fix(guards): read the command from tool_input, not the payload root - #80
Merged
Conversation
Both PreToolUse guards parse the hook payload with `data.get("command")`.
Claude Code delivers the command nested as `tool_input.command`, so every real
invocation fell through to exit 0: guard-lightweight-tag blocked no lightweight
tag and guard-gh-release blocked no `gh release create` in the harness they are
installed into. Every other hook in a standard install reads
`(payload.get("tool_input") or {}).get("command")`.
Found while installing both guards after a release was cut without them
(2026-08-09). Wiring them changed nothing, which is what exposed the parse.
parse_command now prefers tool_input.command and falls back to a top-level
`command`, so direct invocation and other harnesses keep working. A non-dict
payload yields "" instead of raising.
Adds tests/guard-payload-shape.test.sh: 13 assertions over both payload
shapes, both guards, the allow paths (signed tag, annotated tag, notes-file
edit, read-only view) and the malformed-input cases. Green against the fix,
red on exactly the two nested blocking cases against the previous scripts.
Signed-off-by: Sebastian Mendel <info@sebastianmendel.de>
Dependency Review✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.Scanned FilesNone |
SonarCloud flagged 12 code smells and 33.8% duplicated lines on the previous shape: thirteen near-identical check() calls, each passing positional parameters straight through. The assertions and their expected exit codes are unchanged; they now come from a single case table read by one loop. Signed-off-by: Sebastian Mendel <info@sebastianmendel.de>
Both PreToolUse guards are single-file scripts by design: the install recipe copies one of them into ~/.claude/hooks/ and wires that path in settings.json, so a shared helper module would break every single-file copy. The payload parsing is therefore identical in both — 13 lines each, enough to fail new_duplicated_lines_density on any PR touching them (29.5% on this one, with 0 open issues). Scope is duplication detection on those two files only; every other rule keeps applying, and the test file that carried the real smells was restructured rather than excluded. Signed-off-by: Sebastian Mendel <info@sebastianmendel.de>
sonar-project.properties is scanner-only. This project runs SonarCloud Automatic Analysis (sonar.autoscan.enabled=true, verified via the settings API), which reads .sonarcloud.properties — the previous commit's file was never picked up and the API confirmed no cpd exclusion was in effect. projectKey/organization dropped: autoscan supplies both, and neither was read from the file. Signed-off-by: Sebastian Mendel <info@sebastianmendel.de>
|
Member
Author
|
Bot-Review nicht verfügbar — von Hand geprüft und darauf entschieden. Copilot antwortet mit der Quota-Meldung. Die fünf Pflicht-Checks sind grün; Ich bin Autor dieses PRs — keine unabhängige Review. Geprüft:
Zum Sonar-Verlauf, weil zwei Anläufe nötig waren: die erste Fassung des Tests brachte 12 Code Smells und 33,8 % Duplikation, beides aus dreizehn fast gleichen |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Both PreToolUse guards this skill ships are no-ops in Claude Code. They read the command from the payload root; Claude Code nests it under
tool_input. Installing them therefore changes nothing — no lightweight tag is blocked, nogh release createis blocked.Came from
/retrosession on 2026-08-09/10. Found while installing the guards after a release had been cut without them.Symptom: After wiring both guards into
~/.claude/settings.json,git tag v0.28.0(lightweight) andgh release create v9.9.9were both allowed.Cause:
parse_command()doesdata.get("command", ""). The Claude Code PreToolUse payload is{"tool_name": "Bash", "tool_input": {"command": "..."}}, so the lookup misses and the guard exits 0 on every invocation. Measured, same guard, same command, two payload shapes:guard-lightweight-tag{"command": "git tag v0.28.0"}guard-lightweight-tag{"tool_input": {"command": "git tag v0.28.0"}}guard-gh-release{"command": "gh release create v9"}guard-gh-release{"tool_input": {"command": "gh release create v9"}}Every other hook in a standard install reads
(payload.get("tool_input") or {}).get("command").Required behavior: Prefer
tool_input.command, keep the flat shape working for direct invocation and tests.Verification:
scripts/tests/guard-payload-shape.test.sh.Change
parse_command()in both guards: readtool_input.commandfirst, fall back to a top-levelcommand, and return""for a non-dict payload instead of raising. No change to what either guard blocks — only to whether it ever sees the command.New
scripts/tests/guard-payload-shape.test.sh— 13 assertions covering both payload shapes on both guards, the allow paths (git tag -s,git tag -a,gh release edit --notes-file,gh release view) and the malformed-input cases (empty stdin, non-JSON) which must never block.Test plan
--notes-fileedit, read-only view all exit 0SonarCloud
Quality Gate rot, einzige verletzte Bedingung
new_duplicated_lines_density29.5 % > 3. Offene Issues: 0 (die 12 Code Smells der ersten Fassung sind mit dem Umbau des Tests auf eine Falltabelle weg, die Testdatei zählt jetzt 0 duplizierte Zeilen).Die verbleibenden 26 duplizierten Zeilen sind das identische
parse_commandin beiden Guards, 13 Zeilen je Datei — also exakt der Fix, den dieser PR an derselben Stelle in zwei Dateien anbringt:guard-gh-release.pyguard-lightweight-tag.pytests/guard-payload-shape.test.shEin gemeinsames Modul wäre hier falsch: beide Guards werden einzeln nach
~/.claude/hooks/kopiert (so steht es in der Installationsanweisung und so macht essettings.json). Ein Import würde jede Einzelkopie brechen.Damit sind die drei Bedingungen aus
git-workflow/references/merge-gate-watcher.md§ "Sonar gate introspection" erfüllt: einzige rote Bedingung ist eine Zurechnungsmetrik auf berührten Zeilen, offene Issues sind 0, und die Begründung steht hier.