create-backlog

name	create-backlog
description	Creates backlog epics in issue tracker from a specification document
allowed-tools	mcp__*, Bash, Read, Write, Edit, Glob, Grep, Task
argument-hint	<spec-path> [--decompose] [--update] [--dry-run] [--yolo]
disable-model-invocation	true

Create Backlog

Input: $ARGUMENTS = spec path (positional) + optional flags (--decompose, --update, --dry-run, --yolo)

If $ARGUMENTS contains --yolo, activate YOLO mode: auto-approve human gates. Note: Gate 3 confirmation can still be overridden by --dry-run.

Configuration

Read Automation Config from CLAUDE.md section ## Automation Config. Follow ../../core/config-reader.md.

Required:

• Issue Tracker: Type, Instance, Project

Optional:

• Sprint Planning: Epic template (path to custom template file — overrides default Epic Card Template)
• Agent Overrides: Path (default: customization/)
• Decomposition: Max subtasks (default: 7) — used only with --decompose

Flag Parsing

Parse $ARGUMENTS:

• Remove --decompose, --update, --dry-run, --yolo from the arguments string
• Remainder = spec path (file or directory)
• If spec path is empty: STOP with "Usage: /agent-flow:create-backlog [--decompose] [--update] [--dry-run] [--yolo]"
• --decompose and --update are mutually exclusive. If both present: STOP with "Cannot use --decompose with --update."
• --dry-run can combine with any other flag.

Orchestration

0. MCP pre-flight check

If --dry-run, skip MCP check (no tracker writes will occur).

Otherwise, follow ../../core/mcp-preflight.md:

• Read Type from Automation Config (Issue Tracker section)
• Check that at least one mcp__* tool matching the tracker type is accessible
• If not accessible → BLOCK with:

  [agent-flow] 🔴 Pipeline Block
  Agent: create-backlog
  Step: MCP pre-flight check
  Reason: Cannot connect to your {Type} issue tracker.
  Detail: Expected tool prefix: mcp__{Type}__*. No matching tool is registered in this session.
  Recommendation: Run /agent-flow:check-setup for diagnostics, or /agent-flow:setup-mcp to configure the {Type} integration.
  

0b. State initialization

Create .agent-flow/backlog-{YYYYMMDD-HHmmss}/ directory. Initialize state.json with:

{
  "schema_version": "1.0",
  "run_id": "backlog-{YYYYMMDD-HHmmss}",
  "parent_run_id": null,
  "mode": "backlog-creation",
  "pipeline": "create-backlog",
  "status": "running",
  "started_at": "{ISO-8601}",
  "updated_at": "{ISO-8601}",
  "config": {
    "profile": null,
    "flags": [],
    "retry_limits": {
      "fixer_iterations": 5,
      "test_attempts": 3,
      "build_retries": 3
    }
  },
  "backlog": {
    "spec_path": "{spec-path}",
    "epics_total": 0,
    "epics_created": 0,
    "epics_failed": 0,
    "subtasks_created": 0,
    "created_issues": []
  }
}

Follow atomic write protocol from ../../core/state-manager.md.

Step 1: Read specification

Read the spec path provided in $ARGUMENTS:

• Directory: Glob {spec-path}/epics/*.md (spec-based scaffold format). If no epics/ subdir exists, glob {spec-path}/*.md.
• Single file: Read the single file.
• Multiple files (space-separated or glob pattern): Read each matched file in order.

If the path does not exist or is empty: STOP with "Specification path not found or empty: {spec-path}"

Update state.json: write backlog.spec_path. Follow atomic write protocol from ../../core/state-manager.md.

Step 2: Extract epics (backlog-creator agent)

You MUST invoke Task(subagent_type='agent-flow:backlog-creator', model='sonnet'). DO NOT inline-execute.

Context to pass:

• Specification content (all files read in Step 1, concatenated with file boundary markers)
• Epic template path: {sprint_planning.epic_template} if configured — otherwise omit (agent uses built-in template)
• Max epics: 10

Before dispatch, check Agent Overrides: follow ../../core/agent-override-injector.md for backlog-creator overrides.

If the backlog-creator agent Blocks: display the block message and STOP.

Store from backlog-creator output:

• epic_list: structured list of epics (title, scope, AC, size, dependencies, verification)
• epics_total: count of epics in the list

Update state.json: write backlog.epics_total. Follow atomic write protocol from ../../core/state-manager.md.

Step 3: Human gate (preview)

Display the Backlog Summary table from backlog-creator output:

## Backlog Summary

| # | Epic | AC | Size | SP | Dependencies |
|---|------|----|------|----|--------------|
| 1 | {title} | {count} | {XS/S/M/L} | {points} | {deps or "none"} |

If --dry-run:

• Display individual epic cards (full Epic Card Template for each epic)
• STOP with "Dry run complete. No tracker issues created."

