Three durability improvements suggested in the PR #3241 top-level review: ## 1. Golden daemon-event snapshots (mocks/golden/*.events.json + apps/daemon/tests/mocks-golden.test.ts) Smoke-test verified that mocks RUN; that catches crashes but not a parser change that semantically reshapes the events the daemon emits. Commit the daemon-event sequence for 3 representative traces: - claude 314d6833 — median-complexity agent-browser session - codex dcdff3b3 — 14-tool refactor - opencode 9a9522ec — 7-tool data-report apps/daemon/tests/mocks-golden.test.ts spawns the mock, feeds stdout through the real createClaudeStreamHandler / createJsonEventStreamHandler, normalizes per-spawn volatile fields (only sessionId today, only on claude), and deep-equals against the committed snapshot. A parser regression fails the test loudly. After an intentional parser change, regenerate: MOCKS_GOLDEN_UPDATE=1 pnpm --filter @open-design/daemon test mocks-golden git diff mocks/golden/ # eyeball; commit if shapes match intent ## 2. Provenance fields on every manifest entry (mocks/scripts/lib/manifest-utils.mjs + mocks/manifest.json) Augment inspectRecording() to write: captured_at — ISO 8601 from existing meta.timestamp cli_version — null until harvester writes it protocol_version — null until harvester writes it anonymization_version — null until harvester writes it captured_at is now populated for all 179 existing entries from the meta event the harvester already emits. The harvester in nexu-io/agent-pr-explore is the next step for cli_version / protocol_version / anonymization_version — once those are populated, consumers can detect when a recording is older than ~1 minor version behind the live CLI and flag for re-harvest. No matrix of (cli_version × agent) recordings — that explodes maintenance. Just metadata per recording so trust decay is visible. ## 3. Real-CLI contract check (mocks/scripts/contract-check.sh + docs/MOCKS-CONTRACT-CHECK.md) Mocks catch parser regressions against recordings; they do NOT catch recordings drifting away from the live agent CLI as that CLI evolves. The contract check spawns the real CLI alongside the mock with a fixed deterministic prompt + diffs top-level event-type distributions. Deliberately human-driven, not cron-scheduled: - costs real LLM tokens per invocation - requires real CLI auth - maintainer reads the output, not a regex Suggested triggers per doc: real-CLI release notes mentioning "output format" / "stream" / "JSON" / "events"; before a parser refactor; ad-hoc when something looks off. ## Coverage note README updated to position mocks as "deterministic protocol/parser coverage" (not "e2e replacement") per mrcfps framing. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| bin | ||
| golden | ||
| lib | ||
| recordings | ||
| scripts | ||
| manifest.json | ||
| mock-agent.mjs | ||
| README.md | ||
mocks/ — replay-based mock CLIs for OD's supported agents
A drop-in replacement for the real agent CLIs (claude, opencode,
codex, gemini, cursor-agent, deepseek, qwen, grok, the
ACP family devin / hermes / kilo / kimi / kiro / vibe, and
the AMR vela CLI) that replays pre-recorded sessions in each CLI's
native protocol — stdout streaming for most, JSON-RPC over stdio for
ACP and AMR. Zero LLM tokens.
Used by:
- E2E tests in
apps/daemon/tests/— run the full chat-server pipeline against a known agent trace, assert UI events / artifacts. - Local self-tests during development — iterate on
chat-routes.ts,claude-stream.ts,json-event-stream.tsparser changes without burning provider budget. - Demo / onboarding — show what a 17-tool
claudeediting session looks like end-to-end, offline. - Regression harness — replay the same trace before and after a charter / parser change; diff the events the daemon surfaces.
The recordings are anonymized exports from open-design's Langfuse project (179 traces across 9 agents and 5+ skills as of this commit).
tl;dr
# First-time setup — pull the recording corpus from R2 (~30s, 4.5MB):
bash mocks/scripts/fetch-recordings.sh
# Subsequent runs hit the local cache (sha256-verified, instant).
# Make the mock CLIs override the real ones for this shell:
export PATH="$PWD/mocks/bin:$PATH"
# Pick any recording to play back (8-char prefix OK):
export OD_MOCKS_TRACE=04097377
# Speed up replay (skip inter-event sleeps):
export OD_MOCKS_NO_DELAY=1
# Now anything that spawns opencode/claude/codex gets the recording:
echo "any prompt body" | opencode run
echo "any prompt" | claude -p --output-format=stream-json
echo "any prompt" | codex exec
The mock binaries are bash wrappers that exec
node mocks/mock-agent.mjs --as <agent>. Anything fed to stdin is
discarded by the renderer but used by the recording picker (see hash
mode below).
