feat: context-firewall plugin #31
Conversation
|
Note Other AI code review bot(s) detectedCodeRabbit has detected other AI code review bot(s) in this pull request and will avoid duplicating their findings in the review comments. This may lead to a less comprehensive review. 📝 WalkthroughWalkthroughAdds a new "context-firewall" plugin with manifest, documentation, commands, agent specs, schemas, examples, hooks, scripts, and gitignore entries to preprocess large inputs via sub-agents producing strict, auditable JSON results and verification artifacts. Changes
Sequence Diagram(s)mermaid Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes 🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 inconclusive)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing touches🧪 Generate unit tests (beta)
🧹 Recent nitpick comments
📜 Recent review detailsConfiguration used: Organization UI Review profile: CHILL Plan: Pro 📒 Files selected for processing (3)
✅ Files skipped from review due to trivial changes (1)
🧰 Additional context used📓 Path-based instructions (1)**/*.md📄 CodeRabbit inference engine (CLAUDE.md)
Files:
🧠 Learnings (1)📚 Learning: 2026-01-17T04:29:56.872ZApplied to files:
🪛 LanguageToolplugins/context-firewall/commands/cf-run.md[style] ~121-~121: Three successive sentences begin with the same word. Consider rewording the sentence or use a thesaurus to find a synonym. (ENGLISH_WORD_REPEAT_BEGINNING_RULE) 🔇 Additional comments (11)
✏️ Tip: You can disable this entire section by setting Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Fix all issues with AI agents
In `@plugins/context-firewall/agents/cf-fileworker.md`:
- Around line 1-27: The YAML frontmatter in cf-fileworker.md is invalid because
the Examples: block is inside the frontmatter; fix by moving the Examples
section (everything from "Examples:" through the end of examples) below the
closing '---' so only valid YAML keys remain in the frontmatter (retain "name:
cf-fileworker", "model: inherit", "color: cyan"), or alternatively convert the
examples into a YAML block scalar under a single frontmatter key; ensure the
file begins and ends the frontmatter with '---' and that no plain-text example
blocks remain inside it.
In `@plugins/context-firewall/commands/cf-run.md`:
- Around line 27-29: Update the wording in the error-handling description for
extracting `--spec` from `$ARGUMENTS` to match the schema: replace ambiguous
"followup(s)" phrasing with the exact field name "followups" and consolidate
duplicate sentences so the block describing the SubResult.v1 error output reads:
emit a schema-valid SubResult.v1 with status: "tool_error" and a single
"followups" array containing the required usage message. Apply the same
replacement/consolidation to the other identical block that mentions producing
SubResult.v1 with status: "tool_error" (the later occurrence in the file) so
both `--spec` parsing failure descriptions use the exact "followups" field name
and non-redundant wording.
🧹 Nitpick comments (5)
plugins/context-firewall/agents/cf-critic.md (1)
1-18: Add explicitallowed-toolsin frontmatter. Tool access is currently implicit; declare the intended permissions to keep behavior stable and auditable.🔧 Suggested frontmatter addition (adjust list to intended access)
--- name: cf-critic description: Use this agent when a high-risk TaskSpec/SubResult needs an independent second pass to reduce hallucinations, validate coverage, or resolve cross-file causal claims. +allowed-tools: ["Read", "Bash", "Task", "AskUserQuestion", "WebFetch", "WebSearch"] Examples:Based on learnings, please keep tool permissions explicit in agent frontmatter.
