| name | escalate-over-shortcut |
| description | Use when about to commit a draft solution that suppresses a problem (silent fallbacks like || true, [ExcludeFromCodeCoverage] on throw/stub code, lowered thresholds), pulls state from somewhere illegitimate (binaries from an unrelated app's install dir, prod wiring under tests/, cross-worktree file copies), adds configurability with no production caller (hard-coded values keyed to one platform, test-only knobs), ships a placeholder/no-op (return [], empty handler) as the answer, or leaves rationale-free drift (mid-phase dependencies, TODO/FIXME, unexplained changes). Agent-to-self companion to pushback; fires on the agent's own draft, not on the user's request. Most invocations end "no, not a shortcut, ship" - that's healthy; the value is the check. |
Escalate Over Shortcut
Why this skill exists
Agents optimize locally. The work item said "make this phase pass"; the phase passes; deliverable met. When the real environment pushes back - a missing native DLL, a platform-bound class, a coverage gate over untestable code, a flaky emulator - the locally-optimal move is whatever unblocks the task fastest. The globally-optimal move is often to STOP and tell the orchestrator what's weird.
The failure mode this skill prevents: the agent silently ships a shortcut, the build looks green, the integration is actually broken. Escalation is cheap; the agent isn't the right one to absorb that decision alone.
The failure mode this skill must NOT create: a paranoid agent that flags every careful piece of design as a shortcut. The bar is concrete: the draft contains a pattern the agent would be embarrassed to defend in PR review.
This is agent -> self review. Companion to pushback, which is agent -> user.
When this fires
You're holding a draft solution and one of these shapes is present. Each is a class of smell; full catalogue with code excerpts is in references/red-flag-patterns.md.
- Suppressing a problem instead of fixing it. Silent fallbacks (
|| true, empty catch), coverage exclusions on stub/throw code, lowered thresholds to escape a gate. The problem becomes invisible; the gate goes green; the integration is still broken.
- Pulling state from somewhere illegitimate. Reading binaries from an unrelated app's install dir, copying files between concurrent worktrees, placing production-shaped wiring under
tests/. The diff works locally because this environment happens to provide what's needed; it won't elsewhere.
- Configurability that only tests use. Hard-coded values keyed to one platform, parameters with no production caller, "test-only" knobs. Sometimes the right shape - see the next section.
- Unexplained mid-task drift. A dependency added mid-phase with no brief support, TODO/FIXME left in code, "temporary" constants. The rationale is in your head right now; it won't be in PR review. A new third-party dependency is the sharp case: that a library is the obvious or standard tool for the task (a mocking framework to write a test, an HTTP client for one call) is not the brief asking for it. Pulling a dependency into the project is a decision the orchestrator owns; "it's the normal way to do this" is the rationalization to catch, not the license to proceed. A hand-rolled fake or a few lines of stdlib usually costs less than the dependency you were about to add.
- Placeholder-as-answer. A no-op return (
return [], return None, an empty handler body) shipped instead of the real implementation, with reasoning like "the honest state when X isn't wired yet" or "the test's assertions hold vacuously over an empty collection." The test passes - but only because there's nothing to assert against, not because the work was done. The tell: you're defending the placeholder as "not the obvious shortcut, since I didn't fake the data." Avoiding the obvious shortcut isn't the same as doing the work; the missing wiring is the actual problem and the empty return makes it invisible. Right move: BLOCKED, name the missing wiring, ask whether to fake (with explicit acceptance) or stop.
If your draft contains something that rhymes with these but isn't on the list - the gut check catches it.
The gut check
"If a senior engineer saw this PR and it touched X, would I be comfortable defending it?"
If the defense involves "well, it's just for the test" or "this machine happens to have it" or "we can fix this in the next round" - that's the signal. STOP and escalate.
The check applies even when no pattern above matches verbatim. The shapes are the common forms of the failure; the gut check is the underlying discipline.
When the brief or the code offers you the shortcut
The shapes above are easiest to catch when you invented them. The harder case is when the shortcut is handed to you - the brief says "just make the noise stop," the existing code carries a comment like "if you need different wiring for tests, do it from the test project," or a task frames a suppression as "the honest interim state." The temptation is to read that framing as license.
That framing IS the signal. Someone writing down the shortcut shape and offering it to you doesn't make it shaped right - it makes it more likely that the gap you'd be papering over is the exact thing the orchestrator wanted surfaced. A brief or a seed comment that pre-authorizes the smell is describing a symptom, not granting permission. STOP and escalate the same as if you'd reached for the shortcut on your own initiative - name the shortcut the framing invited, and ask whether to take it (with explicit acceptance) or fix the upstream gap.
