Numberly Vault DB Injector Vault DB Injector

Desloppify

1. Your Job

Maximise the strict score honestly. Your main cycle: scan → plan → execute → rescan. Follow the scan output's INSTRUCTIONS FOR AGENTS — don't substitute your own analysis.

Don't be lazy. Do large refactors and small detailed fixes with equal energy. If it takes touching 20 files, touch 20 files. If it's a one-line change, make it. No task is too big or too small — fix things properly, not minimally.

2. The Workflow

Three phases, repeated as a cycle.

Monorepos and multi-project directories

If the workspace contains multiple programs (e.g., frontend + backend in sibling folders), scan each one separately — do not scan the parent directory:

desloppify --lang typescript scan --path ./frontend
desloppify --lang python scan --path ./backend

Each --path target should be a single coherent project. Scanning a parent that contains multiple programs mixes state and path context, producing unreliable results.

Phase 1: Scan and review — understand the codebase

desloppify scan --path .       # analyse the codebase
desloppify status              # check scores — are we at target?

After scanning, always run desloppify next — it tells you exactly what to do, in order. Don't interpret the scan output yourself or ask the user what to do. Just run next and follow its instructions.

The scan will tell you if subjective dimensions need review. Follow its instructions. To trigger a review manually:

desloppify review --prepare    # then follow your runner's review workflow

Phase 2: Plan — decide what to work on

After reviews, triage stages and plan creation appear in the execution queue surfaced by next. Complete them in order — next tells you what each stage expects in the --report:

desloppify next                                        # shows the next execution workflow step
desloppify plan triage --stage observe --report "themes and root causes..."
desloppify plan triage --stage reflect --report "comparison against completed work..."
desloppify plan triage --stage organize --report "summary of priorities..."
desloppify plan triage --complete --strategy "execution plan..."

For automated triage: desloppify plan triage --run-stages --runner codex (Codex) or --runner claude (Claude). Options: --only-stages, --dry-run, --stage-timeout-seconds.

Then shape the queue. The plan shapes everything next gives you — next is the execution queue, not the full backlog. Don't skip this step.

desloppify plan                          # see the living plan details
desloppify plan queue                    # compact execution queue view
desloppify plan reorder <pat> top        # reorder — what unblocks the most?
desloppify plan cluster create <name>    # group related issues to batch-fix
desloppify plan focus <cluster>          # scope next to one cluster
desloppify plan skip <pat>              # defer — hide from next

Phase 3: Execute — grind the queue to completion

Trust the plan and execute. Don't rescan mid-queue — finish the queue first.

Branch first. Create a dedicated branch — never commit health work directly to main:

git checkout -b desloppify/code-health    # or desloppify/<focus-area>
desloppify config set commit_pr 42        # link a PR for auto-updated descriptions

The loop:

# 1. Get the next item from the execution queue
desloppify next

# 2. Fix the issue in code

# 3. Resolve it (next shows the exact command including required attestation)

# 4. When you have a logical batch, commit and record
git add <files> && git commit -m "desloppify: fix 3 deferred_import findings"
desloppify plan commit-log record      # moves findings uncommitted → committed, updates PR

# 5. Push periodically
git push -u origin desloppify/code-health

# 6. Repeat until the queue is empty

Score may temporarily drop after fixes — cascade effects are normal, keep going. If next suggests an auto-fixer, run desloppify autofix <fixer> --dry-run to preview, then apply.

When the queue is clear, go back to Phase 1. New issues will surface, cascades will have resolved, priorities will have shifted. This is the cycle.

3. Reference

Key concepts

• Tiers: T1 auto-fix → T2 quick manual → T3 judgment call → T4 major refactor.
• Auto-clusters: related findings are auto-grouped in next. Drill in with next --cluster <name>.
• Zones: production/script (scored), test/config/generated/vendor (not scored). Fix with zone set.
• Wontfix cost: widens the lenient↔strict gap. Challenge past decisions when the gap grows.

Scoring

Overall score = 25% mechanical + 75% subjective.

• Mechanical (25%): auto-detected issues — duplication, dead code, smells, unused imports, security. Fixed by changing code and rescanning.
• Subjective (75%): design quality review — naming, error handling, abstractions, clarity. Starts at 0% until reviewed. The scan will prompt you when a review is needed.
• Strict score is the north star: wontfix items count as open. The gap between overall and strict is your wontfix debt.
• Score types: overall (lenient), strict (wontfix counts), objective (mechanical only), verified (confirmed fixes only).

