Create an Instance AI workflow eval

Each eval is one JSON case — authored locally as a file in
packages/@n8n/instance-ai/evaluations/data/workflows/ (the disk loader
auto-discovers *.json, no registration step), with a LangTracer suite as its
durable home. Cases validate against
`harness/schema.ts`
(.strict() — unknown keys fail at load). The eval
README is the
exhaustive field reference; this skill is the opinionated how.

Committing new case JSONs into the repo is no longer the recommended
approach. Author the file locally (uncommitted), calibrate it against a real
build, then push it to a lang-tracer suite with eval:langtracer-push
(see Push to a lang-tracer suite) —
--suite baseline for the consolidated corpus n8n CI runs, or a dedicated
capability suite like agents.
The suite is the home for the case; the eval CLI reads it back via
--source langtracer. You still write the JSON file — it's just the input to
the push, not a committed artifact.

Seeded cases. An inline seed pushes with the case — the case-write API
stores it verbatim, so the suite is its home like any other case. Only a
seed.mode: "replay" case is refused (listed under skipped:): it's
reconstructed from a LangSmith trace at run time, so it dies when that trace is
pruned and has no durable home. Don't commit a replay case either — derive a
synthetic case from it. See `case-shapes.md`.

Set the autonomy level first

Before you source, draft, or run anything, decide how hands-on the driver
wants to be — and say it back. This skill runs at one of two autonomy levels.
If the request makes the level clear ("just author and calibrate it yourself" vs.
"stop me at each step", or an explicit mode), adopt it, state it in one line, and
note how to override (e.g. "say 'stop me at calibration' to add a checkpoint").
If it's not clear, ask the driver one question offering the two levels
before doing any work.

The skill has four natural decision gates — selection (which real
failure to encode), shape + expectations (archetype, must-haves, scope
trim), calibration (classify each red and resolve keep/loosen/drop), and
push (kind + tier). The level decides what happens at each gate:

Level	Who decides when to stop	Behaviour
autonomous	agent	Runs all four gates start-to-finish; reports a decision log at the end for the driver to review — with the pushed case, its suite, and the source thread as links (Share links, never bare ids), and a Linear ticket proposal for any kept capability-gap red (Capability gap → propose a Linear ticket).
checkpoint	driver, per gate	Stops at each gate with a compact proposal + recommendation; driver says "go" or redirects. At the calibration gate, hands the driver a link to the just-built thread on the live instance plus login credentials so they can review the real conversation and workflow themselves before confirming (below).

Calibration is special-cased at both levels. A calibration verdict that
flips a case's meaning — a real capability-gap red vs. a harness-caused red, or
any loosening that would let a known-bad build pass — is surfaced explicitly
(interactively in checkpoint; in the decision log in autonomous), never silently
committed. It's the one call where a quiet mistake corrupts the suite, so it
never fully auto-commits.

Checkpoint calibration — review the real thread on the instance. Because the
calibration verdict is trust-critical, in checkpoint mode you don't ask the
driver to trust your reading of the run. You built the case against a live
instance with --keep-workflows (step 4), so the thread and the workflow are
still there — hand the driver a direct link and let them look:

• Thread: <base-url>/assistant/<threadId> — the exact conversation the case
  ran (the run prints the threadId; the built workflow prints as BUILT (<id>)
  and opens at <base-url>/workflow/<id>).
• Login: the email + password the eval signs in with (the owner you seeded on
  the instance — see `running-evals.md`; the default local
  seed is nathan@n8n.io / PlaywrightTest123).

Present, per red: the assertion, whether it went green/red, your proposed
classification (real capability gap / harness limitation / noise) and
keep/loosen/drop, and the review link. The driver logs in, reads the thread and
the workflow, and confirms or redirects before you write the verdict back into
the case description.

Share links, never bare ids

Every lang-tracer entity has a shareable web page, but the CLI and the MCP hand
you numeric ids — eval:langtracer-push prints + created <slug> (#621),
get_eval_run returns a run number, list_conversations returns thread ids.
An id is unclickable: the driver has to go find it. Whenever you name a case,
suite, thread, cluster, or run in anything a human reads — a checkpoint
proposal, the end-of-run decision log, a PR description, a Linear ticket, a Slack
message — render it as a link, keeping the id in the label:

pushed as [#621](https://lang-tracer.n8n-maintenance.workers.dev/test-cases/621)

Build links off the web base (LANGTRACER_URL, in production
https://lang-tracer.n8n-maintenance.workers.dev). Never off the API bases —
${LANGTRACER_URL}/api/v1 and /api/mcp are machine endpoints, and a link into
either 404s for the driver or dumps JSON.

Entity	URL	Where the id comes from
Test case	<base>/test-cases/<id>	push output (#<id>); create_test_case / search_test_cases
Suite	<base>/suites/<suiteId>	push header Suite "<slug>" (#<id>); list_suites
Source conversation	<base>/conversations/<threadId>	list_conversations / get_conversation
Cluster report	<base>/clusters/<id>	list_cluster_runs / get_latest_cluster_run
Eval run (sweep)	<base>/results?sweep=<runId>	list_eval_runs / get_eval_run (runId is sweeps.id, the "run #N")

Two links that are not lang-tracer and don't take this base: the built thread
(<base-url>/assistant/<threadId>) and workflow (<base-url>/workflow/<id>) live
on the n8n instance the eval ran against. When both are relevant — reviewing a
calibration red, writing a capability-gap ticket — give both, labelled, so nobody
has to guess which host a link points at.

