// Drive the full migration for an op family — audit, delegate per-op alignment to align-op, run cross-op cleanup. Two terminal outcomes: SUCCESS opens a PR; CLEANUP_REGRESSION exits blocked without a PR when post-cleanup tests fail.
| name | align-family |
| description | Drive the full migration for an op family — audit, delegate per-op alignment to align-op, run cross-op cleanup. Two terminal outcomes: SUCCESS opens a PR; CLEANUP_REGRESSION exits blocked without a PR when post-cleanup tests fail. |
Family name from tileops/manifest/ (e.g., reduction, norm, attention).
family namealign-op (promoted or blocked), cleanup succeeds, PR opens.CLEANUP_REGRESSION terminal in the state diagram. Distinct from the non-terminal per-op REPORT_BLOCKED state.align-op) and either (a) CLEANUP + CREATE_PR succeed, or (b) a promoted op's tests fail after CLEANUP's dual-path removal, causing the run to exit via CLEANUP_REGRESSION with the regression recorded. (REPORT_BLOCKED is the non-terminal per-op blocked state and never terminates the run.)align-family delegates every per-op stage to align-op, invoked as a separate sub-agent per op. The family orchestrator never runs any atomic per-op skill directly — those live inside align-op's contract.
align-family does not write tileops/manifest/. After the refactor, align-op is the sole manifest writer (at its own FLIP_STATUS step); align-family observes status transitions via align-op's SUCCESS return. No align-family stage edits, modifies, or flips the manifest.
Directly-invoked sub-skills of align-family are exactly two: audit-family (in AUDIT) and align-op (per op).
| Stage | Sub-skill |
|---|---|
| AUDIT | audit-family |
| ALIGN_OP | align-op |
stateDiagram-v2
[*] --> AUDIT
AUDIT --> GROUP_BY_BASE: gap report generated
GROUP_BY_BASE --> ROUTE: ops grouped by base class
ROUTE --> ALIGN_OP: ready (mode=minor) or semantic_gap (mode=redesign)
ROUTE --> REPORT_BLOCKED: classification=blocked (audit-family)
ALIGN_OP --> CLEANUP_GATE: align-op SUCCESS
ALIGN_OP --> REPORT_BLOCKED: align-op BLOCKED
REPORT_BLOCKED --> CLEANUP_GATE: record blocked reason
CLEANUP_GATE --> ROUTE: group incomplete, next op
CLEANUP_GATE --> CLEANUP: all siblings promoted or blocked
CLEANUP --> ROUTE: legacy path removed, next group
ROUTE --> CREATE_PR: all ops processed
CLEANUP --> CLEANUP_REGRESSION: promoted op test fails after cleanup
CLEANUP_REGRESSION --> [*]: terminal blocked, no PR
CREATE_PR --> [*]
After each sub-agent returns and before dispatching the next, verify:
test -z "$(git status --porcelain)"
This catches tracked changes, staged changes, AND untracked files. If not clean: the sub-agent's commit failed (pre-commit hook, staging issue) or left new files uncommitted. Orchestrator commits on behalf, then proceeds. Every agent must start with a clean worktree.
When align-op rewrites a base class during its per-op pipeline, it may create a dual-path __init__ (legacy + spec) to keep unmigrated sibling tests passing. This is correct temporary debt — the cleanup gate removes it.
Dual-path definition: a class __init__ with runtime branching to support two incompatible construction interfaces, and forward dispatching to two execution paths. Not polymorphism — same semantics, temporary interface coexistence.
/audit-family <family>
Gap report written to .foundry/migrations/<family>.json.
Group ops by base_class from the gap report. Each group is a set of sibling ops sharing a base class. Process groups in order; within each group, process ops in order (first op likely fixes the base class, subsequent ops validate).
base_class is a required field in the gap report. audit-family must populate it for every op entry. If an op inherits Op directly (no intermediate base class), its base_class is "Op" — these ops form a single group but are independent (no shared base class to rewrite, so cleanup gate is a no-op for this group).
Track group completion: a group is complete when all its ops are promoted or blocked.
Read the gap report from AUDIT. For each op in the current group, route by classification (the field audit-family populates):
Gap-report classification | Action | Why |
|---|---|---|
ready | → ALIGN_OP with --mode=minor | Existing code already conforms; align-op's minor path runs the spec tests (DONE_SKIP since they pass), then the shared downstream + flip. |
semantic_gap | → ALIGN_OP with --mode=redesign | Family-scoped historical migration treats every semantic_gap op as a structural redesign by default — legacy code is being replaced wholesale. For per-op semantic_gap cases that are actually minor manifest deltas, use single-op align-op <op> --mode=minor directly instead of align-family. A future recommended_mode gap-report field could refine the routing automatically. |
blocked | → REPORT_BLOCKED with audit's reason. Skip align-op. | Audit determined the op cannot be migrated autonomously (no pytorch_equivalent, kernel-layer change required, etc.). No per-op work to do. |
The mapping is deterministic — align-family MUST pass --mode= explicitly so align-op never falls into its interactive prompt branch in a batch family migration.
