Productize the agent-improver routine as a first-class plugin capability

[devantler]

🤖 Generated by the Daily AI Engineer

Evidence

Maintainer direction (2026-07-19, interactive session): productize the agent-improver routine. Live state shows a half-finished extraction: the automated-ai-engineer plugin already bundles an agent-improver agent and the agent-improvement skill ships in agent-skills — but the copy that actually runs daily is the monorepo-local .claude/agents/agent-improver.md, and the two have already diverged (170 vs 154 lines today). A live Claude Code session on the deployment host loads only the local agent; no plugin-prefixed improver appears, so whether the shipped agent even loads for a consumer is unverified. Improvements the routine makes to itself land in the local fork and never reach the product.

Audience & problem

Anyone deploying an autonomous engineer from this plugin (including our own two instances): they get the run-loop engineer and surveyor as a product, but the meta-engineer that keeps such a fleet improving is effectively deployment-private, undocumented as a consumable, and drifting from its shipped copy.

Hypothesis

Making the shipped improver the single source of truth — with the deployment reduced to a thin configuration shim — turns the daily self-improvement loop into product value: every deployment inherits the routine, and the routine's own evolution ships to everyone instead of forking.

Success signal

The monorepo deployment runs the shipped improver (local file is a shim or gone); a consumer following only the plugin README can stand up the routine; divergence between local and shipped copies is structurally impossible or CI-guarded. Guardrail: the deployed routine's daily runs keep working throughout.

Smallest useful change

Reconcile today's two diverged copies into one generic definition plus a deployment shim (first child below).

Design constraints

• Authority defaults to human-gated. This deployment's full-symmetric-authority (incl. the enforcement layer) is a per-deployment maintainer grant made in an interactive session — the productized default must be draft-PR / human-gated, with the grant documented as an explicit opt-in the consuming maintainer makes. Productization must not export a loosened guardrail.
• Telemetry sources are deployment-specific. The generic definition needs a documented consuming-contract section (like the engineer's Portfolio map / Trust gate sections) naming: per-instance telemetry locations, the scorecard store, and the schedule/dispatch mechanism.
• Tool-neutral per the portability principle — the contract lives in AGENTS.md shape, Claude-native files stay thin wrappers.

Acceptance criteria

• One generic improver definition of record; the monorepo deployment consumes it (shim, not fork), with drift CI-guarded or structurally impossible.
• The consuming contract (required sections: telemetry sources, scorecard store, authority grant, schedule) is documented in the plugin, and the scorecard's six-parameter schema ships with the skill.
• Hands-on verification that the bundled improver agent loads and is invocable in each supported consumer (extends the Verify VS Code discovery of bundled plugin agents and settle the agent filename convention #74 verification pattern).
• README/docs present the improver as a headline plugin capability with a worked setup example.

Rough size: M–L (ships as children; spans agent-plugins, agent-skills, and a monorepo convergence PR).

[coderabbitai]

🔗 Related PRs

#34 - docs: describe the live multi-upstream skill-sourcing model [merged]
#36 - feat: bundle in-house agent-skills into category plugins [merged]
#53 - fix: omit skills/agents path fields so plugins install in Claude Code [merged]
#55 - docs(adr): define the Automated AI Engineer plugin boundary [merged]
#63 - feat: add the automated-ai-engineer plugin (agents+skills bundle per ADR-0002) [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!