Where the best cases come from

The strongest cases encode a real failure, not an invented premise. Two
connections help you find and verify one: LangTracer clusters real
conversations into capability-gap themes (discover what actually fails, at
scale), and LangSmith holds the raw traces (verify exactly what happened in a
run). LangTracer is the discovery layer; the durable artifact is almost always a
synthetic case you author from what you learn (use seed.mode: "replay" only per
`case-shapes.md`). See
`sourcing-cases.md` for connecting the MCPs and the
discover → verify → encode workflow.

Pick the case shape first

The corpus is four archetypes. Decide which you're writing before you draft — it
determines the fields, the grading, and how you validate. They compose (a seeded
case can still assert outcome), but the primary shape drives the work.

Archetype	Question it answers	Primary fields
Build (default)	Does the workflow the agent builds actually work?	outcomeExpectations + executionScenarios
Behaviour / process	Does the agent converse correctly (ask the right clarifying question, not re-ask, honour a correction, respect plan approval)?	processExpectations + multi-turn director script; often build-only
Credential	Does the build behave correctly given a specific credential view?	credentials[]
Seeded	Start mid-thread, with prior work already in place, and drive the turn under test	seed (authored mode: "inline"; "replay" for a local check)

Build is documented in full below. The other three, the director-script
vocabulary, and the seeding modes are in `case-shapes.md`.

Core principle (all shapes)

Write expectations from intent, then calibrate against a real build. Decide
up front what makes any correct solution correct — the must-haves implied by
what the user actually said — then build the workflow once for real to calibrate
granularity: loosen what's over-specified, confirm the must-haves are
achievable, and catch requirements the agent legitimately satisfies a different
way. Don't transcribe one observed build into assertions — that overfits the
eval into "did the agent reproduce that run" instead of "did it solve the
problem."

Keep the conversation in the user's voice. State the goal and real
constraints the way a user would — don't name node types, wire up the structure,
or restate your outcomeExpectations in the prompt. If the conversation spells
out the build, the case only tests whether the agent can follow instructions and
the expectations become tautological; the gap between what the user asks for and
how a correct workflow realizes it is the capability under test. Even when the
anchor is honoring a user's stated technical preference, phrase it as their
need + constraint ("I need field X and the built-in node doesn't expose it, so
pull it straight from the API") — not as an implementation spec ("use an HTTP
Request node").

Write the conversation in English unless the user asked otherwise (or the
case exists specifically to test non-English handling). Sourced real threads are
frequently non-English — translate the intent into English when you rewrite the
prompt in the user's voice; the failure mode is the anchor, not the original
language.

Trim to the smallest multi-turn conversation that reproduces the issue.
Real sourced threads are long (dozens of turns of setup, debugging, and
tangents) — do not transcribe them. Distill to the fewest turns that still
drive the build or behaviour under test. Every retained turn must earn its place:
a turn stays only if it is load-bearing — a value the agent must ask for
(withheld until asked, via a director note), a correction/push-back the case
exists to test, or a plan approval that gates the build. If removing a turn
doesn't change what's tested, remove it. Collapse to a single turn whenever
the whole request can be stated at once without a load-bearing exchange; keep it
multi-turn only for those exchanges, and keep each director script in one turn
(don't fabricate assistant "done" turns to sequence steps — see
`case-shapes.md`). A minimal conversation isolates the
capability; a transcribed one buries it in noise and tests instruction-following.

Size the build, not just the assertions. Real sourced prompts are often
kitchen-sink ("production-ready, runs forever, 3 feed posts and 8 stories a
day", "generate 50 articles daily") and reliably blow the ~900s build budget (see
"Known harness limitations"). A faithful trim is a legitimate authoring move:
reduce batch sizes, drop one of several parallel pipelines, or merge adjacent AI
steps so the case builds within budget — then note the reduction in the case
description ("the original request also asked for an 8-stories/day pipeline;
scoped to feed posts so it builds in budget"). Keep the capability under test; cut
the combinatorial bulk. A case that never builds tests nothing.

Workflow

These steps map to the four gates from Set the autonomy level first:
sourcing (before step 1) is the selection gate; steps 1–2 are the shape +
expectations gate; steps 5–6 are the calibration gate; steps 7–8 are the
push gate. In autonomous mode you flow through all of them and summarize in
a decision log; in checkpoint mode you pause at each with a proposal, and at
calibration you hand the driver the thread link + login to review the real build
(see Set the autonomy level first). Calibration
(step 6) always surfaces meaning-flipping verdicts explicitly regardless of level.

