Skip to content

feat(review): read review-digest config from the config-cascade layers - #18082

Open
melodic-automation-lanes[bot] wants to merge 3 commits into
mainfrom
work-resolve-issue/issue-18006-44964496235-1
Open

melodic-automation-lanes[bot] wants to merge 3 commits into
mainfrom
work-resolve-issue/issue-18006-44964496235-1

Conversation

@melodic-automation-lanes

@melodic-automation-lanes melodic-automation-lanes Bot commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

Refs: #18006

Summary

Opened by the work-resolve-issue lane. The run's own summary:

The review-digest resolver now reads the config-cascade layers: six new review plugin options (digest_policy, digest_max_files, digest_max_changed_lines, digest_blast_radius, digest_risk_paths, digest_opt_in_label) as the userConfig layer, the team docs/conventions/review-digest.yaml, and the personal docs/conventions/review-digest.local.yaml. Under the issue's placement rule every key in the old user-global file is a per-user value an option can hold, because none says what a repository's commits, PRs or files must look like. The two list keys are JSON-list string options, a pattern this plugin's manifest already uses (council_panel, loosened_check_patterns). I chose that over multiple: true because the manifest reference does not say how a multiple value renders when substituted into skill text. During the dual-read window, ~/.claude/review-digest.json and .claude/review-digest.local.json are still read, each just below the layer that replaces it, so the new layer wins on each key it sets and the old file fills the rest. Each run prints one warning per old file that names /review:setup apply. That setup gains a digest-migrate.mjs script. Its check names each old file and where each setting goes. Its apply moves the personal file into the YAML after a preview and a yes, verifies that the moved values resolve the same and that the file is untracked and gitignored, then deletes the old file. For the per-user file it lists the options still unset (setup never writes user settings) and deletes the file after a yes once every key is covered.

Remaining work

  • Remove the read of ~/.claude/review-digest.json and .claude/review-digest.local.json (and the .claude path in CONFIG_PATHS) in a later change, once no old file is reported.

Not verified

  • The node test suite (plugins/review/tests/explain-change.test.mjs) was not executed: this lane runs no interpreter. The new and changed tests were walked through the code by hand and by the review subagent. Left to the PR's CI.
  • The README options table was written by hand to match scripts/sync-plugin-options-docs.py output; the generator's --check was not run here and runs in CI.
  • Tidy skipped: the change is 13 files and about 400 changed lines, over the 10-file and 300-line cutoffs, and it touches docs/conventions/.
  • check-affected could not run comment-hygiene (fetch denied) or size-check (script missing in the lane); comment-lint is pull-request only.
  • The migration fixes made after the review subagent's report were not re-reviewed by a second subagent.
  • A userConfig option cannot hold an empty string (empty reads as unset), so an old user-global opt_in_label: "" has no option home. Setup reports it and tells the operator to set it in docs/conventions/review-digest.local.yaml per repository and remove it from the old file, and refuses to delete the…

Fix

  • plugins/review/lib/review-digest.mjs: resolveDigestConfig takes the digest options. New layer order: old user-global, userConfig, team, old personal, personal YAML. One retired-file warning per old file, naming /review:setup apply and any new layer that wins over it (only options holding a valid value are named as winning). Exports userConfigLayer,…
  • plugins/review/skills/explain-change/scripts/digest-policy.mjs: accepts repeated --setting <option>=<value> and passes the values to the resolver.
  • plugins/review/skills/reviewer-media/scripts/decide.mjs: the same --setting flag for the review-digest thresholds it reads.
  • plugins/review/skills/setup/scripts/digest-migrate.mjs (new): check, personal and user actions for moving the old digest files. Before reading, counting or writing docs/conventions/review-digest.local.yaml, personal refuses a target the digest would not read (tracked, not gitignored, or a symlink at it or a parent), so nothing is written through a symlink and the old file is never deleted in favour of a YAML file the resolver ignores. Exit codes: 0 done, 1 refusal, 2 usage, 3 needs --yes, 4 options to set.
  • plugins/review/.claude-plugin/plugin.json: six new digest_* userConfig options with no default, so an unset option stays an unsubstituted placeholder and reads as unset.
  • plugins/review/README.md: the generated options table gains the six rows, written by hand in the generator's format.
  • plugins/review/skills/explain-change/SKILL.md, plugins/review/skills/reviewer-media/SKILL.md: substitute the digest options and pass each set one as --setting, keeping the whole argument in single quotes with each ' in a value written as '''.
  • plugins/review/skills/setup/SKILL.md: new Retired digest files section (check row, apply steps, the same quoting rule, and the refusal of a target the digest would not read); the description and the what-it-does-not-do list name the move.
  • plugins/review/tests/explain-change.test.mjs: new tests for the new layers only, empty or placeholder options, an invalid option, old files only (values apply, one warning each), both present (new wins per key, warnings name both), a malformed old file, and an unknown --setting. The two existing assertions that now meet retired-file warnings match those lines exactly. Migration tests cover a tracked YAML file that sets every old key and a dangling symlink at the YAML path, each refused with the old file kept, beside the earlier migration tests.
  • docs/conventions/review-digest.md: the layers list now names userConfig, team and personal YAML, plus the dual-read window and the setup move.
  • docs/contracts/config-cascade/README.md: the review-digest Implementers rows name the new layers and the dual-read window.
  • .changes/review/work-resolve-issue-issue-18006-44964496235-1-f19b0115.md: minor changelog fragment (Added, Deprecated).

