delegation-kit

User-directed model delegation for Claude Code and Codex. Configure the models you want through existing native adapters or any supported OpenAI-compatible text endpoint. No kit qualification or benchmark is required to use a model.

Every dispatch requires an explicit user request. Authorization never silently extends to retries, fallback, review or more workers. The lead verifies results before integration. Review is optional by default, with required and cross-family policies available.

Quick start

./install.sh
# Inspect and edit the personal configuration, then validate it:
delegation-config show
delegation-config validate
delegation-config apply
delegation-route resolve --lane builder --json

The installer initializes the personal configuration only if absent and backs up a legacy installation before migration. Profiles live outside projects at ${XDG_CONFIG_HOME:-$HOME/.config}/delegation-kit/config.json, with an explicit DELEGATION_CONFIG_FILE override. Never put credential values in profiles.

Configuration examples and migration include an existing provider and a local endpoint, parameter support, model identity, review policies and output schemas. Selection is read-only and grants no permission. Run a selected profile only when the user explicitly requests it:

delegation-run --profile PROFILE --lane builder \
  --prompt-file /tmp/task.txt --output /tmp/result.txt --workdir /path/to/repo

Capabilities and evidence

Native adapters retain sandbox and file-access controls. Text adapters only send the supplied prompt and return text or a patch; assigning a builder role cannot grant worktree access. No automatic provider substitution or retry.

Metrics distinguish requested model identity from provider-reported identity. An absent identity is usable as requested-only; an incompatible response model fails unless its alias was explicitly configured. Provider reports are not independent certification.

The kit ships no routing gates, benchmark evidence, or qualification machinery. config/presets.json holds the starting profiles, adapters own their roles and efforts, and evaluation/ is a frozen archive of earlier qualification runs that no command reads.

Current release: 0.26.1

Version 0.26.1 corrects the 0.26.0 compatibility snapshot. Version 0.26.0 removes every dispatch timeout the kit imposed on its own: a run has no time limit unless the profile sets timeout, which now bounds the whole run on every adapter, native CLIs included. The --allow-provisional no-op is gone. Version 0.25.2 trimmed the kimi-k3 preset to the roles its runner implements. Version 0.25.1 fixed the installer so a key file the uninstaller backed up is restored instead of asked for again, and cleans the last gate-era wording out of the configuration docs. Version 0.25.0 removed the qualification machinery that 0.24.0 had demoted to history: the central and executable routing gates, the executor contract, the evidence snapshot and Epoch importer, the schema compiler, and the runners' --evaluation and --preflight-only modes are gone, together with roughly 13,000 lines of code, records, tests, and documentation. The starting profiles now live in config/presets.json; each runner's check reports the adapter's roles and efforts; delegation-route is a read-only discovery over the personal configuration. Explicit per-dispatch authorization, sandboxes, model identity checks, and the patch trust boundary are unchanged. See CHANGELOG.md for details and full history.

Documentation

• Personal configuration, adapters and migration
• Routing policy (advisory)
• External executors: adapters and their limits
• Compatibility and release verification
• Adapting the kit

Testing

./run-tests.sh
npm test

The suites use isolated fixtures and simulated providers. Tests do not authorize or prove live provider behavior. Releases and live paid tests are separate work.

npm package

npx delegation-kit runs the installer from a verified release checkout. The npm package is an installer wrapper, not a second runtime implementation.

License

MIT. See LICENSE.