1. State the must-haves first. From the conversation alone, list what every
   correct workflow must do (trigger type, essential operations, gating
   condition) — those become draft outcomeExpectations. Required fields:
   conversation (≥1 turn, first user), complexity, tags, and at least
   one of executionScenarios / processExpectations / outcomeExpectations.
2. Draft the case from the template below; validate it loads (see
   "Validate").
3. Smoke-test the environment with one case before any batch. Run a single
   case end-to-end first. This validates auth / model / --base-url / the built
   dist for ~1/Nth the cost — distinct from validating a case. If that one case
   crashes at execution (especially with an identical error you'd expect to hit
   every case), fix the environment before running the batch (see "A red is
   signal" → environment check). Running 15 cases only to discover a stale-dist
   crash on all of them wastes a full run.
4. Build it once against a running instance (see
   `running-evals.md`) with --keep-workflows so the built
   workflow stays for inspection.
5. Inspect — read the built workflow (the run prints BUILT (<id>); fetch
   via GET /rest/workflows/<id>) and the HTML report's transcript to see what
   the agent actually did.
6. Calibrate — sharpen assertions; never dull them to force a green. Fix
   assertions that are genuinely mis-sized: relax one that pins a choice the
   conversation left open (so a valid alternative build wrongly fails), tighten
   one a wrong build would slip past, and phrase executionScenarios to match how
   the workflow runs on mocked data. But when a scenario goes red because the
   build has a real gap, or because the harness can't exercise it, that red is
   the result — keep it and surface why (see "A red is signal", below). Never
   delete a scenario, weaken an assertion, or drop to build-only just to make the
   run green.
7. Push to the suite — do NOT commit the JSON. Once calibrated, push the case
   into its curated lang-tracer suite with eval:langtracer-push (see
   Push to a lang-tracer suite); the suite is the
   case's home, not the repo. Leave the data/workflows/*.json file uncommitted
   (or delete it once it's in the suite). Committing new case JSONs into the repo
   is no longer the approach. (An inline seed pushes with the case; only a
   replay case is refused — it's a local throwaway; see
   `case-shapes.md`.) For a sourced case,
   finish by linking it to its source thread/finding over the MCP — see
   Link the pushed case to its source.
8. Hand back links, and a ticket proposal if the case found a gap. Report the
   pushed case as <base>/test-cases/<id>, not #<id> (Share links, never bare ids), and if calibration kept a real
   capability-gap red, propose a Linear ticket for it (Capability gap → propose a Linear ticket) rather than leaving the
   gap as a red case nobody owns.

--iterations N is available to measure flakiness (pass@k / pass^k) — reach for
it when you suspect a case is non-deterministic or before promoting it to a
gated tier, not as a routine step (each iteration is a full build + execution).

Gut-check: if you can't picture a plausible wrong build that this case
reliably turns red, the assertions are too loose to guard anything.

Confirm the precondition fired, not just the green. For any conditional
assertion — "when X happened, the agent did Y" (most processExpectations, and
any behaviour case) — a pass has two readings: the agent did Y, or X never
happened and the assertion passed vacuously. A behaviour case that hinges on
the mock producing a specific failure (e.g. an AI node simulated to empty so a
downstream parse node fails) is the classic trap: if the mock instead returns
parseable data, the failure never occurs and the case guards nothing while
showing green. Calibration must read the execution trace and the agent's
finalText (buildTrace.finalText in the verifier snapshot, or the HTML report)
and verify X actually materialised — the direct-loop eval-results.json does not
persist per-expectation judge reasoning, so pass/fail alone can't tell you which
reading you got.

A sourced failure that no longer reproduces is still worth keeping — it's now a
regression guard. When you encode a real failure and calibration shows the
current build handling it correctly (behaviour drifts across versions), the case
doesn't lose value: it flips from capability-gap (currently red) to regression
guard (currently green, catches a re-introduction). Keep it — but only after the
non-vacuous check above proves it would turn red on the bad behaviour, else the
"guard" guards nothing.

A red is signal — surface it, don't work around it

Calibration exists to right-size assertions, not to make a case pass. When a
run turns a scenario or expectation red, classify the red first — then keep it.

First rule out the environment. Before reading any red as a signal about a
case, check the shape of the failures across the run. If every case fails the
same way — every scenario with the same execution error while builds succeed,
or every build erroring identically before it starts (an Agent error: Something went wrong… with zero tool calls) — that is almost never the cases;
it's a broken environment, most often a stale dist after a branch or worktree
switch. Two shapes to know:

• Stale packages/core / packages/cli dist — a refactor moved a runtime
  export and the built dist still calls the old one, so builds succeed but every
  execution fails the same way (e.g. (0 , n8n_workflow_1.createDeferredPromise) is not a function after createDeferredPromise moved to @n8n/utils).

• Stale/half-built @n8n/instance-ai dist — every run errors before building
  (Agent error…, zero tool calls) and the instance log shows Cannot find module '@/utils/...' from dist/skills/*.js: the build's tsc-alias step (which
  rewrites @/ path aliases to relative requires) didn't complete, so the dist is
  internally inconsistent.

• Out-of-sync node_modules — pnpm build itself dies early with Cannot find module '@n8n/<pkg>' even though that package is a declared workspace:*
  dependency and has a dist/. The workspace symlink is missing from the
  consumer's node_modules (typical after a branch or worktree switch). Confirm
  with ls -d packages/<consumer>/node_modules/@n8n/<pkg>; fix with a plain
  pnpm install — no need for the heavier pnpm reset --full.

