fix(pydantic): install the pydantic-ai CLI and match the real pai interface

[Copilot]

The Smoke Pydantic AI job fails at the Predownload Pydantic AI CLI step with error: Failed to spawn: pai. uv run pai only resolves commands from an existing uv project, so the CLI was never installed — and the rest of the engine definition was written against a pai interface that does not exist.

Verified against pydantic-ai 2.26.0: pai has no run subcommand (the prompt is positional), the model must be passed as -m provider:model, there is no PAI_MODEL env var, and MCP servers can only be supplied through an agent spec passed with -a.

shared/pydantic.md

• Install via pip install --user "pydantic-ai==$GH_AW_ENGINE_VERSION" and verify with "$HOME/.local/bin/pai" --version, mirroring the Aider engine. uv lives under /opt/hostedtoolcache, which is not mounted into the AWF container; ~/.local/bin is. Engine version pinned to 2.26.0 (was a placeholder 0.1.0).
• Execution changed from uv run pai run <prompt> to pai --no-stream -m "$PAI_MODEL" [-a .pydantic-ai/agent.json] "<prompt>".
• Model is rewritten to openai-chat:<model> so it resolves to OpenAIChatModel behind the AWF OpenAI-compatible proxy (openai: would select the Responses API).
• Replaced the no-op mcp-config-env-var: GH_AW_MCP_CONFIG with an MCP config adapter that renders gateway servers into a Pydantic AI agent spec, passed with -a. Non-HTTP servers are skipped with a log line.
• Added pypi.org / files.pythonhosted.org network defaults.

Generated spec:

{
"model": "openai-chat:gpt-5",
"capabilities": [
{ "MCP": { "id": "safeoutputs", "url": "http://host.docker.internal:9999/mcp/safeoutputs", "headers": { "Authorization": "…" } } }
]
}

The spec-level model is a valid-by-construction fallback; -m takes precedence whenever the workflow declares a model.

pkg/workflow

• New execution.model-env-provider-separator for behavior-defined engines, defaulting to / so existing engines (aider, opencode) are unaffected. Needed to emit openai-chat:claude-sonnet-4-5 rather than openai-chat/claude-sonnet-4-5. Schema, generated frontmatter reference, and a table-driven unit test included.

smoke-pydantic.md

• Test requirements now cover model connectivity and MCP tool availability. The previous file-writing / bash / git log assertions could never pass: pai is a chat CLI and ships no filesystem or shell tools.

Lock file recompiled.

---

Generated by 👨‍🍳 PR Sous Chef · gpt54 · 17.1 AIC · ⌖ 6.55 AIC · ⊞ 6.1K · ◷
_{Comment /souschef to run again}

[github-actions]

Triage

Category: feature | Risk: high | Score: 30/100

Component	Score
Impact	15/50
Urgency	10/30
Quality	5/20

Recommended action:defer

Still a draft/WIP with most of the checklist unchecked and no diff yet (0 changed files). Revisit once implementation is pushed and CI runs.

Generated by 🔧 PR Triage Agent · auto · 39.7 AIC · ⌖ 2.79 AIC · ⊞ 8K · ◷

[pelikhan]

/smoke-pydantic

[github-actions]

⚠️Smoke Pydantic AI failed. Pydantic AI encountered unexpected challenges...

🐍 Smoke Pydantic AI — Powered by Pydantic AI

[pelikhan]

@copilot fix https://github.com/github/gh-aw/actions/runs/31231477754/job/93036140031#step:26:1

[Copilot]

@copilot fix https://github.com/github/gh-aw/actions/runs/31231477754/job/93036140031#step:26:1

Fixed in c99a2d4. The step failed with awf: command not found (exit 127) — BehaviorDefinedEngine.GetInstallationSteps returned nil when the engine declares no installation block and no harness-script, so the Install AWF binary step was never emitted even though the agent runs inside the firewall sandbox. Both that path and the non-npm package-manager path now go through BuildNpmEngineInstallStepsWithAWF(nil, workflowData), which emits only the AWF steps when the firewall is enabled. Lock file recompiled.

[github-actions]

Great work on fixing the Pydantic AI engine integration! 👍

This PR successfully addresses the root cause of the smoke test failures by:

• Installing the pai CLI properly — via pip install rather than relying on uv run, which wasn't resolving commands correctly
• Matching the real pai interface — correcting the model format (openai-chat:model instead of /), MCP config handling through agent specs, and execution flags
• Updating test expectations — focusing on model connectivity and MCP tool availability rather than filesystem/shell operations that pai doesn't support
• Adding comprehensive unit tests — new behavior_defined_engine_model_test.go with table-driven scenarios

The fix is focused, well-documented, and includes the necessary test coverage. This is ready for review and merge. The pr-priority:low and pr-action:defer labels are fine — this can go in at the right time for the maintainers.