Verification

The runner ran each check from the base copy against 502e038:

  • scripts/check-changelog-fragments.sh --check: passed
  • scripts/check-changed-skills.sh --merge-base: passed

The run reports:

  • read https: //code.claude.com/docs/en/plugins/manifest-reference: userConfig types are string, number, boolean, directory and file; multiple allows a string array; options needs a default or required; values reach skill content only through ${user_config.KEY} substitution, and plugin option…
  • check-changelog-fragments: all 3 fragments valid
  • check-changed-skills: explain-change, reviewer-media and setup PASS, 0 errors (warnings only: body growth, setup body length, setup description growth, no Gotchas, no 'Use when:')
  • check-affected: first run failed validate-plugins on the new option descriptions (over 300 characters, one with **). Fixed; the rerun shows 14 passed, 9 skipped, 0 failed
  • cant-fail-scan --file plugins/review/tests/explain-change.test.mjs --check: PASS, 112 test blocks, no gating findings
  • fresh-context review subagent over the whole diff. It found an existing test the change broke (line 333), a migration that would move a tracked or symlinked old personal file the resolver refuses, a verify step that compared against old values instead of the winning layer, a written YAML left…
  • The lane run: https://github.com/melodic-software/claude-code-plugins/actions/runs/44964496235

Left out: the items under Not verified.

Related

#18006

Lane-Activity: work-resolve-issue/resolve-issue
@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-refine/fix-docs

Run 44966576871, attempt 1

Field Value
Model rung 0
Leg handoff
Run job success
Activity step success
Commits pushed none (read lane)
Run 44966576871

Failed GitHub API calls

No failed call recorded by the run job's scripts.

@github-actions github-actions Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

claude-security-review: 7 inline finding(s) on 502e0382a6fc5a0ac4d5753d43a4484df73fd568; the summary is in this lane's pull request comment.

claude-lane: claude-security-review

Comment on lines +139 to +144
"digest_max_files": {
"type": "number",
"title": "Digest max files",
"description": "The digest files trigger fires when a pull request changes more files than this; shipped default 5. Empty (default): unset. The docs/conventions/review-digest YAML files win over this option.",
"min": 0
},

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

manifest userConfig.digest_max_files: flag, new number option; its value reaches digest-policy.mjs, decide.mjs and digest-migrate.mjs through a --setting shell argument, so it is not inert

It has no default, so an unset option widens nothing. userConfigLayer accepts only ^(0|[1-9]\d*)$ as a count. Anything else stays a string and is rejected as invalid.

severity: IMPORTANT

claude-lane: claude-security-review

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: plugins/review/.claude-plugin/plugin.json:139, plugins/review/lib/review-digest.mjs:131

This flag states facts and asks for no change, so it goes to a person to accept. What the code shows: the option is type number with min 0 and no default, so an unset option adds no layer; userConfigLayer turns the value into a number only when it matches ^(0|[1-9]\d*)$, and anything else stays a string that isCount rejects with an 'invalid max_files; key ignored' warning. A number value holds no shell quote character, and the skill text now also says to escape any ' inside the single-quoted --setting argument.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Accepted by the PR lane session. Exposed surface today: none outside the operator. The value comes only from the operator's own plugin settings, the script validates it after parsing, and a quoting slip breaks only the operator's own command. No permission, trust boundary or cost changes. Condition: if automation or untrusted input ever fills these options, pass the settings through a file or stdin instead of a shell argument.