plugins/context-firewall/agents/cf-aggregator.md (1)
5-14: Frontmatter key formatting inconsistency.The
Examples:on line 5 appears as a standalone label rather than a proper YAML key-value pair. If this is intended as metadata, consider usingexamples:as a key with the example content as its value. Otherwise, if examples are meant to be part of the description field or the body, move them outside the frontmatter block.plugins/context-firewall/commands/cf-verify.md (1)
146-150: Clarify status when all sampled claims are "partial".The status rules define three states but the interaction with "partial" claims is ambiguous:
okrequiresfailed == 0and all checked claims passedpartialcovers "unverifiable locators only"If all N sampled claims result in "partial (unverifiable)", the count shows
checked=N, passed=0, failed=0. By line 148, this doesn't qualify asok(not all passed). By line 150, this would bepartial. Consider explicitly stating this case in the rules for clarity.plugins/context-firewall/schemas/sub-result.v1.schema.json (2)
1-5: Consider adding$schemadeclaration.The schema is missing a
$schemaproperty to declare which JSON Schema draft it conforms to. This helps validators and tooling apply the correct semantics.Suggested addition
{ + "$schema": "https://json-schema.org/draft/2020-12/schema", "$id": "context-firewall/SubResult.v1", "type": "object",
200-207: Consider constrainingcountsobject structure.The
countsproperty accepts any object shape ("type": "object"with nopropertiesoradditionalPropertiesconstraints). If specific count fields are expected (e.g.,lines_read,files_processed), defining them would improve validation and documentation. If the shape is intentionally freeform, consider adding a brief comment ordescriptionfield.
📜 Review details
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (24)
.gitignoreREADME.mdREADME.zh-CN.mdplugins/context-firewall/.claude-plugin/plugin.jsonplugins/context-firewall/README.mdplugins/context-firewall/README.zh-CN.mdplugins/context-firewall/agents/cf-aggregator.mdplugins/context-firewall/agents/cf-critic.mdplugins/context-firewall/agents/cf-fileworker.mdplugins/context-firewall/commands/cf-run.mdplugins/context-firewall/commands/cf-spec.mdplugins/context-firewall/commands/cf-verify.mdplugins/context-firewall/examples/logs-timeout-task.jsonplugins/context-firewall/examples/repo-entrypoints-task.jsonplugins/context-firewall/hooks/hooks.jsonplugins/context-firewall/schemas/settings.v1.schema.jsonplugins/context-firewall/schemas/sub-result.v1.schema.jsonplugins/context-firewall/schemas/task-spec.v1.schema.jsonplugins/context-firewall/schemas/verify-report.v1.schema.jsonplugins/context-firewall/scripts/.gitkeepplugins/context-firewall/scripts/settings-frontmatter.mdplugins/context-firewall/scripts/test-plan.mdplugins/context-firewall/skills/context-firewall/SKILL.mdplugins/context-firewall/todos.md
🧰 Additional context used
📓 Path-based instructions (1)
**/*.md
📄 CodeRabbit inference engine (CLAUDE.md)
Markdown linting must ensure compliance with
.markdownlint.jsonconfiguration which disables MD013 and MD041 rules
Files:
plugins/context-firewall/scripts/settings-frontmatter.mdplugins/context-firewall/README.mdREADME.mdplugins/context-firewall/scripts/test-plan.mdplugins/context-firewall/commands/cf-spec.mdREADME.zh-CN.mdplugins/context-firewall/agents/cf-fileworker.mdplugins/context-firewall/todos.mdplugins/context-firewall/agents/cf-aggregator.mdplugins/context-firewall/agents/cf-critic.mdplugins/context-firewall/commands/cf-run.mdplugins/context-firewall/skills/context-firewall/SKILL.mdplugins/context-firewall/README.zh-CN.mdplugins/context-firewall/commands/cf-verify.md
🧠 Learnings (4)
📚 Learning: 2026-01-17T04:29:56.872Z
Learnt from: CR
Repo: lollipopkit/cc-plugins PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-17T04:29:56.872Z
Learning: Applies to */agents/*.md : Define behavior primarily via system prompts in Markdown files using YAML frontmatter to specify tool permissions
Applied to files:
plugins/context-firewall/scripts/settings-frontmatter.mdplugins/context-firewall/agents/cf-critic.md
📚 Learning: 2026-01-17T04:29:56.872Z
Learnt from: CR
Repo: lollipopkit/cc-plugins PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-17T04:29:56.872Z
Learning: Plugins should read per-project configuration from the target project's `.claude/<plugin-name>.local.md` file for local settings
Applied to files:
plugins/context-firewall/scripts/settings-frontmatter.md.gitignore
📚 Learning: 2026-01-17T04:29:56.872Z
Learnt from: CR
Repo: lollipopkit/cc-plugins PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-17T04:29:56.872Z
Learning: Applies to */agents/*.md : Ensure plugins have an `agents/` directory containing AI agent definitions as Markdown files with YAML frontmatter
Applied to files:
plugins/context-firewall/agents/cf-fileworker.mdplugins/context-firewall/agents/cf-critic.md
📚 Learning: 2026-01-17T04:29:56.872Z
Learnt from: CR
Repo: lollipopkit/cc-plugins PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-17T04:29:56.872Z
Learning: Applies to */hooks.json : Ensure plugins have a `hooks/` directory with lifecycle hooks defined in `hooks.json`
Applied to files:
plugins/context-firewall/hooks/hooks.json
🪛 LanguageTool
plugins/context-firewall/commands/cf-run.md
[grammar] ~28-~28: Ensure spelling is correct
Context: ....v1** with status: "tool_error" and a followup describing the required usage. - If th...
