blitz

name

blitz

description

A continuously-firing background work runner for Claude Code. Queue real tasks ("/blitz add"), track long-term goals ("/blitz goal add"), and enable per-project audit rotations ("/blitz audit enable"). Blitz fires on a configurable cadence with a per-project idle gate, decomposes goals into increments on a long-lived branch, and persists every fire's output to ~/blitz/runs/<timestamp>/. Includes commands to manage the backlog, goals, audits, schedules, and kill switches (off/on/skip).

⚡ Blitz — Background Work Runner

A Claude Code skill for people on Claude Max plans. Queue tasks, declare long-term goals, and enable per-project audits. Blitz fires on a configurable cadence with a per-project idle gate, runs work through parallel agents, and writes every agent's output to disk so you wake up to actual artifacts, not a closed terminal.

Parallelism is for throughput, not quota multiplication. Five agents finish five tasks in the time of one — that's the win.

---

Step 0 — Route by Subcommand

Read the first word of arguments after /blitz:

Subcommand	Action
add	Add a task to the backlog (M-ADD)
list, ls	Show the backlog (M-LIST)
remove, rm	Remove a backlog item by number (M-REMOVE)
clear	Clear all backlog (with confirmation) (M-CLEAR)
goal add / goal list / goal review / goal log / goal remove / goal autopush	Goal lifecycle (M-GOAL)
audit enable / audit disable / audit list / audit run	Audit lifecycle (M-AUDIT)
setup	Run the configuration wizard (M-SETUP)
auto	Unattended scheduled run (M-AUTO)
off	Disable auto-schedule firing (M-OFF)
on	Re-enable auto-schedule firing (M-ON)
skip	Skip the next scheduled auto-fire (M-SKIP)
empty / anything else	Run the backlog now (M-RUN)

Config path: ~/.claude/blitz.json (Windows: C:\Users\<user>\.claude\blitz.json) Output path: ~/blitz/runs/<timestamp>/ (Windows: C:\Users\<user>\blitz\runs\<timestamp>\)

If ~/.claude/blitz.json doesn't exist on any subcommand other than setup, create a minimal default with empty backlog: [] so reads don't fail. Don't force setup on first run.

---

M-ADD — Add a Task to the Backlog

Syntax: /blitz add <task description> (optional flag: --project <path>)

1. Parse the task description (everything after add minus any flags).
2. Determine project path:
   • If --project <path> provided, use it.
   • Otherwise, use the current Claude Code working directory.
3. Append to backlog[] in config:

   { "id": <next>, "task": "<description>", "project": "<absolute path>", "added": "YYYY-MM-DD" }
   

4. Confirm:

   ✅ Added #[id] to backlog
      Task:    <description>
      Project: <project path>
      Backlog now has [N] item(s).
   

If the task description is empty, ask the user what they want to add.

---

M-LIST — Show the Backlog

Display every backlog item with its id, task, project (shortened), and date added. Group by project if there are multiple. Show total count at top.

⚡ Blitz backlog — [N] items

  #1  Refactor auth.js for clarity              [myapp]    added 2026-04-29
  #2  Generate tests for billing module         [myapp]    added 2026-04-30
  #3  Write API docs for /v2 endpoints          [api-svc]  added 2026-05-01

Run /blitz to blitz through them.

If empty: Backlog is empty. Add something with /blitz add <task>.

---

M-REMOVE — Remove an Item

Syntax: /blitz remove <id> (or rm).

Remove the item with that id from backlog[]. Confirm:

✅ Removed #[id]: <task>. Backlog now has [N] item(s).

If the id doesn't exist, list the current backlog and ask which to remove.

---

M-CLEAR — Clear Everything

Confirm first: Clear all [N] backlog items? (yes/no). Only proceed on explicit "yes". Then empty backlog: [] and confirm.

