Skip to content

Agent-first delivery control plane

Keep coding agents moving. Keep delivery under control.

dacli keeps critical-path work moving through your chosen coding harness—with isolated execution, exact-tree review, and GitHub delivery.

Read the operator playbook View the pitch deck Inspect the repository
go install github.com/mlnomadpy/dacli/cmd/dacli@latest v0.3.1
Bounded by policycycles, tokens, WIP, timeouts
One durable recordbriefs, runs, reviews, landing
GitHub-first deliveryissues, PRs, CI, merge truth
Human release authorityexceptions, STOP, publication

Authority, separated

Autonomy works when every participant has one job.

dacli does not replace the orchestrator or the coding agent. It keeps their work inside a process whose claims can be checked after the model session is gone.

A

Agent judgment

The orchestrator interprets goals, decomposes work, and judges evidence. Coding agents implement, test, and review inside their assigned scope.

D

What dacli controls

Task graphs, routing, claims, worktrees, budgets, checkpoints, review identity, failure policy, and the exact GitHub landing transaction.

H

What stays human

Product direction, credentials, exceptions, emergency stop, hybrid-provider authority, and every tag or public release decision.

One governed transaction

From the next critical task to evidence on trunk.

Every transition consumes recorded state and produces the next inspectable checkpoint. A refusal is an answer; unknown external state fails closed.

  1. 01
    Plan

    Select work that moves the critical path

    Dependencies, slack, priority, capacity, and path claims determine the ready frontier.

  2. 02
    Execute

    Isolate workers before they write

    Each coding agent gets a scoped brief, bounded budget, declared harness, worktree, and non-overlapping claim.

  3. 03
    Verify

    Run the contracts that changed

    Repository-aware routing chooses structured commands and working directories for the affected paths.

  4. 04
    Review

    Exact-tree review, not a generic approval

    An independent reviewer binds its verdict to identity, branch, commit, and tree. Corrections require a fresh review.

  5. 05
    Land

    Observe GitHub before recording completion

    Checks, reviews, merge state, and fresh trunk evidence must agree before acceptance and issue closure.

The record stays visible

Operate the swarm from evidence, not chat history.

The dashboard is a projection of the same durable workspace the CLI and MCP server use. Supervise loop phase, wave width, token reservations, model routing, outcome windows, worker state, dependency flow, and delivery blockers without creating a second source of truth.

  • Explain every recommendation. Rank, rejected routes, claims, freshness, and the next safe action stay attached.
  • Recover without guessing. Phase journals, handoffs, and exact remote identities make restart behavior explicit.
  • Measure improvement honestly. Adjacent windows retain sample size, task-size context, coverage, and exact drill-down evidence; missing cost stays unknown.
  • Publish the minimum. Public repositories receive public-safe fields; internal evidence needs recorded authority and `--include-internal`.
Read the dashboard guide Open representative evidence
outcome analytics
Representative dacli outcome analytics with adjacent windows, coverage, provenance, and exact evidence links
Throughput, latency, reliability, tokens, and partial provider-reported cost stay bounded by sample size, coverage, provenance, and exact drill-down evidence.
operator attention
Representative dacli operator attention queue with ranked policy state and evidence-linked next actions
Unknown GitHub state, stale evidence, and owner handoffs remain ranked conditions with explicit reasons and safe next actions—not dismissible dashboard alerts.
390px / analytics
Representative dacli analytics route at an exact 390 by 844 pixel viewport
The exact 390px capture preserves navigation, scope, phase, identity, and readable controls without page-level overflow.

Start with a bounded loop

Keep the harness explicit. Preview before launch.

Use Codex throughout this run, negotiate the installed surface, and inspect the resolved policy before any worker starts. A multi-provider run requires an explicit allowlist plus `--hybrid`.

v0.3.1 includes verified archives, checksums, and SBOMs for macOS, Linux, and Windows.

bounded-loop.sh5 commands
  1. 01dacli version --compatibility --json
  2. 02dacli capabilities --json
  3. 03dacli adopt --provision-roles
  4. 04dacli start --project <slug> --profile loop --harness codex --dry-run
  5. 05dacli start --project <slug> --profile loop --harness codex

Boundaries worth asking about

The safe default is explicit.

Does dacli switch coding CLIs when one fails?

No. A Codex run stays on Codex. Cross-CLI routing requires an explicit harness allowlist and `--hybrid`; installing another CLI does not authorize fallback.

Does it publish internal findings to GitHub?

Not by default. Public repositories receive the typed public-safe projection. Internal findings, decisions, and verdicts need separate repository authority and an explicit request.

Can it run forever?

Continuous operation is repeated bounded transactions with durable checkpoints—not an unbounded permission grant. Budgets, idle policy, no-progress halts, STOP, and human release authority remain active.

Direction in. Verified progress out.

Give the swarm room to work.
Keep the delivery process accountable.

Read the operator playbook Inspect the source