Recordings live on R2, not in this repo
The 179-recording corpus (~4.5 MB) is hosted on Cloudflare R2 at
open-design-mocks and fetched on demand — pnpm install does NOT
pull them, and the repo stays small. Recordings only land in
mocks/recordings/ when:
- You run
bash mocks/scripts/fetch-recordings.shdirectly, OR bash mocks/scripts/smoke-test.shruns and the dir is empty (auto- fetch fallback), OR- A mock binary spawn finds no data — it errors with a pointer at the fetch script (no silent failure).
This is by design: contributors who don't touch agent code don't pay
the fetch cost. CI jobs that DO touch agent code (apps/daemon/tests/
parser changes, etc.) run the fetch as a quick pre-step and cache
mocks/recordings/ between runs.
# Fetch everything (parallel, sha256-verified, idempotent):
bash mocks/scripts/fetch-recordings.sh
# Fetch a subset:
bash mocks/scripts/fetch-recordings.sh --agent claude # 57 claude traces
bash mocks/scripts/fetch-recordings.sh --outcome failed # 35 failed-path traces
bash mocks/scripts/fetch-recordings.sh --skill agent-browser
# Override cache location (e.g. share across multiple OD checkouts):
OD_MOCKS_CACHE_DIR=~/.cache/od-mocks bash mocks/scripts/fetch-recordings.sh
Manifest at mocks/manifest.json is the committed source of truth —
it lists every recording's trace_id, sha256, bytes, agent,
outcome, skills, multi_turn, plus histograms over the corpus.
Tooling reads this; you don't have to.
Provenance per recording
Beyond identity (trace_id, sha256), each manifest entry carries
fixture-trust signals so consumers can decide whether the recording
is still meaningful as the real CLIs evolve:
| Field | Meaning |
|---|---|
captured_at |
ISO 8601 timestamp of the original session — populated for all 179 current entries |
cli_version |
The CLI version the trace was captured against (e.g. "claude-code 1.0.65") — populated only on traces the harvester writes it to, null otherwise |
protocol_version |
Stream-format version ("claude-stream-json/v1", "opencode/json-event-stream") — populated by harvester |
anonymization_version |
Which anonymizer pass scrubbed the recording — populated by harvester |
For now most of these are null on the existing 179 — the harvester in
nexu-io/agent-pr-explore is the next thing to teach to
write them. Once a recording's cli_version falls behind the actual
CLI by more than one minor version, treat it as a candidate for
re-harvest.
Golden daemon-event snapshots
mocks/golden/<trace>.events.json holds the exact event sequence the
OD daemon emits when fed each (mock CLI → handler) pipeline. Diffed
on every pnpm --filter @open-design/daemon test run by
apps/daemon/tests/mocks-golden.test.ts.
A parser refactor that semantically changes events (drops a field,
renames sessionId, stops emitting turn_end) fails the diff loudly.
After an intentional parser change, regenerate:
MOCKS_GOLDEN_UPDATE=1 pnpm --filter @open-design/daemon test mocks-golden
git diff mocks/golden/ # eyeball the new shapes
git add mocks/golden/ && git commit -m "mocks: refresh goldens for <parser change>"
Per-spawn volatile fields (currently just claude's generated
sessionId) are stripped to "<normalized>" so the snapshot stays
stable. See mocks/golden/README.md for the coverage rationale.
Real-CLI contract check
The mocks catch parser regressions against the recordings; they do
not catch the recordings themselves drifting away from the live
agent CLIs. For that, mocks/scripts/contract-check.sh spawns a real
CLI alongside the mock with a fixed prompt and prints a side-by-side
event-type distribution.
This is human-driven and costs real LLM tokens — run on a real-CLI
release or before a parser refactor, not on a cron. Full doc:
docs/MOCKS-CONTRACT-CHECK.md.