---

M-GOAL — Goal Lifecycle

Long-term goals are user-defined objectives that compound over many fires. Each goal has a durable plan (~/blitz/goals/<id>/plan.md) and a long-lived branch (blitz/goal-<id>). Goal items live in backlog[] with kind: "goal".

M-GOAL-ADD — Add a goal

Syntax: /blitz goal add <description> (optional flag: --project <path>)

1. Parse the description.
2. Determine project path: --project flag, else current cwd.
3. Append to backlog[]:

{
  "id": <next>,
  "kind": "goal",
  "task": "<description>",
  "project": "<absolute path>",
  "added": "YYYY-MM-DD",
  "branch": "blitz/goal-<id>",
  "autoPush": false,
  "lastTouched": null,
  "incrementCount": 0
}

4. Confirm:

✅ Added goal #[id] to backlog
   Goal:    <description>
   Project: <project path>
   Branch:  blitz/goal-<id> (will be created on first increment fire)
   Plan:    ~/blitz/goals/<id>/plan.md (will be written on first fire)

The plan is not written at queue time. The first fire on this goal writes it.

M-GOAL-LIST — Show goals

Display every kind: "goal" item with id, description, project, branch, increment count, and last-touched date. If none, print: No goals. Add one with /blitz goal add <description>.

M-GOAL-REVIEW — Review a goal's branch

Syntax: /blitz goal review <id>

1. Look up the goal item; load branch field.
2. In the goal's project directory, run:

git log <branch> ^main --oneline
git diff main..<branch> --stat

3. Print plan.md status: count of [x], [ ], and [~] items.
4. End with a hint: Merge with: git checkout main && git merge <branch>.

If the branch doesn't exist, print: No commits yet on this goal.

M-GOAL-LOG — Show increment history

Syntax: /blitz goal log <id>

Read every result file in ~/blitz/runs/*/goal-<id>-*.md (sorted by timestamp), print the top-level heading and the first paragraph of each.

M-GOAL-REMOVE — Remove a goal

Syntax: /blitz goal remove <id>

Remove the item from backlog[]. Do not delete the branch or plan.md. Print:

✅ Removed goal #[id]: <description>
   The branch blitz/goal-<id> and plan.md were left in place.
   Clean up manually: git branch -D blitz/goal-<id> && rm -rf ~/blitz/goals/<id>

M-GOAL-AUTOPUSH — Toggle autopush

Syntax: /blitz goal autopush <id> on|off

Set autoPush on the goal item. When true, increment fires also git push origin <branch> after committing. Default is off.

Confirm:

✅ Goal #[id] autopush: <on|off>

---

Goal Fire Flow

When the auto-fire pulls a kind: "goal" item, route to one of two flows based on plan.md state.

Planning fire (no plan, or empty plan)

This is the first fire for a goal. No code is written. The agent only produces a plan.

1. Spawn one general-purpose agent with this prompt template:

Working directory: <goal.project>

You are the planning agent for a long-term goal. Your sole job is to write
a plan.md file. You will NOT write any code.

Goal: <goal.task>

Read the project state (git log -30, top-level file tree, README.md if it
exists, package.json or equivalent) and produce a numbered Markdown
checklist of small, mergeable, demoable increments toward the goal.

Each increment must:
- Be self-contained (no item depends on incomplete prior items)
- Be mergeable on its own (could ship as a single PR)
- Be small enough that one focused agent could finish it in one fire
- Have a clear "done" criterion

Write the checklist to: ~/blitz/goals/<goal.id>/plan.md

Format:
  # Plan: <goal.task>
  Project: <goal.project>
  Created: <date>

  ## Increments
  - [ ] 1. <one-line summary>: <2-3 sentence detail>
  - [ ] 2. ...

Also write a result file to <output_dir>/goal-<goal.id>-plan.md containing
the plan and a short paragraph explaining your reasoning (why these
increments, in this order).

Do NOT make any commits. Do NOT modify any project files.

2. After the agent finishes, update the goal item: lastTouched = now, incrementCount stays at 0.
3. The goal will hit the increment-fire flow on its next fire.

