Add create-issue skill with Haiku delegation and cross-shell fix

[oto-macenauer-absa]

Summary

Resolves #12

• Add create-issue skill — two-phase architecture: Haiku subagent discovers repo templates (cheap), primary model drafts and posts
• Leaner than initial prototype: 68% fewer instruction lines (~950 fewer input tokens per invocation)
• Broader trigger matching: "log bug", "report problem", "submit ticket", etc.
• Eval suite with 3 test cases (bug report, deep context, cross-repo feature request)
• Fix shell compatibility: write body to temp file instead of stdin heredocs that break on PowerShell/special chars

Release Notes

Two-phase architecture with Haiku delegation

Template discovery (repo scanning, API calls, YAML parsing) is delegated to a Haiku subagent in Phase 1. Only drafting and posting runs on the primary model.

• Template discovery is mechanical work — perfect for a cheaper, faster model
• Saves 25-30% cost on template-heavy repos where discovery involves multiple API round-trips

Instruction footprint reduced by 68%

Skill definition went from 347 lines across 4 files down to 109 lines across 3 files (net -238 lines):

Change	Lines removed
Deleted --help/SKILL.md sub-skill	-29
Deleted references/gh-commands.md	-112
Condensed references/templates.md	-124
Leaner SKILL.md (74 to 57 lines)	-17

Every line of skill definition is injected into context when the skill triggers. 238 fewer lines = ~950 fewer input tokens per invocation.

Broader trigger matching

Description now catches natural language variations: "log a bug", "report a problem", "submit a ticket", "track as an issue" — not just "create an issue".

Behavioral guardrails

• Fix suggestions only when root cause is clear from context (prevents hallucinated fixes)
• Template structure respected — no extra sections added beyond what the template defines
• Body written to temp file instead of stdin heredocs — works across PowerShell and Bash

Eval suite

3 test cases covering distinct scenarios:

• Login blank screen — bug report with TypeError context
• Rate limiter units mismatch — deep technical context capture
• VS Code feature request — cross-repo issue with -R flag and remote template discovery

Measured results

Metric	With skill	Without skill	Delta
Pass rate	100% +/- 0%	93% +/- 9%	+7%
Consistent --body-file	100%	0%	Eliminates quoting bugs
Template discovery	Yes (remote repos)	No	Critical for repos with issue bots

Token cost analysis:

• 950 tokens/invocation saved by instruction reduction (v1 to v2)
• Haiku delegation moves ~50% of tool calls to a model ~10x cheaper
• Template compliance prevents rejected issues on repos that enforce templates via bots — avoiding a retry that would cost 2x tokens

Architecture

User prompt -> Skill triggers
  |-- Phase 1 (Haiku): discover templates -> structured JSON
  |-- Phase 2 (primary model): draft from JSON + post via gh CLI

Test plan

• Verified skill triggers on natural language issue requests
• Confirmed gh issue create --body-file tempfile works across PowerShell and Bash
• Eval pass rate: 100% with skill vs 93% baseline
• Cross-repo template discovery tested (microsoft/vscode feature_request.md)

[miroslavpojer]

Nice idea.

[lsulak]

LGTM, although, why not just just adjust the existing create-issue skill we have on the other place?