Explore agent-driven workflow automation

[antonis]

Summary

Exploratory research issue (not a committed feature) to investigate how much of the SDK's issue-handling and incident-response workflow we can automate with agents. Goal: map the design space, ship the cheapest safe slice, and decide what — if anything — is worth productionizing.

There are (at least) two entry points, or triggers, that both converge on the same downstream outcome: a triaged issue, a reproduction, and ultimately a fix PR.

Prior art — how sentry-javascript does it

sentry-javascript already runs this as a two-stage pipeline (.agents/skills/triage-issue, plus triage-issue.yml / auto-fix-issue.yml):

	Stage 1 — triage	Stage 2 — auto-fix
Trigger	Auto on issues: [opened], gated to Bug/Feature labels	Manual dispatch only (auto-run commented out: "until we verified how that works")
GitHub perms	Read-only	Write — branch, commit, draft PR
Output	Triage report → Linear / job summary	Draft fix PR

Mechanics: thin GitHub Action → anthropics/claude-code-action@v1 (via OpenRouter, claude-opus-4-8) running a /triage-issue slash command. The intelligence is a small skill (a SKILL.md workflow + a mandatory prompt-injection gate + a few helper scripts). Issue content is treated as untrusted data, never instructions; tools are a tight allowlist; secrets are fenced in a dedicated CI environment. Notably, even JS's stage 2 only edits code from a report — it does not reproduce anything.

Trigger 1 — GitHub Action (issue-driven)

1. Triage — port the JS skill, adapted to RN's reality (hybrid SDK: JS/native/bridge layers, New vs Old arch, iOS/Android, bare vs Expo, bundled native-SDK version). Read-only, low risk.
2. Reproduction — the hard, RN-specific part: a faithful repro often needs a real device/simulator across the New/Old × iOS/Android × Expo/bare matrix, which a cheap ubuntu Action can't stand up.
3. Fix PR (draft) — for well-localized JS-layer bugs, a draft PR + test, human-gated.

Trigger 2 — Sentry-native (crash-driven)

1. A Sentry alert/webhook fires on an SDK-attributed crash or a crash-free-rate regression (SDK Crashes / SDK health).
2. Automation opens or enriches a GitHub issue with stack trace, affected SDK/native versions, device/arch breakdown, frequency.
3. Feeds the same repro → fix-PR pipeline as Trigger 1.

Open questions: identifying a crash as SDK-caused vs app-caused with acceptable precision; trigger mechanism (Sentry alert webhook / Seer / internal tooling); noise & rate-limiting to avoid PR spam; overlap with existing SDK-health monitoring.

Both paths converge

Trigger 1 (GitHub issue) ─┐
                          ├─→ Triage → Reproduce → Draft Fix PR → human review
Trigger 2 (Sentry crash) ─┘

The reproduction + fix stage is shared and is where most of the risk and value sits.

Feasibility & proposed phasing

• Phase 1 — read-only triage (high feasibility, ~1–2 days). Near-verbatim port of the JS skill; static analysis only (grep + reasoning), never runs the app, can't write anything. This repo already has the plumbing: agents.toml trust config, the local .agents/skills/ convention, and the OpenRouter access model. A local spike is scaffolded — see checklist.
• Phase 2 — JS-layer-only auto-fix (research). Draft PRs, manual trigger, well-localized bugs only — mirrors where JS is today.
• Phase 3 — automated reproduction on the arch matrix (open question). The genuinely unsolved part, and the natural home for Trigger 2.

Scope of this research

• Evaluate porting the JS triage skill; identify RN-specific gaps (arch matrix, native bridge, Expo, bundled native
• Scaffold Phase-1 triage skill + workflow locally as a spike (not merged; secrets/ci-triage env still needed to run)
• Dry-run the triage skill against real RN issues; assess report quality
• Spike the hardest unknown — automated reproduction against the sample apps — before anything downstream
• Investigate Sentry-side trigger mechanics for SDK Crashes (alert/webhook/Seer) and SDK-vs-app attribution
• Define human-in-the-loop checkpoints and guardrails (draft-only PRs, prompt-injection gate, DSN/secret hygiene)

[linear-code]

RN-785