What gets emitted
Each renderer matches the EXACT event shapes the OD daemon expects, as
verified line-by-line against the parsers in apps/daemon/src/:
| CLI | OD streamFormat | Parser source |
|---|---|---|
opencode |
json-event-stream (opencode kind) |
json-event-stream.ts:handleOpenCodeEvent |
codex |
json-event-stream (codex kind) |
json-event-stream.ts:handleCodexEvent |
claude |
claude-stream-json |
claude-stream.ts:createClaudeStreamHandler |
gemini |
json-event-stream (gemini kind) |
json-event-stream.ts:handleGeminiEvent |
cursor-agent |
json-event-stream (cursor-agent kind) |
json-event-stream.ts:handleCursorEvent |
deepseek qwen grok |
plain |
server.ts (raw stdout = final assistant text) |
devin hermes kilo kimi kiro vibe |
acp-json-rpc |
acp.ts:attachAcpSession |
vela (AMR) |
acp-json-rpc + login / models subcommands |
runtimes/defs/amr.ts + apps/daemon/tests/fixtures/fake-vela.mjs (sibling stub) |
Note on
geminiandcursor-agent: OD's parsers for these two agents do NOT recognize tool-call events — only init / assistant text / usage. The renderers therefore emit ONLY the final assistant text wrapped in the expected init/text/usage envelope. Tool calls present in the source recording are silently dropped (which matches the real CLI's UI behavior — these agents don't surface tools in OD's chat view).
Note on ACP agents (
devin/hermes/kilo/kimi/kiro/vibe): These do NOT stream stdout — they speak JSON-RPC v2 over stdio. OD's daemon sendsinitialize→session/new→ (optionalsession/set_model) →session/prompt; the mock responds in order, streams text viasession/updatenotifications carryingagent_message_chunkparts, then responds to the prompt request with usage stats. Tool calls aren't part of the ACP protocol on this path (tools surface via MCP or other side channels), so they're dropped from playback.
Note on
vela(AMR): vela is the bin OD's AMR runtime spawns. It extends the generic ACP shape withagentCapabilities+modelsblocks ininitialize/session/new, plus a strict set_model gate —session/promptis rejected with -32602 untilsession/set_model(orsession/set_config_option) has been called for the current sessionId, mirroring real vela 0.0.1 contract.vela also has two non-ACP subcommands:
vela login→ writes~/.amr/config.jsonwith a fake profile so OD's daemon login route +AmrLoginPillpoller see the same on-disk projection production produces.vela models→ prints the production-shapedpublic_model_* velacatalog.Error injection envs (kept in sync with
apps/daemon/tests/fixtures/fake-vela.mjs):FAKE_VELA_SESSION_NEW_ERROR/FAKE_VELA_SET_MODEL_ERROR/FAKE_VELA_PROMPT_ERROR/FAKE_VELA_LOGIN_FAIL/FAKE_VELA_REQUIRE_SET_MODEL=0.
Each tool call from the recording is rendered with the original input arguments and tool output. The agents' assistant text is rendered as the final message.
Recording selection
Driven by env vars, in priority order:
| Env | Behavior |
|---|---|
OD_MOCKS_TRACE=<id> |
Always play this trace. 8-char prefix OK. |
OD_MOCKS_BY_PROMPT_HASH=1 + stdin prompt |
Deterministic by sha256(prompt) % len(all). Same prompt → same trace. Useful for "stable answer per question" tests. |
OD_MOCKS_POOL=<tag> |
Random within the tag pool. Examples: agent:claude, skill:agent-browser, outcome:failed. |
OD_MOCKS_SEED=<str> |
Makes "random" picks reproducible across runs. |
OD_MOCKS_NO_DELAY=1 |
Skip inter-event waits. |
OD_MOCKS_RECORDINGS_DIR=<path> |
Override the recordings dir. |
If none are set, a uniformly random recording is played each invocation.
The mock binary announces the picked trace id on stderr:
[mock-opencode] picked 04097377… via fixed
This line is invisible to OD's stdout parser but useful for "wait, why did my test get the FAQ-fix trace?" debugging.
Recording catalog
The recordings live as one JSONL file per Langfuse trace under
recordings/. Each file starts with a meta event carrying:
{
"type": "meta",
"source": {"provider": "langfuse", "trace_id": "...", "project_id": "..."},
"agent": "claude" | "codex" | "opencode" | "gemini" | "cursor-agent" | "qwen" | "copilot" | "deepseek" | "antigravity",
"model": "...",
"outcome": "succeeded" | "failed" | "errored" | "interrupted",
"duration_ms": 33620,
"tool_call_count": 17,
"error_count": 0,
"total_tokens": 12345,
"tags": ["agent:claude", "skill:agent-browser", "open-design", ...],
"user_input": "...",
"session_id": "..."
