Dale

A compact development toolbox for Codex: think one-on-one, learn from recent Codex task history, reframe hard problems through rigorous lenses, build a trusted repository index, synthesize a task graph that adapts while it runs, keep repeatable agent work moving in a verified loop, put consequential work through a resident review loop, prove technical claims against direct evidence, or turn complex material into an effective standalone HTML visualization.

Install

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

dale is an interactive TUI installer built on the grok-build pager stack; headless flags (install, update, list --json, uninstall, with --yes, --codex-home, --update-url, --agents-md) exist for agents and CI — see tui/README.md.

Skills

$dale-brainstorm

A focused conversation between you and the current Codex agent. It keeps the work in exploration mode instead of rushing into a plan or implementation.

• stays one-on-one by default;
• separates evidence, assumptions, preferences, and unknowns;
• challenges the strongest option instead of agreeing automatically;
• proposes at most one visible evidence task, only when a separable research question emerges and you explicitly approve it;
• routes an approved evidence task to the cheapest sufficient GPT-5.6 model.

$dale-coach

Runs a read-only retrospective over recent Codex tasks using visible GPT-5.6 Luna tasks with max thinking.

• defaults to Codex tasks updated during the rolling last 30 days, newest-first;
• automatically surveys up to the app's current 50-task discovery limit;
• accepts up to 100 explicit task ids and splits them across two Luna Max batches when needed;
• reads every selected task at least once, then deep-reads the sources needed to support or reject candidate patterns;
• separates user-controlled choices from Codex behavior, platform limits, and unknown causes;
• returns at most three evidenced changes, one reusable move, and one bounded experiment — or honestly reports that no material change is supported.

$dale-lenses

Applies one primary reasoning lens and, only when it can change the conclusion, one challenging lens to a bounded difficult problem.

• selects causal, systems, temporal, conservation, inversion, constraint, contradiction, or decision-theoretic reasoning by problem signature;
• requires a method-specific artifact instead of a tour of frameworks;
• fixes a falsifier or stop condition before interpreting new evidence;
• exposes the assumptions and observations that can reverse the conclusion;
• stays in the current task and does not silently plan or execute.

Its result is a compact crux map: model, evidence boundary, risky prediction, decision impact, and remaining unknowns.

$dale-index

Turns the current repository into six evidence-backed project primitives:

Primitive	Source of truth for
REQ.md	Product intent, scope, requirements, and acceptance criteria
CONTEXT.md	Architecture, repository map, contracts, and constraints
STATE.md	Current objective, active work, blockers, risks, and next actions
TDD.md	Test strategy, commands, quality gates, and missing coverage
DESIGN.md	Visual language, primitives, states, and accessibility
DECISIONS.md	Durable decisions, alternatives, evidence, and consequences

Dale launches three parallel, read-only discovery tasks, then sends their raw reports through an adversarial evidence verifier. The calling task is the sole integrator and writes only claims that survive verification.

Each visible task is routed independently across GPT-5.6 Luna, Terra, and Sol. Repository inventory does not default to frontier-level reasoning.

$dale-graph

Builds a unique execution graph directly from the request. It does not choose a fixed diamond, pipeline, debate, or other remembered template.

1. Derives observable outcomes, artifacts, evidence boundaries, and ownership.
2. Creates only the nodes and edges justified by the work.
3. Forks the calling Codex task when available, inheriting its completed history and project context.
4. Sends the active request and exact node contract into every fork because an unfinished turn is not part of inherited history.
5. Routes every node to the cheapest sufficient GPT-5.6 model and reasoning effort.
6. Lets every visible node orchestrate one bounded layer of node-local subagents inside its own authority and write scope.
7. Rebuilds the graph when runtime evidence changes dependencies, proof needs, or required work.
8. Verifies, rejects, revises, and integrates surviving artifacts into one result.

The outer graph remains visible and user-owned. Node-local subagents cannot create Codex tasks, expand authority, or spawn another agent layer.

$dale-loop

Designs the smallest closed loop that can keep agent work moving without making the user carry state between steps:

observe -> decide -> dispatch -> verify -> persist -> wait or stop

The core skill chooses a focused variant when the work has a clear shape:

Skill	Loop shape
$dale-loop-project	Split a multi-PR project, review each head, and carry clean changes through merge
$dale-loop-pr	Watch and repair one pull request until it is merge-ready or blocked
$dale-loop-repo	Run a conservative recurring repository-maintenance heartbeat
$dale-loop-watch	Monitor a changing source and react only to meaningful changes
$dale-loop-goal	Keep one linear task moving until its completion predicate passes

Every variant defines objective verification, persistent state, resource limits, and human gates. Live systems outrank stale task reports, and an agent's own completion claim is never sufficient proof.

