candid-loop（坦诚循环）

Candid Loop

Run candid-review repeatedly until your code is clean. This skill automates the fix-review-fix cycle, applying fixes and re-running reviews until no issues remain (or max iterations is reached).

Workflow

Execute these steps in order:

Step 1: Load Configuration

Load loop configuration from CLI flags and config files.

Precedence (highest to lowest):

1. CLI flags
2. Project config (.candid/config.json → loop field)
3. User config (~/.candid/config.json → loop field)
4. Defaults

Check CLI Arguments

Parse CLI arguments for loop options:

Valid modes:

• auto - Automatically apply all fixes
• review-each - Go through each fix one by one (Yes/No for each)
• interactive - Full control with skip, ignore, and batch options

Valid categories: critical, major, standards, smell, edge_case, architectural, all

If CLI flags are provided, use them and skip config file checks for those specific options.

Check Project Config

Read .candid/config.json and extract the loop field:

jq -r '.loop // null' .candid/config.json 2>/dev/null

If the loop field exists:

• Extract mode if not set by CLI
• Extract maxIterations if not set by CLI
• Extract enforceCategories if not set by CLI
• Extract ignored object (categories, patterns, ids)

Output when loading from config: Using loop settings from project config

Check User Config

Same procedure for ~/.candid/config.json if project config doesn't have loop field.

Output when loading from user config: Using loop settings from user config

Apply Defaults

For any options not set by CLI or config:

mode = "auto"
maxIterations = 5
enforceCategories = ["all"]
ignored = { categories: [], patterns: [], ids: [] }

Step 2: Initialize Loop State

Set up tracking variables:

iteration = 0
totalFixesApplied = 0
allFixedIssues = []
registerQuestionsRaised = 0
registerQuestionsResolved = 0
registerPriorDecisionsApplied = 0

Display mode banner:

Auto mode:

[Auto Mode] Running candid-loop with max [N] iterations...

Review-each mode:

[Review-Each Mode] Running candid-loop with max [N] iterations...
You will review each fix one by one.

Interactive mode:

[Interactive Mode] Running candid-loop with max [N] iterations...
Full control: skip, ignore, or batch process fixes.

Step 3: Run Review Loop

Execute the main loop:

WHILE iteration < maxIterations:
    iteration++

    Execute Step 3.1 through Step 3.6

    IF exitLoop == true:
        BREAK

IF iteration >= maxIterations AND remainingIssues > 0:
    Output warning: "Max iterations ([N]) reached. [M] issues remain."
    List remaining issues

Step 3.1: Run candid-review

Display iteration header:

[Iteration [N]/[MAX]]
Running candid-review...

Invoke the candid-review skill to analyze current code:

• Use the Skill tool to run /candid-review
• Pass through tone settings if configured
• Wait for review to complete and save state to .candid/last-review.json

Note: candid-review will present its own fix selection prompt. In auto mode, we need to handle this by selecting "Apply all fixes". In interactive mode, we defer to the user's choices within candid-review.

Decision Register: If the decision register is enabled in config, each candid-review iteration reads and updates the register file independently. This means:

• Questions raised in iteration 1 are recorded in the register
• In iteration 2+, candid-review checks the register before raising the same question again
• If a question was answered in a previous iteration, the answer is reused automatically
• In auto mode, Clarification Needed issues with prior answers from the register are applied without prompting

Step 3.2: Read Review Results

After candid-review completes, read the saved review state:

cat .candid/last-review.json 2>/dev/null

Parse the JSON to extract:

• issues array with all found issues
• Each issue has: id, file, line, category, title, description

If file doesn't exist or is empty:

No review state found. candid-review may not have completed.

Exit with error.