Increment fire (plan.md exists with at least one unchecked item)

1. Read ~/blitz/goals/<goal.id>/plan.md. If every item is [x], the goal is done — set a marker comment in plan.md (> Goal complete: YYYY-MM-DD), update the goal item with lastTouched = now, and exit. Skip auto-promoting a "review and merge" task since the user knows from /blitz goal list.
2. Pick the lowest-numbered unchecked item.
3. Spawn one general-purpose agent with this prompt template:

Working directory: <goal.project>

You are an increment agent for goal #<goal.id>: "<goal.task>"

The full plan is at ~/blitz/goals/<goal.id>/plan.md. The increment you must
implement is item #<N>: <item summary>.

Steps:
1. Switch to (or create from main) the long-lived branch: <goal.branch>.
   - git fetch origin; git checkout <goal.branch> 2>/dev/null || git checkout -b <goal.branch> main
   - git rebase main (only if no conflicts; otherwise stay on the branch as-is and note in result)
2. Implement item #<N>. Read enough of the codebase to do this correctly.
3. If the project has a test command and it runs in under 60 seconds, run
   it and confirm it passes. Detect the command from package.json scripts,
   pytest presence, or similar. If detection fails or tests are slow, skip
   them and note in the result.
4. Commit: git add -A && git commit -m "blitz(goal-<goal.id>): <item summary>"
5. If <goal.autoPush> is true, push: git push origin <goal.branch>
6. Update plan.md: change item #<N> from [ ] to [x], append a one-liner
   with the commit SHA: "      ✓ <commit-sha-short>: <one-line summary>"
7. Write a result file to <output_dir>/goal-<goal.id>-<slug>.md with:
   - Heading: # Goal #<goal.id> increment: <item summary>
   - What was done (bullet list of changes)
   - Why this approach
   - Branch + commit SHA + any push status
   - What's next (the next unchecked plan item, if any)

If the picked item is no longer applicable (e.g., already done by user, or
the codebase has changed in a way that invalidates it), mark it [~]
(skipped) with a one-line reason in plan.md and pick the next unchecked
item.

Hard rules:
- Only commit on <goal.branch>. Never on main.
- Never push unless <goal.autoPush> is true.
- Never edit plan.md items above the one you picked.

4. After the agent finishes, update the goal item: lastTouched = now, incrementCount += 1.

---

M-AUDIT — Audit Lifecycle

Audits are per-project maintenance rotations that produce read-only reports. Audit items live in backlog[] with kind: "audit". Audits never modify project files.

M-AUDIT-ENABLE — Enable an audit rotation

Syntax: /blitz audit enable <project> (project may be a path or . for cwd)

1. Resolve the project to an absolute path. If a kind: "audit" item already exists for that project, print Already enabled and exit.
2. Append to backlog[]:

{
  "id": <next>,
  "kind": "audit",
  "project": "<absolute path>",
  "rotation": ["security","deadcode","tests","deps","docs","todos"],
  "rotationIdx": 0,
  "lastSwept": null,
  "promotedFindings": []
}

3. Confirm:

✅ Audit rotation enabled for <project>.
   Audits will rotate: security → deadcode → tests → deps → docs → todos.
   Run /blitz audit run <project> to trigger one immediately.

M-AUDIT-DISABLE — Disable an audit

Syntax: /blitz audit disable <project>

Remove the matching kind: "audit" item. Confirm.

M-AUDIT-LIST — Show enabled audits

Print one row per audit item: project, last-swept date, next audit type (rotation[rotationIdx]), promoted-finding count.

M-AUDIT-RUN — Run the next audit immediately

Syntax: /blitz audit run <project>

Bypass the auto-fire scheduler and run the next audit type for this project right now. Same flow as Audit Fire Flow below.