$dale-max

Runs a fixed high-assurance control loop for consequential work. Unlike Dale Graph, which creates its topology on the fly, Dale Max preserves the topology below while generating reviewer lenses, proof obligations, and model routes from the actual task.

1. The coordinator plans the outcome and direct proof.
2. A persistent forked Codex Worker thread produces the artifact, receives every revision, and may spawn its own task-local subagents.
3. A persistent forked Codex Reviewer 1 / Lead Reviewer thread creates fresh Reviewer 2..N subagents with spawn_agent only for material, task-specific failure modes.
4. A fresh Plan Reviewer checks the initial plan, then another fresh instance checks plan drift against each actual result.
5. The coordinator synthesizes evidence and applies a PASS, REVISE, REJECT, or BLOCKED gate.
6. Failed criteria loop back to the same Worker; a passed result is compressed into the shortest complete handoff.

Green nodes in the diagram are resident forked Codex threads; white nodes are ephemeral phases or spawn_agent subagents. Every green thread receives explicit authority to spawn one level of task-local subagents and must report that the capability is available before its artifact can pass. “Turn into haiku” means concise editorial compression, not a literal poem unless requested. Dale Max is explicit-only because it runs an exhaustive review loop with no preset cost, token, reviewer-count, or cycle budget.

$dale-proof

Verifies an existing implementation, fix, migration, deployment, runtime behavior, or compatibility claim without turning verification into a remediation or independent-review loop.

• splits broad success statements into atomic material claims;
• works backward from each claim to the shortest decisive primary signal;
• distinguishes declared, implemented, exercised, integrated, and observed behavior;
• reconciles conflicting evidence by directness, target, revision, freshness, reproducibility, and coverage rather than majority vote;
• returns scoped PROVEN, PARTIAL, UNPROVEN, or CONTRADICTED verdicts.

Missing runtime or user-visible evidence stays missing: source inspection and green secondary checks cannot manufacture a full pass.

$dale-visualize

Turns code, architecture, research, plans, comparisons, reports, incidents, design systems, concepts, and editable decisions into a self-contained HTML artifact.

• selects one primary pattern and at most two supporting patterns from Anthropic's MIT-licensed html-effectiveness examples;
• adapts the information shape instead of copying fictional sample content;
• defaults to standalone HTML with inline CSS, SVG, and minimal JavaScript;
• uses interaction only when it improves understanding or enables a decision;
• preserves source evidence and labels unknowns instead of inventing data;
• verifies desktop, mobile, keyboard, interactions, and runtime errors in a real browser before delivery.

The bundled gallery includes patterns for comparisons, code review, module maps, design systems, prototypes, SVG figures, flowcharts, explainers, plans, reports, incidents, slide decks, and small editing interfaces.

GPT-5.6 routing

Model	Default role
Luna	Bounded discovery, inventory, extraction, and narrow checks
Terra	Everyday implementation, debugging, integration, and normal proof
Sol	Difficult cross-layer synthesis and consequential ambiguity

The live Codex tool schema is authoritative. Dale shows the selected model, reasoning effort, and concrete routing reason before dispatch. Sol at xhigh, max, or ultra is never inherited merely because the coordinator uses it.

Dale Lenses and Dale Proof stay in the current task by default. They do not create visible tasks or hidden subagents merely to multiply perspectives or make a verdict appear independent.

Principles

• visible Codex tasks for top-level ownership;
• node-local subagents for bounded inner parallelism;
• runtime graph synthesis instead of predefined orchestration shapes;
• explicit read/write scopes and conflict edges;
• evidence before synthesis;
• falsifiable models before confident conclusions;
• verdicts no stronger than their direct evidence;
• verify first, integrate second;
• no Git mutation unless the user requested it.

Repository layout

.codex-plugin/plugin.json
skills/
  dale-brainstorm/
  dale-coach/
  dale-lenses/
  dale-index/
  dale-graph/
  dale-loop/
  dale-loop-project/
  dale-loop-pr/
  dale-loop-repo/
  dale-loop-watch/
  dale-loop-goal/
  dale-max/
  dale-proof/
  dale-visualize/
assets/
  dale-banner.png
  dale-icon.png
  dale-logo.png
  dale-brainstorm.png
  dale-lenses.png
  dale-index.png
  dale-graph.png
  dale-loop.png
  dale-proof.png
  dale-visualize.png

Select a skill from Codex or invoke it explicitly with $dale-brainstorm, $dale-coach, $dale-lenses, $dale-index, $dale-graph, $dale-loop (or one of its five focused variants), $dale-max, $dale-proof, or $dale-visualize.

License

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