Dale

A focused toolbox for Codex: think through a direction, map a repository, delegate independent work, verify the result, and resume work when needed.

Install and update

curl -fsSL https://borkiss.net/dale-install.sh | sh

Run dale for the interactive skill picker. For an existing installation:

dale update --yes
dale list --json

Restart Codex after updating. The updater supports plugin-cache and standalone skill installations and preserves the previous plugin version for rollback. Headless options and the installer implementation are documented in tui/README.md. Updating skills does not require replacing your global AGENTS.md.

Execution in 0.5

Dale uses the calling agent and subagents by default. Delegate when a work unit can progress independently and saves time or supplies meaningful independent evidence. Small or tightly coupled work stays local. The coordinator works on a complementary part, integrates accepted artifacts, and verifies the final target.

A graph is a set of outcomes, dependencies, and evidence obligations. It does not require a sidebar task for every node. Workers inherit the chosen model and reasoning by default, including Astra when it is the calling model. Explicit user routes override inheritance; economical routing is optional and uses the live model catalog rather than a hard-coded generation or maximum effort.

Each worker has an observable output, minimal inputs, a read/write boundary, and an evidence requirement. Overlapping mutations are serialized or explicitly isolated. Subagents share the checkout by default. The coordinator schedules one delegation level by default and checks the integrated result.

Skills

Select a skill in Codex or name it in your request. Coach and Max remain explicit-invocation workflows. Each skill is independently installable; local references belong to that skill and do not require another Dale skill.

Repository knowledge

Index completes or refreshes an established primitive directory in place, or starts a new set in the repository root:

• REQ.md: intent, scope, requirements, acceptance criteria.
• CONTEXT.md: architecture, contracts, repository map, operations.
• STATE.md: current work, blockers, risks, next actions.
• TDD.md: actual test strategy, commands, quality gates, coverage gaps.
• DESIGN.md: existing UI language and states, or explicit non-applicability.
• DECISIONS.md: evidenced decisions, alternatives, consequences, and dates.

An incremental refresh checks changed claims and affected consumers, including uncommitted work. Missing baselines trigger broader inspection. Discovery is read-only; the coordinator writes and checks the final documents. There is no mandatory worker count or saved-project requirement.

Evidence and lifecycle

Max retains independent review as an acceptance gate. Its worker may be the calling agent or a scoped subagent; additional reviewers need a concrete uncovered risk. If independence is unavailable, useful work can continue but the Max gate cannot be reported as passed. Review and testing stop when the required evidence is sufficient; changed artifacts invalidate affected evidence.

Proof distinguishes declared, implemented, tested, and observed behavior. A worker’s completion message or green local checks alone do not prove a deployed or integrated result. Report what changed, decisive evidence, and limitations; full orchestration ledgers are available when requested.

Loops persist source identities, actions, evidence, and next conditions. They re-observe on wake and do not assume subagents survive between runs. Unchanged monitoring stays quiet. Existing user authorization persists; the workflow does not invent authority for publication, merge, deployment, or external messages.

Design basis and validation

The Astra adaptation follows the official prompting guidance: explicit delegation conditions, proportionate verification, focused instructions, and continuation through already-authorized work. Skills retain outcome and ownership contracts while avoiding compulsory orchestration machinery.

Validate skill frontmatter and references, then exercise realistic scenarios: small direct changes, independent work with a shared contract, incremental indexing, unavailable review, monitoring-only requests, and mid-run corrections. Compare correctness, rework, unnecessary questions, elapsed time, and usage when available. Static validation alone does not establish speed or cost improvements.

License

Dale is licensed under the Mozilla Public License 2.0. Bundled third-party examples retain their source license notices.