For each op routed here, invoke align-op as a separate sub-agent with the mode determined in Step 3:
align-op <op_name> --mode=<minor|redesign>
align-op owns the entire per-op pipeline internally — its internal stages (classify, dispatch on case, test / implement / bench, revalidate, flip status, cleanup, report) are align-op's contract, not align-family's. See align-op/SKILL.md for the authoritative stage list and the conditional-IMPLEMENT rule. align-family does not manage or observe align-op's internal stages; the only interface between them is align-op's SUCCESS / BLOCKED return.
Per-op outcome:
align-op returns SUCCESS → op is promoted (manifest status already flipped by align-op's FLIP_STATUS). Record the returned report. Proceed to CLEANUP_GATE.align-op returns BLOCKED → op is blocked. Capture align-op's BLOCKED reason verbatim as the per-op result. Proceed to CLEANUP_GATE.align-family MUST NOT re-flip manifest status or re-run per-op validation on its own; align-op's SUCCESS return is the single source of truth that the manifest was flipped.
The flip performed by align-op MUST stay within the Status flip carve-out; align-family MUST NOT batch contractual-field edits across ops in the same PR.
After each per-op outcome — whether align-op returned SUCCESS / BLOCKED or ROUTE recorded an audit-classified blocked op directly to REPORT_BLOCKED — check group completion:
promoted or blocked? → trigger CLEANUPRemove dual-path legacy code from the base class. This step fires once per base-class group, after all siblings have gone through align-op (SUCCESS or BLOCKED).
Actions:
__init__ branch (if M is not None and N is not None path)_legacy flag and _forward_legacy methodM, N keyword-only parameters from __init__--check-op for promoted ops only (blocked ops' tests may legitimately fail)If any promoted op's test fails after cleanup → transition to CLEANUP_REGRESSION (terminal): record the regression, skip CREATE_PR, exit with blocked status. Do not proceed with a broken state.
Two distinct blocked states, split so the diagram has one meaning per state:
REPORT_BLOCKED — per-op align-op returned BLOCKED. Records the reason, returns to CLEANUP_GATE, continues with sibling ops. Non-terminal.CLEANUP_REGRESSION — a promoted op's tests fail after CLEANUP's dual-path removal. Terminal: the family migration exits without opening a PR. The blocked report becomes the run's final artefact.Timeout policy for blocked ops: if a group has blocked ops that prevent the cleanup gate from firing for an extended period, the orchestrator may force cleanup — remove legacy path and mark blocked ops' tests as xfail. This is a human decision, not automatic.
After all ops processed:
align-opalign-op reports)align-op (e.g., needs_kernel_work, needs_human_decision) for human doc reviewalign-op's BLOCKED reasonsSingle-shot review of a tile-ai/TileOPs PR as a separate GitHub identity from the PR author. Manual / interactive — for autonomous multi-round review until APPROVE, run `bash .claude/skills/review-tileops/loop.sh <PR>` instead.
Per-round driver of stateful agent-driven review-resolution on a tile-ai/TileOPs PR (developer side). Designed for /loop dynamic mode — re-fires until a terminal action. Outer caller must run preflight.sh once before starting the loop. State persists in `.foundry/runs/{issue-<N> | pr-<PR>}/resolve/`.
Per-op orchestrator that brings a single op into alignment with its manifest entry. Classifies the op into one of three cases (green field / interface redesign / minor delta), dispatches to the right path (scaffold-op for new, archive+rescaffold+port for redesign, implement-op for minor), then runs the shared downstream (test → bench → validate → flip status → report). Complements the family-scoped `align-family`; per-op entry when you know the op you want to touch.
Modify op code to match the manifest-declared interface, making spec tests pass.
Generate or re-align one `tileops/manifest/` entry from a reference-API docs URL. Caller provides the manifest key (`op_name`); skill writes that one entry. Idempotent.
Patch one missing structural field (kernel_map, static_dims) on an existing tileops/manifest/ entry. Auto-detects the field via the validator or takes `--field=<name>`. Reference-derivable fields (signature.*, shape_rules, dtype_combos, roofline) belong to add-manifest, not here.