Jordan Ritter
3b1f628266
AG2's ConversableAgent runs every user message through
``autogen.code_utils.content_str``, which only accepts content-part
types in {"text", "input_text", "image_url", "input_image", "function",
"tool_call", "tool_calls"}. CopilotChat / the AG-UI runtime emits image
and document attachments as the modern shape
{"type": "image" | "document", "source": {...}}
and the demo page's legacy-converter-shim.tsx ALSO appends a legacy
{"type": "binary", mimeType, data | url}
mirror alongside it (to keep the @ag-ui/langgraph converter happy on
LangChain-based integrations — it rides through on the ag2 path too).
Both shapes trip autogen's allowed-types gate with
ValueError("Wrong content format: unknown type image within the
content")
…BEFORE the request reaches the vision model — observed live in the
D6 multimodal probe (commit d8a0a25db, which originally quarantined
the feature as NSF).
Fix
---
Add ``agents/_multimodal_normalize.py``: a ``NormalizingAGUIStream``
subclass of ``AGUIStream`` that overrides ``dispatch()`` to normalize
AG-UI image/document/binary content parts to OpenAI Chat Completions
``image_url`` parts AFTER ``RunAgentInput`` Pydantic parsing and BEFORE
``AgentService`` serialises the messages for autogen.
This is the only correct interception point:
- Too early (ASGI body rewrite before Pydantic): ``RunAgentInput``
rejects ``image_url`` because it is not an AG-UI standard type —
the discriminated union only accepts image/document/binary/text.
- Too late (inside ConversableAgent): requires patching autogen
internals.
The override works by calling ``normalize_messages_for_autogen()`` on
the dict-serialised messages (same form as ``run_stream`` produces via
``model_dump()``) and re-injecting them via a ``_PatchedRunAgentInput``
wrapper that overrides only ``.messages``, delegating all other
attribute access to the original ``RunAgentInput``.
Conversions:
- {"type": "image", "source": {"type": "data", value, mime_type}} →
{"type": "image_url", "image_url": {"url": "data:<mime>;base64,<value>"}}
- {"type": "image", "source": {"type": "url", value}} →
{"type": "image_url", "image_url": {"url": value}}
- {"type": "document", "source": ...} → image_url with the document's
mime preserved (data:application/pdf;base64,...). The vision model
still can't natively read PDFs, but the request reaches the model
instead of being rejected upstream, which is the failure mode this
fix targets.
- {"type": "binary", mimeType, data | url} → image_url (the
legacy-shim parts ride through cleanly).
- {"type": "text", ...} and already-normalised image_url parts pass
through unchanged (identity-preserved on no-op turns).
Failure path: any normalization error is logged at WARNING and the
original messages are forwarded unchanged — autogen's own ValueError
fires verbatim with its error surface intact.
Manifest + fixture
------------------
- showcase/integrations/ag2/manifest.yaml: remove multimodal from
not_supported_features (with its now-stale comment) and add it back
to the features list next to voice.
- showcase/aimock/d6/ag2/multimodal.json: add the D6 fixture pair
using the actual autoPrompt strings from sample-attachment-buttons.tsx
("can you tell me what is in this demo image I just attached" /
"can you tell me what is in this demo pdf I just attached").
TDD evidence (red-green)
------------------------
showcase/integrations/ag2/tests/python/test_multimodal_normalize.py
contains 14 unit tests, pinned at three layers:
1. RED/GREEN against autogen's actual content gate:
* test_autogen_rejects_raw_agui_image_part — confirms
content_str([{type: image, source: ...}]) raises the verbatim
ValueError the D6 probe surfaced. This is the regression pin: if
autogen ever relaxes the gate, this test fails and we know to
revisit the normalizer.
* test_normalized_content_is_accepted_by_autogen — after
normalize_messages_for_autogen(...), content_str accepts every
part and renders "<image>" for the image_url part.
2. Shape coverage: modern image data/url, modern document, legacy
binary data/url, mimeType camelCase alias, plain-text passthrough,
plain-string content, assistant/tool messages untouched,
unrecognised source → text placeholder, idempotency.