Step 3.3: Filter Issues by Category

If enforceCategories is NOT ["all"]:

Filter the issues array to only include issues matching enforceCategories:

filteredIssues = issues.filter(issue =>
    enforceCategories.includes(issue.category)
)

Category mapping:

• critical → issues with category "critical"
• major → issues with category "major"
• standards → issues with category "standards"
• smell → issues with category "smell"
• edge_case → issues with category "edge_case"
• architectural → issues with category "architectural"

Output: Enforcing categories: [list]. Filtered to [N] issues.

Step 3.4: Filter Out Ignored Issues

Apply the ignored filters from config:

1. Filter by ignored categories:

filteredIssues = filteredIssues.filter(issue =>
    !ignored.categories.includes(issue.category)
)

2. Filter by ignored patterns (regex match on title):

filteredIssues = filteredIssues.filter(issue =>
    !ignored.patterns.some(pattern =>
        new RegExp(pattern, 'i').test(issue.title)
    )
)

3. Filter by ignored IDs:

filteredIssues = filteredIssues.filter(issue =>
    !ignored.ids.includes(issue.id)
)

If any issues were filtered:

Filtered out [N] ignored issues ([M] remaining)

Step 3.5: Check for Completion

If filteredIssues.length === 0:

[Iteration [N]/[MAX]] No issues found!

exitLoop = true

Skip to Step 4 (Summary).

Step 3.6: Handle Issues Based on Mode

If mode == "auto":

Display found issues:

[N/MAX] Found [M] issues. Applying fixes...

The candid-review skill will have already applied fixes if the user selected "Apply all fixes" in its prompt. Since we're in auto mode, we should have configured candid-review to auto-apply.

For each issue that was fixed, log:

      ✓ [icon] Fixed: [title] in [file]:[line]

Track fixes:

totalFixesApplied += appliedCount
allFixedIssues.push(...appliedIssues)

Decision Register in auto mode:

If the decision register is enabled, candid-review handles register consultation during its Step 6 (before raising Clarification Needed questions). In auto mode:

• Prior decisions from the register are applied automatically by candid-review — log: Applied prior decision (#N) for [file]
• New Clarification Needed issues (no prior answer) are NOT auto-applied — they require human input and are recorded as open in the register
• These are treated as "skipped" in the loop — they do not count as remaining issues that block loop completion
• Output: Applied [N] prior decisions, skipped [M] new questions requiring clarification

Track register activity:

registerPriorDecisionsApplied += priorDecisionsCount
registerQuestionsRaised += newQuestionsCount
registerQuestionsResolved += resolvedCount

Where resolvedCount comes from candid-review's Step 10.5 output — it includes questions answered by the user during Phase 8b and auto-resolutions from re-review. Read the register file after each iteration to count newly resolved entries compared to the previous read.

Continue to next iteration.

If mode == "review-each":

Go through each fix one by one with simple Yes/No prompts.

Display found issues:

[N/MAX] Found [M] issues. Reviewing each...

For each issue, use AskUserQuestion:

[1/M] [icon] [title]
File: [file]:[line]
Problem: [description]

Question: "Apply this fix?"

Options:

1. "Yes, apply this fix"
2. "No, skip this fix"

Track user choices:

• If "Yes" → Apply the fix, increment totalFixesApplied
• If "No" → Continue to next issue

After processing all issues, if any fixes were applied, continue to next iteration.

If mode == "interactive":

Full control mode with additional options for skipping, ignoring, and batch processing.

Display found issues:

Found [M] issues:

For each issue, use AskUserQuestion:

[1/M] [icon] [title]
File: [file]:[line]
Problem: [description]

Question: "How would you like to handle this issue?"

Options:

1. "Apply fix" - Apply this fix and continue
2. "Skip" - Skip this issue and continue to next
3. "Add to ignore list" - Add this issue ID to ignored.ids and skip
4. "Skip all remaining" - Exit the loop with remaining issues listed

Track user choices:

• If "Apply fix" → Apply the fix, increment totalFixesApplied
• If "Skip" → Continue to next issue
• If "Add to ignore list" → Add issue.id to ignored.ids in config, continue
• If "Skip all remaining" → Set exitLoop = true, break inner loop

After processing all issues in interactive mode, if any fixes were applied, continue to next iteration.

