@retry-now/core

The engine behind retry-now — an autonomous self-improvement loop (윤회 / reincarnation) whose context is reborn at 0 every iteration.

This package is the shared runtime that the CLI and every agent integration build on. It owns the scaffold, the agent ↔ driver protocol, prompt synthesis, the agent adapters, and the reincarnation loop driver itself.

Most users don't install this directly — reach for @retry-now/cli or one of the agent integrations. Install @retry-now/core only when building your own tooling on top of the engine.

Install

bun add @retry-now/core   # or: npm install @retry-now/core

What's inside

| Area | Responsibility | |---|---| | scaffold | Creates and maintains the git-ignored .retry-now/ runtime directory | | signal / state protocol | The one-way agent → driver signal.json and the driver-owned state.json convergence counters | | prompt synthesis | Builds each life's analyze / improve prompts from the three intent prompts in the config | | agent adapters | Spawns a fresh, headless, zero-context session for opencode | codex | claude | | loop driver | Runs one life end-to-end (analyze → improve → record) and decides when the loop has consummated (converged) | | frontends | Installs the /retry-now (or $retry-now) trigger for each agent |

Public API

import {
  runDriverCli,          // entrypoint used by the CLI / agent driver-entry scripts
  installFrontend,       // install the /retry-now trigger for an agent
  buildFrontendBody,     // synthesize the trigger command body
  recoverProject,        // reconstruct a loop whose driver was killed mid-batch
  createCommandRunner,   // wrap a spawn into the test/lint runner `recoverProject` requires
  spawnVerifyCommand,    // the default shell spawn for configured verification commands
  type RecoverReport,
  type FrontendInstallResult,
} from '@retry-now/core'

recoverProject requires a commandRunner rather than defaulting one, so the module never creates a child process itself:

const { reports, code } = await recoverProject(projectRoot, {
  commandRunner: createCommandRunner(spawnVerifyCommand),
})

The engine is dependency-light and runs on Bun ≥ 1.1.

How one life runs

Each iteration starts with a brand-new ANALYZE session. It reads the code, produces a batch plan of independently revertible items, then the driver runs a fresh implementation session and a separate fresh review session for each item. A review's kept verdict is a recommendation: the driver independently re-runs configured test/lint before advancing the approved snapshot and reverts only that item on failure. Only driver-gated outcomes are recorded. Analysis, implementation, and review may each select a different CLI agent, model, and variant. The loop stops only when several consecutive lives honestly find nothing left to improve.

Repository transactions cover Git-visible files only: tracked files, non-ignored untracked files, symlinks, and modes. The approved index is restored from its exact captured bytes, then verified by HEAD, the staged tree, and file contents, modes, and symlinks; post-restoration raw index bytes may differ because Git regenerates its stat cache. Git-ignored files are outside the restore boundary. ANALYZE mutations are restored by the driver, ordinary mid-batch aborts restore the full iteration start, and an agent-created commit is left untouched while a project-level HEAD quarantine blocks reruns. Repositories containing submodule/gitlink entries are rejected before an agent is launched.

Those guarantees assume the driver is alive to enforce them, which a host restart breaks. A driver.lock still present at startup proves the previous driver was killed, so the next run transitions that run's stale status: running to interrupted and reports any uncommitted residue. recoverProject then reconstructs the batch from the per-item review signals and backups: it rolls the first unreviewed item back, re-runs the configured test/lint, proves commit attribution, and commits only the reviewed-kept prefix — refusing, with a reason, at any step it cannot prove.

See the main README for the full model, the convergence rules, and configuration.

License

MIT