Fix it, don't calibrate around it: run a full ordered pnpm build (a targeted
--filter build can fail on unrelated stale-dep type errors; for the instance-ai
shape, cd packages/@n8n/instance-ai && pnpm build runs tsc && tsc-alias), then
restart the instance — the running node process holds the old dist in memory,
so rebuilding on disk changes nothing until restart (and kill by env-var pattern
misses it — kill the actual lsof -t -iTCP:<port> PID). Re-probe one case, confirm
it builds and executes, then re-run the batch. Only once uniform environment
failures are excluded do the three categories below apply:

• Real build / capability gap — the agent's workflow is wrong or missing
  something the user asked for (a miswired branch, a missing retry, wrong field
  keys). This is exactly what the eval is for. Keep it red. Don't loosen the
  assertion or drop the scenario; a currently-red gap is the capability signal
  today, and a re-introduction guard once the builder improves. Then propose a
  Linear ticket for the gap — see Capability gap → propose a Linear ticket.
• Harness limitation — the build is correct but the mock/execution layer
  can't exercise the path (see "Known harness limitations", below). Keep the
  scenario and say so in its description — that this red is harness-caused,
  not a build defect — so nobody misreads it as a product bug. Keep it out of
  gated tiers if it hard-fails every run; when the harness gains the capability it
  starts earning its keep with no re-authoring.
• Genuine non-determinism — the same build flips green/red across runs.
  This is the only real "noise". Confirm it with --iterations N before calling
  it flaky, then de-tier and note it; deletion is the last resort.

Annotate every kept red in the case description with a scannable prefix so a
future reader tells the two apart at a glance. Use Harness note: … for a
harness-caused red (name the limitation and why the build is still correct), and
Capability-gap finding: current build reds because <X> — a real builder bug (flips to a regression guard once fixed) for a real gap. Consistent prefixes keep
the corpus greppable and stop harness reds from being misread as product bugs.

The one move to never make is working around a red by weakening what the case
checks — deleting a failing scenario, loosening an assertion until a wrong
build would pass, or quietly converting to build-only. That makes the suite look
greener than the product is, which is the opposite of the eval's job: bugs and
harness gaps are the deliverable, so highlight them, don't engineer around
them. If you catch yourself editing a case so that a known-bad build would now
pass, stop.

Who confirms the classification depends on the autonomy level. In
checkpoint mode the keep/loosen/drop decision is the driver's to confirm: you
hand them the thread link + login (see Set the autonomy level first) so they can review the real conversation
and workflow, then you write the agreed Harness note: / Capability-gap finding: prefix back into the case description. In autonomous mode the agent
proposes it explicitly in the end-of-run decision log. Either way the
classification is stated in the open, never silently committed — misreading a
harness red as a real gap (or the reverse) is the one calibration mistake that
quietly corrupts the suite.

Capability gap → propose a Linear ticket

A kept capability-gap red is a product bug you just characterised better than
any bug report would. But a red case in a suite doesn't assign itself to anyone:
without a ticket the gap sits in CI as permanent noise, and the next person to
read the run assumes someone already owns it. So once a red is classified as a
real gap (and the driver has confirmed it, per the autonomy level), propose a
Linear ticket for it.

Propose, don't create. Per AGENTS.md, never open a
Linear ticket unasked. Put the draft in front of the driver — interactively in
checkpoint mode, in the decision log in autonomous mode — with a title, a team,
and the body, and let them say go. Skip the proposal in two cases:

• The gap already has a ticket. Check the case's linked issues on its page,
  and run get_linear_ticket_context <TEAM-N> on any candidate identifier the
  driver or the source thread mentions, before you draft a duplicate.
• The red isn't a capability gap. A Harness note: red is a
  lang-tracer/harness issue, and genuine non-determinism is a case-hygiene chore.
  Neither belongs in the builder's queue.

The draft body should carry what makes the gap actionable, all of it already in
hand from calibration:

• The eval case, as a link — <base>/test-cases/<id> (see Share links, never bare ids). This is the reproducer; it's the most
  valuable line in the ticket.
• What failed, verbatim — the failing outcomeExpectation /
  processExpectation or scenario name, plus the judge's stated reason. Not a
  paraphrase: the exact text is what the fixer will grep for.
• What the build did instead — the specific defect (miswired branch, wrong
  field key, missing gate), and links to the real evidence: the source conversation
  (<base>/conversations/<threadId>) and the built thread + workflow on the eval
  instance (<base-url>/assistant/<threadId>, <base-url>/workflow/<id>).
• Blast radius, if you know it — the cluster theme or the number of real
  conversations behind the gap (<base>/clusters/<id>) is what turns "one red
  case" into a prioritisable bug.

File it from the case page so the link is made. The <base>/test-cases/<id>
page has a Create Linear issue dialog that creates the ticket and links it to
the case; that link is what makes get_linear_ticket_context <TEAM-N> later return
the case, its scenarios, the source conversation, and its analysis in one call. The
case↔ticket link is only writable from that UI — lang-tracer's MCP and /api/v1
are read-only for it — so if the ticket gets created some other way (a Linear MCP,
if your harness has one, or Linear directly), say plainly that it isn't linked, ask
the driver to link it on the case page, and meanwhile put the identifier + URL in
the case description via update_test_case so the provenance isn't lost.

Then extend the description prefix with the ticket, so the corpus stays greppable
in both directions: Capability-gap finding: current build reds because <X> — a real builder bug (flips to a regression guard once fixed). Tracked in [<TEAM-123>](<ticket url>). And when the build itself is wrong — not just a
scenario red under a correct build — push it with --set-kind capability_gap into
a suite of that kind (see
Push to a lang-tracer suite).

