A verified build workflow for AI coding agents.
Claude Code:
/plugin marketplace add tyroneross/build-loop
/plugin install build-loop@build-loop
Codex:
codex plugin marketplace add tyroneross/build-loop
codex plugin add build-loop@build-loopRestart the host after installation or updates.
Claude Code:
/build-loop:run add billing settings with tests
/build-loop:feedback describe what went wrong
Codex:
$build-loop add billing settings with tests
$build-loop tests pass locally but fail in CI
Build Loop assesses the repository, plans the work, executes it, reviews the result, fixes failures, and records reusable lessons. It resumes relevant incomplete work automatically and preserves unrelated work before starting fresh.
It asks for confirmation only before a production release, an irreversible deletion, or a major user-impacting decision.
Use npm only for an exact pin or a host without plugin marketplace support:
npm install -g @tyroneross/build-loop@0.47.0
build-loop-install --host allDurable memory (decisions, lessons, retrospectives) resolves through one function: memory_store_root(). Resolution order is $BUILD_LOOP_MEMORY_STORE_ROOT (aliases $BUILD_LOOP_MEMORY_ROOT, $AGENT_MEMORY_ROOT), else the existing ~/dev/git-folder/build-loop-memory checkout if present, else ~/.build-loop-memory.
Set $BUILD_LOOP_MEMORY_STORE_ROOT to a private directory per run or benchmark attempt — this is the sandbox mechanism, and it always wins even from a throwaway workdir.
Throwaway workdirs never write to the canonical store automatically: a run under an OS/CI temp directory, a benchmark harness run under data/local/bench-runs/, or a generic-named git repo (ws, tmp, test, scratch, ...) with no remote configured are all detected and redirected to .build-loop/memory-sandbox/ inside that workdir. Setting $BUILD_LOOP_MEMORY_DISABLE=1 forces the same redirect from anywhere. Every skipped canonical write is logged to skipped.jsonl in the sandbox; the build still completes.
Surface counts in this release: 57 skills, 29 agents, and two Claude commands.
Groundwork exchange validates .designdoc/build-request.json and records verified implementation evidence in .designdoc/implementation-map.json.
Rally verifies nothing on its own. Hosts derive agent identity with scripts/rally_point/actor_identity.py before rally enter --tool "$RALLY_TOOL".
Apache-2.0. See LICENSE.