caspar-rewrite-subagent

name

caspar-rewrite-subagent

description

Rewrite Caspar subagent definitions into compact, contract-driven agents with clear routing, least-privilege tools, output schemas, and an auditable rewrite trail. Use when reviewing or rewriting `plugins/caspar/agents/*.md`, when the user asks to improve subagents/agent definitions, or when checking generated `plugins/caspar-codex/agents/*.toml`. Do not use for rewriting workflow skills; use `caspar-rewrite-skill` for `SKILL.md` compression.

Caspar Subagent Rewrite

Rewrite one or more Caspar subagent definitions from broad persona prompts into focused worker contracts. Produce a reviewable proposal by default; apply changes to the live plugin only when the user explicitly asks.

Inputs

• Target agent names from $ARGUMENTS (finder, analyst, patterns, reviewer, tester, dev, web-research) or explicit file paths.
• Source path: plugins/caspar/agents/<agent>.md.
• Generated Codex path: plugins/caspar-codex/agents/<agent>.toml.
• Mode:
  • proposal by default: write review artifacts only.
  • apply only if the user explicitly asks to update live definitions.

Assume cwd is the Caspar repo root. If no target is provided, ask which subagent to rewrite.

Grounding

Read these before drafting:

1. docs/tasks/main/research/subagent_definition_best_practices_053126.md - subagent best practices and practitioner findings.
2. The source agent file(s) in full.
3. scripts/translators/agents.cjs only when applying or checking Codex TOML generation.
4. docs/caspar-design-research.md only if deciding whether a block is load-bearing or generic bloat.

Source of truth is plugins/caspar/agents/*.md; generated TOML is derived output.

Procedure

1. Inventory The Agent

Record:

• name, description, tools, codex_sandbox_mode, model, and body line/token size.
• Whether the role is read-only, write-capable, or ambiguous.
• Current generated Codex name, description, sandbox_mode, and developer_instructions if present.
• Known risks from the research doc: chatty description, write access on read-only role, generic capability catalog, mandatory artifact writes, missing return contract, or context-sharing assumptions.

2. Classify The Role

Choose one primary role and make the rewrite fit it:

• locator: finds files/areas; returns grouped paths only.
• analyst: explains how code works; returns cited implementation facts.
• pattern-miner: finds reusable examples; returns 2-5 examples and recommendation.
• reviewer: identifies risks/findings; returns severity-ranked findings first.
• tester: writes or verifies tests; decide whether it is write-capable or read-only.
• implementer: writes code; may edit files and run commands.
• web-briefer: verifies current/third-party facts; returns sourced findings.

If a source agent mixes roles, flag the split decision instead of silently broadening it.

3. Block Triage

Walk the source top to bottom. For each block, assign one verdict:

• KEEP: changes behavior measurably.
• CUT: persona filler, generic expertise, tutorials, marketing, jokes, or capability catalogs the model already knows.
• CONTRACT: restate as mission, inputs, boundaries, method, or return contract.
• PERMISSION: move to tools/sandbox/model decision.
• REFERENCE: preserve only by linking to an existing reference doc.

Keep specifics that define behavior: routing triggers, scope boundaries, least-privilege tools, citations, exact output schema, empty-result behavior, verification commands, artifact paths, and "do not write" constraints.

4. Draft The Agent Contract

Use this shape unless the existing agent has a stronger local convention:

---
name: <agent>
description: <one or two sentences: trigger, boundary, return type>
tools: <least-privilege tool list>
codex_sandbox_mode: <read-only|workspace-write>
model: <model if intentionally set>
---

You are a focused <role>. <One sentence mission.>

## Mission
<What this agent produces, not a broad capability list.>

## Inputs Expected
- <files/diff/range/task/error/source constraints expected from parent>

## Boundaries
- <what to do>
- <what not to do,  /    >


 

Description rules:

• Put all trigger and "do not use when" information in the frontmatter description.
• Be specific, not long.
• Remove jokes, sales language, and second-person coaching.

Tool rules:

• Read-only roles default to Read, Grep, Glob, LS and codex_sandbox_mode: read-only.
• Web research adds web tools only.
• Implementers/test authors may use write/edit/bash and workspace-write.
• If a role sometimes writes and sometimes verifies, flag whether to split it.

Return rules:

• Require compact, structured output.
• Require file:line for code claims.
• Require links/version/date notes for web claims.
• State what to return when no findings or no matches are found.

5. Write The Observability Trail

In proposal mode, write one artifact per agent:

docs/tasks/caspar-light/proposals/agents/<agent>.rewrite.md

Create the directory if needed. The artifact must contain:

1. # Subagent Rewrite Proposal: <agent>
2. ## Inventory
3. ## Role Classification
4. ## Block Triage with | source block | verdict | reason |
5. ## Permission Decision
6. ## Proposed Agent Definition as a fenced Markdown block
7. ## Stripped Block Ledger with | stripped block | this prevented/claimed | replacement |
8. ## Load-Bearing Coverage with | item | where it lives now |
9. ## Self-Check
10. ## Open Questions

Do not overwrite live agent files in proposal mode.

6. Apply Only On Explicit Request

When applying:

1. Edit only plugins/caspar/agents/<agent>.md source files.
2. Regenerate Codex output with npm run sync-codex.
3. Verify generated output with npm run sync-codex -- --check --quiet.
4. Run targeted tests if translator/config behavior changed; otherwise run the smallest relevant verification.
5. Do not hand-edit plugins/caspar-codex/agents/*.toml except to inspect diffs.

Self-Check

For each rewritten agent, report pass/fail:

• Description names trigger, boundary, and return type.
• Body has Mission, Inputs Expected, Boundaries, Method, Return Contract.
• Read-only roles have no Write, Edit, or Bash, and use codex_sandbox_mode: read-only.
• Write-capable roles justify write access.
• No generic capability catalog remains.
• No casual/promotional language remains.
• No nested subagent delegation is requested.
• Claims require citations or source links as appropriate.
• Empty-result behavior is specified.
• Generated Codex TOML matches source when apply mode was used.

Output To User

Return:

1. Files written or changed.
2. Key rewrite decisions, especially permissions and role splits.
3. Self-check result.
4. Open questions that need human review before applying.