Example

Minimal build case:

{
  "description": "What this case tests.",
  "conversation": [{ "role": "user", "text": "<the build prompt>" }],
  "complexity": "medium",
  "tags": ["build", "<nodes>", "<concepts>"],
  "triggerType": "schedule",
  "outcomeExpectations": ["<a must-have any correct workflow satisfies>"],
  "executionScenarios": [
    {
      "name": "happy-path",
      "description": "<what this run exercises>",
      "dataSetup": "<what the external services return>",
      "successCriteria": "<observable proof the run succeeded>"
    }
  ]
}

A fuller case with a multi-turn director script (withhold a value until asked,
push back on a wrong plan):

{
  "description": "Scheduled GitHub-bugs digest to Slack. Repo and channel are withheld until the agent asks; the plan must filter to the 'bug' label before it's approved.",
  "conversation": [
    { "role": "user", "text": "Every weekday at 9am, fetch this week's open bugs from our GitHub repo and post a short summary to Slack." },
    { "role": "assistant", "text": "Which repo and which Slack channel should I use?" },
    { "role": "user", "text": [
        "[Withhold the repo and channel until the agent asks; then say the repo is 'acme/widgets' and the channel is '#eng-bugs'.",
        "When the agent shows a plan or setup card, reject it unless it filters issues to the 'bug' label — a digest of ALL issues is wrong. Once it filters to bugs, approve.]"
    ] }
  ],
  "messageBudget": 8,
  "complexity": "medium",
  "tags": ["behaviour", "build", "schedule", "http-request", "slack"],
  "triggerType": "schedule",
  "processExpectations": [
    "The agent asked for the repo and Slack channel before building, since the prompt named neither.",
    "The agent's final plan filtered issues to the 'bug' label — if its first attempt didn't, it corrected after the user pushed back rather than summarizing all issues."
  ],
  "outcomeExpectations": [
    "A Schedule Trigger runs the workflow on a recurring weekday-morning cadence.",
    "Open issues are fetched from GitHub (HTTP Request or GitHub node) and filtered to the 'bug' label before the summary is built.",
    "One Slack message summarizing the fetched bugs is posted to the #eng-bugs channel the user gave."
  ],
  "executionScenarios": [
    {
      "name": "posts-bug-digest",
      "description": "Three open bugs are returned; a summary is posted to Slack",
      "dataSetup": "The GitHub issues request returns three open issues labelled 'bug' ('Login 500', 'Timezone off by one', 'CSV export truncates'). The Slack postMessage call returns { \"ok\": true, \"ts\": \"1700000000.0003\" }.",
      "successCriteria": "The run completes without errors and posts one Slack message to #eng-bugs that references the three fetched bug titles."
    }
  ]
}

What each piece is doing:

• conversation[0] is sent to the builder raw. The opening turn is the real
  prompt — never put a [director note] in it (it would leak verbatim).
• The [bracketed] turn is a director script for the user-proxy — behaviour,
  never spoken. Here it withholds values until asked and rejects a plan that
  misses the label filter. Keep the whole script in one turn and encode ordering
  inside it (don't fabricate assistant "done" turns to sequence steps — see
  `case-shapes.md`). applies-each-change-when-asked (in the
  baseline LangTracer suite) is a good real example.
• dataSetup describes only what external services return. That's the layer
  the harness controls (below).

dataSetup and the mock layer

The harness mocks by intercepting outbound HTTP requests to external services
and having an LLM answer them from the node's config and API docs. It does not
let you set a node's output directly, and it does not mock n8n internals
(Code/Set/Merge/IF/Switch run for real on the mocked data; triggers and DB
nodes get LLM-generated pin data). So:

• Write dataSetup as what each external service returns ("the GitHub
  request returns three issues labelled bug"), not as node outputs or internal
  state.
• The strongest scenarios exercise external-service responses — that's what
  the harness reproduces most faithfully.
• Data Table reads are pinned to the scenario. A read op (get /
  rowExists / rowNotExists) is treated as the scenario's "stored state" and
  pinned with data derived from your dataSetup, so change-detection / dedup /
  "last seen" scenarios can be exercised — describe the stored rows in
  dataSetup. Two caveats: the pinned rows are LLM-generated (steered, not
  byte-exact — don't assert exact values off them), and writes/inserts aren't
  pinned (they hit the real per-thread table, recreated schema-only with no
  rows), so read-after-write within one run isn't faithful — the read reflects
  dataSetup, not what the run just wrote. A third caveat: only Data Table
  reads are seedable this way — dedup / change-detection built on workflow
  static data (removeItemsSeenInPreviousExecutions, $getWorkflowStaticData)
  is not seedable, because static data starts empty every run, so such a
  scenario reds vacuously (it sees everything as "new"). To get a seedable
  change-detection scenario, steer the build toward a Data Table; otherwise
  accept the static-data red as a harness limit and carry the logic in
  outcomeExpectations. Note the agent may choose static-data dedup on its own.
• Don't assert exact counts that depend on mock generation ("exactly 7 posts").
  Say "fewer than the original 10".

