Skip to content

refactor(trace-analysis): delete the dead and repeated code in six over-gate TraceLens functions - #2069

Merged
ZhengGong-amd merged 10 commits into
mainfrom
lean/shafeng/cc20-tracelens-gate
Oct 21, 2026
Merged

refactor(trace-analysis): delete the dead and repeated code in six over-gate TraceLens functions#2069
ZhengGong-amd merged 10 commits into
mainfrom
lean/shafeng/cc20-tracelens-gate

Conversation

@fengshaoyi-amd

@fengshaoyi-amd fengshaoyi-amd commented Oct 20, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

What this does

Six functions on the TraceLens trace-analysis path were over the code-metrics gate: five in tracelens_analysis.py and aggregate_by_source_function in tracelens_skill_runner.py. This PR brings each one under every limit it broke, mostly by deleting what it re-checked: guards on inputs whose only caller already guarantees them, branches no producer can reach, a parameter only tests passed, two lazily filled caches, and local copies of a shared env parser and a shared numeric coercion. _inject_collective_candidates also hands its workload matching, which was a set of closures inside it, to named module functions.

Production code: +347 / −473 (net −126) across 4 files, of which +339 / −472 is the two TraceLens modules. Tests: +6 / −10 in one file. The code-metrics baseline loses 13 entries, lowers the two module-length entries and gains none.

origin/main is merged into this branch (merge-base 958874f). The PR also carries three constraint comments that #2057 dropped and that came back after it merged: the _TraceBuffers.graph_coverage docstring and _finalize in _bypass_trace_reader.py, and build_verdict in trace_selfcert.py.

This continues the complexity cleanup of #2024, #2026, #2052, #2057 and #2067. Each commit message lists what it deleted and why that is safe.

Commit Function Production lines Baseline file
5bf288d one collective-op table in _enrich_kernel_contract _enrich_kernel_contract +29 / −41
d10be8c drop unreachable branches in classify_patchability classify_patchability +13 / −31
0283c4f one grep job in find_benchmark_files find_benchmark_files +37 / −67
8dc55b7 split CSV invocation recovery out of _finalize_candidates _finalize_candidates +51 / −58
89caf82 one P-item prose rule in aggregate_by_source_function aggregate_by_source_function +49 / −107
9f169ca drop test-only roots and dead reads in collective injection _inject_collective_candidates (also the test file) +160 / −168
a7d5cfa drop the baseline entries the reworked functions cleared baseline only 13 removed, 2 lowered
90b3efd merge origin/main resolved by key
1aedcc2 restore the comments that state constraints _TraceBuffers, build_verdict +8 / −1
fea0dbe record the TraceLens module lengths after the merge with main baseline only 2 module lengths re-measured

All four production files are under src/hyperloom/orchestrator/trace_analysis/.

Gate numbers per function

Measured with the gate's own collectors (ruff C901, complexipy, function lines), origin/main → this branch. Limits are 20, 30 and 80; bold values were over the limit at base.

Function Cyclomatic Cognitive Lines
_enrich_kernel_contract 12 → 11 35 → 29 91 → 62
classify_patchability 20 → 19 32 → 29 97 → 79
find_benchmark_files 19 → 10 44 → 20 90 → 47
_finalize_candidates 18 → 7 58 → 18 115 → 60
aggregate_by_source_function 14 → 6 38 → 17 158 → 79
_inject_collective_candidates 53 → 13 131 → 24 150 → 69

New functions, cyclomatic / cognitive / lines:

  • tracelens_analysis.py: _attach_collective_workloads 12 / 27 / 63, _attach_workload 12 / 25 / 41, _recover_csv_invocations 10 / 21 / 46, _files_mentioning 3 / 6 / 18, _first_shape_record 7 / 14 / 14, _workload_family 2 / 1 / 6, _workload_dtypes 1 / 4 / 4, _has_workload 1 / 1 / 3, _is_all_reduce_name 1 / 1 / 3, _trace_name 1 / 0 / 3.
  • tracelens_skill_runner.py: _pitem_prose 3 / 8 / 23.

Deleted functions: the seven closures inside _inject_collective_candidates (_name, _workload_dtypes, _has_workload, _is_all_reduce_workload, _workload_family, _first_shape_record, _merge_workloads) and find_benchmark_files._is_multigpu.