(QB_NEW_EN_ORTHOGRAPHY_ERROR_IDS_1)
[style] ~121-~121: Three successive sentences begin with the same word. Consider rewording the sentence or use a thesaurus to find a synonym.
Context: ...rative text before or after the JSON. - Do not paste large raw inputs into the out...
(ENGLISH_WORD_REPEAT_BEGINNING_RULE)
🔇 Additional comments (24)
.gitignore (3)
3-7: Good addition for plugin local configuration.The ignore patterns align well with the plugin architecture mentioned in the learnings, ensuring local configuration and state files are not committed.
5-6: Current patterns are correct; plugins use flat.claude/structure for local config files.The patterns
.claude/*.local.mdand.claude/*.local.jsonare sufficient. All plugins follow the documented format of.claude/<plugin-name>.local.md(e.g.,.claude/context-firewall.local.md,.claude/devloop.local.md). Subdirectories under.claude/(such as.claude/context-firewall/) are used for persistence data (task-specs, results), not for local configuration files. No expansion to recursive patterns is needed.Likely an incorrect or invalid review comment.
7-7: The.claude/context-firewall/directory is correctly ignored.This directory is the runtime output location for the context-firewall plugin (as defined in
plugins/context-firewall/schemas/settings.v1.schema.json). User configuration is stored separately in.claude/context-firewall.local.mdand is not affected by this gitignore pattern, allowing per-project settings to be version-controlled as intended.plugins/context-firewall/.claude-plugin/plugin.json (1)
1-19: Manifest looks consistent and complete.
Clear metadata and a dedicated commands entrypoint; no issues spotted.README.md (1)
9-12: Doc update is clear and aligned with the plugin list.
The new bullet is concise and fits the existing format.README.zh-CN.md (1)
9-12: Chinese README update mirrors the English entry well.
Looks consistent and clear.plugins/context-firewall/scripts/test-plan.md (1)
1-18: Test checklist is clear and actionable.
Good coverage of core commands and persistence checks.plugins/context-firewall/scripts/settings-frontmatter.md (1)
1-30: Settings template is well-structured and self-describing.
No issues spotted.plugins/context-firewall/README.md (1)
1-59: Comprehensive README with clear structure.
Covers commands, agents, settings, schemas, and examples succinctly.plugins/context-firewall/schemas/verify-report.v1.schema.json (1)
1-28: Looks solid and strict.
The schema is tight (additionalProperties: false) and the required fields match the intended VerifyReport payload.plugins/context-firewall/examples/logs-timeout-task.json (1)
1-31: Clear, realistic TaskSpec example.
The structure and constraints read well for a large-log workflow.plugins/context-firewall/commands/cf-spec.md (1)
1-48: Doc is clear and operational.
The steps and output rules are concise and easy to follow.plugins/context-firewall/examples/repo-entrypoints-task.json (1)
1-28: Good example for repo entrypoint discovery.
The inputs/questions/must_cover are aligned with the stated objective.plugins/context-firewall/todos.md (1)
1-12: Nice lightweight roadmap list.
