Real output from a Pi run — not a mock dashboard:

⊗ taskflow self-improve  6/7 · blocked · $0.095
    ✓ discover            agent   deepseek-v4-flash  10t ↑38k ↓6.7k $0.011
  ┌ ✓ write-runner-tests  agent   claude-sonnet-4-6  10t ↑13  ↓6.6k $0.020
  ├ ✓ write-store-tests   agent   claude-sonnet-4-6  10t ↑11  ↓10k  $0.018
  ├ ✓ write-agents-tests  agent   claude-sonnet-4-6  10t ↑28  ↓13k  $0.030
  └ ✓ fix-stability       agent   claude-sonnet-4-6  10t ↑13  ↓3.9k $0.012
    ✓ verify              gate    BLOCK 3 type errors in test files
    ⊘ report              reduce  skipped · Gate blocked  ↳ fix-stability

The layout is the DAG: parallel rails are concurrency, long edges are dependencies, and the gate states why downstream work stopped. No separate control plane needed to read it.

Nine operations answer questions about a graph for zero model calls — plan · verify · compile · lint · ir · trace · replay · why-stale · analytics. You can price a run before spending on it, and re-ask what happened after it ends.

The compounding layer for multi-agent work. Declarative DAGs, statically verified, executed in isolated subagents, resumable across sessions, replayable without tokens, recomputed from the smallest stale frontier.

verify before spend  ·  replay without tokens  ·  recompute the stale frontier

repo · docs · examples · changelog

One project, gone deep. Everything below is smaller by design.

Following me is a subscription, so here is what it buys.

Watch taskflow releases for the shipping version of this list.

Getting something to run once is only principal. Real completion leaves interest: a contract for what done meant, evidence of what actually happened, and an honest note about where uncertainty remains. Chat is excellent at momentum, and momentum evaporates — two weeks later, why the fifteenth revision happened is buried under the first fourteen.

So I put the compounding part in artifacts instead of transcripts: the interface that creates the work should not be the only place that can understand it.

Two rulers I actually use when choosing:

01  Evidence outlives confidence.
02  Constraints give a tool its character — omissions as deliberate as features.

More: Work should outlive chat · Trust comes from boundaries · Small tools, sharp edges

Things I built for myself. Listed because they're where the ideas got tested first — not as products, and not all of them public.

• Hermit (private, personal daily driver) — a governed local kernel for long-running work: permissions with shape, receipts for what happened, recovery after failure. It's how I found out which parts of "bounded authority" survive contact with my own impatience; taskflow's gates and effect model owe it a lot. Built for one user, so the source stays closed by choice — not a product, and not a stalled open-source plan.
• home-compass — Beijing housing scorecards where "affordable on paper" ≠ "comfortable to live with."
• selffield — a personal mirror grown from evidence and tension, with no personality labels.
• cli-lab (private) — local CLI experiments consolidated into one monorepo. The standalone snapshots in my repo list are the archived predecessors: kept for history, not maintained.

I work at MiniMax in Beijing, between agent infrastructure and full-stack products. Day languages: TypeScript · Python · Node.js · Vue — but the language I care about most is the one between a system and the person trying to understand it.

Engineering is practice; writing is necessity. What I ship publicly, I back with evidence; what I keep private, I say is private.