Known harness limitations that turn scenarios red regardless of the build

These produce a reliable red on a correct build. Don't engineer around them
— write the scenario for the behaviour you want and note in description that the
red is harness-caused (per "A red is signal", above):

• Resource-locator fields left empty for setup (Google Sheets / Drive /
  Calendar and similar node pickers). The agent legitimately leaves the
  document/folder/calendar ID blank for the user to pick at setup; the mock
  substitutes __evalMockResource, and the node then crashes looking it up
  ("Sheet with ID __evalMockResource not found", or "Cannot read properties of
  undefined"). Any scenario whose success path runs through such a node
  hard-fails before anything downstream executes.
• Trigger and Data-Table-read pin data is LLM-generated, so not byte-exact.
  Both are steered by your dataSetup (see the mock-layer section above — you
  can influence what a trigger emits or what a stored-row read returns), but
  because the values are generated, a scenario that asserts exact values or counts
  off them is flaky. Assert shape/branch/relative facts, not exact figures. (The
  residual hard red here: polling / form triggers still occasionally fail to load
  entirely — "workflow not found".)
• Mock response shape — the LLM-generated mock response can omit the real
  envelope, crashing a downstream parse/format node. Recurring, reproducible
  shapes to expect (all produce a red on a correct build):
  • OpenAI structured output — historically the mock returned a plain
    {content: "..."} instead of the Responses envelope
    (output[0].content[0].text), so a Structured Output Parser /
    Information Extractor / Text Classifier got nothing and errored with
    Model output doesn't fit required format. The Responses-envelope
    normalizer (PR #33578, merged) fixes the flat-envelope case, so many of these
    now execute cleanly. A narrower residual red remains for structured-output
    schemas declared with strict additionalProperties: false: the normalized
    output wrapper (and any extra fields the mock invents, e.g. subject/date)
    violate the strict schema, so the node still rejects the mock output. Both the
    old and residual forms are the mock, not the build — carry correctness in
    outcomeExpectations and note the red as harness-caused.
  • Gmail mock returns headers as top-level capitalized fields (From,
    Subject) instead of under payload.headers, so a Code/Filter node reading
    the sender/subject gets empty strings (e.g. a "drop no-reply senders" safety
    gate lets everything through). Assert the wiring/ordering of such a gate in
    outcomeExpectations, not its runtime effect in a scenario.
  • A less-common API (e.g. Gemini's top-level candidates) can omit its envelope
    the same way.
  • Google Drive resumable upload — the initiate-upload mock omits the
    Location header carrying the session URL, so a Drive file-upload node fails
    with a 400. Any build that uploads a generated image/file to Drive can red on
    this.
• Agent-tool nodes can't be executed standalone. An AI-Agent tool node
  (toolHttpRequest and other supplyData-only LangChain nodes with no execute
  method) only runs when the agent invokes it; the harness executing it directly
  fails with has a "supplyData" method but no "execute" method. A near-universal
  red for chat-trigger / AI-agent build cases whose tool is an HTTP-request tool —
  the build is correct, so carry correctness in outcomeExpectations (agent wired
  to trigger + model + tool) and note the execution red as harness-caused.
• Poll/wait loops can't be fast-forwarded. A workflow that submits an async
  job then polls for completion (generate → poll status until ready → download)
  can't advance the mocked status deterministically, and a Wait node runs in
  real time, so the scenario reds with an execution timeout (framework_issue).
  The build is correct — carry it in outcomeExpectations and note the red as
  harness-caused.
• A build can time out and produce no scored result at all — the run reports
  BUILD FAILED: Run timed out and zero graded expectations. Don't assume "spec
  too big": the more common cause is a single-prompt case where the agent asks
  a clarifying ask-user question and the build hangs on the unanswered
  question until the per-iteration timeout (see `case-shapes.md`
  — only confirmations auto-approve). Before treating a timeout as spec size,
  classify it: read the agent's final response in the report / trace (did it
  ask a question? flag an infeasibility? or genuinely churn through a huge
  build?), and re-run the case solo (--concurrency 1) — concurrency both
  masks a stalled build (it hits the cap) and can time out a perfectly healthy
  build purely by queueing it behind the per-instance build cap (default 4), so a
  solo run either surfaces the real reason in seconds or simply passes outright.
  Fix per cause: a solo run that now passes → it was concurrency contention,
  not the case (split big batches across lanes — see
  `running-evals.md`); a clarifying-question stall → author
  multi-turn with a director note that pre-answers it; a genuine infeasibility → it's an
  infeasibility/honesty behaviour case (processExpectations), not a build case;
  a true oversized spec → the timeout is itself a finding, but note it so the
  zero isn't mistaken for a scored failure.

outcomeExpectations vs processExpectations

Both are natural-language assertions graded by the same Sonnet judge, and each
counts as a pass-rate unit. They judge different surfaces:

• outcomeExpectations — the resulting workflow, judged from the
  workflow JSON. Assert node choices and configuration, connection topology and
  branch wiring, data/expression references, trigger cadence, gating conditions.
  They run everywhere, including prebuilt/MCP runs (no transcript needed).
• processExpectations — how the agent behaved during the build, judged
  from the transcript. Assert clarifying questions asked (or not re-asked),
  tool-call behaviour, plan/approval handling, batching, honouring a correction,
  ordering. They need a transcript, so they're skipped in prebuilt/MCP runs.

Rule of thumb: an assertion about the artifact is an outcome expectation; an
assertion about the conversation or the agent's choices along the way is a
process expectation. A case with no executionScenarios is a valid
build-only case, graded by these expectations plus the workflow checks.