No issues spotted in the checklist formatting.plugins/context-firewall/skills/context-firewall/SKILL.md (1)
1-130: Well-structured skill workflow spec. Clear evidence contract, verification steps, and operational defaults; reads consistent with the rest of the plugin docs.plugins/context-firewall/schemas/settings.v1.schema.json (1)
1-42: Schema looks solid. Clear bounds and defaults for warnings, sampling, constraints, and persistence.plugins/context-firewall/README.zh-CN.md (1)
1-59: 内容清晰、结构完整。 与其他文档一致，覆盖了命令、代理、设置与示例。plugins/context-firewall/schemas/task-spec.v1.schema.json (1)
1-174: Schema is comprehensive and strict. Input variants, constraints, and optional policy blocks are well-specified.plugins/context-firewall/agents/cf-aggregator.md (1)
21-57: LGTM!The agent documentation is well-structured with clear input/output specifications, deduplication logic, clustering behavior, and conflict detection rules. The evidence signature formats align correctly with the
SubResult.v1schema definitions.plugins/context-firewall/commands/cf-verify.md (2)
86-92: Verify Read tool offset/limit semantics.The formula
offset = startandlimit = end-start+1assumes the Read tool uses 1-indexed line numbers whereoffsetis the starting line. Ensure this matches the actual Read tool behavior in your environment—some tools use 0-indexed offsets or byte-based positioning.
1-180: LGTM!The command specification is comprehensive, covering argument parsing, claim sampling, locator-specific verification logic, and report generation. The deterministic sampling approach and persistence options are well-documented.
plugins/context-firewall/schemas/sub-result.v1.schema.json (2)
68-77: LGTM!The
Locatordiscriminated union usingoneOfis well-designed, covering all documented locator types (line_range, symbol_range, tool_call, byte_range, json_path, stack_signature). Each variant correctly usesconstfor thetypediscriminator.
116-155: LGTM!The
RerunHintvariants are well-structured with appropriate constraints—web_fetchrequires url/prompt,web_searchrequires query with optional domain filters, andunknownis intentionally extensible withadditionalProperties: truefor forward compatibility.plugins/context-firewall/hooks/hooks.json (1)
1-53: LGTM!The hooks configuration correctly implements advisory lifecycle hooks for PreToolUse, PostToolUse, and UserPromptSubmit. The prompts are well-crafted with clear rules, actionable guidance, and references to local configuration options. The consistent 30-second timeout and advisory-only pattern align with the plugin's non-blocking design philosophy. Based on learnings, this satisfies the requirement for plugins to have a
hooks/directory with lifecycle hooks defined inhooks.json.
✏️ Tip: You can disable this entire section by setting review_details to false in your review settings.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: dd5d19e8ab
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| model: inherit | ||
| color: cyan | ||
| --- |
There was a problem hiding this comment.
The FileWorker agent frontmatter does not declare any tools, yet the instructions require it to use Grep/Glob/Read to scan large inputs. In Claude Code, agents without an explicit tools list cannot access those tools, so /cf-run will spawn FileWorkers that are unable to read or search inputs and will return empty/failed SubResults. Add the necessary tools (at least Read/Grep/Glob) to the frontmatter to make the workflow functional.
Useful? React with 👍 / 👎.
| model: inherit | ||
| color: red | ||
| --- |
There was a problem hiding this comment.
The Critic agent is expected to independently re-check evidence and detect contradictions, but its frontmatter omits a tools list entirely. Without Read/Grep/Glob access, the Critic cannot inspect source inputs, so any requested critic pass would be unable to verify claims. Add at least Read/Grep/Glob (and any other needed tools) to the frontmatter so the critic can perform its stated role.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Fix all issues with AI agents
In `@plugins/context-firewall/agents/cf-critic.md`:
- Around line 31-44: The example JSON in cf-critic.md uses pipe-delimited
literals like "ok|partial|tool_error" which can be copied verbatim and break
JSON; update the example to use a concrete value (e.g., set "status": "ok") and
add an explicit rule line(s) that enumerate allowed values for the "status"
field (allowed: "ok", "partial", "tool_error") and for the
"discrepancies[].type" field (allowed: "missing_coverage", "weak_evidence",
"conflict") so readers see both a valid sample and the permitted alternatives;
target the JSON example and the rule text around the "status" and
"discrepancies.type" fields in cf-critic.md.