Prompt: "Create {N} epics in {tracker_type} tracker? [Y/n]" If --yolo: auto-approve (display "[auto-approved]"). If rejected (user enters n): STOP with "Cancelled. No issues created."

Step 4: Create tracker issues

Accumulator pattern — NON-BLOCKING:

SET success_count = 0
SET failure_count = 0
SET created_issues = []  // list of {epic_index, tracker_issue_id, title}

Update mode (--update flag):

Execute the update matching algorithm (see Update Matching section below):

1. Fetch all open Feature/Epic issues from the tracker project (limit: 100)
2. For each epic in epic_list, compute match against existing issues using:
   • Prefix match: do the first 40 normalized characters match? (boolean)
   • Token overlap: Jaccard similarity of word-token sets >= 0.7
3. Display Update Preview table:

   ## Update Preview
   
   | # | Epic | Match | Tracker Issue | Similarity |
   |---|------|-------|---------------|------------|
   | 1 | {title} | MATCHED | {ID} | {score} |
   | 2 | {title} | NEW | -- | -- |
   
   Update {M} existing issue(s) and create {N} new issue(s)? [Y/n]
   

4. If --yolo: auto-approve. Otherwise wait for confirmation.
5. For matched epics: update issue description via MCP (preserve title, update body with rendered Epic Card).
6. For unmatched epics: proceed to per-tracker creation (same as create mode below).

Create mode (default, and for unmatched epics in update mode):

For each epic in epic_list (or unmatched epics in --update mode):

Build the Epic Card content from the Epic Card Template:

## {Epic Title}

**Type:** feature
**Size:** {XS|S|M|L} ({N} SP)
**Dependencies:** {comma-separated epic titles, or "none"}

### Scope
{2-3 sentences describing what needs to be built}

### Acceptance Criteria
1. {Testable criterion}
2. {Testable criterion}
3. {Testable criterion}

### Verification
- Unit: {what to test with unit tests}
- Integration: {what to test with integration tests}
- E2E: {what to test end-to-end}

If sprint_planning.epic_template is configured and the file exists: use that template instead.

Per-tracker epic creation dispatch:

TRY:
    IF tracker_type == "youtrack":
        result = mcp__youtrack__create_issue(
            project: {issue_tracker.project},
            summary: {epic.title},
            description: {epic_card_content},
            type: "Feature"
        )
        SET new_id = result.id

    ELSE IF tracker_type == "jira":
        // Attempt Epic issue type; fall back to Story if Epic unavailable
        TRY:
            result = mcp__jira__create_issue(
                project: {issue_tracker.project},
                summary: {epic.title},
                description: {epic_card_content},
                issuetype: "Epic"
            )
        CATCH issuetype_error:
            LOG WARN "Epic issue type unavailable in Jira project {issue_tracker.project}. Falling back to Story."
            result = mcp__jira__create_issue(
                project: {issue_tracker.project},
                summary: {epic.title},
                description: {epic_card_content},
                issuetype: "Story"
            )
        SET new_id = result.key

    ELSE IF tracker_type == "linear":
        result = mcp__linear__create_issue(
            teamId: {issue_tracker.project},
            title: {epic.title},
            description: {epic_card_content},
            labelNames: ["feature"]
        )
        SET new_id = result.id

    ELSE IF tracker_type == "github":
        result = mcp__github__create_issue(
            owner: {owner from issue_tracker.project},
            repo: {repo from issue_tracker.project},
            title: {epic.title},
            body: {epic_card_content},
            labels: ["epic"]
        )
        SET new_id = result.number

    ELSE IF tracker_type == "gitea":
        // Gitea: use Bash curl REST API (MCP Gitea does not guarantee epic label support)
        owner = {owner from issue_tracker.project}
        repo  = {repo from issue_tracker.project}
        result = Bash(
            curl -s -X POST "{issue_tracker.instance}/api/v1/repos/{owner}/{repo}/issues"
              -H "Authorization: token $GITEA_TOKEN"
              -H "Content-Type: application/json"
              -d '{"title":"{epic.title}","body":"{epic_card_content_escaped}","labels":[]}'
        )
        // If GITEA_TOKEN is not set, fall back to:
        // result = mcp__gitea__create_issue
        //   with parameters: owner, repo, title, body
        SET new_id = result.number

    ELSE IF tracker_type == "redmine":
        result = mcp__redmine__create_issue(
            project_id: {issue_tracker.project},
            subject: {epic.title},
            description: {epic_card_content},
            tracker_id: "Feature"
            // If "Feature" tracker unavailable, omit tracker_id (use project default)
        )
        SET new_id = result.id

    // --- Write to state ---
    ADD {epic_index: {N}, tracker_issue_id: new_id, title: epic.title} to created_issues
    success_count += 1

    // Update state.json per epic (atomic, immediate)
    UPDATE state.json: increment backlog.epics_created, append to backlog.created_issues
    Follow atomic write protocol from ../../core/state-manager.md