Comment on lines +145 to +150
"digest_max_changed_lines": {
"type": "number",
"title": "Digest max changed lines",
"description": "The digest changed-lines trigger fires when additions plus deletions exceed this; shipped default 200. Empty (default): unset. The docs/conventions/review-digest YAML files win over this option.",
"min": 0
},

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

manifest userConfig.digest_max_changed_lines: flag, new number option; its value reaches digest-policy.mjs, decide.mjs and digest-migrate.mjs through a --setting shell argument, so it is not inert

It has no default, so an unset option widens nothing. It gets the same whole-number check in userConfigLayer as digest_max_files.

severity: IMPORTANT

claude-lane: claude-security-review

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: plugins/review/.claude-plugin/plugin.json:145, plugins/review/lib/review-digest.mjs:131

This flag states facts and asks for no change, so it goes to a person to accept. What the code shows: the option is type number with min 0 and no default, so an unset option adds no layer; it gets the same whole-number check in userConfigLayer as digest_max_files, and a value that fails it is rejected by isCount with an 'invalid max_changed_lines; key ignored' warning. A number value holds no shell quote character, and the skill text now also says to escape any ' inside the single-quoted --setting argument.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Accepted by the PR lane session. Exposed surface today: none outside the operator. The value comes only from the operator's own plugin settings, the script validates it after parsing, and a quoting slip breaks only the operator's own command. No permission, trust boundary or cost changes. Condition: if automation or untrusted input ever fills these options, pass the settings through a file or stdin instead of a shell argument.

@github-actions

github-actions Bot commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

claude-security-review reviewed c4bf31a20ab84a27ef5598a75d2abaad3d9d4baf: read 13 of 13 changed file(s); 6 finding(s), 4 as inline comments.

Security review: PR #18082 (review-digest settings read from the config-cascade layers)

I read all 13 changed files in .claude-lane/pr.diff. This lane already reviewed an earlier head (502e0382), so this review reports only CRITICAL and IMPORTANT findings (blocking only).

No CRITICAL finding and no hold. The change touches no trust control: no trusted-actor list, no token request or broker rule, no kill switch, no merge hold, no .github/actions/ gate, and no workflow.

What I checked at this head

  • The fixes for the earlier findings hold. personalPlan now calls personalOverrideApplies(root, to, …, { requireIgnored: true }) before it reads, counts or writes docs/conventions/review-digest.local.yaml. That check runs lstat on the file and on every parent directory and refuses a tracked file, a tracked parent or submodule, a symlink, and a file that is not gitignored (view-settings.mjs L83-L113). So the migration no longer writes through a dangling symlink, and it no longer deletes the old file in favour of a YAML file the digest ignores.
  • Resolver trust boundary. The team layer is still read from the base commit through baseReader. The new personal YAML is read only after personalOverrideApplies passes, so a pull request that commits docs/conventions/review-digest.local.yaml is ignored as tracked. addSetting accepts only the six digest_* names, so --setting cannot inject medium or any other key. The test "a --setting naming no digest option exits 2" covers this.
  • Deletes. digest-migrate.mjs deletes nothing without --yes (it exits 3 instead). rmSync(from) runs only on an old personal file that passed personalOverrideApplies, which rules out a symlink.
  • Instruction-surface lens. The setup skill's "What this skill does not do" list now allows writing docs/conventions/review-digest.local.yaml and deleting the two old digest files. Judged by consequence, this is not a finding: the skill keeps disable-model-invocation: true, every delete needs --yes after a preview, and the moved values are read back and verified first. I did not consult the protected-class list in docs/contracts/instruction-exception-register/README.md. The file is present, but I checked this one change by consequence only.

Findings

Six manifest userConfig.digest_*: flag lines, all IMPORTANT. Six new userConfig keys are added to an existing plugin. None is a selector: four are free-form strings with no options list, and two are numbers. None is inert, because the model puts each value into a --setting '<name>=<value>' shell argument. None has a default, so none widens anything when unset, and none installs or runs code on update.

