AGENT SKILL

agentic-coding-loop

Use when implementing code from a spec through a bounded build-test-fix loop with explicit acceptance criteria, automated checks, live verification when needed, and clear stop conditions.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
3

01 前言 Front Matter

name
agentic-coding-loop
description
Use when implementing code from a spec through a bounded build-test-fix loop with explicit acceptance criteria, automated checks, live verification when needed, and clear stop conditions.

03 SKILL.md 正文

Agentic Coding Loop

Use this when a coding task should move through a closed loop: understand the spec, create or identify checks, implement the smallest slice, run evidence, fix the root cause, and repeat until the acceptance criteria are met or the loop hits a real blocker.

This is not a general instruction to "keep trying." The loop must be bounded, observable, and grounded in checks that would catch the behavior users or systems depend on.

Loop Contract

Before changing code, name:

  • Goal: the user-visible or system-visible behavior that should change.
  • Acceptance criteria: what must be true before the work can be called done.
  • Evidence path: the tests, type checks, lint, API checks, browser checks, or runtime probes that will prove the criteria.
  • Budget: the maximum useful attempts, time, or escalation threshold.
  • Authority boundary: actions the agent may take autonomously and actions that require a human decision.

Derive acceptance criteria from the current governing spec, user request, and observed behavior; use code and existing tests as evidence, not as the authority for what the product should do. If criteria are missing, infer only a small, unambiguous set. Ask for a high-impact missing decision while continuing independent authorized work. A regression check should fail for the reported bug and retain a valid positive control, rather than mirror the proposed fix.

The Loop

  1. Orient. Read the relevant code, docs, tests, issue, or spec. Identify the existing patterns before adding a new one.
  2. Ratchet first when possible. If the task fixes a bug or adds observable behavior, create the cheapest check that should fail before the change and turn green after it. Use [[regression-ratchet]] and [[feature-coverage-not-just-line-coverage]] when they apply.
  3. Implement the smallest coherent slice. Prefer one behavior change over a broad refactor. Keep mechanical cleanup separate from logic changes.
  4. Run the nearest checks. Start with fast local checks, then escalate to integration, browser, API, or live verification when the behavior depends on those surfaces.
  5. Inspect failures. Treat failing checks as signal. Read the actual error, identify the root cause, and make the smallest correction that explains it.
  6. Repeat with memory of prior attempts. Each iteration must use evidence from the previous one. Do not rerun the same failing command without a code, data, environment, or hypothesis change.
  7. Close with evidence. Report what changed, which artifacts were checked, which commands or live paths verified it, and any remaining gaps.

Verification Ladder

Use the cheapest layer that can actually prove the claim:

Claim Minimum useful evidence
Pure code behavior Unit or behavior test that names the behavior
Type or interface contract Type check, schema check, or contract test
Cross-module flow Integration test or fixture that spans the boundary
UI behavior Browser check, screenshot, or interaction trace
API behavior Request/response check against the target environment
Runtime or deployment state Real artifact check, rollout state, or service probe

Passing a lower layer does not prove a higher-layer claim. A green unit test does not prove a browser flow works; a successful build does not prove the deployed artifact changed. Use [[verify-real-artifact]] and [[process-aware-done]] for final claims.

Continuous Backlog Handoff

In a multi-worker backlog, hand the tested PR, exact head, acceptance evidence, and gaps to the existing PR shepherd. That role owns review, CI, conflicts, qualified merges, and post-merge verification; the worker may take the next disjoint authorized slice within the WIP limit. One PR's CI wait does not stall unrelated workers. Preserve holds and repository-required checks, and use [[verify-delegated-work]] for policy-permitted exact-artifact evidence reuse. PR handoff alone does not complete acceptance that requires merge or runtime proof.

Unattended Worker Loops

When a script restarts a worker until its unit of work is done, keep the judgment in the model and the control flow in deterministic code:

  • Directive vs feedback. Give the worker a persistent directive, re-read every iteration, and a separate feedback channel consumed once. Rules that must outlive one iteration (decisions, scope, blocker rules) go in the directive; individual review comments go in feedback.
  • Mechanical exit. A script, not the model, decides completion: the PR exists, required checks are green on its exact head commit, no feedback or unresolved review thread is pending, and the branch contains its base's current head. A model's claim of "done" changes nothing.
  • Narrow blocker contract. Preserve all Stop Conditions below, including missing credentials or access, bounded failures, authority or scope limits, and verification that cannot be independent. Within those boundaries, repair merge conflicts, missing pins or tools, scaffolding, and failing gates; a missing contract owned by another unit of work is an external blocker. A blocked worker writes one line with a fixed prefix and stops; the coordinator verifies the claim before accepting it.
  • Isolation. One worktree, branch, and state directory per loop. Bound every external call with a timeout, and stop processes only by the exact process IDs the loop recorded.
  • Move judgment left. When the same correction recurs across loops, promote it into the shared directive, then into a script check or required gate.

Stop Conditions

Stop and report success only when:

  • acceptance criteria are satisfied;
  • evidence came from the right layer for the claim;
  • the final diff is scoped to the task; and
  • any follow-up is genuinely separate from the shipped behavior.

Stop and escalate as blocked when:

  • required credentials, dependencies, data, or environment access are missing;
  • the same failure repeats after a bounded number of distinct hypotheses;
  • the task requires a product, security, legal, financial, or operational decision outside the agent's authority;
  • the implementation would exceed the agreed scope; or
  • verification cannot be made independent of the implementation.

Call a partial result "partial" or "blocked." Do not upgrade it to done because time ran out or the latest command looked promising.

Output Shape

Use this compact closeout:

Change: <what changed>
Evidence: <checks or live paths run>
Result: <succeeded, partial, or blocked>
Gaps: <none or explicit remaining uncertainty>
Next: <only if a separate follow-up is warranted>

Anti-Patterns

  • Iterating without an acceptance criterion or budget.
  • Treating retries as progress when no hypothesis changed.
  • Adding broad refactors inside a behavior loop.
  • Accepting proxy evidence for a user-visible claim.
  • Creating tests that only assert the implementation detail you just wrote.
  • Continuing autonomously after the loop crosses a human decision boundary.

04 同仓库更多 Skills