}
Subsequent events are tool_call, tool_result, and report (the
final assistant text).
Indexed metadata
mocks/manifest.json is a flat manifest with one entry per recording
plus histograms over all recordings, committed to the repo. It's also
mirrored to R2 alongside the .jsonl files so consumers can fetch the
current catalog without cloning. Query with jq:
# All multi-turn claude sessions about HTML editing
jq '.entries[] | select(.agent=="claude" and .multi_turn==true)' \
mocks/manifest.json | head -50
# Failed codex traces (negative-path tests)
jq '.entries[] | select(.agent=="codex" and .outcome=="failed") | .trace_id' \
mocks/manifest.json
# Agent-browser skill, sorted by tool count desc
jq '[.entries[] | select(.skills | index("agent-browser"))] | sort_by(-.tool_count)' \
mocks/manifest.json
Headline stats (current dataset)
| Dimension | Distribution |
|---|---|
| Agents | claude 57 · opencode 41 · codex 38 · gemini 25 · cursor-agent 11 · qwen/copilot/deepseek 2 each · antigravity 1 |
| Outcomes | succeeded 144 · failed 35 |
| Skills | default 71 · ad-creative 50 · algorithmic-art 30 · agent-browser 22 · video-hyperframes 2 · magazine-web-ppt / brainstorming / data-report / penpot-flutter 1 each |
| Multi-turn | 124 traces tied to a session with ≥2 turns |
| Artifact | 18 traces produce <artifact> output |
Anonymization
User-specific data has been scrubbed from every recording:
/Users/<name>/…,/home/<name>/…,C:\Users\<name>\…→${HOME}/…/%USERPROFILE%\…- Project UUIDs → stable
proj-001,proj-002, … per recording - meta tag
project:<uuid>rewritten too
The anonymizer is idempotent. Tool input/output payloads (HTML, code, etc.) are preserved verbatim — they're templated UI without cell-level PII; if a future audit finds otherwise, add specific scrubs in the harvester repo (see "Adding more recordings" below) and re-run.
Adding more recordings
Local maintainer flow — the .jsonl never enters the repo. Only the manifest delta (≈200 B per entry) gets committed.
Step 1 — produce an anonymized .jsonl
The harvester that produced the current 179-trace set lives in a
separate repo, nexu-io/agent-pr-explore. See its README
for how to authenticate against your trace store, filter by skill /
agent / outcome, and anonymize the result. Output is one
<trace-id>.jsonl file per recording.
Step 2 — one-shot upload + manifest update
# prereq, once: wrangler login (OAuth, no token to manage)
bash mocks/scripts/upload-recording.sh /path/to/<trace-id>.jsonl
The script validates the file, prints the manifest entry it will add,
uploads the .jsonl to R2, rewrites mocks/manifest.json locally, then
uploads the updated manifest to R2 too (so consumers see the new entry
without waiting for the next git push).
Step 3 — commit the manifest delta
git add mocks/manifest.json
git commit -m "mocks: add recording <trace-id>"
git push # or open a PR — your call
The only thing in the commit is a ~200-byte JSON edit listing the new
entry's trace_id, sha256, bytes, agent, outcome, skills,
etc. The .jsonl itself stays in R2.
Trust model
- R2 write is wrangler-OAuth gated. Maintainers do
wrangler loginonce. The bucket is on the powerformer Cloudflare account (pinned in the script). No long-lived tokens in repo secrets, no Action to hijack — just account access. - Repo stays small forever. No .jsonl files ever land in git; the manifest grows by ~200 B per recording.
- Read stays public. Anyone can fetch via the r2.dev URL — see Recordings live on R2, not in this repo.
Removing a recording
# 1. delete from R2
export CLOUDFLARE_ACCOUNT_ID=64ad4569ffd912432d6b86d5656484c4
wrangler r2 object delete open-design-mocks/recordings/v1/<trace-id>.jsonl --remote
# 2. drop the entry from manifest.json (edit by hand, or use `jq`)
# 3. re-upload manifest
wrangler r2 object put open-design-mocks/recordings/v1/manifest.json \
--file mocks/manifest.json --remote
# 4. git add mocks/manifest.json && git commit && git push
There's no automation for delete because (a) it's rare and (b) you
want a human to think about whether removing a recording would
invalidate any test fixtures that pin it via OD_MOCKS_TRACE=<id>.