Step 3.7: Update Ignore List (If Requested)

If user chose "Add to ignore list" for any issues:

1. Read current .candid/config.json (or create if doesn't exist)
2. Add issue IDs to loop.ignored.ids array
3. Write updated config back to file

# Read, modify, write
jq '.loop.ignored.ids += ["issue-id-1", "issue-id-2"]' .candid/config.json > tmp.json
mv tmp.json .candid/config.json

Output: Added [N] issues to ignore list in.candid/config.json

Step 4: Display Summary

After loop exits (success or max iterations), display final summary:

Success (no issues remaining):

Candid Loop Complete

Summary:
- Iterations: [N]
- Issues fixed: [M]
- Status: PASS

Your code is clean!

Max iterations reached:

Candid Loop Stopped

Summary:
- Iterations: [N] (max reached)
- Issues fixed: [M]
- Issues remaining: [P]
- Status: INCOMPLETE

Remaining issues:
  [icon] [title] in [file]:[line]
  [icon] [title] in [file]:[line]
  ...

Consider:
- Increasing --max-iterations
- Adding persistent ignores for false positives
- Manually reviewing complex issues

User cancelled (interactive mode):

Candid Loop Cancelled

Summary:
- Iterations: [N]
- Issues fixed: [M]
- Issues skipped: [P]
- Status: CANCELLED

Skipped issues:
  [icon] [title] in [file]:[line]
  ...

Decision Register section (add to all summary variants when register is enabled):

If decisionRegister.enabled == true in config, append this section to whichever summary is displayed:

Decision Register:
- Prior decisions applied: [N]
- New questions raised: [M]
- Questions resolved: [P]
- Open questions: [Q]

If open questions remain:

Open questions can be reviewed at: [registerPath]/review-decision-register.md

Configuration

Config File Schema

Add to .candid/config.json:

{
  "version": 1,
  "tone": "harsh",
  "loop": {
    "mode": "auto",
    "maxIterations": 5,
    "enforceCategories": ["all"],
    "ignored": {
      "categories": [],
      "patterns": [],
      "ids": []
    }
  }
}

Field Descriptions

Examples

Ignore all edge cases and architectural issues:

{
  "loop": {
    "ignored": {
      "categories": ["edge_case", "architectural"]
    }
  }
}

Ignore issues mentioning Unicode or timezone:

{
  "loop": {
    "ignored": {
      "patterns": ["Unicode", "timezone", "DST"]
    }
  }
}

Only enforce critical and major issues:

{
  "loop": {
    "enforceCategories": ["critical", "major"]
  }
}

CLI Examples

# Run with defaults (auto mode, all categories, max 5 iterations)
/candid-loop

# Review-each mode - go through each fix one by one (Yes/No)
/candid-loop --mode review-each

# Interactive mode - full control with skip, ignore, batch options
/candid-loop --mode interactive

# Limit iterations
/candid-loop --max-iterations 3

# Only fix critical issues
/candid-loop --categories critical

# Fix critical and major issues
/candid-loop --categories critical,major

# Combine options
/candid-loop --mode review-each --max-iterations 10 --categories critical,major

Issue Categories Reference

Remember

The goal of candid-loop is to automate the review-fix cycle so you can quickly get your code to a clean state.

Mode selection:

• auto - Fast iteration, applies all fixes automatically
• review-each - Go through each fix one by one with simple Yes/No
• interactive - Full control with skip, ignore list, and batch options

Best practices:

• Start with auto mode for quick cleanup
• Use review-each mode to understand what's being fixed
• Use interactive mode when you need to ignore or skip specific issues
• Add patterns to ignore list for known false positives
• Keep maxIterations reasonable (5-10) to avoid infinite loops
• Review the summary to understand what was fixed