Audit Fire Flow

When the auto-fire pulls a kind: "audit" item:

1. Pick the next audit type: audit_type = item.rotation[item.rotationIdx].
2. Spawn one general-purpose agent with this prompt template:

Working directory: <audit.project>

You are an audit agent. Run the <audit_type> audit on this project. You
MUST NOT modify any project files. Output is read-only.

Audit type: <audit_type>

Definitions:
  security  - OWASP-style issues. Look for hardcoded secrets, SQL/command
              injection, unsanitized inputs, weak crypto, missing auth
              checks. Output: file:line + severity (high/medium/low).
  deadcode  - Unused exports, unreachable branches, orphaned files.
              Output: file:line + reason.
  tests     - Public functions/classes with no test coverage. Output:
              file:line + symbol name.
  deps      - Outdated packages, breaking-change risk. Output: package
              name + current → latest version + risk note.
  docs      - Public APIs (exports, public methods) without docstrings or
              JSDoc. Output: file:line + symbol name.
  todos     - TODOs/FIXMEs grouped by age (from git blame) and severity.
              Output: file:line + age + content.

Write a markdown report to <output_dir>/audit-<project_slug>-<audit_type>.md
with this structure:
  # <audit_type> audit — <project_slug>
  Date: <YYYY-MM-DD>
  Total findings: <N>

  ## Findings
  ### Finding 1: <one-line summary>
  Severity: <high|medium|low>
  Location: <file>:<line>
  Hash: <sha256 of: file_path + ":" + line + ":" + audit_type + ":" + finding_summary>
  Detail: <2-4 sentences>

  ### Finding 2: ...

Sort findings by severity (high first), then by file path.

3. After the agent finishes, the skill (not the agent) does: a. Read the report. Parse Hash: lines from each finding. b. Find the highest-severity finding whose hash is not in item.promotedFindings. c. If one exists, append a new kind: "task" item to backlog[]:

{
  "id": <next>,
  "kind": "task",
  "task": "<finding summary> (<file>:<line>)",
  "project": "<audit.project>",
  "added": "<YYYY-MM-DD>",
  "source": { "auditId": <audit.id>, "auditType": "<audit_type>", "findingHash": "<sha256>" }
}

d. Append the finding's hash to item.promotedFindings. e. Update item.rotationIdx = (item.rotationIdx + 1) % len(item.rotation). f. Update item.lastSwept = now. 4. Save the config.

If every finding's hash is already in promotedFindings, the report is still written but no task is auto-promoted. The skill prints All <audit_type> findings already promoted; nothing new to triage.

---

M-RUN — Run the Backlog (default /blitz)

This is the main path. Takes the top items off the backlog, spawns parallel agents, persists output.

R1 — Check Backlog

If backlog is empty:

Backlog is empty. Add tasks with /blitz add <task description>.

Or, if you want to run anyway against the current project, type "auto" and I'll
generate generic improvement tasks for the current directory.

Wait for input. If the user says "auto", generate 3 generic tasks scoped to the current cwd (security audit, test gen, doc gen — concrete, not vague). Otherwise stop.

R2 — Optional Urgency Read

Ask in one message:

Quick: usage stats? session: X% / Yh | weekly: X% / Yh — or press Enter to skip.

If skipped, default to running 3 tasks (medium intensity). If provided, compute:

session_remaining = 100 - session_used
weekly_remaining  = 100 - weekly_used
session_rate = session_remaining / session_hours
weekly_rate  = weekly_remaining / weekly_hours
priority = max(session_rate, weekly_rate * 2)

Map score → number of tasks to pull from backlog:

Score	Tasks to run
> 40	5–6
20–40	3–4
10–20	2–3
< 10	1–2

If backlog has fewer items than that, run all of them.

R3 — Show the Plan

╔══════════════════════════════════════════════════════╗
║                   ⚡ BLITZ — RUN PLAN                 ║
╠══════════════════════════════════════════════════════╣
║  Pulling [N] of [M] backlog items                    ║
║  Output: ~/blitz/runs/<timestamp>/                   ║
╚══════════════════════════════════════════════════════╝

  #1  <task>          [project]
  #2  <task>          [project]
  ...

