Orchestrator injects full issue body into skill_command instead of passing only the URL

[Trecek]

Problem

During pipeline execution, the orchestrator reads GitHub issue content (via get_issue_title or by having it in context from earlier steps), then when constructing the skill_command for downstream skills like make-plan, rectify, or investigate, it embeds the full issue body into the prompt string rather than just the issue URL/number.

The downstream skills (make-plan, investigate, etc.) independently fetch the issue themselves via fetch_github_issue — they are designed to do this. The orchestrator injecting the full content is:

1. Wasteful — the issue body consumes orchestrator prompt context for no reason
2. Redundant — the downstream skill re-fetches the same content
3. Potentially stale — if the issue is edited between the orchestrator read and the skill read, the skill gets the fresh version but the orchestrator's injected copy is stale

Root Cause

The recipe step notes instruct the orchestrator to append GitHub Issue: {inputs.issue_url} to the skill_command, but the orchestrator (being an LLM with the issue content already in context) goes beyond the instruction and pastes the full issue body into the prompt instead of just the URL reference.

Affected Recipes

• implementation.yaml — plan step (skill_command construction)
• remediation.yaml — investigate step (skill_command construction)
• implementation-groups.yaml — plan step (skill_command construction)
• Any other recipe step where the orchestrator has issue content in context and constructs a skill_command that references an issue

Expected Behavior

The orchestrator should pass only the issue URL or number to downstream skills via the skill_command string. It should never copy issue body content into the prompt. The downstream skill is responsible for fetching the issue content it needs.

Proposed Fix

1. Strengthen the recipe step notes to explicitly prohibit injecting issue content — e.g., "Append ONLY the URL. Do NOT copy issue body content into skill_command."
2. Consider whether get_issue_title should be restructured to avoid giving the orchestrator access to content it doesn't need (it currently returns only title/number/slug, but the orchestrator may still have the body from earlier context).
3. Audit all recipe steps that pass issue references to downstream skills for the same pattern.