Separate cognitive work from mechanical API execution to reduce skill-related token consumption by ~76-83% during sprint workflows. Changes: - Add batch-execution.md skill with 4-phase protocol - Promote mcp-tools-reference and batch-execution to frontmatter for planner and orchestrator agents (auto-injected, zero re-read) - Replace "Skills to Load" with phase-based "Skill Loading Protocol" - Restructure planning-workflow.md Steps 8-10 for batch execution - Update agent matrix in CLAUDE.md and docs/CONFIGURATION.md - Add Phase-Based Skill Loading documentation section - Clean up .gitignore (transient files, dev symlinks) Token impact: - 6-issue sprint planning: ~76% reduction - 10-issue sprint planning: ~80% reduction - 8-issue status updates: ~83% reduction Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
4.4 KiB
name, description
| name | description |
|---|---|
| batch-execution | Batch MCP API execution to eliminate redundant skill reloading between repetitive operations |
Batch Execution Pattern
Purpose
Separate cognitive work (analysis, planning, decision-making) from mechanical API execution (issue creation, dependency setup, status updates). Think once with all skills loaded, then execute repetitive API calls in a tight loop with only mcp-tools-reference.md needed.
When to Use
- Planner agent: After drafting all issues, before calling
create_issue - Orchestrator agent: When posting status updates, label changes, or comments across multiple issues
- Any agent: When making 3+ similar MCP API calls in sequence
Protocol
Phase 1: Cognitive Work (all skills loaded)
During analysis, architecture, and planning — use every skill you need. Read files, think deeply, ask questions, resolve ambiguity. This is where the full skill set pays for itself.
Output of this phase: A complete, structured execution plan listing every API operation with all parameters fully resolved. No ambiguity, no placeholders that require re-analysis.
Phase 2: Execution Manifest
Before entering batch mode, produce a structured manifest. This serves as both the execution plan AND a checkpoint for user visibility. Format:
## Execution Manifest
### Issue Creation (N items)
1. `create_issue` — title: "[Sprint XX] feat: ..." | labels: [...] | milestone: N
2. `create_issue` — title: "[Sprint XX] fix: ..." | labels: [...] | milestone: N
...
### Dependency Setup (N items)
1. `create_issue_dependency` — issue: $1 depends_on: $2
...
### Milestone Assignment (N items)
1. `update_issue` — issue: $1 milestone: N
...
### Status/Label Updates (N items)
1. `update_issue` — issue: #N labels: [add "Status/In-Progress"]
...
Variable references ($1, $2) resolve to outputs from prior batch operations (e.g., $1 = issue number returned by first create_issue).
Phase 3: Batch Execute
Execute the manifest items in order, grouped by operation type:
- All
create_issuecalls — collect returned issue numbers - Resolve variable references — map
$1→ actual issue #45,$2→ #46, etc. - All
create_issue_dependencycalls — using resolved numbers - All milestone assignments — if not done during creation
- All status/label updates — using resolved numbers
Rules during batch execution:
- Do NOT re-read any skill files between calls
- Do NOT re-analyze or second-guess the plan — it was finalized in Phase 1
- Do NOT add commentary between individual API calls — batch them silently
- DO track success/failure per operation
- DO continue on individual failures (log and report at end)
- DO post a progress summary after each batch group completes
Phase 4: Batch Report
After all operations complete, report results:
## Batch Execution Complete
### Issue Creation: 6/6 ✓
- #45: [Sprint 19] feat: JWT generation
- #46: [Sprint 19] feat: Login endpoint
- #47: [Sprint 19] feat: Token refresh
- #48: [Sprint 19] test: Auth unit tests
- #49: [Sprint 19] docs: API documentation
- #50: [Sprint 19] chore: CI pipeline
### Dependencies: 4/4 ✓
- #46 depends on #45
- #47 depends on #45
- #48 depends on #46, #47
- #49 depends on #46
### Milestone Assignment: 6/6 ✓
- All issues assigned to Sprint 19
### Failures: 0
Error Handling
| Error | Action |
|---|---|
| Single API call fails | Log error, continue with next item |
| Dependency target missing (prior create failed) | Skip dependency, log as blocked |
| All creates fail | STOP batch, report connection/auth issue |
| Partial milestone assignment | Report which issues weren't assigned |
After batch completes, if any failures: present failure summary and ask user whether to retry failed operations or continue.
Anti-Patterns
| Wrong | Why | Right |
|---|---|---|
Re-reading mcp-tools-reference.md before each API call |
Wastes tokens; you already know the tool signatures | Read once, execute many |
| Interleaving analysis with API calls | Forces full context per call | Finish ALL analysis first, THEN batch execute |
| Calling create_issue one at a time with commentary between | Token overhead per turn | Queue all creates, execute in tight loop |
| Stopping the batch to ask user about individual items | Defeats batching purpose | Complete batch, report results, ask about failures |