Sizing each assertion

Right-size against what the agent was actually told. An assertion is
well-sized when every correct build passes it and a wrong or lazy build fails
it — and it holds the agent only to what the conversation specified, not to one
run's arbitrary choices. Two failure modes:

• Too tight — pins a choice the conversation left open. If the prompt never
  named a vendor, "calls flightaware.com" fails a valid build that used a
  different source. But if the conversation specified it, pin it — when the
  user said "email me via Gmail," "sends via a Gmail node" is correct and
  required, not too tight.
• Too loose — a non-solution would also pass. "Fetches data from somewhere"
  passes a workflow that fetches but never compares — it doesn't prove the
  change-detection the prompt asked for.

Quick check — the substitution test: would a reasonable alternative
implementation of what the user asked for still pass? Examples (flight-status
case, where the source and channel were left unspecified):

Verdict	Assertion	Why
❌ too tight	"Has an HTTP Request node calling flightaware.com"	Vendor was unspecified; a valid AeroDataBox build fails. (If the user had said "scrape FlightAware", this would be correct.)
❌ too tight	"Publishes via HTTP Request nodes"	Pins the transport when a first-party node is the idiomatic path — e.g. the Facebook Graph API node is the correct way to reach the Instagram Graph API, so a valid build using it fails. Assert the capability ("publishes to Instagram via the Graph API, through the Facebook Graph API node or HTTP Request"), not the mechanism.
❌ too loose	"Fetches flight data from somewhere"	A workflow that fetches but never compares passes — doesn't prove change-detection.
✅ right	"Persists the previously-seen status and compares it to the freshly-fetched one"	The defining behaviour; substitution-proof across vendors and storage choices.
✅ right	"Alert is sent only on the change-detected branch, gated by a conditional"	Proves the gate without pinning node or channel.

Put intent the conversation only implied (a preferred but unstated channel) in
processExpectations, not outcomeExpectations.

Robust design vs harness flakiness

Two different things — keep them apart:

• Robust assertion design (always do this). The agent's unspecified choices
  vary run to run. Source-agnostic outcomeExpectations for an unspecified
  source aren't a concession to flakiness — they're the correct assertion.
• Harness limitations (surface them, don't hide them). Some paths hard-fail
  on a correct build regardless of dataSetup — empty resource-locator fields
  that crash Sheets/Drive/Calendar nodes, polling triggers failing to load (see
  "Known harness limitations" above). (State-bearing Data Table reads are no
  longer in this bucket — they're pinned from dataSetup; only the write path and
  exact-value assertions stay unreliable.) The fix is to document, not to work
  around: note the limitation in description and keep a hard-failing scenario
  out of gated tiers.
  Only when a scenario flips non-deterministically run to run is it genuine
  noise worth removing — a scenario that reliably fails for a documented harness
  reason is a standing record of what the harness can't yet test, and stays.

Negative execution scenarios

Don't stop at the happy path — but only assert graceful handling the prompt
actually implied. Most agent-built workflows don't add error handling by
default, so "the workflow crashes on bad input" is a legitimate builder finding,
not a test-case bug. Where graceful handling is expected, phrase
successCriteria as the absence of the wrong action ("no alert is sent", "run
completes without error") as much as the presence of the right one: empty /
not-found source, source error / timeout, malformed response.

Relationship to the always-on workflow checks

Every successful build is also graded by ~28 always-on binary checks across 7
dimensions (structure, topology, parameter correctness, intent, AI wiring,
craftsmanship, security) —
`binaryChecks/checks/`.
Those are broad and low-visibility. Writing a targeted expectation for your
specific case is still worth it even when a binary check nominally covers it —
a named case-level assertion gives far better visibility into this behaviour
than one row buried in a 28-check rubric. Don't skip an assertion just because a
generic check exists.

When a scenario fails, the verifier tags a failure category (builder_issue,
mock_issue, framework_issue, verification_failure, build_failure). Treat
it as a hint, not ground truth — we've seen a genuine node misconfiguration
tagged mock_issue, and a real mock problem tagged as a build error. Open the
HTML report and check the actual execution and the generated workflow before
concluding whether the failure is your case, the build, or the harness.

Outputs of a run

• workflow-eval-report.html (in the run's .data/ dir) — the highest-value
  view: full conversation transcript with tool calls, per-node execution traces,
  the exact intercepted requests and the mock responses, Phase-1 hints, verifier
  reasoning, and the workflow-check rubric. Human-oriented; start here when
  debugging.
• eval-results.json — structured results (the machine-readable artifact;
  the direct loop produces this even with no LangSmith). Good for an LLM or
  script to parse. For per-case attribution under concurrency, parse this, not
  the streamed verbose log — with more than one lane the log lines interleave
  across cases, so a [scenario] FAIL line in the stream can't be reliably tied to
  its case. Authoritative fields: testCases[].buildSuccessCount,
  buildExpectationResultsPerRun[][].{pass,reason}, and
  scenarios[].runs[].{passed,failureCategory,rootCause,execErrors}.
• eval-pr-comment.md — the rendered PR comment (aggregate + regression
  comparison), always written.