  • Four are posted inline again: digest_policy, digest_blast_radius, digest_risk_paths and digest_opt_in_label. Their earlier threads were settled by resolved alone, with no owner decision, so the flag lines stand. The quoting fix in 65d93a1 narrows the exploit path but leaves the escaping to the model.
  • Two are already posted and unchanged, so they are not posted again.

Spend: none. The change adds no lane runs, models or billed resources.

critical-findings: 0

Manifest lines (6)
  • manifest userConfig.digest_policy: flag, new free-form string option (no options list); the model puts its value into a single-quoted --setting shell argument
  • manifest userConfig.digest_max_files: already posted, feat(review): read review-digest config from the config-cascade layers #18082 (comment)
  • manifest userConfig.digest_max_changed_lines: already posted, feat(review): read review-digest config from the config-cascade layers #18082 (comment)
  • manifest userConfig.digest_blast_radius: flag, new free-form JSON string option that the model puts into a single-quoted --setting shell argument
  • manifest userConfig.digest_risk_paths: flag, new free-form JSON string option (path globs) that the model puts into a single-quoted --setting shell argument
  • manifest userConfig.digest_opt_in_label: flag, new free-form string option that the model puts into a single-quoted --setting shell argument

@github-actions github-actions Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

claude-review: 3 inline finding(s) on 502e0382a6fc5a0ac4d5753d43a4484df73fd568; the summary is in this lane's pull request comment.

claude-lane: claude-review

Comment thread plugins/review/lib/review-digest.mjs Resolved
@github-actions

github-actions Bot commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

claude-review reviewed c4bf31a20ab84a27ef5598a75d2abaad3d9d4baf: read 13 of 13 changed file(s); 0 finding(s), 0 as inline comments.

Code review: review-digest config-cascade layers (#18082), head c4bf31a2

Blocking only. This lane already reviewed an earlier head (502e0382), so only blocking findings were in scope this round. Advisory findings were dropped.

No blocking findings. I read all 13 changed files in full. The three findings from the earlier round are fixed at this head:

  • Old personal file deleted although its replacement is never read. personalPlan now runs personalOverrideApplies(..., { requireIgnored: true }) on docs/conventions/review-digest.local.yaml before it reads, counts or writes anything. Both delete paths therefore refuse a tracked or symlinked target, or one that is not gitignored. A test covers each of these refusals: a tracked YAML file and a dangling symlink.
  • Warning assertions widened to match the code. The two assertions in plugins/review/tests/explain-change.test.mjs (lines 301 and 335) are exact again. They check the count and match each retired-file line in full.
  • Warning names invalid options as winning. The retired user-global warning now names only options whose value passes isValidSetting. This matches how digest-migrate.mjs counts an option as set.

The settled threads about quoting --setting values in the three SKILL.md files are unchanged at this head and stay settled.

What the change does now. The layer order in resolveDigestConfig is: old user-global, userConfig, team, old personal, personal YAML. That matches the convention doc and the config-cascade row. --setting parsing goes through the shared addSetting, and a name outside the digest options exits 2.

Closing issue. The PR body says Refs: #18006 and uses no closing keyword. The PR context has no closing-issue section, so I checked no acceptance criteria.

Not verified. The author says the node test suite was not run and was left to CI. I ran nothing either, because this lane has no shell. I traced the new tests through the code by hand only.

Cost. No billing surface is touched. The change spends nothing beyond normal lane runs.

Paved paths. The standards index has no paved-path row, so I checked nothing against one.

Loosened checks. The prompt has no loosened_check_verdict:, loosened_check_patterns: or loosened_check_decision_record: line. I therefore used fail with no consumer patterns and did not check the owner-decision citation requirement. Neither changed assertion is loosened at this head.
Loosened checks would be marked: fail (default; no verdict line in the prompt); patterns: 0

Scope. A security lane runs on this repository, so security findings are left to it.

Advisory and manifest lines (3)
  • manifest userConfig (the six new digest_* fields): pass. Each new key is read through ${user_config.digest_*} in the explain-change, reviewer-media and setup SKILL.md files. None has a default, on purpose: an unset option falls back to the lower layers and to DEFAULTS.
  • manifest version: pass. The changelog fragment's bump: minor matches the added options plus a deprecation.
  • manifest description/keywords: pass. No capability is removed, and the description does not list the new options.

Verifier dropped 0 of 0 candidates.