3. NormalizingAGUIStream class surface tripwire.
Control-plane D6 RED→GREEN:
RED (no normalizer, pre-fix container): d6:ag2/multimodal → red
(HTTP 500 agent_run_error_event from content_str ValueError)
GREEN (NormalizingAGUIStream applied): d6:ag2/multimodal → green
Showcase aimock
Tagline: aimock fixture-directory layout, context-routing semantics, schema
validation, drift-risk surface, and the manual add/update fixture flow. For
service reconstruction (Railway image / startCommand / fixture URLs) see
./RAILWAY.md. For matcher semantics / fixture-authoring
gotchas (sequenceIndex, hasToolResult, context mirroring) see
../GOTCHAS.md.
Deterministic LLM fixture server for showcase E2E testing. Replaces real LLM API calls (OpenAI, Anthropic, Gemini) with pre-recorded responses so Playwright tests can run PR-gated in CI without API keys and without rate limits or non-determinism.
Railway pulls ghcr.io/copilotkit/aimock:latest directly (no wrapper image). The fixtures in this directory are loaded at boot via GitHub raw URLs configured in the Railway service's startCommand.
What aimock is
aimock (@copilotkit/aimock) is a general-purpose LLM mock server. It speaks the OpenAI, Anthropic, and Gemini REST shapes (including SSE streaming), loads fixtures from disk at startup, and responds to incoming chat completions by matching the user's message text against fixture match criteria.
The showcase deployment runs aimock in proxy mode — --proxy-only with real upstream URLs configured for each provider. Unmatched requests are forwarded to the real API; matched requests short-circuit with the fixture response. This makes the sidecar safe to deploy as a general-purpose smoke-test aid: tests that hit fixture-matched prompts get deterministic responses, and anything else just falls through.
Directory Structure
showcase/aimock/
shared/ Fixtures loaded by ALL integrations (smoke, universal prompts)
d4/ D4-depth fixtures — per-integration, single-demo coverage
<slug>/ One directory per integration slug (e.g. langgraph-python/)
d6/ D6-depth fixtures — per-integration, all-pills coverage
<slug>/ One directory per integration slug
feature-parity.json Legacy flat fixture file (pre-context-routing)
smoke.json Minimal smoke fixture
README.md This file
Context routing. D4 and D6 fixtures use aimock's --context-field flag to scope fixture matching by integration. Each integration's dev server passes its slug as the context value in LLM requests (via X-AIMock-Context header or request body field). Aimock only considers fixtures whose match.context equals the incoming context value, so d4/langgraph-python/ fixtures never interfere with d4/mastra/ fixtures even if they share the same userMessage pattern.
Per-integration isolation. Every <slug>/ directory contains fixtures specific to that integration. This prevents cross-contamination: if mastra needs a different tool name than langgraph-python for the same demo, each has its own fixture file. The shared/ directory holds fixtures that apply regardless of context (e.g., smoke checks, universal greeting prompts).
Fixtures in this directory
feature-parity.json— 35+ fixtures covering the nine showcase demos across 17 packages: agentic chat (weather, backgrounds, themes), tool rendering (pie/bar charts, weather cards), HITL (plans, steps, approvals), Sales Dashboard (deals, pipelines, todos), and assorted meeting/flight/greeting prompts. Consumed by the per-packagetest_e2e-showcase-on-demandPlaywright suites and loaded at Railway boot via GitHub raw URL.smoke.json— a single minimal fixture (userMessage: "Respond with exactly: OK"→content: "OK"). Used by/api/smokeendpoints in each package to verify the aimock → package → UI round-trip without depending on a real agent.
Fixture match semantics: userMessage is a substring match against the last user turn. First fixture to match wins, so more specific prompts should appear before more generic ones (see the "Based on the following context, write a concise" entry that precedes the generic report / plan fixtures to protect CrewAI's startup probe).