Across the 2 touched production files: C901 total 781 → 740, cognitive total 1,362 → 1,269, functions 150 → 153. Functions over a gate limit: cyclomatic 2 → 1, cognitive 11 → 5, length 11 → 5. What is left is outside this work: main, write_reports, enrich_candidates_with_runtime_metadata, _apply_vendor_operator_playbook_grouping, _check_chunk_quality, _drive_sdk_stream and run_tracelens_skill.

Module lengths: tracelens_analysis.py 5,352 → 5,277, tracelens_skill_runner.py 1,380 → 1,322.

New names

  • _COLLECTIVE_OPS: the kernel-name tag → collective-op table that _enrich_kernel_contract reads for both the membership test and the lookup.
  • _BENCHMARK_GREP_INCLUDES: the grep --include globs of find_benchmark_files.
  • _MULTIGPU_PATH_TAGS: the path tags that sort multi-GPU benchmarks last.
  • _files_mentioning: one grep for one keyword under one directory.
  • _recover_csv_invocations: fills a candidate's shapes, invocation cases and raw arg spec from ops_unique_args.csv.
  • _pitem_prose: one prose entry per TraceLens P-item, P1 first.
  • _attach_collective_workloads: gives each source-resolved collective its traced workload and returns the rows to append.
  • _attach_workload: one donor lends its whole workload, several lend their distinct first inputs; it replaces the inline one-donor copy and _merge_workloads.
  • _trace_name, _workload_dtypes, _has_workload, _workload_family, _first_shape_record: the injector's closures moved to module level.
  • _is_all_reduce_name: whether a trace name carries all-reduce semantics; it is _is_all_reduce_workload without its unreachable kernel_contract branch.

Removed

  • The test-only source_roots parameter of _inject_collective_candidates, and the notes parameter of its _skip helper.
  • Two lazily filled fused-MoE caches in _finalize_candidates.
  • The _pitem_prose_seen set that deduplicated P-item prose while rows arrived and had to be popped before return.
  • The _COLL tuple that repeated _OPMAP's tags, and two of three reads of the TP size.
  • Local copies of shared helpers: the boolean env parser (now hyperloom.common.env.env_flag) and the _safe_float alias (now safe_float).
  • Unreachable branches: the tensile arm of the aten:: check, the source-type condition on the aiter device-source promotion, the kernel_contract arm of the all-reduce check, the dtypes reads and the empty-family check.
  • Guards on inputs the only caller already guarantees: the injector's input validation, model_params or {}, the empty-input return, isinstance checks and try/except blocks in aggregate_by_source_function, proc.stdout or '', if source_file: and not _sf or.
  • A hand-written dedupe loop (now dict.fromkeys), the nested _is_multigpu, and a 13-line grep argv.

Why the deletions are safe

Line numbers are on this branch.