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

The unattended merge path is about to arm native auto-merge at head 502e0382a6fc5a0ac4d5753d43a4484df73fd568, the head its arming checks judged. The arm fails if the head moves first.

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

Merge-Risk: max=HIGH demote=true head=502e0382a6fc5a0ac4d5753d43a4484df73fd568 rubric=10.4.6 rater=claude-opus-5-5 checker=claude-opus-5-5

@github-actions

github-actions Bot commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/address-feedback

Run 44970097612, attempt 1

Field Value
Model rung 0
Leg handoff
Run job success
Activity step success
Commits pushed none (read lane)
Run 44970097612

Failed GitHub API calls

No failed call recorded by the run job's scripts.

Lane-Activity: pr-unblock/address-feedback
Lane-Thread: PRRT_kwDOTCGFQM66eNQy
Lane-Thread: PRRT_kwDOTCGFQM66eNQ4
Lane-Thread: PRRT_kwDOTCGFQM66eNQ7
Lane-Thread: PRRT_kwDOTCGFQM66eNQ_
Lane-Thread: PRRT_kwDOTCGFQM66eNRD
Lane-Thread: PRRT_kwDOTCGFQM66eOLu
Lane-Thread: PRRT_kwDOTCGFQM66eOLy
Lane-Thread: PRRT_kwDOTCGFQM66eOL0
@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/apply-feedback-patch

Run 44967575837, attempt 1

Field Value
Model rung 0
Leg base
Run job success
Activity step success
Commits pushed 65d93a1
Run 44967575837

Failed GitHub API calls

No failed call recorded by the run job's scripts.

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

pr-unblock / address-feedback finished round 1 of max 4 on 65d93a1.

@github-actions

github-actions Bot commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/publish-feedback

Run 44970097612, attempt 1

Field Value
Model rung 0
Leg base
Run job success
Activity step success
Commits pushed none recorded
Run 44970097612

Failed GitHub API calls

No failed call recorded by the run job's scripts.

@melodic-automation-lanes melodic-automation-lanes Bot added the needs-session A pr-unblock stop a session builder takes: a change the lane cannot make. Not an owner stop. label Nov 22, 2026
@melodic-automation-lanes

melodic-automation-lanes Bot commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor Author

The pr-unblock lane stopped address-feedback on this pull request: the lane left a review finding open and no adjudicator verdict settled it: none was recorded, it found the finding does not hold where the lane had not disputed it (an UNCERTAIN or VALID-defer reply), which only an INCORRECT reply lets a verdict resolve, it found the finding asks only for tests whose covering check runs had not all passed when the round began, or it raised a security question on a finding the security-review lane rated below CRITICAL; so a session builder takes it. The lane labels it needs-session for a session builder, who makes the change this lane cannot, and the run ends in a failing pr-unblock/address-feedback status on the head the lane gated.

When a retry on this head can change the outcome, the lane runs one on its own, up to its per-head cap, and edits this comment if it stops again; a run that does not stop clears the status. When no retry can (a review thread the lane does not take or settle, or a failing check after the retries are spent), a person resolves the threads or gets the checks green, then dispatches the workflow below: that run finds the cause gone and clears the status. On a pull request the lanes App opened, the dispatchers start that run once per stop; on any other, a person does. A new commit to the branch starts clean.

To run it again on this head now, dispatch the pr-unblock workflow:

gh workflow run pr-unblock.yml -R melodic-software/claude-code-plugins -f pr-number=18082 -f activity=address-feedback

@github-actions

github-actions Bot commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/report-escalation

Run 44970097612, attempt 1

Field Value
Model rung 0
Leg base
Run job success
Activity step success
Commits pushed none recorded
Run 44970097612

Failed GitHub API calls

No failed call recorded by the run job's scripts.

@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/resolve-conflicts

Run 44969251244, attempt 1

Field Value
Model rung 0
Leg handoff
Run job success
Activity step success
Commits pushed none (read lane)
Run 44969251244

Failed GitHub API calls

No failed call recorded by the run job's scripts.

Lane-Activity: pr-unblock/resolve-conflicts
@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-unblock/apply-conflict-patch

Run 44969251244, attempt 1

Field Value
Model rung 0
Leg base
Run job success
Activity step success
Commits pushed c4bf31a
Run 44969251244

Failed GitHub API calls

No failed call recorded by the run job's scripts.