Launch [N] parallel agents? (go / adjust / cancel)

If the user says "adjust", let them edit the list (skip a task, swap an id, etc.). On "go", proceed.

R4 — Prepare Output Directory

Create ~/blitz/runs/<timestamp>/ where timestamp is YYYY-MM-DD_HHMM-<trigger> (trigger = manual, weekly, or session).

R5 — Spawn Agents in Parallel

In one Agent tool call block (multiple <invoke> blocks in the same message), spawn one agent per pulled item. Route by kind:

For kind: "task" — use this prompt template:

Working directory: <task.project>

Task: <task.task>

Produce thorough, long-form output. Show full work — code diffs, complete
implementations, line-by-line analysis where appropriate. Do not summarize
or truncate.

When you finish, write your full output to:
  <output_dir>/<task_id>-<short-slug>.md

Use the Write tool to create that file. Include a top-level heading with
the task description, then your full work.

Use subagent_type: "general-purpose" for code/writing tasks, "Explore" for pure research, "Plan" for architecture-only tasks. Default to general-purpose.

For kind: "goal" — see Goal Fire Flow above. Choose planning-fire or increment-fire prompt based on plan.md state.

For kind: "audit" — see Audit Fire Flow above. Use the audit agent prompt template.

After all agents finish, run kind-specific post-processing (R6 covers this).

R6 — Write Run Summary

After all agents complete, write <output_dir>/_summary.md:

# Blitz run — <timestamp>

Trigger: <manual|weekly|session>
Tasks: <N> run in parallel

## Results
- [#1 task title](1-slug.md) — <one-line status from agent>
- [#2 task title](2-slug.md) — <one-line status>
...

## Backlog state after run
Removed: #1, #2, #3
Remaining: <N> items

Kind-aware backlog updates (after summary is written):

• For each completed kind: "task" item: remove from backlog[] (v0.1 behavior).
• For each completed kind: "goal" item: update lastTouched = now and incrementCount += 1 (planning fire keeps incrementCount at 0). Goal stays in backlog.
• For each completed kind: "audit" item: run the audit-finding promotion logic (see Audit Fire Flow step 3). Audit stays in backlog.
• For each item that was skipped by the idle gate: leave it untouched.

R7 — Final Report

╔══════════════════════════════════════════════════════╗
║                  ⚡ BLITZ COMPLETE                    ║
╠══════════════════════════════════════════════════════╣
║  Tasks done:  [N]                                    ║
║  Output:      ~/blitz/runs/<timestamp>/              ║
║  Backlog:     [N] items remaining                    ║
╚══════════════════════════════════════════════════════╝

Open the summary: <output_dir>/_summary.md

Tell the user the absolute path so they can click it.

---

M-SETUP — Configuration Wizard

Seven steps. Show progress (Setup 1/7).

S1 — Weekly reset

When does your weekly Claude quota reset? (e.g. "Mondays 9am EST")

Parse to: weeklyReset.dayOfWeek (0=Sun…6=Sat), hour, minute, timezone. (This is informational for the user; the cadence model does not anchor to it.)

S2 — Cadence

How often should blitz check for work? (default: every 3h) Lower values catch more idle windows but burn more setup overhead.

Parse to integer hours, store as firing.cadenceHours.

S3 — Idle threshold

Skip a fire if you've committed in the project within the last X minutes. (default: 30)

Store as firing.idleMinutes.

S4 — Blackout windows (optional)

Want to suppress firing during specific hours? (e.g. "weekdays 9am–6pm") Press Enter to skip.

Parse user input into one or more blackout windows. Each window has days (array of mon/tue/wed/thu/fri/sat/sun) and start/end (HH:MM). Append to firing.blackoutWindows.