  • classify_patchability, tensile arm. library is always a lowercased string (tracelens_analysis.py:1218), and a tensile library returns at the vendor-backend check (:1219-1222; tensile is in _VENDOR_BACKEND_LIBRARIES at :1109) before the aten:: check (:1223) can see it. not library and library == "" are the same test on a string.
  • classify_patchability, aiter promotion. For a source_type already in {hip_cpp, triton, python, flydsl} the old code skipped the promotion and then returned True, "" at the end (:1243-1246); the promotion (:1241-1242) now returns the same value one step earlier.
  • _inject_collective_candidates, source_roots and input validation. The only production caller is main (:5074-5079); it passes no source_roots and hands over load_report_tasks rows (:5070), which are always a list of dicts (_analysis_json.py:61-71). The roots come from _aiter_csrc_root() (:1733-1734), as they always did when no roots were passed.
  • _inject_collective_candidates, dtypes and kernel_contract reads. _member_to_candidate (_analysis_json.py:90-114) writes neither key, and neither do the collective rows (_nccl_summary_candidates.py:230-256). The only writer of kernel_contract is _enrich_kernel_contract (tracelens_analysis.py:3722), reached from write_reports (:3879 → :3641), which main calls at :5184, after injection at :5074. The dtypes writers in the tree are in the bypass trace reader (_bypass_trace_reader.py:591, :724), a different route that does not reach the injector. A family key can no longer be empty: a row joins a family only if _is_all_reduce_name finds an all-reduce tag in its name (:1580), and _workload_family (:1585-1590) never strips a name to nothing.
  • _inject_collective_candidates, _skip now flushes messages on every path. The first skip (:1760) runs before anything is appended. At the second (:1775), extract_collective_candidates raised ValueError, and all of its raises (_nccl_summary_candidates.py:177-197, through _prorated_totals at :113-151) come before its first log_fn call (:209, :225), so messages is empty. The third skip (:1793) already flushed them. The log file text is the same on every path.
  • env_flag. env_flag reads the same truthy tokens the deleted parser did: stripped, lowercased, one of 1/true/yes/on; anything else is off (hyperloom/common/env.py:27, :44-61, :89-96).
  • aggregate_by_source_function. Its only production caller is build_task_groups (tracelens_analysis.py:3725), which keeps only dict rows and returns [] before calling it when none is left (:3738-3741). build_task_groups is called only from write_reports (:3888), whose rows come from _finalize_candidates (:5098 → :5139 → :5184; the dry-run-only raw-trace path at :5162 also ends in _finalize_candidates, :2399). Those rows carry numbers or None in the fields read here: _analysis_json.py:92-94 and :107, _nccl_summary_candidates.py:233-239, and _finalize_candidates rounds duration_us and fills a missing gpu_pct (tracelens_analysis.py:2863, :2867-2868). Every group is created with the row that opens it, so rows is never empty and legacy_task_group_keys is always set. P-item prose still takes rows in arrival order (first row per P-item wins) and is then sorted by rank and title. The _safe_float alias has no other reader in the tree.
  • _enrich_kernel_contract. Its only caller is enrich_candidates_with_runtime_metadata (:3641), which passes the result of load_model_kernel_params (:3576); that function always returns a dict (:3482, :3483). The only writer of kernel_contract is this function (:3722), and it writes a dict. _COLLECTIVE_OPS (:3645) iterates in the same order the old _COLL tuple did, so the first match, and the bare reduce last, are unchanged.
  • find_benchmark_files. Its only production caller is _stamp_candidate_metadata (:2763-2765). It returns before the source-stem code when repo_root is empty (:2065-2066). repo_root is the candidate's kernel_repo, which _finalize_candidates sets with find_repo_root(source_file) (:2872), and find_repo_root returns "" for an empty source file (:1822-1823). So with an empty source_file, the code that if source_file: guarded is never reached, and Path("").stem is empty anyway. With text=True and capture_output=True, proc.stdout is always a string.
  • _finalize_candidates. kernel_repo is set at :2872 before the FlyDSL fallback reads item["kernel_repo"] (:2882), and os.path.isfile("") is false, so not _sf or was redundant (:2878). The fused-MoE resolvers read ops_unique_args.csv on every call and return new lists (:2573-2606, :2609-2626); nothing caches them, so no candidate shares a list with another.

Observable effect

Nothing an operator sees changes. Candidate rows, task groups, the analysis JSON, trace-health warnings and log lines are the same for every input the repo's own callers produce.

The differences below are reachable only by input that no caller in the repo produces:

  • _inject_collective_candidates.
    • It no longer takes source_roots=; only tests passed it. It always scans the aiter csrc root, as production already did.
    • Input that is not a list of dicts no longer raises ValueError("Collective candidates must be a list of mappings"). Its only caller passes load_report_tasks rows.
    • A trace row whose dtypes sit only under a dtypes key no longer counts as a workload, and a row that is all-reduce only through a pre-set kernel_contract no longer counts as all-reduce. At injection time no row carries either key.
  • aggregate_by_source_function. None input, non-dict rows, and non-numeric duration_us, call_count, gpu_pct or tracelens_pitem_rank now raise. The base returned [], skipped the row or the metric, or read the rank as 0. Its only caller passes _finalize_candidates rows, which carry those fields as numbers or None.
  • _enrich_kernel_contract. model_params=None and a non-dict kernel_contract now raise. The only caller passes the dict that load_model_kernel_params returns, and only this function writes kernel_contract.
  • find_benchmark_files. A non-empty repo_root with source_file=None now raises TypeError. The only caller derives repo_root from the source file, so a missing source file means an empty repo_root and an early return.
  • _finalize_candidates. Fused-MoE candidates no longer share one CSV read: each reads ops_unique_args.csv itself, as every other candidate in the loop already did, and gets its own invocation-case dicts instead of shared ones. The CSV is the same file within one run, so the written JSON is identical.