Reviews

Four paths to get subjective scores:

• Local runner (Codex): desloppify review --run-batches --runner codex --parallel --scan-after-import — automated end-to-end.
• Local runner (Claude): desloppify review --prepare → launch parallel subagents → desloppify review --import merged.json — see skill doc overlay for details.
• Cloud/external: desloppify review --external-start --external-runner claude → follow session template → --external-submit.
• Manual path: desloppify review --prepare → review per dimension → desloppify review --import file.json.

Batch output vs import filenames: Individual batch outputs from subagents must be named batch-N.raw.txt (plain text/JSON content, .raw.txt extension). The .json filenames in --import merged.json or --import findings.json refer to the final merged import file, not individual batch outputs. Do not name batch outputs with a .json extension.

• Import first, fix after — import creates tracked state entries for correlation.
• Target-matching scores trigger auto-reset to prevent gaming. Use the blind-review workflow described in your agent overlay doc (e.g. docs/CLAUDE.md, docs/HERMES.md).
• Even moderate scores (60-80) dramatically improve overall health.
• Stale dimensions auto-surface in next — just follow the queue.

Integrity rules: Score from evidence only — no prior chat context, score history, or target-threshold anchoring. When evidence is mixed, score lower and explain uncertainty. Assess every requested dimension; never drop one.

Review output format

Return machine-readable JSON for review imports. For --external-submit, include session from the generated template:

{
  "session": {
    "id": "<session_id_from_template>",
    "token": "<session_hmac_from_template>"
  },
  "assessments": {
    "<dimension_from_query>": 0
  },
  "findings": [
    {
      "dimension": "<dimension_from_query>",
      "identifier": "short_id",
      "summary": "one-line defect summary",
      "related_files": ["relative/path/to/file.py"],
      "evidence": ["specific code observation"],
      "suggestion": "concrete fix recommendation",
      "confidence": "high|medium|low"
    }
  ]
}

findings MUST match query.system_prompt exactly (including related_files, evidence, and suggestion). Use "findings": [] when no defects found. Import is fail-closed: invalid findings abort unless --allow-partial is passed. Assessment scores are auto-applied from trusted internal or cloud session imports. Legacy --attested-external remains supported.

Import paths

• Robust session flow (recommended): desloppify review --external-start --external-runner claude → use generated prompt/template → run printed --external-submit command.
• Durable scored import (legacy): desloppify review --import findings.json --attested-external --attest "I validated this review was completed without awareness of overall score and is unbiased."
• Findings-only fallback: desloppify review --import findings.json

Reviewer agent prompt

Runners that support agent definitions (Cursor, Copilot, Gemini) can create a dedicated reviewer agent. Use this system prompt:

You are a code quality reviewer. You will be given a codebase path, a set of
dimensions to score, and what each dimension means. Read the code, score each
dimension 0-100 from evidence only, and return JSON in the required format.
Do not anchor to target thresholds. When evidence is mixed, score lower and
explain uncertainty.

See your editor's overlay section below for the agent config format.

Plan commands

desloppify plan reorder <cluster> top       # move all cluster members at once
desloppify plan reorder <a> <b> top        # mix clusters + findings in one reorder
desloppify plan reorder <pat> before -t X  # position relative to another item/cluster
desloppify plan cluster reorder a,b top    # reorder multiple clusters as one block
desloppify plan resolve <pat>              # mark complete
desloppify plan reopen <pat>               # reopen
desloppify backlog                          # broader non-execution backlog

Commit tracking

desloppify plan commit-log                      # see uncommitted + committed status
desloppify plan commit-log record               # record HEAD commit, update PR description
desloppify plan commit-log record --note "why"  # with rationale
desloppify plan commit-log record --only "smells::*"  # record specific findings only
desloppify plan commit-log history              # show commit records
desloppify plan commit-log pr                   # preview PR body markdown
desloppify config set commit_tracking_enabled false  # disable guidance

After resolving findings as fixed, the tool shows uncommitted work, committed history, and a suggested commit message. After committing externally, run record to move findings from uncommitted to committed and auto-update the linked PR description.