S5 — Output directory

Where should agent output be written? (default: ~/blitz/runs)

Store as outputDir. Set goalsDir to a sibling: <parent>/goals (default: ~/blitz/goals). Create both directories.

S6 — Auto-schedule (opt-in, asked explicitly)

Final question: enable auto-fire? I'll create one cadence-based scheduled job that runs /blitz auto every h.

Anytime: /blitz off (disable), /blitz skip (skip next), /blitz on (re-enable)

[Y] Enable auto-fire [N] Save config without scheduling

If N: save config, exit setup with confirmation.

If Y: detect OS, create the cadence-based scheduled job, then run S7.

Windows scheduled task (cadence)

Use a single PowerShell wrapper script. Direct cmd /c claude.exe ... invocation fails silently from Task Scheduler; a wrapper is required.

$claudePath = (Get-Command claude -ErrorAction SilentlyContinue).Source
if (-not $claudePath) { throw "claude not found in PATH — install Claude Code first" }

$wrapperPath = "$env:USERPROFILE\.claude\blitz-runner.ps1"
$wrapperContent = @"
`$ErrorActionPreference = 'Continue'
`$logDir = `"`$env:USERPROFILE\blitz`"
New-Item -ItemType Directory -Force -Path `$logDir | Out-Null
`$log = `"`$logDir\blitz.log`"
`"=== Cadence fire `$(Get-Date) ===`" | Out-File `$log -Append -Encoding UTF8
& '$claudePath' -p '/blitz auto' 2>&1 | Out-File `$log -Append -Encoding UTF8
`"=== Done `$(Get-Date) ===`" | Out-File `$log -Append -Encoding UTF8
"@
Set-Content -Path $wrapperPath -Value $wrapperContent -Encoding UTF8

$action = New-ScheduledTaskAction -Execute "powershell.exe" `
  -Argument "-NoProfile -ExecutionPolicy Bypass -File `"$wrapperPath`""

# Cadence trigger: every <cadenceHours> hours, starting now
$startAt = (Get-Date).AddMinutes(2)  # let setup finish first
$trigger = New-ScheduledTaskTrigger -Once -At $startAt `
  -RepetitionInterval (New-TimeSpan -Hours <cadenceHours>) `
  -RepetitionDuration ([TimeSpan]::FromDays(365))

# Remove any v0.1 tasks first
Unregister-ScheduledTask -TaskName "Blitz-Weekly" -Confirm:$false -ErrorAction SilentlyContinue
Unregister-ScheduledTask -TaskName "Blitz-Session" -Confirm:$false -ErrorAction SilentlyContinue

Register-ScheduledTask -TaskName "Blitz-Cadence" -Action $action -Trigger $trigger -Force

Mac/Linux crontab

# Convert cadenceHours to a cron expression: every N hours, on the hour.
# For cadenceHours=3, the expression is: 0 */3 * * *

(crontab -l 2>/dev/null | grep -v '# Blitz'; \
  echo "# Blitz cadence"; \
  echo "<cron-expression> claude -p '/blitz auto'") | crontab -

autoScheduleEnabled: true in config.

S7 — Verify auth

After creating the scheduled job, immediately run a test:

Testing: I'll trigger /blitz auto right now to confirm authentication works
in a non-interactive context. This will run with whatever's queued, or
defer cleanly if all gates skip.

On Windows: Start-ScheduledTask -TaskName "Blitz-Cadence" then check Get-ScheduledTaskInfo -TaskName "Blitz-Cadence" for last-run result.

On Mac/Linux: directly run claude -p "/blitz auto" in a subshell.

If it fails (auth not picked up in non-interactive context), tell the user clearly:

⚠ Auto-fire test failed: <error>

The scheduled job was created, but `claude` couldn't authenticate when run
non-interactively. Auto-mode won't work until this is fixed.