Checklist

  • Description: above.
  • Linked issue(s): none.
  • Tests:
    • One test file changes. ExtractCollectiveCandidatesTests.setUp in test_nccl_summary_candidates.py patches tracelens_analysis._aiter_csrc_root to return the test's source root (stopped in addCleanup), and eleven injection calls drop the deleted source_roots= argument. No assertion changed or was removed, so every contract and failure-mode assertion of these tests still runs, now through the production root lookup. test_injection_without_a_source_root_is_also_reported still patches _aiter_csrc_root to "" itself, so it still pins the missing-root health warning. The keyword it passed, source_roots=[], already fell back to that lookup on main.
    • No test pinned any other deleted guard. The remaining assertion on the deleted set (test_tracelens_csv.py:3202, "_pitem_prose_seen" not in g) still holds.
    • Locally on Windows, every test file that imports tracelens_analysis fails at collection on this branch and on the base alike: tracelens_analysis.py imports TraceLens at module level (unchanged since 2026-09), and TraceLens is not installed on that machine. For the same reason the nine ExtractCollectiveCandidatesTests fail there on this branch only, because their setUp now patches tracelens_analysis._aiter_csrc_root; every other local result matches the merge-base. Linux CI on this PR, which installs TraceLens, is the test authority.
    • Both modules are in [tool.coverage.run] omit (pyproject.toml:372-373), so the diff-coverage gate has no changed statements to count.
  • Size/complexity triggers crossed:
    • 820 changed production lines: above the 400-line warning, under the 1,000-line budget.
    • No added or changed function crosses a gate limit. classify_patchability (79 lines, cyclomatic 19, cognitive 29) and aggregate_by_source_function (79 lines) sit close to the triggers; each is one ordered decision ladder or one loop over its rows.
    • Over the cyclomatic-10 trigger: classify_patchability (19, from 20), _inject_collective_candidates (13, from 53), _attach_collective_workloads (12), _attach_workload (12) and _enrich_kernel_contract (11, from 12). The first is the routing gate's ordered rejections; the two new ones are the one match-and-attach loop and the one-donor or many-donor copy.
    • tracelens_analysis.py (5,277 lines) and tracelens_skill_runner.py (1,322) stay over the 1,200-line module limit the baseline carries for them; each gets shorter.
  • If this simplifies or refactors:
    • Mechanisms removed: listed under Removed above.
    • Contract preserved: every in-repo caller and all persisted data.
    • Tests: as above.
  • Observable effect: none; see the section above for why, and for the out-of-contract inputs that now behave differently.
  • Breaking changes: no. The dropped source_roots= keyword belongs to a private function that only tests called with it.
  • PR addresses single concern: yes, bringing these six TraceLens functions under the gate.
  • Root cause is upstream: no.

Known merge interactions

Trial merges used git merge-tree --write-tree, against this branch and against origin/main alone. None of the PRs below adds a call to a function or name this PR deletes or re-signs.

Gate

scripts/code_metrics.py --base-ref 958874f20 (the merge-base) passes on the 6 touched files: 0 new, 0 worse and 0 out-of-date entries, no new or worse duplication or dead code, and 0 findings from the comment, English-only, repeated-literal and A1–A12 checks. Baseline: −13 entries, +0; the two TraceLens module-length entries are set to the merged lengths, 5,277 and 1,322 (main's entries are 5,347 and 1,366).

The gate lists two units under Already over at the base, both untouched here: main (898 lines) and write_reports (204) in tracelens_analysis.py, each two lines over the 896 and 202 that main's baseline carries for them. It warns that _bypass_trace_reader.py is 883 lines, over the 800-line warning and within the 1,200-line limit; the three restored comments add six of those lines. ruff check and ruff format --check are clean on the touched files.

fengshaoyi-amd and others added 7 commits October 20, 2026 22:05
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): one collective-op table in _enrich_kernel_c…
5bf288d 
…ontract

Delete the local _COLL grouping tuple: it listed the same tags in the
same order as _OPMAP, so the membership test and the first-match lookup
now both read one module table, _COLLECTIVE_OPS, with the identical
match order ("reduce" still last).

Read TP_SIZE / TENSOR_PARALLEL_SIZE once instead of three times, and
replace the nested seqlen_regime ternary with a first-match lookup over
("prefill", "decode") that yields the same value.