Usage from OD's test code
From a test (Vitest / Jest)
import { spawn } from 'node:child_process';
import { join } from 'node:path';
const MOCK_BIN = join(__dirname, '../../mocks/bin');
it('parses an opencode session with 4 tool calls into 4 UI events', async () => {
const child = spawn('opencode', ['run'], {
env: {
...process.env,
PATH: `${MOCK_BIN}:${process.env.PATH}`,
OD_MOCKS_TRACE: '06a9324a', // 4-tool claude session
OD_MOCKS_NO_DELAY: '1',
},
stdio: ['pipe', 'pipe', 'pipe'],
});
child.stdin.write('test prompt');
child.stdin.end();
// ... assert events parsed from child.stdout
});
From a manual playback
# See what claude's 17-tool "delete v2" session emits to OD:
export PATH=$(git rev-parse --show-toplevel)/mocks/bin:$PATH
export OD_MOCKS_TRACE=04097377
export OD_MOCKS_NO_DELAY=1
echo "anything" | claude -p --output-format=stream-json | jq .type | uniq -c
Files
mocks/
├── README.md ← you are here
├── mock-agent.mjs ← entry; routes --as <agent> to format renderer
├── lib/
│ ├── recording-picker.mjs ← env-driven trace selection
│ ├── format-opencode.mjs ← matches handleOpenCodeEvent
│ ├── format-codex.mjs ← matches handleCodexEvent
│ ├── format-claude.mjs ← matches createClaudeStreamHandler
│ ├── format-gemini.mjs ← matches handleGeminiEvent
│ ├── format-cursor-agent.mjs ← matches handleCursorEvent
│ ├── format-acp.mjs ← JSON-RPC server matching attachAcpSession
│ ├── format-vela.mjs ← AMR vela: ACP + models block + set_model gate
│ ├── vela-subcommands.mjs ← `vela login` + `vela models` handlers
│ └── format-plain.mjs ← raw stdout (deepseek/qwen/grok)
├── bin/
│ ├── opencode claude codex
│ ├── gemini cursor-agent
│ ├── deepseek qwen grok
│ ├── devin hermes kilo kimi kiro vibe
│ └── vela ← 15 bash wrappers, PATH-overlay
├── manifest.json ← committed: 179 entries' metadata + sha256 + provenance + R2 storage hints
├── golden/ ← committed: daemon-event regression snapshots
│ ├── README.md
│ └── *.events.json ← 3 representative traces (claude/codex/opencode)
├── scripts/
│ ├── smoke-test.sh ← 21 checks; auto-fetches recordings if empty
│ ├── fetch-recordings.sh ← pull from R2 (parallel, sha256-verified, idempotent)
│ ├── upload-recording.sh ← maintainer-local: validate + wrangler put + manifest update
│ ├── contract-check.sh ← real-CLI vs mock protocol drift check (manual)
│ └── lib/
│ └── manifest-utils.mjs ← shared sha256 / meta-parse / manifest-rebuild logic
└── recordings/ ← populated at runtime, gitignored .jsonl
└── .gitignore ← recordings come via fetch
No external dependencies. Pure node:fs/crypto/child_process. Works
under any Node ≥18.
Limitations
copilot,qoder,pi(the nichecopilot-stream-json/qoder-stream-json/pi-rpcformats) are recorded but not yet rendered as their native protocols — they fall back to the plain renderer for now. If you need them, add aformat-<agent>.mjsfollowing the same pattern asformat-codex.mjs; the parsers are inapps/daemon/src/{copilot-stream,qoder-stream}.tsand the pi-rpc handler insideapps/daemon/src/server.ts.- The mock does not honor CLI flags that change semantics (
--model,--permission-mode,--allowed-tools). They're silently ignored.
Provenance / safety
All recordings come from open-design's own Langfuse project (the
open-design project under the powerformer org). Users opted into
telemetry when they installed the desktop client. The anonymizer
removed user-identifying paths and project UUIDs before checking in.
If you find a recording that includes content that should be redacted, follow the Removing a recording flow above.