CATCH error:
    LOG WARN "Could not create tracker issue for epic '{epic.title}': {error}"
    failure_count += 1
    CONTINUE  // NON-BLOCKING — proceed to next epic

Per-Tracker Epic Creation Parameters:

Tracker	MCP Tool Prefix	Title Param	Description Param	Type / Label	Notes
YouTrack	mcp__youtrack__*	summary	description	type: "Feature"	Top-level issue, no parent
Jira	mcp__jira__* or mcp__atlassian__*	summary	description	issuetype: "Epic"	Fallback to "Story" if Epic type unavailable
Linear	mcp__linear__*	title	description	labelNames: ["feature"]	No native Epic type; use label
GitHub	mcp__github__*	title	body	labels: ["epic"]	Uses REST via MCP
Gitea	Bash curl REST or mcp__gitea__*	title	body	labels: ["epic"]	Bash preferred; MCP fallback
Redmine	mcp__redmine__*	subject	description	tracker_id: "Feature"	Fallback to project default tracker

Step 5: Display result

Created {success_count}/{success_count + failure_count} epic issues.

If --decompose and subtasks were created:

Created {subtasks_created} sub-tasks across {epic_count} epics.

If failure_count > 0:

({failure_count} failures. Check warnings above.)

Update state.json: set top-level status to "completed". Follow atomic write protocol from ../../core/state-manager.md.

Step 6 (--decompose): Subtask decomposition

Only executed if --decompose flag is present.

Run AFTER Step 4 for each successfully created epic issue (i.e., every issue in created_issues).

For each epic in created_issues:

1. You MUST invoke Task(subagent_type='agent-flow:architect', model='opus'). DO NOT inline-execute.

   • Context: Epic: {epic.title}\nSpec content:\n{epic_card_content}\nParent tracker issue: {tracker_issue_id}
   • Before dispatch, check Agent Overrides: follow ../../core/agent-override-injector.md for architect overrides
   • Expected output: architectural task tree with subtasks (each subtask includes title, scope, files, estimated_lines, maps_to)

2. If architect blocks: LOG WARN "Architect blocked for epic '{epic.title}': {reason}". Continue to next epic — NON-BLOCKING.

3. From architect output, extract subtask list. For each subtask:

   Build subtask description:

   {subtask.scope}
   
   Addresses: {subtask.maps_to[0]}, {subtask.maps_to[1]}, ...
   
   Files: {subtask.files[0]}, {subtask.files[1]}, ...
   
   Parent issue: {tracker_issue_id}
   

   (Omit "Addresses:" line if maps_to is empty. Omit "Files:" line if files is empty. "Parent issue:" always present.)

4. Create sub-issues using the same per-tracker dispatch table from implement-feature Step 5a (with parent parameter pointing to the epic's tracker_issue_id). Follow the same accumulator pattern (NON-BLOCKING on individual subtask failures).

5. Update state.json: increment backlog.subtasks_created by the count of successfully created sub-issues. Follow atomic write protocol from ../../core/state-manager.md.

Update mode (--update) matching algorithm:

Normalize title: lowercase, strip leading/trailing whitespace, collapse multiple spaces.

For each epic in epic_list:
  For each open tracker issue:
    prefix_match  = (normalized_epic_title[:40] == normalized_issue_title[:40])
    jaccard       = |token_intersection| / |token_union|   // tokens = split on whitespace+punctuation
    match         = prefix_match OR jaccard >= 0.7
  IF exactly one match: pair epic <-> issue
  IF multiple matches: select highest Jaccard. If tied, select most recently updated. WARN.
  IF no match: add to unmatched_epics list.

Edge cases:

• Empty tracker (0 open issues): all epics are unmatched, behave as create mode.
• Closed/resolved issues: not included in fetch (filtered by open state only).
• Title changed significantly: no match, new issue created.

Rules

• NON-BLOCKING epic creation: a single epic failure NEVER stops the batch; accumulate counts and continue
• NON-BLOCKING subtask creation (--decompose): same rule applies to each sub-issue
• Epic issues MUST NOT have the On start set state transition applied (they represent planned work, not active execution)
• Language fidelity: preserve all diacritics and non-ASCII characters from spec content without escaping
• Agent Overrides: follow ../../core/agent-override-injector.md for backlog-creator and architect invocations
• If sprint_planning.epic_template is set but the file is missing: WARN and use the built-in Epic Card Template — do not block
• Max 10 epics per invocation (enforced by backlog-creator agent; display note if spec contains more)
• --dry-run skips MCP pre-flight, skips all tracker writes, and always stops after the preview gate
• All tracker writes use the SAME MCP tool conventions as implement-feature Step 5a — no divergence from established patterns
• Block Comment Template for fatal errors:

  [agent-flow] 🔴 Pipeline Block
  Agent: create-backlog
  Step: {step where failure occurred}
  Reason: {max 2 sentences}
  Detail: {technical output}
  Recommendation: {what the human should do}