Workarounds:
  1. Manually run /blitz before each reset (still useful with the backlog)
  2. Disable auto-schedule: /blitz off

If it works, confirm:

✅ Auto-fire verified.
   Cadence:  every <cadenceHours>h
   Idle gate: <idleMinutes>m
   Blackouts: <count> windows configured
   Output:   <outputDir>

---

M-AUTO — Unattended Scheduled Run

Triggered by the cadence-based scheduled job. No interactive prompts.

1. Load ~/.claude/blitz.json. If missing or autoScheduleEnabled is false, exit silently.
2. If skipNextFire is true, set it to false, save, log ⚡ Auto-fire skipped (per /blitz skip) to ~/blitz/blitz.log, exit.
3. Run migration check (see "Migration v1 → v2"). If migrating, save and continue.
4. Blackout window check: get current weekday + time-of-day in the user's timezone. For each window in firing.blackoutWindows, if today's weekday is listed and current time falls in [start, end), log ⚡ Blackout active, exit.
5. Pull items by tier, respecting firing.caps:
   • First, walk backlog[] looking for kind: "goal" items. Take up to caps.goalsPerFire (default 1), preferring the goal with the oldest lastTouched.
   • Then walk for kind: "task" items. Take up to caps.tasksPerFire (default 3), in order added.
   • Then walk for kind: "audit" items. Take up to caps.auditsPerFire (default 1), preferring the audit with the oldest lastSwept (null sorts oldest).
6. Per-item idle gate: for each pulled item, run the gate check on its project:

# In the item's project directory:
RECENT_COMMIT=$(git log --since="<idleMinutes> minutes ago" --oneline | head -1)
DIRTY=$(git status --porcelain | head -1)

• If RECENT_COMMIT is non-empty OR DIRTY is non-empty, skip this item. Leave it in the backlog; it will be eligible on the next fire.
• If the project is not a git repo, the gate is bypassed (treated as idle).

7. If every pulled item was skipped: log ⚡ All targets active, deferring (next fire in <cadenceHours>h) to ~/blitz/blitz.log. Exit clean. Backlog unchanged.
8. Otherwise, prepare the output dir: ~/blitz/runs/<YYYY-MM-DD_HHMM>-cadence/.
9. Spawn agents in one parallel block, one per non-skipped item, routed by kind:
   • kind: "task" → existing R5 prompt template.
   • kind: "goal" → Goal Fire Flow (planning or increment, based on plan.md).
   • kind: "audit" → Audit Fire Flow.
10. After all agents finish, run kind-specific post-processing:
    • For tasks: remove the task from backlog[] (current v0.1 behavior).
    • For goals: update lastTouched and incrementCount. Goal item stays in backlog[].
    • For audits: parse the report, auto-promote one finding per Audit Fire Flow step 3. Audit item stays in backlog[].
11. Write <output_dir>/_summary.md listing each item's result file with kind tag and one-line status.
12. Append a one-liner to ~/blitz/blitz.log:

[YYYY-MM-DD HH:MM] cadence-fire: ran <K> items (goals=<g>, tasks=<t>, audits=<a>); skipped <S>; output=<output_dir>

---

M-OFF / M-ON / M-SKIP — Schedule Controls

/blitz off — set autoScheduleEnabled: false in config. Also disable the OS job:

• Windows: Disable-ScheduledTask -TaskName "Blitz-Cadence" (also handle legacy v0.1 tasks: Blitz-Weekly, Blitz-Session)
• Mac/Linux: comment out the # Blitz cadence line in crontab

Confirm: 🛑 Auto-fire disabled. Run /blitz on to re-enable, or /blitz manually anytime.

/blitz on — reverse of off. Re-enable the OS job and set autoScheduleEnabled: true.

/blitz skip — set skipNextFire: true in config. The next M-AUTO invocation will exit and reset the flag. Confirm: ⏭ Next scheduled fire will be skipped. Subsequent fires will run normally.

