fix(integrations): dispatch Forge via prompt flag only

[hmohammadi]

Description

Fixes #4666. Part of the audit in #2416.

ForgeIntegration never overrode build_exec_args(), so it inherited the generic implementation from MarkdownIntegration (base.py), which emits -p <prompt> --model <model> --output-format json. Forge accepts -p/--prompt, but the other two flags do not exist in its CLI, so every workflow command:/prompt: step targeting Forge aborted at argument parsing before the agent ran:

$ forge -p "hello" --model gpt-4o --output-format json
error: unexpected argument '--model' found

Usage: forge.exe --prompt <PROMPT>
$ echo $?
2

Unlike the Amp case (#4581), both inherited extras are invalid, so both dispatch paths broke:

• stream=False (output_json=True) appends --output-format json → exit 2.
• stream=True (the default) drops --output-format but still appends --model <model> whenever a model is configured → exit 2.

The only surviving combination was stream=True with no model, which reduces to plain forge -p <prompt>.

This keeps -p and drops the two invalid flags:

$ uv run python -c "from specify_cli.integrations import get_integration; print(get_integration('forge').build_exec_args('/speckit-plan add OAuth', model='gpt-4o', output_json=True))"
['forge', '-p', '/speckit-plan add OAuth']

Verified against forge --help (v2.13.21):

Flag previously sent	In the Forge CLI?	Notes
-p <prompt>	Yes	kept as-is; also accepts the /speckit-… slash invocations
--model <model>	No	no model flag exists (see below)
--output-format json	No	no top-level equivalent; --porcelain is subcommand-only

On model: it is deliberately dropped rather than remapped. Forge exposes no model-selection flag — model choice is a persisted setting (forge config set model <provider> <model>, or interactive forge select model). --agent <AGENT> takes an agent ID, 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_FORGE_EXTRA_ARGS is applied before -p, matching the opencode / goose / codex ordering. Confirmed Forge parses its global flags ahead of the prompt flag — forge --verbose -p "x" --bogus rejects only --bogus.

Same fix shape as the existing one-off overrides for Amp (#4581), opencode (#2409) and goose (#3781); follows the AmpIntegration.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_forge.py, which previously had no build_exec_args coverage at all.

Red-green verified. With the source change reverted but the tests kept, all 4 fail; with the fix restored, all 4 pass:

# fix stashed, tests kept
4 failed, 26 deselected

# fix restored
4 passed, 26 deselected

Worth noting one of these caught a real trap during development: a first draft of the "uses -p" test used output_json=False with no model, which is exactly the one combination that already passes on main. It was rewritten to assert the real dispatch shape (model set, output_json=True).

End-to-end against the real Forge CLI (npx -y @antinomyhq/forge@latest, v2.13.21) on Windows 11:

# before — main's argv
$ forge -p "/speckit-plan add OAuth" --model gpt-4o --output-format json
error: unexpected argument '--model' found          (exit 2)

# after — this branch's argv, with a deliberately bogus probe flag to stop
# before execution rather than at parsing
$ forge -p "/speckit-plan add OAuth" --zz-probe-stop
error: unexpected argument '--zz-probe-stop' found

Only the probe flag is rejected, confirming the generated argv itself parses clean.

Targeted suites on this branch:

• tests/integrations/test_integration_forge.py — 30 passed
• tests/integrations/test_extra_args.py, test_base.py, test_integration_base_markdown.py (the shared machinery this override calls into) plus tests/test_agent_config_consistency.py — 177 passed, 1 skipped
• tests/integrations — 17 failed, 2634 passed, 14 skipped. All 17 failures are symlink tests that abort in fixture setup with OSError: [WinError 1314] A required privilege is not held by the client, because this Windows account has no symlink-creation privilege. None of them is a Forge test, and collection is 2661 on main versus 2665 here — exactly the four tests added, both figures measured. CI is the authority on pass/fail rather than my machine.

ruff reports the same 10 pre-existing findings on these two files before and after the change; this diff introduces none.

AI Disclosure

• I did not use AI assistance for this contribution
• I did use AI assistance (fill in the disclosure below)

AI disclosure: Agent/tool: Claude Code (interactive CLI session). Model: Claude Opus 5, 1M-token context window, high reasoning effort. Mode: human-supervised, not autonomous — I reviewed the diff and the tests line by line, approved each step, and ran the quoted commands on my own machine. Extent: code generation (the build_exec_args override and the four regression tests) plus drafting of this description, the linked issue #4666, and any comments I post on this PR. The Forge flag surface was verified against forge --help from the installed binary (@antinomyhq/forge v2.13.21) and against the upstream clap Cli struct, not from model recall.

[hmohammadi]

Follow-up on the Testing section: I now have a Linux environment for this repo, so the Windows caveat above is resolved. Full suite on this branch:

8359 passed, 17 skipped, 0 failed

Ubuntu 26.04.1 LTS, Python 3.14.4, PowerShell 7.6.6, uv run pytest against 5bf5cb0 with a clean working tree.

The failures reported earlier were symlink tests aborting in fixture setup with OSError: [WinError 1314] A required privilege is not held by the client — my Windows account lacks symlink-creation privilege. That was an artifact of the machine, not of this change, and it does not occur here.

This comment was drafted with Claude Code (Claude Opus 5); I ran the suite and verified the numbers on my own machine.