/internal/docs/CONTINUOUS_IMPROVEMENT.md

Continuous Improvement Loop

Go Micro is an agent harness. This file defines the autonomous loop that builds it — the framework's own thesis (an agent operating a system) pointed at itself. Claude Code drives the loop; Codex executes scoped tasks; the human sets direction and can stop or revert anything at any time.

North Star. Every increment must advance the thesis in THESIS.md: a holistic agent harness and service framework encapsulating the lifecycle of services → agents → workflows. Judge each change against it — work that doesn't move toward that lifecycle isn't an improvement, however clean.

Autonomy

Full autonomy, no approval gates. Each increment: Claude Code picks the work, implements it (or dispatches Codex), opens a PR, and merges it — including reviewing and merging Codex's PRs. The only gate is correctness: go build, go test, and golangci-lint must be green (that's not an approval, it's not shipping broken code).

Transparency replaces approval: every increment ends with a one-line digest, and every change is a small, reversible, single-concern PR the human can revert.

What counts as an improvement

Grounded in real signal, never speculative rewrites. Each cycle draws from:

1. Roadmap — the Now/Next items in ROADMAP.md (harness depth: durable runs, observability, streaming, human-in-the-loop; hardening: resilience, conformance).
2. Open issues — the scoped backlog (e.g. #3010–#3014).
3. Improvement radar — a scan each cycle for: missing/weak tests, lint or quality issues, docs/code drift, and DX friction.
4. Dogfooding — actually build with the harness (micro new → run → chat, an agent + a flow) and fix what hurts. Friction found here is high-signal.

The cycle (one increment)

1. Sync master.
2. If a Codex PR is open and CI-green → review (diff + gates + correctness vs its issue) and merge it.
3. Else pick the single highest-value item from the sources above.
4. Implement it, or dispatch to Codex (@codex <instruction> on the issue) if it's a well-scoped chunk and Codex is free. Codex is serial — one task at a time.
5. Verify build/test/lint locally.
6. Open a PR (one concern) and merge it.
7. Post a one-line digest; refresh the backlog from the radar.

Roles

• Claude Code — orchestrator, implementer, reviewer, integrator, merger.
• Codex — serial builder for well-scoped chunks, dispatched via @codex.
• Human — sets direction; owns brand/positioning copy and breaking public-API decisions; can stop or revert anything.

Guardrails

• One concern per PR; small and reversible.
• Stay on claude/* branches (Codex on codex/*); never two agents on one branch; base PRs on master (don't stack on an in-flight branch). See CODEX.md.
• Off-limits without the human: brand/positioning/marketing copy, breaking public API changes, product-default changes with broad behavioral impact, new dependencies, architectural rewrites. The loop proposes these in the digest; it does not merge them autonomously.

Scheduling

• In-session cron (CronCreate) — runs increments while this Claude session is alive. Convenient, but the remote environment is reclaimed on inactivity and recurring jobs expire after 7 days, so it is not a durable scheduler.
• GitHub Actions (durable) — a scheduled workflow that runs the loop independently of any session. This is the real backbone; it opens a fresh tracking issue for each increment and dispatches Codex there. It needs a CODEX_TRIGGER_TOKEN repo secret from a user account Codex responds to; without that secret the workflow deliberately no-ops to avoid ignored bot comments. See .github/workflows/continuous-improvement.yml and the mechanics below.

How the durable loop works (mechanics)

Hard-won wiring — change any one piece and the loop silently stops producing merged PRs. Each scheduled run:

1. Opens a fresh issue per increment (Continuous improvement increment #N) and posts the @codex instruction on it. Why a fresh issue: Codex derives its branch name from the triggering issue's context, so re-using one tracker issue collapses every run onto one branch name and only the first PR opens — the rest collide and silently fail.

2. Posts as a user, not the Actions bot. Codex ignores @codex comments authored by github-actions[bot], so the dispatch uses CODEX_TRIGGER_TOKEN (a PAT for a user account Codex follows). No token → the step no-ops.

3. Codex opens the PR itself with gh — never make_pr. In the Codex Cloud sandbox the make_pr tool is a no-op stub: it records the PR title/body for the manual "Create PR" button and never pushes a branch or calls the API. So the dispatch and AGENTS.md tell Codex to do it by hand:

   git switch -c codex/increment-<issue>     # unique branch, codex/ prefix
   git push -u origin codex/increment-<issue>
   gh pr create --base master --label codex --title "…" --body "… Closes #<issue>"
   gh pr merge  --squash --auto --delete-branch
   

   This requires the Codex setup script to install gh and run gh auth setup-git (so git push is authenticated) with a write-scoped token.

4. Merges via GitHub native auto-merge, gated by branch protection. master requires the CI status checks (build, tests, golangci-lint) and 0 approving reviews. gh pr merge --auto enables auto-merge; GitHub lands the PR the moment checks pass and deletes the branch. Closes #<issue> auto-closes the tracking issue. There is no merge sweep workflow — branch protection is the gate.

Do-not-break list

• Don't re-add required approvals to master — it blocks every autonomous merge. The intended gate is green CI only.
• Don't point the dispatch at one standing tracker issue — one issue per run.
• Don't tell Codex to use make_pr (or imply a token "isn't a substitute"): it cannot open a PR. gh is the only path.
• Don't manually re-implement a Codex increment during the summary→PR lag (Codex posts an optimistic "opened a PR" comment ~30–45 min before the PR actually appears). Re-doing it creates duplicate PRs and stale branches that then block the next run. Wait for the PR, or let it ride.

Stop / redirect

• In-session: CronDelete <id> (or end the session).
• Durable: disable/delete the workflow.
• Or just tell Claude Code to pause or change focus — direction always wins over the loop.