fix(integrations): dispatch Amp via execute mode - #4581

Description

Fixes #4580. Part of the audit in #2416.

AmpIntegration never overrode build_exec_args(), so it inherited the generic implementation from MarkdownIntegration (base.py:1062), which emits -p <prompt> --model <model> --output-format json. None of those three flags exist in the Amp CLI, so every workflow command:/prompt: step targeting Amp aborted at argument parsing before the agent ran:

$ amp -p "hello" --model gpt-5 --output-format json
Error: error: unknown option '-p'
$ echo $?
1

Amp's non-interactive entry point is -x/--execute, and its structured-output flag is --stream-json (valid only alongside --execute), so this dispatches through those instead:

$ uv run python -c "from specify_cli.integrations.amp import AmpIntegration; print(AmpIntegration().build_exec_args('hello', model='gpt-5', output_json=True))"
['amp', '--execute', 'hello', '--stream-json']

Verified against amp --help (v0.0.1789372854-ge954c8):

On model: it is deliberately dropped rather than remapped. Amp exposes no model-selection flag. -m/--mode takes an agent mode (low/medium/high/ultra, or a plugin mode) which "controls the model, system prompt, and tool selection" — it is not a model identifier, so forwarding the caller's model onto it would silently select the wrong thing. This is the one judgement call in the diff; happy to change it if maintainers prefer a different mapping.

On flag order: SPECKIT_INTEGRATION_AMP_EXTRA_ARGS is applied before --execute, because --execute [message] takes the prompt as an optional inline value — appending between the two would consume the prompt as the operator's flag value.

Same fix shape as the existing one-off overrides for opencode (#2409) and goose (#3781); follows the OpencodeIntegration.build_exec_args() pattern.

Testing

• Tested locally with uv run specify --help
• Ran existing tests with uv sync && uv run pytest
• Tested with a sample project (if applicable)

Added 4 regression tests to tests/integrations/test_integration_amp.py (previously a stub with no build_exec_args coverage). All 4 fail on main with '-p' != '--execute' and pass with this change; the file's 22 tests pass in full.

tests/test_agent_config_consistency.py: 28 passed.

tests/integrations/test_extra_args.py, test_base.py, test_integration_base_markdown.py — the shared machinery this override calls into (_apply_extra_args_env_var, _resolve_executable, the MarkdownIntegration base): 119 passed, 1 skipped.

Full-suite context, stated plainly: on this machine (Windows 11, Python 3.14.7) main already reports 182 failed / 7694 passed / 189 skipped before this change. Every failure is environmental — 105 bash/parity and 63 symlink tests, on a box with no symlink privilege — and none are in tests/integrations/test_integration_amp.py. Of the 18 baseline failures under tests/integrations, all 18 are symlink tests. A full tests/integrations run against this branch was still in progress at the time of writing; I will post the before/after comparison as a comment when it completes.

End-to-end against the real Amp CLI (npx @sourcegraph/amp, v0.0.1789372854-ge954c8) on Windows 11: the previously generated argv exits 1 at parsing with unknown option '-p', while the argv this PR generates is accepted and proceeds to Amp's device-login prompt (this machine has no AMP_API_KEY, so execution stops there rather than at flag parsing).

AI Disclosure

• I did use AI assistance (describe below)

Investigation, patch and tests were written with Claude Code; I reviewed them and ran every command quoted above on my own machine. The Amp flag surface was verified against amp --help from the installed binary, not from model recall.