Generated by ✅ Contribution Check · auto · 89.7 AIC · ⌖ 2.91 AIC · ⊞ 8.7K · ◷

[pelikhan]

/smoke-pydantic

[github-actions]

⚠️Smoke Pydantic AI failed. Pydantic AI encountered unexpected challenges...

🐍 Smoke Pydantic AI — Powered by Pydantic AI

[pelikhan]

@copilot fix https://github.com/github/gh-aw/actions/runs/31232615654/job/93039364484#step:27:1

[pelikhan]

/smoke-pydantic

[github-actions]

⚠️Smoke Pydantic AI failed. Pydantic AI encountered unexpected challenges...

🐍 Smoke Pydantic AI — Powered by Pydantic AI

[pelikhan]

@copilot Fix https://github.com/github/gh-aw/actions/runs/31238152428/job/93054524987

[Copilot]

@copilot Fix https://github.com/github/gh-aw/actions/runs/31238152428/job/93054524987

Fixed in 3a7cb0f.

Two issues, both confirmed from the run artifacts and reproduced locally against pydantic-ai 2.26.0.

1. The copilot/ model prefix was wrong. The previous commit added it on the theory that the api-proxy steers by model name. The artifacts disprove that — comparing the api-proxy otel spans:

run	model sent	span	status
31237213483	claude-sonnet-4.5	api_proxy.copilot.request	200, gen_ai.response.model: claude-sonnet-4.5, 3057/397 tokens
31238152428	copilot/claude-sonnet-4.5	api_proxy.copilot.request	400 model_not_supported

The bare name already steers to Copilot (steering is by port, fixed in 0f8ba8d), and models.json from the proxy lists claude-sonnet-4.5, not copilot/claude-sonnet-4.5. The harness now strips the provider segment again.

2. The real remaining failure was the response shape. Run 31237213483 got a clean 200 and still failed, with ValidationError: choices.{0..3}.index → None, object → None. That is the proxy's aggregated non-streaming body, which pydantic-ai validates strictly via _ChatCompletion.model_validate. Reproduced locally with a mock server returning a body missing object/index: --no-stream raises exactly that error, while the streaming path parses the same content fine. --no-stream is therefore dropped so pai uses the streaming path — which is also the shape Copilot natively produces.

Verified end-to-end by extracting the harness from the compiled lock file and running it against a mock proxy: it sends model=claude-sonnet-4.5, stream=true, and exits 0.

[pelikhan]

/smoke-pydantic

[github-actions]

⚠️ Security scanning failed for Smoke Pydantic AI. Review the logs for details.

🐍 Smoke Pydantic AI — Powered by Pydantic AI

[github-actions]

Smoke Test: Pydantic AI - ✅ PASS

1. Model Connectivity: ✅
2. MCP Tool Availability: ✅

Overall Status: PASS

Run: https://github.com/github/gh-aw/actions/runs/31242043047

🐍 Smoke Pydantic AI — Powered by Pydantic AI · sonnet45 · 7.23 AIC · ⊞ 4.3K · ◷
_{Comment /smoke-pydantic to run again}

[pelikhan]

@copilot resolve the merge conflicts on this branch.

[Copilot]

Pull request overview

Fixes Pydantic AI CLI installation and execution through the AWF proxy with MCP support.

Changes:

• Installs and invokes pai 2.26.0 correctly.
• Adds a Pydantic AI harness and MCP agent-spec adapter.
• Updates smoke-test expectations and generated workflow.

Show a summary per file

File	Description
pkg/workflow/behavior_defined_engine.go	Adds AWF installation for non-npm engines.
.github/workflows/shared/pydantic.md	Defines installation, proxy routing, and MCP integration.
.github/workflows/smoke-pydantic.md	Revises smoke-test requirements.
.github/workflows/smoke-pydantic.lock.yml	Recompiled generated workflow.

Review details

Tip

Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

• Files reviewed: 4/4 changed files
• Comments generated: 2
• Review effort level: Balanced

[gh-aw-bot]

@copilot Quick triage nudge for this PR.

Please refresh the branch if GitHub can update it cleanly, address the remaining review feedback, and run the pr-finisher skill before handing back to maintainers.

Open items (newest first):

• merge conflicts are currently reported on this branch; please resolve them before final hand-off
• no recorded failed checks in compact candidate data

Run: https://github.com/github/gh-aw/actions/runs/31242800863

Generated by 👨‍🍳 PR Sous Chef · gpt54 · 17.1 AIC · ⌖ 6.55 AIC · ⊞ 6.1K · ◷
_{Comment /souschef to run again}

[github-actions]

🎉 This pull request is included in a new release.

Release: v0.86.2