Test seam vs. escape hatch
Adding configurability is not, by itself, a shortcut. The distinguishing question is not "does the default have a non-null value" - every parameter has some default. The question is about call sites:
Does any production call site pass a value other than the default?
Equivalent gut-check: if you deleted the parameter and inlined the default as a constant, would any production code break?
- Yes, real production callers pass non-default values -> injection seam. A
Clock parameter that defaults to Clock.systemUTC() and is wired with a timezone-aware clock from the composition root is real configurability.
- No, every production caller uses the default; only the test passes something else -> escape hatch. The parameter exists solely so the test can override it. That's a test-only knob dressed up as a constructor param. Escalate to ask whether the production code should be re-shaped instead - use the test framework's virtual time (
runTest + advanceTimeBy), expose a Scheduler the production composition actually wires, or inject a real abstraction with a meaningful production value.
When in doubt - escalate.
The BLOCKED escalation
When STOP triggers, emit BLOCKED with five fields. This is a complete outcome - not a "starting point" - and the orchestrator handles it.
## BLOCKED
**Task:** Phase 12, task 3 - wire FFmpeg native DLLs for integration test
**About to commit:** `WindowStream.Integration.Tests.csproj` step that copies
ffmpeg-*.dll out of `$(ProgramFiles)\obs-studio\bin\64bit\` into the
test output directory.
**Smell:** reading binaries from an unrelated app's install dir. OBS happens
to ship FFmpeg 7 DLLs; that's not a supported acquisition path.
**Tried:**
- NuGet: no `FFmpeg.AutoGen.runtime.*` for v7 published.
- Manual download: BtbN release builds available, no automation in place.
- vcpkg: not configured in this repo.
**To proceed cleanly:** orchestrator pre-places `ffmpeg-*.dll` (e.g.,
via a small downloader invoked by the build), or accepts the OBS copy
explicitly as a knowingly-fragile starting point.
- Task - phase + task number + brief context.
- About to commit - the actual diff: file paths and one-line behaviour.
- Smell - which shape, and why specifically it's wrong-shaped.
- Tried - alternatives considered and what blocked each. Shows the search space.
- To proceed cleanly - what the orchestrator would need to provide.
Do not ship the shortcut alongside the BLOCKED. The orchestrator's decision (accept the shortcut, fix the upstream gap, narrow scope) is clean only if the diff doesn't exist yet.
When the user is the orchestrator
In a foreground session, the user IS the orchestrator. Same five fields, inline:
"Heads up before I commit X - looks like a [smell]. Tried A, B, C; each blocked by [reason]. Cleanest path forward is [orchestrator action]; otherwise [the shortcut] as known-fragile. Which?"
One line back from the user resolves it. The point is awareness before the shortcut lands, not litigation of the design.
What this is NOT
Not pushback. Pushback is agent -> user, fired on the user's requests. This is agent -> self, fired on the agent's drafts. If the user asks for something that looks like a shortcut, that's pushback. If you're about to commit something on your own initiative that looks like a shortcut, that's this skill.
Not smoke-test. Smoke-test asks "does the change work?" This skill asks "is the change shaped right?" A smoke-pass on a shortcut-shaped solution is the exact failure mode this skill prevents - the test passes, the integration is broken.
Not pessimism. Most drafts ship cleanly. Most invocations end in "no, not a shortcut, ship." That's healthy; the value is the check.
Not paranoia about test code. Mocks, fakes, injectable parameters with sensible defaults - these are good design. The test-seam-vs-escape-hatch question above tells you which is which.
Self-check before shipping
Before any "phase done" or commit, silently answer:
- Does the diff contain a pattern from the shapes above, or something rhyming?
- Would I defend the shape of this change to a senior engineer without "just for the test" or "this machine happens to have"?
- If a parameter or knob is new, does any production call site pass a non-default value? (If every production caller uses the default and only the test overrides it, it's a test-only knob - escalate.)
- If I added a dependency, did the brief ask for it? (A library being the obvious or standard tool for the job is not the brief asking for it - a new third-party dependency is the orchestrator's call.)
- If something is suppressing a failure (
|| true, empty catch, [ExcludeFromCodeCoverage]), is the suppression itself the thing I should escalate?
- If the answer I'm about to ship is empty / zero / no-op (
return [], return None, empty body): does the test pass because the work was done, or because there's nothing for the assertions to fail against? If the latter, the work isn't done - escalate the missing wiring.
If any check trips: STOP, emit BLOCKED, do not commit the diff. The orchestrator handles it.
If all checks pass: ship.