Sync policy
Fixtures are hand-maintained. There is no automated capture, no scheduled re-recording, and no drift-detection job that compares fixture responses against what a real LLM would say. The authoritative behavior is whatever is checked in.
The safety net is two-layered load-time validation, not behavioral verification:
- Load-time schema validation (
--validate-on-loadin the RailwaystartCommandand in every test entrypoint that boots aimock) — the container refuses to start if any fixture uses an unrecognized response key (e.g.textinstead ofcontent). See #3973. - CI schema validation (
showcase/scripts/__tests__/aimock-fixtures.test.ts) — theshowcase_validateworkflow runsloadFixtureFile+validateFixturesfrom@copilotkit/aimockagainst everyshowcase/aimock/*.jsonon every PR. A broken fixture fails the PR before merge.
Neither layer catches behavioral drift — if a package's agent code changes what it asks the LLM (new prompt, new tool, renamed tool), the existing fixture keeps matching and keeps returning the old response. The test either keeps passing (wrong assertion) or fails at the UI-assertion layer (missing tool call, missing text), and a human has to trace it back to the fixture.
Adding or updating a fixture
The process is manual. There is no CLI for this directory specifically — aimock's upstream --record mode can proxy real API calls and write fixtures, but the showcase repo does not wire it up and does not commit recorded fixtures.
- Identify the user prompt your test issues and decide what response you need (plain text, a tool call, an error).
- Add an entry to
feature-parity.jsonunderfixtures. Keep more specificuserMessagematches above more generic ones. Valid response keys:content,toolCalls,error,embedding. - Run the fixture-validation suite locally:
pnpm --filter @copilotkit/showcase-scripts test aimock-fixtures - Run the per-package E2E against the new fixture:
./showcase/scripts/run-e2e-with-aimock.sh <slug> [test-filter] - Ship it. Fixture changes take effect on the next Railway service restart (aimock fetches fixtures from GitHub raw URLs at boot).
When a package's agent code changes in a way that changes its LLM calls, the person making the change is responsible for updating the corresponding fixture. There is no automation to remind you.
Drift risk
Drift surfaces as flaky or silently-wrong E2E tests, not as a dedicated signal. Symptoms and how to respond:
- Playwright assertion fails on a UI element that depends on a tool call (
WeatherCardmissing, chart not rendering) → the agent is now calling a differently-named tool than the fixture has; update the fixture'stoolCalls[].name/arguments. - Assertion on assistant text fails → the agent's prompt changed; either update the fixture's
match.userMessageto the new prompt substring or update the fixture'scontent. smoke.jsonhealthcheck fails against a deployed package (/api/smokereturns non-OK) → either the package's smoke route changed or aimock is down; check the Railway service and the smoke-monitor workflow.- Container fails to start post-deploy → load-time validation caught a broken fixture; CI should have caught it first, investigate why it didn't.
There is no scheduled drift-detection job that compares fixture responses against live LLM output. If this becomes a problem, the path forward is to wire aimock's --record mode into a periodic workflow that re-captures against real providers and diffs against checked-in fixtures — but that's not built today.
Related workflows
test_e2e-showcase-on-demand.yml(historicallyshowcase_aimock-e2e.yml) — triggered by/test-aimock <slug>PR comments orworkflow_dispatch. Installs@copilotkit/aimock@latest, boots it withfeature-parity.json, spins up the target package's dev server againstOPENAI_BASE_URL=http://localhost:4010/v1, and runs the package's Playwright suite. Posts pass/fail back to the PR.showcase_validate.yml— runs fixture schema validation (aimock-fixtures.test.ts) on every PR that touchesshowcase/**.showcase_deploy.yml— builds and deploys all showcase services. aimock is no longer in this workflow's matrix (Railway pulls the upstreamghcr.io/copilotkit/aimock:latestimage directly).showcase_smoke-monitor.yml— every 15 minutes, polls/api/smokeon all deployed showcase packages. Those smoke endpoints internally hit aimock'ssmoke.jsonfixture to verify the full stack.