Datacode · engineering workflow

How a goal becomes merged code

Claude plans and judges. Cheap open-model coders on Devin, Cline and Command Code write the code, each slice in its own worktree. Opus checks the whole result once at the end, and nothing merges until that check passes.

Opus: plan, advise, verdict Sonnet: lead, scout, fallback builder Coding lanes via agent-try

The lean pipeline

Goal /dev · /lean · /ggoal · Conductor SONNET · LEAD Step 0 · triage trivial LANES agent-try --tier mechanical SONNET · LEAD Review diff, test, commit nonzero exit → lead does it real work SONNET · OPTIONAL dev-scout · recon map OPUS dev-spec · plan once criteria · non-goals · tiered slices slices by parallel_group EACH SLICE · OWN WORKTREE · GROUPS IN ORDER · 2 RUN SLOTS SCRIPT wt-slice add agent/<task>/<slice> LANES agent-try ladder fast --verify only exit 0 SONNET · LEAD Review slice diff vs its criteria exit 10/11/20/2 SONNET dev-builder last-resort coder accept SCRIPT wt-slice merge → _integration branch commit all slices merged OPUS · ONCE dev-verify full suite · every criterion · diff FAIL · blocking list PER-SLICE ESCALATION 1 re-run agent-try + blocking 2 dev-advisor (opus) consult 3 dev-builder (sonnet) 12 rounds → STUCK report re-enter that slice PASS DONE PR from _integration headless: merge main + push HUMAN GATE NEEDS-APPROVAL prod deploy · spend · destructive escalates
One run, start to finish, with no pauses between slices. The lead only routes and reads verdicts; it never reads source or runs the suite itself. The full test suite runs once, inside dev-verify, on the integrated branch.

The coding ladder inside agent-try

One ladder for every tier. Rungs alternate lanes so a capped lane gets skipped without wasting a call. After two real task failures (--max-failures 2) the ladder stops and Sonnet builds the slice. A mechanical slice stops at rung 6.

free
0devinswe-2-highfree promo
1clinecline-free/deepseek-v4.1-flashfree, daily cap per model
2clinez-ai/glm-5.3-flashfree, daily cap per model
paid quota · cheap open coders
3cmddeepseek/deepseek-v4.1-flashboosted credits
4devinglm-5-3-high1× quota
5cmdzai-org/glm-5.3standard
6devindeepseek-v4-flash-high0.5× quota · mechanical stops here
7cmdmoonshotai/kimi-k2.7-codepremium
8devinkimi-k2-71× quota
9cmdmoonshotai/kimi-k3premium
10devinkimi-k3-low4× quota · heaviest rung
last resort
—claudeSonnet via dev-builderlogged as claude-fallback

Who runs what

RoleModelWhen
Lead orchestratorsonnetEvery run. Routes and reads verdicts.
dev-scoutsonnetBroad or unfamiliar goals, before the spec
dev-specopusOnce per run
Slice buildslanesEvery slice, all tiers
dev-buildersonnetLadder exhausted, or third failure
dev-advisoropusSecond failure, merge conflict, before STUCK
dev-verifyopusOnce, on the integrated tree

There is no Opus builder. Model names are aliases only, never dated ids.

agent-try exit codes

ExitMeaningNext
0Done, fast check green, committed on the slice branchLead reviews diff, merges
10Ladder exhausted on a usage capSonnet builds
11Two models failed the taskSonnet builds, advisor if the check looks wrong
20Refused: unsafe directoryClaude does it
2Config: not logged in, not a repoFix setup

Tiers

TierEffort · turnsFast check
mechanicallow · 40lint/analyze changed files
standardmedium · 100the slice's test file
complexhigh · 150slice tests + typecheck
essentialhigh · 150slice tests + typecheck; Opus verify is the safety

The tier changes effort and turn budget. It never changes the model class.

Safety floors

  • Coding agents run with full permissions, but only inside their own isolated git worktree. The live checkout is never written.
  • Deny rules block force-push and rm -rf ~. Lanes never push; the lead does.
  • A failed rung is reset to its base commit before the next model starts, and gets the previous failure output as evidence.
  • dev-verify diffs the integrated branch against BASE_SHA to catch weakened tests and other reward hacking.

Other entry points

  • FULL tier (opt-in: ultra, panel, full pipeline, /orchestrate): dev-pm → dev-planner → dev-builder → dev-qa → three dev-judges who must agree. Use it for auth, payments, migrations or releases.
  • /ggoal splits a goal into backlog tasks across projects and starts conductor-drive. The Conductor runs this same lean pipeline headless and messages a human only for a gate, STUCK, or DONE.
  • Explicit /lean, /dev or /orchestrate always runs the named pipeline, even for small tasks.