Graduate student at USTC building reliable execution, orchestration, and evaluation infrastructure for AI coding agents.

Prompts determine what an agent attempts; execution, orchestration, and evaluation boundaries determine what a team can trust after concurrent work, interruption, partial side effects, or protocol change. I build those boundaries in my own systems and test the same design principles through upstream work across Pi, Codex, MCP/A2A, OpenHands, Microsoft Agent Framework, and LiteLLM.

Design rule: agent output is a proposal. Authority comes from explicit state ownership, bounded scope, and fresh evidence.

简体中文

A transaction and recovery layer for Pi Coding Agent.

I built PCH around a strict separation: models propose work; the runtime alone decides what becomes durable. A small, passive TypeScript Bridge communicates with an on-demand Host over authenticated, bounded RPC. SQLite WAL, immutable event chains, and content-addressed storage preserve the Goal -> Route -> WorkCell -> Operation -> Evidence lifecycle across restarts.

A single-agent run works directly in the canonical workspace. Multi-agent runs place role-isolated workers in credential-excluding scoped mirrors and treat every hash-bound PatchSet as an untrusted proposal. Before serial integration can change the canonical workspace, PCH checks the lease, fencing token, preimage, postimage, and a freshly executed oracle. After interruption, recovery reconciles unknown side effects before authorizing exactly one next action.

Core mechanisms: TypeScript · authenticated HMAC RPC · SQLite WAL · immutable events · CAS · leases and fencing · fault injection

Architecture · Verification evidence

A native-agent orchestration protocol for software work.

AOG encodes the workflow boundary I use when building with Codex Agents. Primary clarifies a software initiative, proposes a schema-validated module DAG with non-overlapping repository scopes, and dispatches approved modules as native Codex tasks. A bounded standard-library validator rejects cycles, ambiguous paths, duplicate keys, and scope overlap before dispatch.

It is also practical evidence of host-level familiarity: approval, task creation, worktree isolation, Goal/wait lifecycle, native subagents, and commit assembly are mapped onto Codex's actual primitives instead of simulated in a parallel runtime.

The design choice is deliberate: task, worktree, Goal, wait, and native subagent are treated as Codex-owned lifecycle primitives. AOG owns approval, dependency, write-scope, and review policy; accepted module commits are assembled topologically on a dedicated local delivery branch. It adds supervision and ownership rules without rebuilding a second Agent runtime or lifecycle ledger.

Core mechanisms: Python · stateless JSON validation · module DAGs · disjoint write scopes · native Codex tasks and worktrees · bounded subagents · topological commit assembly

Workflow · Validator tests · CI evidence

A runtime-neutral engineering decision layer for AI coding agents.

Engineer Software keeps one canonical workflow source, exposes it as a Codex plugin, and projects it into DeepSeek Harness's filesystem skill contract. Projection, loader, resource, and source-identity validators turn cross-runtime drift into a failing check instead of a documentation problem.

Rather than forcing every task through a fixed pipeline, a deterministic router selects one evidence-gated engineering mode. Routing fixtures cover activation, bypass, and legal transitions. Paired behavior evaluations run baseline and treatment conditions in matched worktrees, preserve raw events and patches, and report rubric scores, timing, and exact sign-test results without turning a small pilot into a general performance claim.

Core mechanisms: Python · Codex plugins · DeepSeek Harness skills · generated projections · deterministic routing · paired behavior evaluation

Evaluation design · Runtime compatibility

• MudBlazor #13622 preserved the non-negative page-index invariant for empty tables and added regression coverage. Merged.

Microsoft Agent Framework #7679 - deterministic parallel Foreach. The implementation adds opt-in, bounded fan-out to declarative .NET workflows while preserving sequential behavior by default. Each iteration runs as an isolated sub-workflow, buffers state and events, and commits in source-index order. Failure or timeout cancels peers and discards buffered commits. Forty focused parallel tests pass on each target framework; the 915-test unit suite passes on both net10.0 and net472.

LiteLLM #37027 - bounding an OOM-prone admin API. I traced the failure to unbounded history reads across legacy spend-log query branches. The change applies stable take/skip ordering everywhere, rejects invalid pagination before querying, and preserves legacy list and summary contracts while making deprecation and pagination explicit. The mapped endpoint suite passes 188 tests, and the current GitHub check suite is green.

MCP C# SDK #1816 - protocol evolution without shared-state leakage. A single typed response-emission boundary strips unsupported fields for legacy clients, supplies defaults for current clients, and edits freshly serialized nodes so shared result instances cannot leak state across protocol versions. Concurrent compatibility warnings are deduplicated. The validated non-Docker scope covers 2,366 core and 617 ASP.NET Core tests.

• Runtime authority: the PCH architecture documents process ownership, durable state, mutation authority, recovery precedence, and the serial integration boundary.
• Orchestration contract: the AOG workflow and plan validator define explicit approval, non-overlapping write scopes, bounded native delegation, and local topological assembly. Validator tests cover cycles, path safety, duplicate keys, input bounds, scope overlap, and deterministic normalization.
• Workflow evaluation: the Engineer Software evaluation suite combines deterministic routing contracts with matched-worktree behavior comparisons and preserves raw events, patches, timing, exclusions, and scoring boundaries.

When a unit-test boundary is narrower than the failure, I exercise the real integration path. For OpenHands browser discovery #4502, I combined focused selection-order regressions with the broader browser-tool suite, four real-browser E2E cases, and an isolated Windows smoke covering the selected executable, CDP connection, DOM state, screenshot capture, and process cleanup.

I am interested in research-oriented open-source collaboration where correctness depends on concurrency, recovery, lifecycle, protocol, or evaluation boundaries.