFAULT_MODEL outside the gateway means no --model
  4. --resume (if requested — launch subcommand only)
  5. --continue (if requested — launch subcommand only)
  6. All extra_args in the order they were passed on the command line

The following examples show the full assembled command for each launch subcommand with no extra args:

amplihack claude
# → claude --model claude-opus-5[1m]

amplihack claude --skip-permissions
# → claude --dangerously-skip-permissions --model claude-opus-5[1m]

amplihack copilot
# → copilot

amplihack codex
# → codex

amplihack amplifier
# → amplifier --model claude-opus-5[1m]

amplihack launch --skip-permissions
# → claude --dangerously-skip-permissions --model claude-opus-5[1m]

AMPLIHACK_DEFAULT_MODEL=claude-sonnet-4-5 amplihack claude
# → claude --model claude-sonnet-4-5

AMPLIHACK_DEFAULT_MODEL= amplihack claude
# → claude

Python launcher parity

The Rust launcher's injection behaviour is designed to match the Python launcher in amplihack/launcher/core.py. The following table documents the parity contract:

Behaviour Python launcher Rust launcher
--dangerously-skip-permissions always injected conditional: Claude-compatible tool AND --skip-permissions
--model <default> opus[1m] unless AMPLIHACK_DEFAULT_MODEL set intentional divergence (#1421): the concrete id claude-opus-5[1m] unless AMPLIHACK_DEFAULT_MODEL is set; an empty value passes none; a dotted Claude id is rewritten to hyphens (#1527); Claude-compatible tools only
--model suppressed when user provides it yes yes
--resume passthrough yes launch subcommand only
--continue passthrough yes launch subcommand only
extra_args forwarded verbatim yes yes

Intentional divergences (not bugs) are documented in Parity Test Scenarios.