Review subagent fragments / instructions

[strawgate]

Let's review how the subagent fragments are merged and make sure that if a workflow is using subagents it has a model prompt with a default and we tell copilot explicitly to call the subagent using that exact model

[github-actions]

3 related issues

• Review agent jobs for prompt anomalies #465 — it targets prompt-level anomalies in agent jobs, which overlaps with this issue’s request to validate subagent instruction composition and runtime behavior.
• Update Claude prompts to have best practices from our gh-aw workflows #132 — it tracks aligning prompt guidance across workflows, which can inform how subagent instructions should consistently enforce model selection and defaults.
• Create a Copilot Agent instructions file #176 — it focused on deriving a Copilot instruction file from prompt fragments, which provides prior context on fragment-driven instruction assembly.

---

What is this? | From workflow: Duplicate Issue Detector

Give us feedback! React with 🚀 if perfect, 👍 if helpful, 👎 if not.

[github-actions]

tl;dr: The model default is consistently present on workflows that use subagents, but subagent-call instructions are still inconsistent; at least the self-review fragments tell agents to call runSubagent without passing a model, so they do not explicitly enforce the workflow’s exact configured model.

Recommendation

Normalize all subagent invocation guidance to always pass model: "${{ inputs.model }}" explicitly (and explicit agentType), including the self-review fragments; then recompile lockfiles and spot-check generated prompts.

Findings

1. Pick Three fragments are partially explicit but still conditional:

   • .github/workflows/gh-aw-fragments/pick-three-keep-one.md:5
   • .github/workflows/gh-aw-fragments/pick-three-keep-many.md:5
   • Both say: call runSubagent with workflow-specified values, defaulting to model: "${{ inputs.model }}" only if unspecified.

2. Pick Three, Keep All is explicit/unconditional:

   • .github/workflows/gh-aw-fragments/pick-three-keep-all.md:5
   • It directly says to call runSubagent with agentType: "general-purpose" and model: "${{ inputs.model }}".

3. Self-review fragments currently omit model in subagent call guidance:

   • .github/workflows/gh-aw-fragments/safe-output-create-pr.md:71
   • .github/workflows/gh-aw-fragments/safe-output-push-to-pr.md:73
   • Both instruct: “Spawn a code-review sub-agent via runSubagent ...” with no explicit model argument.

4. The omission is present in merged lockfile prompts:

   • .github/workflows/gh-aw-issue-fixer.lock.yml:1042 contains the same runSubagent guidance without model.

5. Workflows using these fragments already define model defaults (so this is an instruction consistency gap, not missing defaults):

   • Example: .github/workflows/gh-aw-issue-fixer.md:18 (engine.model: ${{ inputs.model }}), and .github/workflows/gh-aw-issue-fixer.md:24-28 (inputs.model default "gpt-5.3-codex").

Verification

I ran targeted repository checks (no runtime tests needed for this prompt/content triage):

$ rg -n "gh-aw-fragments/(pick-three-keep-one|pick-three-keep-many|pick-three-keep-all|safe-output-create-pr|safe-output-push-to-pr)" .github/workflows/gh-aw-*.md
# showed all workflows importing these subagent-related fragments

$ python - <<'PY'
# enumerated every gh-aw-*.md importing those fragments and checked model input defaults
PY
# result: all listed workflows reported model_input_default=yes

$ rg -n "Spawn a `code-review` sub-agent via `runSubagent`" .github/workflows/gh-aw-issue-fixer.lock.yml
# hit at line 1042, confirming merged prompt still omits model

Related prior work checked: closed PR #445 and issue #444; PR #445 states subagent guidance alignment work happened, but current fragment text still has the inconsistency above.

Detailed Action Plan

1. Update subagent instruction fragments for consistency:

   • .github/workflows/gh-aw-fragments/safe-output-create-pr.md (line ~71)
   • .github/workflows/gh-aw-fragments/safe-output-push-to-pr.md (line ~73)
   • Change wording to require explicit model, e.g. call runSubagent with agentType: "code-review" and model: "${{ inputs.model }}".

2. Tighten conditional wording in:

   • .github/workflows/gh-aw-fragments/pick-three-keep-one.md:5
   • .github/workflows/gh-aw-fragments/pick-three-keep-many.md:5
   • Prefer explicit instruction matching pick-three-keep-all.md style.

3. Recompile and validate:

   • make compile
   • make lint

4. Spot-check compiled outputs for at least one workflow per fragment group:

   • verify lockfiles now include explicit model: "${{ inputs.model }}" in runSubagent instructions.

5. Optional hardening: add a prompt-audit check (grep-based) that fails CI if runSubagent appears without an adjacent model argument in source fragments intended for direct invocation instructions.

Related Items

Type	Link	Relevance
Issue	#498	Current triage target
PR	#445	Prior subagent-guidance alignment work
Issue	#444	Parent change set that led to #445
File	.github/workflows/gh-aw-fragments/pick-three-keep-one.md:5	Conditional model wording
File	.github/workflows/gh-aw-fragments/pick-three-keep-many.md:5	Conditional model wording
File	.github/workflows/gh-aw-fragments/pick-three-keep-all.md:5	Explicit model wording baseline
File	.github/workflows/gh-aw-fragments/safe-output-create-pr.md:71	Missing explicit model in runSubagent instruction
File	.github/workflows/gh-aw-fragments/safe-output-push-to-pr.md:73	Missing explicit model in runSubagent instruction
File	.github/workflows/gh-aw-issue-fixer.lock.yml:1042	Merged output confirms omission
File	.github/workflows/gh-aw-issue-fixer.md:18,24-28	Workflow model wiring/default evidence

---

What is this? | From workflow: Issue Triage

Give us feedback! React with 🚀 if perfect, 👍 if helpful, 👎 if not.