Drop the "model_params or {}" fallback and the isinstance check on an
existing kernel_contract: the only caller passes the dict that
load_model_kernel_params always returns, and kernel_contract is only
ever written by this function, as a non-empty dict.

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): drop unreachable branches in classify_patch…
d10be8c 
…ability

The aten:: check tested library in {"tensile", "pytorch native"}, but a
"tensile" library already returned through the _VENDOR_BACKEND_LIBRARIES
rejection above it, so the check is now one condition on {"",
"pytorch native"} with the same reason text.

The aiter device-source promotion required source_type outside the
patchable set, and the next statement accepted every source_type inside
it, so an aiter .cu/.cuh/.hip source returns True either way; the
promotion now runs first and the source_type set is tested once.

is_runtime_api_name reuses the name already read from the candidate.
The vendor-playbook comment loses its history and keeps the ordering
constraint, the verdict comment no longer repeats the one on
_RESOLVER_VERDICT_METHODS, and the cpp_itfs comment repeated its own
return message.

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): one grep job in find_benchmark_files
0283c4f 
Move the per-keyword grep into _files_mentioning, which returns the
existing files under one directory that contain one keyword, and an
empty list when grep cannot start, times out or exits with an error;
the caller only walks the benchmark directories and keywords.

Delete what the loop no longer needs:
- the "if source_file:" wrapper around the stem keywords: every
  producer of source_file writes a str, and for "" the stem is "" so
  neither keyword is added;
- "proc.stdout or ''": text=True with capture_output always yields str;
- the hand-written seen/unique loop, replaced by dict.fromkeys with the
  same first-seen order;
- the nested _is_multigpu with its docstring, replaced by a sort key over
  the module tuple _MULTIGPU_PATH_TAGS;
- the thirteen-line grep argv, whose include globs are now the module
  tuple _BENCHMARK_GREP_INCLUDES;
- the history reference in the same-library comment.

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): split CSV invocation recovery out of _final…
8dc55b7 
…ize_candidates

Move the shape, invocation-case and raw-arg-spec recovery from
ops_unique_args.csv into _recover_csv_invocations(item, csv_dir), so
_finalize_candidates keeps the per-candidate ids, GPU share and source
classification.

Delete the two lazily filled fused-MoE caches and their None checks.
The fused-MoE resolvers always return a fresh list, and every other
candidate in the same loop already re-reads the CSV for its own
lookups, so a fused-MoE candidate now reads the sidecar like the rest;
the rows, their order and the grafted provenance are unchanged.

Also drop two guards that could not change the result: "not _sf or"
before os.path.isfile, which is False for "", and the
item.get("kernel_repo", "") default for a key assigned a few lines
earlier. The raw-arg-spec lookup reuses the operation name instead of
reading it again.

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): one P-item prose rule in aggregate_by_sourc…
89caf82 
…e_function

The P-item prose was deduplicated while rows were added, through a
_pitem_prose_seen set that had to be popped before return, then sorted
and filtered after grouping. _pitem_prose(rows) now owns the whole rule:
one entry per (rank, title), supplied by the first row naming it, P1
first, entries with no rank and no prose dropped. It reads the rows in
arrival order, before the duration sort, so the same row wins.

Delete guards the only production caller, build_task_groups, already
makes unnecessary: the empty-input early return (an empty input yields
no groups anyway), the isinstance check (it passes dicts only), and the
try/except blocks around the P-item rank and the three metric sums
(every producer writes numbers or None, as the bypass route's
_build_task_groups already assumes). Every group holds at least one
row, so the "if group['rows']" check before reading the primary goes
too.

The three identity calls share one argument dict instead of repeating
four keywords each, the legacy-key merge reads the key every bucket is
created with, and the module-private _safe_float alias is replaced by
the safe_float it named.

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): drop test-only roots and dead reads in coll…
9f169ca 
…ective injection

_inject_collective_candidates (53/131/150 -> 13/24/69) loses:
- the isinstance validation of its input; the only caller passes
  load_report_tasks' list of dicts.
- the source_roots parameter, which only tests set; they now patch
  _aiter_csrc_root, the path production always takes.
- the kernel_contract branch of the all-reduce check; write_reports
  stamps kernel_contract after injection, so no row carries it here.
- the "dtypes" reads; nothing writes that key on a candidate.
- the empty-family check; a name with an all-reduce tag always yields
  a non-empty family.
- the local copy of the boolean env parser, replaced by
  hyperloom.common.env.env_flag (same tokens, same default).
- the notes argument of the skip path; extraction logs nothing before
  the first two skips, so flushing the shared messages is the same
  write.

