Engineer Software

A runtime-neutral, evidence-driven software engineering workflow for AI coding agents.

Choose the smallest trustworthy engineering move before changing code.

Six workflows · How it works · Real examples · Quick start · Validation · 简体中文

Engineer Software gives Codex and DeepSeek Harness two runtime entries into one canonical workflow.
For substantive software work, it selects exactly one bounded engineering mode and defines the
evidence required before an agent can change direction or claim completion.

At a glance: 6 bounded workflows · 25 deterministic routing cases · 2 runtime paths · 1
canonical source

Use it when requirements, failure mechanisms, design choices, implementation scope, structural
ownership, or acceptance evidence materially affect the result. Bypass it when the request is an
ordinary explanation, translation, simple code reading, formatting change, or already-specified
mechanical operation.

Six workflows

The modules are alternative starting points, not a pipeline that every task must complete:

Workflow	Start here when	Evidence required to leave
Shape Work	behavior, compatibility, scope, or acceptance is unclear	smallest sufficient contract and explicit exclusions
Trace Failure	a symptom exists but its cause is unknown	reproduction plus causal evidence
Probe Choice	one named design decision needs a disposable experiment	observed result and decision consequence
Deliver Change	the outcome and edit boundary are closed	focused check, implementation result, and final-state evidence
Inspect Structure	ownership, duplication, or boundaries are the question	traced owners and callers plus a boundary recommendation
Manage Work Items	the requested output is a local PRD, task set, or acceptance list	local artifact with dependencies and acceptance criteria

Example: “Checkout sometimes creates a duplicate order under load.” The skill starts with
Trace Failure, requires a reproduction and causal evidence, and only then allows a transition
to implementation and final verification.

How it works

The workflow is a small decision loop, not a ceremony-heavy pipeline:

1. The router checks whether the request is ordinary work or has material engineering uncertainty.
2. It starts one primary module and makes that module's exit evidence explicit.
3. A later module is entered only when fresh evidence closes the current module and identifies a
   different need.
4. Codex and DeepSeek Harness load the same canonical SKILL.md, references, and routing cases.

The Harness tree is generated and checked from the Codex source; it is not a second hand-maintained
workflow. See runtime compatibility for the official Harness sources,
loader contract, and compatibility limits.

Real examples

Each example follows the same shape: prompt → route → evidence required to proceed.

• “Checkout sometimes creates a duplicate order under load. Find the cause and fix it.” →
  Trace Failure → reproduce the symptom, establish the cause, then add a focused regression.
• “Build a disposable experiment to compare two state-transition models before we choose one.” →
  Probe Choice → observe the named trade-off and record the decision consequence.
• “Add the documented --json flag to the existing status command and verify the specified output
  contract.” → Deliver Change → implement the closed contract and verify the final state.
• “Explain what this function does and why it returns null here.” → Bypass → answer directly
  without adding workflow overhead.

These examples are represented in evals/routing-cases.json.
Run the deterministic routing checks without model access:

python scripts/validate_evals.py
python scripts/validate_harness.py --check
python scripts/run_routing_eval.py --limit 5

Optional live Codex evidence is read-only and environment-dependent:

python scripts/run_routing_eval.py --live --public-submission `
  --output evals/runs/local-routing-results.json

The repository checks routing, projection identity, and evidence contracts. Task-level A/B runs are
sampled behavioral evidence, not a general benchmark claim. This README does not publish a single
speedup percentage; use the paired evaluator and optional latency gate only for like-for-like reruns.
See the behavior A/B guide for the raw format and scoring
limits.

Quick start

Codex

codex plugin marketplace add KirschBluteX/engineer-software
codex plugin add engineer-software@engineer-software
codex plugin list

Start a new task after installation, then ask for a substantive software change or invoke
$engineer-software. Upgrade with:

codex plugin marketplace upgrade engineer-software
codex plugin add engineer-software@engineer-software

Remove it with the installed Codex plugin manager and confirm the result:

codex plugin remove engineer-software@engineer-software
codex plugin list

The existing Codex marketplace manifest and plugin path remain unchanged.

DeepSeek Harness

DeepSeek Harness is an official open-source project, currently marked developer preview. Its
official local skill provider scans project .dsh/skills roots. This checkout includes a generated
projection of the canonical skill. Check it, start Harness, then choose this repository as the
workspace:

python scripts/sync_harness_skill.py --check
python scripts/validate_harness.py --check
npx @deepseek-ai/dsh web

After updating a reviewed checkout or editing the canonical skill, regenerate and verify the same
project entry:

git pull --ff-only
python scripts/sync_harness_skill.py --write
python scripts/validate_harness.py --check

Remove only this generated project entry after confirming the target:

Get-Item .dsh/skills/engineer-software
Remove-Item -LiteralPath .dsh/skills/engineer-software -Recurse

A user-global copy can target $DSH_HOME/skills/engineer-software. Exact target commands,
troubleshooting, official contract sources, and the recorded loader smoke are in
runtime compatibility.

Validation

Use Python 3.9 or newer. The repository is standard-library-first; the development-only
requirements-dev.txt contains the YAML parser used by the validators.

python -m pip install -r requirements-dev.txt
python scripts/validate_project.py
python -m unittest discover -s tests -v
python -m compileall -q scripts tests

validate_project.py aggregates the plugin package, routing fixtures, Harness projection, and
documentation contracts. For a focused failure, run python scripts/validate_plugin.py plugins/engineer-software, python scripts/validate_evals.py, or python scripts/validate_harness.py --check directly. CI keeps the Python 3.9/3.12/3.13 matrix plus the
aggregate validation, unittest, and compile checks. It leaves setup-python's pip cache disabled;
the development file is installed explicitly.

Compatibility, limits, and security

Read docs/compatibility.md for the matrix, install/upgrade/remove paths,
official DeepSeek Harness links, troubleshooting, and the static-contract and loader-smoke evidence.
The short version:

• DeepSeek Harness is a rapidly changing developer preview; compatibility-breaking changes are
  possible.
• The .dsh/skills tree is a generated projection. Edit the Codex canonical source and regenerate;
  drift fails validation.
• Engineer Software is not an official DeepSeek plugin, partnership, or endorsement, and it does not
  invent a Harness manifest outside the documented filesystem skill contract.
• This project does not ship an MCP server, hook, telemetry, credential store, or background
  service. Tool permissions, API keys, and model configuration remain the user's runtime policy.
• Never commit API keys, .env files, session logs, profile state, generated temporary assets, or
  unreviewed screenshots.

GitHub is a distribution target, not a runtime route. This repository performs no issue-tracker,
telemetry, or remote workflow action when a skill is used. See PRIVACY.md,
SECURITY.md, and TERMS.md.

Contributing and roadmap

Start with CONTRIBUTING.md. Keep plugins/engineer-software/skills/engineer-software/
as the only editable workflow source, run the projection check after changes, and add routing
fixtures for new transitions. ROADMAP.md records the deliberately small next steps;
it does not promise a long-lived adapter framework.

License

Engineer Software is released under the MIT License.