---

Config Schema (v2)

{
  "version": 2,
  "weeklyReset": {
    "dayOfWeek": 1,
    "hour": 9,
    "minute": 0,
    "timezone": "America/New_York"
  },
  "firing": {
    "mode": "hybrid",
    "cadenceHours": 3,
    "idleMinutes": 30,
    "blackoutWindows": [
      { "days": ["mon","tue","wed","thu","fri"], "start": "09:00", "end": "18:00" }
    ],
    "caps": { "goalsPerFire": 1, "tasksPerFire": 3, "auditsPerFire": 1 }
  },
  "outputDir": "~/blitz/runs",
  "goalsDir": "~/blitz/goals",
  "autoScheduleEnabled": true,
  "skipNextFire": false,
  "backlog": [
    {
      "id": 1,
      "kind": "task",
      "task": "Refactor auth.js for clarity",
      "project": "C:\\Users\\Adam\\repos\\myapp",
      "added": "2026-05-01"
    },
    {
      "id": 2,
      "kind": "goal",
      "task": "Ship pluginproof MVP",
      "project": "C:\\Users\\Adam\\MCP\\pluginproof",
      "added": "2026-05-04",
      "branch": "blitz/goal-2",
      "autoPush": false,
      "lastTouched": "2026-05-04T14:00:00Z",
      "incrementCount": 0
    },
    {
      "id": 3,
      "kind": "audit",
      "project": "C:\\Adam\\claude_skills",
      "rotation": ["security","deadcode","tests","deps","docs","todos"],
      "rotationIdx": 0,
      "lastSwept": null,
      "promotedFindings": []
    }
  ],
  "nextId": 4
}

Every backlog item has a kind field. Three kinds:

• "task" — a unit of work to run-and-remove (v0.1 behavior)
• "goal" — a long-term objective that produces increments via a durable plan
• "audit" — a per-project maintenance rotation that produces read-only reports

When adding any new item, increment nextId and use it as the new item's id.

Migration v1 → v2

When the skill loads ~/.claude/blitz.json, check version:

1. If version is 2, do nothing.
2. If version is missing or 1, run the migration:
   • Add "kind": "task" to every existing item in backlog[].
   • Add a firing block with defaults: { "mode": "hybrid", "cadenceHours": 3, "idleMinutes": 30, "blackoutWindows": [], "caps": { "goalsPerFire": 1, "tasksPerFire": 3, "auditsPerFire": 1 } }.
   • Drop preResetMinutes if present (functionality removed in v0.2).
   • Drop sessionResetIntervalHours and sessionResetAnchorTime if present (cadence model replaces session-anchored triggers).
   • Add "goalsDir": "~/blitz/goals" if missing.
   • Set "version": 2.
   • Save the file.
3. If on Windows and autoScheduleEnabled is true, replace existing Blitz-Weekly and Blitz-Session scheduled tasks with a single Blitz-Cadence task triggered every cadenceHours. On Mac/Linux, replace cron entries similarly.

Migration is idempotent — re-running it on a v2 config is a no-op.

---

Principles

• Real work only. Backlog holds work the user defines: tasks queued explicitly, goals declared explicitly, audits enabled per-project. Never invented filler.
• Empty backlog + nothing enabled = silent exit. Auto mode never invents work to fill quota.
• Persist everything. Every fire writes artifacts. A fire that produces no file is a bug.
• Parallelism = throughput, not multiplier. It finishes more tasks in the same wall-clock time. Do not oversell it.
• Auto-fire is opt-in and reversible. Always ask, always provide off/skip/on.
• Goals decompose at run-time, not at queue-time. The first fire writes a plan; subsequent fires execute one plan item each. Plans are durable on disk, not in config.
• Code modifications stay on blitz/* branches. Never main, never push without explicit per-goal autoPush.
• Audits never touch project files. Reports go to the output dir. Findings get auto-promoted to the backlog as tasks; the user reviews and discards.
• Per-project idle gate prevents collisions. If you committed in a project recently or have a dirty tree, blitz defers that item.
• Verify auth on setup. A scheduled job that silently fails auth is the worst outcome — test it during S7.