hw-test-dedup

name	hw-test-dedup
description	Find and remove redundant FBOSS agent_hw_tests without losing coverage. Use when asked to reduce/combine/dedup HW tests, retire non-dataplane (set/get) tests where dataplane tests exist, or cut HW test runtime while keeping coverage. Report-first; applying edits is a separate opt-in step.
metadata	{"oncalls":"fboss_agent","strict":true,"apply_to_path":"fbcode/fboss/agent/test/agent_hw_tests/.*\\.cpp$","apply_to_user_prompt":"(?i)(dedup|redundant|fold|combine|reduce).*(hw[ _]?test|agent_hw_test)"}

HW Test De-duplication

Given a set of FBOSS agent_hw_tests, find test cases whose hardware coverage is fully contained in other tests, and recommend removing them — fewer tests, identical coverage. Default output is a report; turning recommendations into code edits is a separate, confirmation-gated step.

IMPORTANT: report-first, never auto-post

• Default run produces a report only. It never edits tests or creates diffs unless --apply is passed AND the user confirms.
• It never submits/publishes to Phabricator on its own.
• When a removal cannot be proven safe, the test is kept and the reason is written down. Bias is always toward keeping coverage.

When to use

• "dedup / combine / reduce these HW tests", "retire set/get tests where we have dataplane tests", "cut HW test count without losing coverage".
• Scope: fbcode/fboss/agent/test/agent_hw_tests/ (this version). Other test trees are out of scope.

Invocation

Form	Meaning
/hw-test-dedup <file1> <file2> ...	Analyze an explicit set of test files.
/hw-test-dedup <dir>	Discover + cluster the test files under a dir, then analyze per cluster.
--apply	After analysis, apply the folds/removals as draft diffs (gated; see Apply rules).
--deep-verify	Opt-in empirical SAI call-trace check for top candidates (needs build/run).
--self-test	Run against the pinned eval corpus and diff verdicts vs references/eval/.
--gap-report <D1,D2,...>	Compare this skill's recommendations to already-landed reduction diffs; emit a gap report.
--resume	Continue from the last completed phase checkpoint.

Steps the agent follows

The orchestration is implemented in scripts/dedup-workflow.js (run via the Workflow tool). Follow these steps:

1. Resolve inputs. Expand the file list (or discover files under the dir). Confirm the cluster grouping with the user when in <dir> mode.
2. Pre-filter (deterministic, no LLM). Run scripts/prefilter.js to cluster by fixture base + feature area and prune pairs that obviously can't overlap. Pruned pairs are recorded for the report.
3. Profile each test (fingerprint). Per the Fingerprinter persona (references/personas.md) and references/fingerprint-schema.md, emit a coverage profile per test. Resolve helper code transitively. Pass the roster by file reference, not inlined.
4. Count-check completeness (deterministic). Run scripts/roster-count.js; the fingerprint count per file must equal the TEST_F/TEST_P/TYPED_TEST count. A mismatch hard-fails the file.
5. Detect overlaps + apply gates. Per the Subset-detector persona over pre-filtered pairs, under the 6 universal gates (references/universal-gates.md) plus the active coverage pack (references/coverage-packs/<area>.md).
6. Second opinions (adversarial). Per the Skeptic persona: 1 lens in report mode, 3 diverse lenses (unanimous-safe) in --apply mode. Refute-by-default; ≥1 reads raw source.
7. Reconcile (deterministic). Build the subset→superset graph; keep a maximal antichain; never remove a node whose only superset is itself removed; last-feature-owner is never removable.
8. Merge-plan. Per the Merge-planner persona, list the unique assertions to port into each survivor first; portability check.
9. Report (and, with --apply, perform merge → delete → arc f/arc lint → review pass → one commit per file → draft diffs).

Hard safety rules (enforced in code by the workflow)

• A test that is the last owner of any ProductionFeature is never removable.
• A removal requires the surviving test(s) to satisfy all universal gates (references/universal-gates.md).
• --apply is allowed for an area only once that area has (a) a validated coverage pack, (b) its own eval corpus with a passing --self-test, and (c) a human spot-check of a pilot report. ACL is the first such validated area; every other area is report-only until it meets the bar.

Reference material

File	What
references/personas.md	The 6 agent personas + their output schemas.
references/universal-gates.md	The 6 universal safety gates.
references/fingerprint-schema.md	The per-test coverage-profile schema.
references/warmboot-semantics.md	verifyAcrossWarmBoots behavior across the 4 base classes.
references/coverage-packs/*.md	Per-area extra checks (acl validated; ecmp/mirror/trunk/srv6 drafted, report-only; qos/copp stubs).
references/eval/	Ground-truth corpora + how --self-test / --gap-report use them.
references/contributing.md	How to add a coverage pack / gate / eval corpus.
references/design.md	The design of record.