Sentry Skills boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run examines Sentry Skills as a package of model-consumed workflows, reference material, subagent prompts and deterministic helper scripts for Sentry employees. Its intended use is engineering and writing assistance, including maintaining the skills themselves. It is an extension or tool mechanism, not an enclosing agent runtime. The functional boundary includes the complete repository artifact, including its shipped authoring and evaluation paths. The operational boundary is complete artifact, partial loop: the package supplies instructions and helpers, while external agent hosts interpret them, issue model calls and execute tools. The run does not claim to cover those hosts' execution loops or enforcement behavior. SRC-1, README.md, agents/code-simplifier.md, skills/skill-writer/SKILL.md.

Agent skills for Sentry employees, following the Agent Skills open format. --- README.md

Restart Claude Code after installation. Skills activate automatically when relevant. --- README.md

The shipped entry chain is the marketplace entry with source ./, the root plugin manifest, and repo-root skills/ and agents/. The README also declares installation through skills.sh for compatible agents. Local .agents/skills mirrors the canonical skill tree; CLAUDE.md is a symlink to AGENTS.md. These are distribution and discovery surfaces, not independently implemented runtimes. SRC-1, .claude-plugin/marketplace.json, .claude-plugin/plugin.json, README.md, .agents/skills, CLAUDE.md.

The canonical skill source for the sentry-skills plugin lives at repo-root skills/. .agents/skills mirrors that tree for local agent tooling, and the root plugin manifest loads repo-root skills/ and agents/. --- README.md

The included maintenance loop routes repository changes through skill-writer, which specifies synthesis, iteration, authoring, description optimization, registration and structural validation. Its registration guidance and Python validator are shipped responsibilities, not excluded callers. The evaluator configuration and scenario also belong to the artifact: they install the local skill for Codex, copy repository material into an isolated workspace and judge generated artifacts. AXIS and Codex themselves are external dependencies, not implementations frozen by this run. SRC-1, AGENTS.md, skills/skill-writer/SKILL.md, skills/skill-writer/references/registration-validation.md, skills/skill-writer/scripts/quick_validate.py, skills/skill-writer/EVAL.md, skills/skill-writer/evals/axis.config.json, skills/skill-writer/evals/scenarios/small-inline-workflow.json.

ALWAYS use /skill-writer — it handles requirements, writing, registration, and validation. --- AGENTS.md

The validator is a structural check. It should fail only for invalid skill format or missing referenced files. The size warning requires author judgment, not a machine gate. --- skills/skill-writer/references/registration-validation.md

A representative consumption loop is iterate-pr: the skill tells the host agent to fetch feedback and checks, fix code, verify locally, commit and push, then monitor again. The bundled monitor implements GitHub CLI polling and terminal markers; it does not implement the model's code-repair decisions. The agents-md skill similarly supplies instructions to persist project guidance in AGENTS.md and optionally a CLAUDE.md symlink, but subsequent host loading remains external. SRC-1, skills/iterate-pr/SKILL.md, skills/iterate-pr/scripts/monitor_pr_checks.py, skills/agents-md/SKILL.md.

1. if changed, verify, commit, push, restart loop --- skills/iterate-pr/SKILL.md

2. If a Claude-compatible entrypoint is required, symlink CLAUDE.md to AGENTS.md. --- skills/agents-md/SKILL.md

Artifact coverage at boundary declaration

The tracked tree was inventoried before classification. Included means inside the functional evidence boundary, not that every file has already been read. Initial inspection traces installation, model-consumed entry points, maintenance, deterministic helpers and an evaluator; leaf-by-leaf behavioral analysis remains later work.

All coverage claims above refer to SRC-1 and the corresponding commit-relative paths. No material repository subtree is deliberately excluded. The partial-loop designation concerns absent enclosing-host implementations and external service execution, not a narrowed selection of repository skills.

Frozen evidence and limits

The frozen source is https://github.com/getsentry/skills at commit d18b7aa8ba878354e5c348310230e652f7690f9c; its read-only access root is the absolute checkout path in frontmatter. HEAD was checked against the acquired pin. The applicability cutoff is 2026-10-14, the supplied run date; it is not a claim that the commit was authored or released on that date. The overall tier is code-grounded, supported by inspected Python implementation and configuration, alongside prompt and maintenance doctrine. No target code, tests, evaluators or examples were executed.

External host runtimes, model providers, plugin/skills installers, AXIS, Warden, GitHub and Sentry/Notion or other service integrations are not separately captured. Their absence prevents conclusions about actual host loading, model compliance, permission enforcement, service-side effects, operational reliability or evaluator effectiveness. Host/tool prerequisites vary by skill; for the inspected PR workflow, authenticated gh, uv and a target repository are explicitly required. SRC-1, README.md, .claude/settings.json, warden.toml, skills/iterate-pr/SKILL.md, skills/skill-writer/EVAL.md. Linked upstream documentation supplies no independent frozen evidence here. Unread available repository files are not access gaps. No execution traces, stored evaluator results or causal experiments have been inspected or registered.

Source register