In `@plugins/context-firewall/commands/cf-run.md`:
- Around line 30-33: The spec detection currently treats `--spec` as inline JSON
only if it literally starts with '{', which fails when there is leading
whitespace; update the detection logic (the code that inspects the `--spec` flag
value in the cf-run command) to trim leading whitespace before checking for '{'
(e.g. use a trimmed/trimStart string for the starts-with test) while preserving
the original `spec` value when you need to treat it as a file path and pass it
to Read; ensure the branch that calls Read still resolves non-absolute paths
relative to the project root after trimming/inspecting.
🧹 Nitpick comments (1)
plugins/context-firewall/schemas/sub-result.v1.schema.json (1)
57-67: Require some evidentiary payload inEvidence.
As written,Evidencecan omit bothquoteandquote_hash, yielding an evidence entry with no verifiable content. Consider enforcing at least one.♻️ Proposed schema tightening
"Evidence": { "type": "object", "additionalProperties": false, "required": ["source", "locator"], "properties": { "source": { "type": "string", "minLength": 1 }, "locator": { "$ref": "#/$defs/Locator" }, "quote": { "type": "string" }, "quote_hash": { "type": "string", "pattern": "^sha256:" } - } + }, + "anyOf": [ + { "required": ["quote"] }, + { "required": ["quote_hash"] } + ] },
📜 Review details
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (8)
plugins/context-firewall/agents/cf-aggregator.mdplugins/context-firewall/agents/cf-critic.mdplugins/context-firewall/agents/cf-fileworker.mdplugins/context-firewall/commands/cf-run.mdplugins/context-firewall/commands/cf-spec.mdplugins/context-firewall/commands/cf-verify.mdplugins/context-firewall/schemas/sub-result.v1.schema.jsonplugins/context-firewall/scripts/test-plan.md
🚧 Files skipped from review as they are similar to previous changes (4)
- plugins/context-firewall/commands/cf-verify.md
- plugins/context-firewall/agents/cf-aggregator.md
- plugins/context-firewall/agents/cf-fileworker.md
- plugins/context-firewall/scripts/test-plan.md
🧰 Additional context used
📓 Path-based instructions (1)
**/*.md
📄 CodeRabbit inference engine (CLAUDE.md)
Markdown linting must ensure compliance with
.markdownlint.jsonconfiguration which disables MD013 and MD041 rules
Files:
plugins/context-firewall/commands/cf-spec.mdplugins/context-firewall/agents/cf-critic.mdplugins/context-firewall/commands/cf-run.md
🧠 Learnings (2)
📚 Learning: 2026-01-17T04:29:56.872Z
Learnt from: CR
Repo: lollipopkit/cc-plugins PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-17T04:29:56.872Z
Learning: Applies to */agents/*.md : Define behavior primarily via system prompts in Markdown files using YAML frontmatter to specify tool permissions
Applied to files:
plugins/context-firewall/agents/cf-critic.mdplugins/context-firewall/commands/cf-run.md
📚 Learning: 2026-01-17T04:29:56.872Z
Learnt from: CR
Repo: lollipopkit/cc-plugins PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-17T04:29:56.872Z
Learning: Applies to */agents/*.md : Ensure plugins have an `agents/` directory containing AI agent definitions as Markdown files with YAML frontmatter
Applied to files:
plugins/context-firewall/agents/cf-critic.md
🪛 LanguageTool
plugins/context-firewall/commands/cf-run.md
[style] ~121-~121: Three successive sentences begin with the same word. Consider rewording the sentence or use a thesaurus to find a synonym.
Context: ...rative text before or after the JSON. - Do not paste large raw inputs into the out...
(ENGLISH_WORD_REPEAT_BEGINNING_RULE)
🔇 Additional comments (1)
plugins/context-firewall/commands/cf-spec.md (1)
39-54: Clear, enforceable JSON-only contract.
The steps and output rules are concise and align with the schema-driven workflow.
✏️ Tip: You can disable this entire section by setting review_details to false in your review settings.
Fixes #30
Summary by CodeRabbit
New Features
Documentation
Chores
✏️ Tip: You can customize this high-level summary in your review settings.