Validate (before running)

cd packages/@n8n/instance-ai
npx tsx -e "import {loadWorkflowTestCasesWithFiles} from './evaluations/data/workflows/index.ts'; console.log(loadWorkflowTestCasesWithFiles('<slug>')[0].fileSlug)"

Push to a lang-tracer suite

Once a case is calibrated, push it (and any others) up into a curated lang-tracer
suite instead of committing the JSON. eval:langtracer-push upserts over the
REST API: it creates cases missing from the suite, updates ones whose content
drifted, leaves the rest unchanged, and never prunes. It's the inverse of
--source langtracer (which pulls a suite down).

cd packages/@n8n/instance-ai
# preview first — no writes (use `npx dotenvx`; the bare `dotenvx` binary is usually not on PATH):
npx dotenvx run -f .env.eval -- pnpm eval:langtracer-push --suite baseline --dry-run --changed
# then push (drop --dry-run):
npx dotenvx run -f .env.eval -- pnpm eval:langtracer-push --suite baseline --changed

• Selectors (at least one required — no accidental push-all): positional
  <slugs...> (exact file slugs), --changed (new/untracked + staged + modified
  data/workflows/*.json, ideal right after authoring an uncommitted case),
  --filter/--tier (with --exclude as a modifier).
• Multiple positional slugs? Skip pnpm — call the script directly. pnpm eval:langtracer-push … slugA slugB forwards the slugs as one joined argument
  ("slugA slugB"), so no case file matches and nothing is pushed. Either use a
  no-positional selector through pnpm (--changed), or run the script directly so
  each slug is its own argv: npx dotenvx run -f .env.eval -- npx tsx evaluations/cli/langtracer-push.ts --suite <slug> <slug1> <slug2> ….
• Env: LANGTRACER_URL + LANGTRACER_API_KEY (an lt_ bearer; one key works
  for MCP + REST) — put them in .env.eval and run under npx dotenvx.
• Options: --set-kind regression|capability_gap (default regression, must
  match the suite's kind), --contains-user-data (default is synthetic). A case
  whose build is correct (outcome expectations green) but that carries a
  currently-red execution scenario from a builder bug is still a regression
  case — it guards the fix; reserve capability_gap for cases where the build
  itself is wrong.
• Scenarios sync on update too: PATCH /cases/:id reconciles
  executionScenarios by name (update in place, insert new, delete missing —
  lang-tracer #48), so scenario edits re-push like any other field. A lang-tracer
  deployment predating that change silently ignores the key; if a pushed scenario
  edit doesn't land, update the scenario in the lang-tracer UI.
• Report what was pushed as links, not #ids. The CLI prints + created <slug> (#621) / ~ updated <slug> (#621, rev 3) and a suite header — that's the
  id, and nothing more. Turn each one into <base>/test-cases/<id> (and the suite
  into <base>/suites/<suiteId>) in whatever you hand the driver, so they can open
  the case they just authored instead of hunting for it. See Share links, never bare ids.
• An inline seed pushes with the case: the case-write API stores it
  verbatim, so a seeded case lives in a suite like any other. Only a replay
  case is refused — the push lists it under skipped:, because it's
  reconstructed from a LangSmith trace at run time and dies when that trace is
  pruned or deleted. Don't commit a replay case either; derive a durable
  synthetic case as the artifact instead.

Link the pushed case to its source (provenance step — always do this)

A sourced case that isn't linked back to the conversation/finding it encodes is
an orphan: six months later nobody can tell what real failure it guards. The
push CLI doesn't carry provenance, so after pushing, link the case over the
lang-tracer MCP with one update_test_case call on the new case id (the
push prints it):

1. sourceThreadId — the source conversation's thread id (plus
   sourceRunId when the case anchors to one specific run/step within it).
   This is the DB-level link every by-version rollup, conversation float, and
   ?sourceThreadId= query joins on — the tags/description below are the
   human-readable layer on top, not a substitute. The thread must already be
   imported into lang-tracer (running get_conversation_analysis on it, as the
   sourcing flow does, is enough); source_kind is derived server-side, and
   the link is only editable on authored cases — promotion-recorded provenance
   is immutable.
2. expectedBehavior — the rule the case enforces, one paragraph — and
   failurePattern — what actually happened in the source thread, with
   turn references. Copy/adapt these from the analysis's extractedCases
   entry when the case came from get_conversation_analysis.

Then add_case_tags (additive; targets the LT-side tags array, not
evalTags, so nothing round-trips into eval runs): add a capability tag (e.g.
instruction-persistence). Tag normalization is aggressive (lowercase, kebab);
colon-form tags get silently dropped. And keep the thread id + turn refs in
the case description too (the drafter's habit of "Sourced from thread "
is the convention) — the description is the only field shown everywhere.

Running

You need a running n8n instance with Instance AI enabled and a working sandbox;
point the eval at it. The harness runs in three modes — direct loop (no
LangSmith; eval-results.json only), LangSmith (also records an experiment

• regression comparison), and prebuilt (--prebuilt-workflows, score
  existing workflows). Narrow a run with --filter <slug> / --tier <name> /
  --exclude. See `running-evals.md` for the run recipes,
  parallel lanes, tiers, and baselines, and the
  README for the full
  flag list. Run with --keep-workflows when you want to review a build by hand —
  in checkpoint mode calibration this is how the driver opens the built thread
  (<base-url>/assistant/<threadId>) and workflow on the instance.

Other eval harnesses (not this skill)

This skill is for data/workflows/ cases. Three siblings exist with their own
data dirs and CLIs: eval:subagent (workflow-build compatibility corpus,
binary-check scored), eval:discovery (asserts first-hop tool/dispatch
routing, no n8n server), eval:pairwise (head-to-head build comparison vs
ai-workflow-builder.ee). Authoring them is out of scope here — see the README
sections of the same names.