Saltar al contenido
AKOS
Todos los packs
StableAI Engineeringcoding-agents

Coding Agents — The Discipline of the Edit

Operationalizes what an agent must actually do when it reads and modifies a real repository. The governing idea: plausible is not correct. A system that writes code by predicting what code should look like produces well-formed output whether or not it is right, which destroys the single cheapest error-detection signal software engineering ever had — the visible hesitation of a person who does not know. Nothing replaces that signal except execution. A coding task is therefore done only when a command has been run against the changed artifact and its real output and exit code have been read; a description of what should happen is a hypothesis, and an unchecked success claim is not a shortcut but a defect in the highest-leverage part of the deliverable.

coding-agentsexecuted-evidenceexit-codesrepository-mappingsearch-before-editread-before-write

Principios

  • · Aplica los principios documentados en el pack coding-agents.

Patrones

  • · Find the project's own commands before writing your own. The manifest, the Makefile, and the CI config say how this repository builds and tests. A hand-rolled equivalent will differ in flags and give a different answer.
  • · Read the repository's contributor and agent-instruction files first. They are short, they are authoritative, and skipping them is how a change arrives in the wrong style with the wrong test framework.
  • · Look at three neighbours before adding a fourth. The nearest three files in the target directory tell you naming, imports, error handling, and test placement faster than any documentation.
  • · Check the working tree is clean before you start. Editing on top of someone else's uncommitted work produces a diff nobody can untangle later.
  • · Match the installed version, not the current one. Conventions and APIs come from what is in the lockfile, not from what the library's front page documents today.

Riesgos

  • · Ver anti-patrones específicos del dominio en el pack.

Checklist

  • [ ] ★ Transcript check: every claim of a verifiable state ("tests pass", "builds clean", "no other callers") is backed by a command that actually ran this session. (CA2, CAE52)
  • [ ] ★ Transcript check: the exit code of each verification command was observed, and no nonzero exit was reported as success. (CA12, CAE53)
  • [ ] ★ Transcript check: no verification was piped into another command that masks its status (cmd | tail, a wrapper whose last statement is an echo), or where one was, the failure was propagated explicitly. (CAE54)
  • [ ] ★ Verification postdates the final edit; no green result from before a later change is being offered as this change's evidence. (CAE51, CAE57)
  • [ ] ★ Every new or modified check was seen failing once on purpose before its passing result was trusted. (CA3, CAE41)
  • [ ] ★ No file was written whose current contents were not read in this session. (CA6, CAE14)

Fuentes

En la bibliografía AKOS