Agent directives

Directives are messages shown to agents at lifecycle phase transitions — use them to switch models, set constraints, or give context-specific instructions.

desloppify directives                     # show all configured directives
desloppify directives set execute "Switch to claude-sonnet-4-6. Focus on speed."
desloppify directives set triage "Switch to claude-opus-4-6. Read carefully."
desloppify directives set review "Use blind packet. Do not anchor on previous scores."
desloppify directives unset execute       # remove a directive

Available phases: execute, review, triage, workflow, scan (and fine-grained variants like review_initial, triage_postflight, etc.).

Quick reference

desloppify next --count 5                         # top 5 execution items
desloppify next --cluster <name>                  # drill into a cluster
desloppify backlog --count 5                      # top 5 backlog items outside execution
desloppify show <pattern>                         # filter by file/detector/ID
desloppify show --status open                     # all open findings
desloppify plan skip --permanent "<id>" --note "reason" --attest "..." # accept debt
desloppify exclude <path>                         # exclude a directory from scanning
desloppify config show                            # show all config including excludes
desloppify scan --path . --reset-subjective       # reset subjective baseline to 0

4. Fix Tool Issues Upstream

When desloppify itself appears wrong or inconsistent — a bug, a bad detection, a crash, confusing output — fix it and open a PR. If you can't confidently fix it, file an issue instead.

Fix and PR (preferred)

Clone the tool repo to a temp directory, make the fix there, and verify it works against the project you're scanning before pushing.

git clone https://github.com/peteromallet/desloppify.git /tmp/desloppify-fix
cd /tmp/desloppify-fix
git checkout -b fix/<short-description>

Make your changes, then run the test suite and verify the fix against the original project:

python -m pytest desloppify/tests/ -q
python -m desloppify scan --path <project-root>   # the project you were scanning

Once it looks good, push and open a PR:

git add <files> && git commit -m "fix: <what and why>"
git push -u origin fix/<short-description>
gh pr create --title "fix: <short description>" --body "$(cat <<'EOF'
## Problem
<what went wrong — include the command and output>

## Fix
<what you changed and why>
EOF
)"

Clean up after: rm -rf /tmp/desloppify-fix

File an issue (fallback)

If the fix is unclear or the change needs discussion, open an issue at https://github.com/peteromallet/desloppify/issues with a minimal repro: command, path, expected output, actual output.

Prerequisite

command -v desloppify >/dev/null 2>&1 && echo "desloppify: installed" || echo "NOT INSTALLED — run: uvx --from git+https://github.com/peteromallet/desloppify.git desloppify"

If uvx is not available: pip install desloppify[full] && desloppify setup

Claude Code Overlay

Use Claude subagents for subjective scoring work. Do not use --runner codex — use Claude subagents exclusively.

Review workflow

Run desloppify review --prepare first to generate review data, then use Claude subagents:

1. Prepare: desloppify review --prepare — writes query.json and .desloppify/review_packet_blind.json.
2. Launch subagents: Split the review across N parallel Claude subagents (one message, multiple Task calls). Each agent reviews a subset of dimensions.
3. Merge & import: Merge agent outputs, then desloppify review --import merged.json --manual-override --attest "Claude subagents ran blind reviews against review_packet_blind.json" --scan-after-import.

How to split dimensions across subagents

• Read dimension_prompts from query.json for dimensions with definitions and seed files.
• Read .desloppify/review_packet_blind.json for the blind packet (no score targets, no anchoring data).
• Group dimensions into 3-4 batches by theme (e.g., architecture, code quality, testing, conventions).
• Launch one Task agent per batch with subagent_type: "general-purpose". Each agent gets:
  • The codebase path and list of dimensions to score
  • The blind packet path to read
  • Instruction to score from code evidence only, not from targets
• Each agent writes output to results/batch-N.raw.txt (matching the batch index). Merge assessments (average overlapping dimension scores) and concatenate findings.

Subagent rules

1. Each agent must be context-isolated — do not pass conversation history or score targets.
2. Agents must consume .desloppify/review_packet_blind.json (not full query.json) to avoid score anchoring.

Triage workflow

Orchestrate triage with per-stage subagents:

1. desloppify plan triage --run-stages --runner claude — prints orchestrator instructions
2. For each stage (observe → reflect → organize → enrich):
   • Get prompt: desloppify plan triage --stage-prompt <stage>
   • Launch a subagent with that prompt
   • Verify: desloppify plan triage (check dashboard)
   • Confirm: desloppify plan triage --confirm <stage> --attestation "..."
3. Complete: desloppify plan triage --complete --strategy "..." --attestation "..."

Source: numberly/vault-db-injector — distributed by TomeVault.