Workload matching leaves the closure as _attach_collective_workloads,
and the one-donor copy and the multi-donor merge become one
_attach_workload. Rows, health warnings, log file text and log.warning
records match base over 9,984 input combinations. The injection tests
patch _aiter_csrc_root in setUp instead of passing source_roots.

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): drop the baseline entries the reworked func…
a7d5cfa 
…tions cleared

The six trace-analysis functions reworked on this branch are within
every code-metrics limit, so their 13 baseline entries go:
- tracelens_analysis.py: _enrich_kernel_contract (cognitive, lines),
  _finalize_candidates (cognitive, lines), _inject_collective_candidates
  (cyclomatic, cognitive, lines), classify_patchability (cognitive,
  lines), find_benchmark_files (cognitive, lines).
- tracelens_skill_runner.py: aggregate_by_source_function (cognitive,
  lines).

The module-lines entries fall to the measured lengths:
tracelens_analysis.py 5347 -> 5272, tracelens_skill_runner.py
1366 -> 1308. No entry is added or raised.

Co-authored-by: Cursor <cursoragent@cursor.com>
@github-actions

github-actions Bot commented Oct 20, 2026 •
edited
Loading

Copy link
Copy Markdown

Code metrics gate: PASSED

Judged: the 6 files this change touches. Files it does not touch never fail it; anything found there is listed under Outside this change for information.

Dimension Limit New Worse Baseline out of date Baseline entries vs base
Cyclomatic complexity fails > 20 0 0 0 54 -1
Cognitive complexity fails > 30 0 0 0 207 -6
Function length (lines) fails > 80 0 0 0 350 -6
Nested block depth fails > 5 0 0 0 21 +0
Parameter count fails > 5 0 0 0 95 +0
Module length (lines) warns > 800, fails > 1200 0 0 0 49 +0
Duplicated lines per file fails > 0 0 0 0 96 +0
Dead code (vulture) fails > 0 0 0 0 2 +0
Check Findings Effect
Comments 0
English only 0
Production code imports test code 0
Repeated literals 0
A1 Private name of another package 0
A8 Flag argument 0
A9 Tuple return 0
A10 Raw vocabulary comparison 0
A11 Hidden global write 0
A12 Pure forwarder 0
A12 Single-implementation abstraction 0

Module length warning: over 800 lines, within the 1200-line failure limit (not failing): 1

Module Lines Limit
src/hyperloom/orchestrator/trace_analysis/_bypass_trace_reader.py 883 warns > 800

Already over at the base (backlog, not this change's): 2

Unit Dimension Base Now
src/hyperloom/orchestrator/trace_analysis/tracelens_analysis.py:4348 main Function length (lines) 898 898
src/hyperloom/orchestrator/trace_analysis/tracelens_analysis.py:3848 write_reports Function length (lines) 204 204
107 findings in files this change does not touch

Outside this change (informational): 107

Unit Dimension Baseline Now
src/hyperloom/orchestrator/state/shared_state.py resolve_graded_comparison Cognitive complexity absent 34
src/hyperloom_kb/collect/mapping.py _experience Cognitive complexity absent 31
src/hyperloom_kb/schema.py FieldDeclaration.__post_init__ Cognitive complexity absent 32
src/kernelforge/cli/forge_loop_command.py _prepare_measurement_driver Cognitive complexity absent 36
src/kernelforge/cli/forge_loop_command.py _read_kb_warm_start Cognitive complexity absent 40
src/kernelforge/cli/forge_loop_command.py forge_loop Cognitive complexity absent 31
src/kernelforge/loop/runner.py IterationLoop._collect_session_candidate Cognitive complexity absent 44
src/kernelforge/loop/runner.py IterationLoop._rule_on_attempt Cognitive complexity absent 38
src/hyperloom/inference_optimizer/cli/__init__.py _reexport_session_environment Function length (lines) absent 111
src/hyperloom/inference_optimizer/cli/__init__.py _resolve_agent_lineup Function length (lines) absent 89
src/hyperloom/inference_optimizer/cli/__init__.py _resolve_run_topology Function length (lines) absent 92
src/hyperloom/inference_optimizer/cli/__init__.py _resume_session Function length (lines) absent 170
src/hyperloom/inference_optimizer/cli/__init__.py _seed_fresh_session Function length (lines) absent 164
src/hyperloom/inference_optimizer/multi_node/cli.py _restart_multinode Function length (lines) absent 107
src/hyperloom/inference_optimizer/multi_node/scripts/launch_multinode.py _build_parser Function length (lines) absent 84
src/hyperloom/inference_optimizer/multi_node/scripts/launch_multinode.py _spawn_rank_actors Function length (lines) absent 84
src/hyperloom_kb/sync.py GlobalSync.pull Function length (lines) absent 84
src/kernelforge/cli/forge_loop_command.py _RunReporter.build_result Function length (lines) absent 92
src/kernelforge/cli/forge_loop_command.py _prepare_measurement_driver Function length (lines) absent 136
src/kernelforge/cli/forge_loop_command.py _read_kb_warm_start Function length (lines) absent 149
src/kernelforge/loop/runner.py IterationLoop._close_iteration Function length (lines) absent 117
src/kernelforge/loop/runner.py IterationLoop._collect_session_candidate Function length (lines) absent 185
src/kernelforge/loop/runner.py IterationLoop._implementer_history Function length (lines) absent 112
src/kernelforge/loop/runner.py IterationLoop._rule_on_attempt Function length (lines) absent 102
src/kernelforge/loop/runner.py IterationLoop._source_attempt Function length (lines) absent 175

...and 82 more; the job log lists every row.

Verdict computed by the gate scripts in /home/runner/work/_temp/code-metrics-gate, not by this tree's copy.

Thresholds and sources: [tool.hyperloom.code_metrics] in pyproject.toml. Tools: complexipy 8.0.1, jscpd 5.4.1, ruff 0.16.2, vulture 2.16.

@fengshaoyi-amd fengshaoyi-amd mentioned this pull request Oct 21, 2026
Merged
fengshaoyi-amd and others added 3 commits October 21, 2026 12:18
@fengshaoyi-amd
Merge remote-tracking branch 'origin/main' into lean/shafeng/cc20-tra…
90b3efd 
…celens-gate

Co-authored-by: Cursor <cursoragent@cursor.com>

# Conflicts:
#	scripts/code_metrics_baseline.txt
@fengshaoyi-amd @cursoragent
refactor(trace-analysis): restore the comments that state constraints
1aedcc2 
Put back the three comments the rework dropped that say what the code
cannot: that roctracer's activity buffer overflows under continuous graph
replay, why a graph-replayed kernel without a cpu_op is not an attribution
failure, and that the verdict's measures are reported but never compared.
Comments that only narrated the code stay deleted.

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd @cursoragent
chore(code-metrics): record the TraceLens module lengths after the me…
fea0dbe 
…rge with main

Co-authored-by: Cursor <cursoragent@cursor.com>
@fengshaoyi-amd
fengshaoyi-amd marked this pull request as ready for review October 21, 2026 06:45
ZhengGong-amd
ZhengGong-amd approved these changes Oct 21, 2026
View reviewed changes

@ZhengGong-amd ZhengGong-amd left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewed: gate numbers re-measured (ruff C901, complexipy, function lines) and match the description; each removed guard / "producer always sets it" claim checked against its producers and call sites; base vs head differential runs show no undisclosed behaviour change; tests covering the touched modules pass.

LGTM.

@ZhengGong-amd
ZhengGong-amd merged commit 01b6ca6 into main Oct 21, 2026
40 of 42 checks passed
@ZhengGong-amd
ZhengGong-amd deleted the lean/shafeng/cc20-tracelens-gate branch October 21, 2026 09:27
@ZhengGong-amd ZhengGong-amd mentioned this pull request Oct 21, 2026
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@ZhengGong-amd ZhengGong-amd ZhengGong-amd approved these changes
@tsrikris tsrikris Awaiting requested review from tsrikris tsrikris is a code owner
@devalshahamd devalshahamd Awaiting requested review from devalshahamd devalshahamd is a code owner

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@fengshaoyi-amd @ZhengGong-amd