[axelray-dev]

Summary

• Preserve complete JSON output from ruff analyze graph instead of applying the text passthrough cap.
• Add a regression test that parses a graph larger than the cap.

Fixes #3929

Test plan

• cargo fmt --all -- --check
• cargo clippy --all-targets
• cargo test --all
• Manual test: Ruff 0.16.6 analyze graph output through rtk remained 17,403 bytes and parsed as JSON.

[CLAassistant]

All committers have signed the CLA.

[Protocol-zero-0]

[P2] Recognize analyze graph after Ruff global options

Verified head 4c51f02f with real Ruff 0.17.0 on Linux. The bare ruff analyze graph big case is fixed: a 301-file package produces 12,628 bytes of valid JSON, identical through native Ruff and this head; the base truncates it to 2,001 bytes.

However, these supported forms still return the same invalid, truncated JSON on this head:

rtk ruff --isolated analyze graph big
rtk ruff --color never analyze graph big
rtk ruff --config ruff.toml analyze graph big
rtk ruff --config=ruff.toml analyze graph big

Native Ruff succeeds and emits the full graph for each. Both native and RTK exit 0, so a caller only discovers the corruption when parsing stdout. Moving --isolated after analyze graph makes the head preserve the graph correctly. A small graph below the cap also passes, which explains why routing-only/small-output tests would miss this.

Reproducer in an empty directory:

from pathlib import Path
p = Path("big")
p.mkdir(exist_ok=True)
(p / "__init__.py").write_text("")
for i in range(300):
    (p / f"m{i:03d}.py").write_text(
        f"from big import m{(i + 1) % 300:03d}\n"
    )
Path("ruff.toml").write_text("line-length = 88\n")

ruff --isolated analyze graph big | python3 -m json.tool >/dev/null
rtk ruff --isolated analyze graph big | python3 -m json.tool >/dev/null

The first parses successfully; the second fails. is_analyze_graph_invocation only matches arguments 0 and 1, so the new exemption is bypassed by global options. Please identify the subcommand after global options, including value-taking --config/--color and their equals forms, using the shared argument-tokenization approach, and add a large-graph regression for those forms.

This is an inherited corruption case that the new fix still leaves open, rather than a regression introduced by this patch. Independent validation covered nine scenarios / 27 native-base-head executions: the forms above, a global option after the action, a small graph, a directory named format, and an invalid option. Exit statuses 0/2 matched native Ruff throughout. Focused Ruff tests, fmt and warnings-denied clippy passed. Full-suite, current-develop integration and other platforms were not tested.

[KuSh]

Round 1 — APPROVED, follow-ups filed

Claim: rtk ruff analyze graph returns ruff's complete JSON instead of cutting it at the 2000-char passthrough cap.
Scope: accept (round 1, frozen)
Ran: ruff 0.16.6 on a 301-file package, 20 argument forms run under native ruff, merge-base f749546, this head, develop 83fea35 and this head merged on develop; invariants: non-default targets (--direction dependents, . default with a 27 KB graph, a package dir named format), over-cap input byte-compared with native, under the hook, every decision; savings 0% on analyze graph by design ("Correctness VS Token Savings" in CONTRIBUTING.md), every other path unchanged vs develop
Pins: 3 decisions (the analyze and graph conjuncts, the exemption limited to the graph) plus the run() wiring; on 4c51f02 6 mutated, 4 survived (both conjuncts, always exempt, wiring). With the fixup, 9 mutated (adding an uncapped else arm, an extra forwarded argument and the arguments forwarded as one), 1 survives: dropping the analyze conjunct, which no real invocation can tell apart (every ruff <X> graph with X ≠ analyze exits 2 with empty stdout)
Matrix: executed this round; pin check: executed this round
CI @5917f7ab5: all green (15 checks, test on ubuntu, macos and both windows targets)

Blocks merge (0)

Filed as follow-ups

• #4654 rtk ruff subcommand detection ignores ruff's global options — outside this PR's claim, same on develop. --isolated, --config X, --color X, -q before analyze, or -v/--isolated/--config X between analyze and graph, still return 2001 B of invalid JSON (8 forms). Same class as #3927 and #3928; the fix is one tokenized subcommand finder for all three detectors.
• #4655 rtk ruff cuts SARIF, JUnit, GitLab, JSON Lines and rule --output-format=json at the passthrough cap — outside this PR's claim, same on develop. rule E501 from #3929's body is listed there too.

Checked and correct — no need to re-verify

• analyze graph big, analyze graph (27 KB), --direction dependents, -v after the files: byte-identical to native and valid JSON; develop returns 2001 B of invalid JSON.
• Small graph, missing directory, --bogus (exit 2), analyze --help, analyze graph format, rule, check: same output and exit codes as develop. Exit codes match native on every row.
• Hook: rtk rewrite leaves ruff analyze … unrewritten, as on develop, so the path is reached by an explicit rtk ruff.
• Head merges cleanly onto develop 83fea35; fmt and clippy clean, tests green on the head and on the merge.

Fixup

Two commits on the branch:

• a merge of develop (clean, no conflicts): the branch base predates tests/common/, which the new integration test needs;
• test(ruff): pin the analyze graph exemption on captured output: replays captured ruff analyze graph and ruff rule E501 output through run() with a stub ruff that checks the exact argv it receives; the unit test uses the captured graph instead of a hand-built one, and a table pins which invocations are exempt. The rule E501 test pins where the exemption stops today; #4655 would change that deliberately.

Your questions

None from the author. The other review on this PR: its global-options finding reproduces (and extends to options between analyze and graph). It is the same on develop, so it is tracked in #4654 rather than held here.

Next

fixup 5917f7a pushed, code-reviewed: merges develop and adds tests pinning the exemption and the run() path on captured ruff output — approving.
Threads: 0 open, 0 resolved this round.

[KuSh]

@Protocol-zero-0 thank you for the global-options repro. It reproduces exactly on 4c51f02, and the same miss also covers -q before analyze and -v, --isolated, --config X between analyze and graph: 8 forms in all.

Every one of them returns the same 2001 bytes on develop, so this PR leaves those forms as they were rather than breaking them. is_check_invocation (#3927) and is_format (#3928) share the root cause, and the fix you describe, finding the subcommand past ruff's global options with tokenize_grammar, is one finder that all three detectors should read. I opened #4654 for it, credited to your review, and approved this PR for the forms it fixes.