[RaycarlLei]

Description

PiIntegration does not override build_exec_args(), so it inherits MarkdownIntegration.build_exec_args() (base.py#L1139-L1158), which appends --output-format json when output_json is true.

Pi has no --output-format option. It selects JSONL events with --mode json, and --mode accepts only text, json and rpc (packages/coding-agent/src/cli/args.ts#L96-L110, documented in docs/cli.md#L43-L48). Any other long option is kept as a possible extension flag (args.ts#L249-L262). Once extensions are loaded, a flag that no extension registers becomes an Unknown option error (core/agent-session-services.ts#L99-L125), and pi exits 1 (main.ts#L916-L925) before it sends the prompt. 71023ae is the v1.3.0 tag, published as @earendil-works/pi-coding-agent@1.3.0, the current npm latest. This reproduces without a provider account. With HOME and PI_CODING_AGENT_DIR pointing at empty directories (not logged in) and PI_OFFLINE=1 PI_SKIP_VERSION_CHECK=1 PI_TELEMETRY=0:

$ pi --version
1.3.0
$ pi -p x --output-format json
Error: Unknown option: --output-format
$ echo $?
1
$ pi -p x --mode json
{"type":"session","version":3,"id":"01a15c0d-ddf6-7299-9d2f-cb8c4e2c9555","timestamp":"2026-10-21T01:22:55.094Z","cwd":"<empty dir>"}
No API key found for the selected model.

Use /login to log into a provider via OAuth or API key. See:
  <pi package>/docs/providers.md
  <pi package>/docs/models.md

The --mode json run gets past option parsing and writes its first JSONL record to stdout. It then exits 1 only because that empty agent directory has no credentials.

The same failure through Spec Kit on main (505781d), with pi 1.3.0 on PATH:

$ .venv/bin/python -c "from specify_cli.integrations.pi import PiIntegration as P; r = P().dispatch_command('speckit.plan', 'x', stream=False); print(r['exit_code'], r['stderr'].strip())"
1 Error: Unknown option: --output-format

Reachability. The built-in workflow steps do not hit this. The command step either streams, where output_json is left as None and resolves to not stream, i.e. False (base.py#L462-L463), or captures with an explicit output_json=False (step/command/__init__.py#L313-L319). The prompt step passes output_json=False (step/prompt/__init__.py#L248-L253). It is reached through the public dispatch_command(..., stream=False) when output_json is left unset (it then resolves to True), through direct build_exec_args() calls (default output_json=True), and by extensions or other callers of those APIs.

Fix. Override build_exec_args() in PiIntegration so structured output requests --mode json, the flag the OMP integration already passes (omp/__init__.py#L29-L53). The rest matches the base implementation: the requires_cli guard, -p <prompt>, SPECKIT_INTEGRATION_PI_EXTRA_ARGS before --model, and --model, which Pi accepts (args.ts#L117-L118). Unlike OMP, the prompt can stay right after -p: Pi's -p takes the next argument as the message when it does not start with - or @ (args.ts#L177-L183), and Spec Kit's dispatch prompts start with /. The change stays inside the Pi integration package, its test file, and its agent page docs/integrations/pi.md, which named --output-format json.

Part of #2416, which lists Pi's inherited pi -p "prompt" --model X --output-format json argv as unverified. Same shape as the Kimi Code fix in #5031.

Testing

• Tested locally with uv run specify --help

• Ran existing tests with uv sync && uv run pytest: run as .venv/bin/python -m pytest, results below

• Tested with a sample project (if applicable): see the first bullet below

• Sample project: specify init proj --integration pi --script sh --non-interactive, then dispatch_command('speckit.plan', 'x', stream=False) against pi 1.3.0. No provider account was used: PI_CODING_AGENT_DIR points at a models.json for a local OpenAI-compatible stub model, and its settings.json sets defaultProjectTrust: "always" so Pi loads the project's .pi/prompts. On main it exits 1 with the error above. On this branch it exits 0 and stdout holds 15 JSONL events, from session to agent_settled; the user message in them is the rendered speckit.plan template with x as its arguments. Without that trust setting (Pi's default is ask), a non-interactive run skips the project's prompt templates and the model receives the literal /speckit.plan x; that is independent of this argv fix and not changed here.

• tests/integrations/test_integration_pi.py gets four argv tests: text mode, --mode json, --model, and SPECKIT_INTEGRATION_PI_EXTRA_ARGS. Run against main, the --mode json and extra-args cases fail with '--output-format' != '--mode'. All four pass on this branch.

• .venv/bin/python -m pytest tests/integrations tests/workflows tests/test_agent_config_consistency.py -q: 3620 passed, 8 skipped.

• uvx ruff@0.15.0 check src tests: all checks passed.

• Full suite, .venv/bin/python -m pytest tests -q (macOS, Python 3.14.3): 12517 passed, 376 skipped, 15 failed. All 15 are in 6 files this PR does not touch, and running those files on unmodified main on the same machine fails the same 15 test IDs. 11 are composed-template cases in test_*_python_parity.py, whose bash variant runs python3 from PATH, and that interpreter has no PyYAML on this machine; they pass with the venv's python3 first on PATH. The other 4 are in test_workflow_step_submission.py, where the fake curl gets a #!<sys.executable> shebang and the interpreter path on this machine contains a space.

AI Disclosure

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

AI disclosure: Claude Code opened this PR on behalf of @RaycarlLei. Claude Code (model: Claude Opus 5.5, reasoning effort xhigh, autonomous agent mode with tool approvals bypassed) reproduced the bug against pi 1.3.0, read the Pi source for the accepted --mode values and its unknown-option handling, wrote the override, the four tests and the docs/integrations/pi.md update, ran the checks above, wrote this description, and pushed the branch. A read-only Claude Code subagent on the same model cross-checked the description's claims and the diff, and its corrections were applied before posting. @RaycarlLei reviewed the diff and a draft of this description and approved posting, but did not re-run the tests or edit the text. The commit carries Assisted-by: Claude Code (model: Claude Opus 5.5, autonomous).

🤖 Generated with Claude Code