@github-actions github-actions Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

claude-security-review: 4 inline finding(s) on c4bf31a20ab84a27ef5598a75d2abaad3d9d4baf; the summary is in this lane's pull request comment.

claude-lane: claude-security-review

Comment on lines +134 to +138
"digest_policy": {
"type": "string",
"title": "Digest policy",
"description": "When /review:explain-change offers a change digest: off, offer (the shipped default: when a trigger fires) or always (build at the ready flip). Empty (default): unset. The docs/conventions/review-digest YAML files win over this option."
},

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

manifest userConfig.digest_policy: flag, new free-form string option (no options list); the model puts its value into a single-quoted --setting shell argument in explain-change and setup

This is still not a selector, because it has no options list. It is not inert either, because the value reaches a command the session runs. Since the last review, the skill text tells the model to write each ' as '\'' (explain-change/SKILL.md, setup/SKILL.md). That escaping is still left to the model, though, and digest-policy.mjs checks the value against off/offer/always only after the shell has parsed the command. The value comes from the user's own settings, so the impact is self-inflicted. The key has no default, so an unset option widens nothing. The earlier thread was resolved by a bot, not answered by an owner, so this flag line stays.

severity: IMPORTANT

claude-lane: claude-security-review

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: plugins/review/.claude-plugin/plugin.json:134, plugins/review/skills/explain-change/SKILL.md:29, plugins/review/skills/setup/SKILL.md:65, plugins/review/skills/explain-change/scripts/digest-policy.mjs:93

The flag is accurate as written. digest_policy is a free-form string with no options list. The skill texts tell the model to place the value inside a single-quoted --setting argument and to escape each single quote as the four-character sequence quote-backslash-quote-quote. digest-policy.mjs checks the value against off, offer and always only after the shell has parsed the command. The quoting therefore depends on the model following the skill text, and reading the code cannot settle whether that is acceptable. The value comes from the user's own settings, so any impact is self-inflicted. The reviewer asks for an owner's answer rather than another bot resolution. Two possible fixes are adding an options list (off, offer, always) to this key, or passing the settings through a file or stdin instead of shell arguments. Either is a design choice for the owner, and the second spans three skills and their scripts. Leaving this open for a person.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Accepted by the PR lane session. Exposed surface today: none outside the operator. The value comes only from the operator's own plugin settings, the script validates it after parsing, and a quoting slip breaks only the operator's own command. No permission, trust boundary or cost changes. Condition: if automation or untrusted input ever fills these options, pass the settings through a file or stdin instead of a shell argument.

Comment on lines +151 to +155
"digest_blast_radius": {
"type": "string",
"title": "Digest blast radius (JSON)",
"description": "Assessed blast radius levels that fire the digest offer, as a JSON list of LOW, MEDIUM, HIGH and CRITICAL; shipped default [\"HIGH\",\"CRITICAL\"]. Empty (default): unset. The docs/conventions/review-digest YAML files win over this option."
},

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

manifest userConfig.digest_blast_radius: flag, new free-form JSON string option that the model puts into a single-quoted --setting shell argument

The key has no default. userConfigLayer parses the value as JSON, and the VALID check rejects anything outside LOW/MEDIUM/HIGH/CRITICAL (plugins/review/lib/review-digest.mjs). Both checks run only after the shell has parsed the command the model wrote. The new '\'' rule in the three skill texts narrows the quoting gap, but the model still has to apply it. The value comes from the user's own settings. The earlier thread was resolved by a bot, not answered by an owner, so this flag line stays.

severity: IMPORTANT

claude-lane: claude-security-review

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: plugins/review/.claude-plugin/plugin.json:151, plugins/review/skills/explain-change/SKILL.md:29, plugins/review/skills/reviewer-media/SKILL.md:43, plugins/review/lib/review-digest.mjs:143

The flag is accurate as written. digest_blast_radius is a free-form JSON string. review-digest.mjs parses it and rejects any level outside LOW, MEDIUM, HIGH and CRITICAL, but only after the shell has parsed the single-quoted --setting argument the model writes. Safe quoting relies on the model applying the escape rule in the skill texts. The value comes from the user's own settings. This is a security flag that reading the code cannot settle, and the reviewer asks for an owner's answer, not a bot resolution. One fix is passing the settings through a file or stdin instead of shell arguments, which spans three skills and their scripts. That is a design decision for the owner, so I'm leaving this open for a person.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Accepted by the PR lane session. Exposed surface today: none outside the operator. The value comes only from the operator's own plugin settings, the script validates it after parsing, and a quoting slip breaks only the operator's own command. No permission, trust boundary or cost changes. Condition: if automation or untrusted input ever fills these options, pass the settings through a file or stdin instead of a shell argument.

Comment on lines +156 to +160
"digest_risk_paths": {
"type": "string",
"title": "Digest risk paths (JSON)",
"description": "Path globs that fire the digest offer when a changed path matches, as a JSON list of at most 64 globs of at most 256 characters each, e.g. [\"db/migrations/*\"]. Empty (default): unset. The docs/conventions/review-digest YAML files win over this option."
},

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

manifest userConfig.digest_risk_paths: flag, new free-form JSON string option (path globs) that the model puts into a single-quoted --setting shell argument

The key has no default. Inside the script, isGlobs allows at most 64 globs of at most 256 characters each. The value is still free text that the model places inside '...' in a bash command. A glob such as ["it's/**"] relies on the model applying the new '\'' rule. The value comes from the user's own settings, so the impact is self-inflicted. The earlier thread was resolved by a bot, not answered by an owner, so this flag line stays.

severity: IMPORTANT

claude-lane: claude-security-review

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: plugins/review/.claude-plugin/plugin.json:156, plugins/review/skills/explain-change/SKILL.md:29, plugins/review/skills/setup/SKILL.md:65

The flag is accurate as written. digest_risk_paths is a free-form JSON list of globs. The script limits it to at most 64 globs of at most 256 characters each, but the raw value still goes inside a single-quoted bash argument the model writes. A glob containing a single quote is only safe if the model applies the escape rule in the skill texts. The value comes from the user's own settings, so any impact is self-inflicted. This is a security flag that reading the code cannot settle, and the reviewer asks for an owner's answer. Removing the dependence on model-applied quoting, for example by passing settings through a file or stdin, is a cross-skill design change for the owner. Leaving this open for a person.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Accepted by the PR lane session. Exposed surface today: none outside the operator. The value comes only from the operator's own plugin settings, the script validates it after parsing, and a quoting slip breaks only the operator's own command. No permission, trust boundary or cost changes. Condition: if automation or untrusted input ever fills these options, pass the settings through a file or stdin instead of a shell argument.

Comment on lines +161 to +165
"digest_opt_in_label": {
"type": "string",
"title": "Digest opt-in label",
"description": "The pull request label that fires the digest offer; shipped default explain-change. Empty (default): unset, so to turn the label trigger off, set opt_in_label: \"\" in a docs/conventions/review-digest YAML file. Those files win over this option."
},

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

manifest userConfig.digest_opt_in_label: flag, new free-form string option that the model puts into a single-quoted --setting shell argument

The key has no default. The opt_in_label check accepts any string, so the script narrows nothing. A label containing a ' stays safe only if the model applies the '\'' rule added to the skill texts. The value comes from the user's own settings. The earlier thread was resolved by a bot, not answered by an owner, so this flag line stays.

severity: IMPORTANT

claude-lane: claude-security-review

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: plugins/review/.claude-plugin/plugin.json:161, plugins/review/skills/explain-change/SKILL.md:29, plugins/review/skills/setup/SKILL.md:65

The flag is accurate as written. digest_opt_in_label is a free-form string, and the opt_in_label check accepts any string, so the script does not narrow it. A label containing a single quote stays safe only if the model applies the escape rule in the skill texts when it builds the --setting argument. The value comes from the user's own settings. This is a security flag that reading the code cannot settle, and the reviewer asks for an owner's answer, not a bot resolution. Possible fixes are constraining the label's characters in the script, or passing settings through a file or stdin. Both are design decisions for the owner, so I'm leaving this open for a person.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Accepted by the PR lane session. Exposed surface today: none outside the operator. The value comes only from the operator's own plugin settings, the script validates it after parsing, and a quoting slip breaks only the operator's own command. No permission, trust boundary or cost changes. Condition: if automation or untrusted input ever fills these options, pass the settings through a file or stdin instead of a shell argument.

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

pr-unblock / address-feedback finished round 2 of max 4 on c4bf31a.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

needs-session A pr-unblock stop a session builder takes: a change the lane cannot make. Not an owner stop.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant