feat: add debug symbols workflow

[thymikee]

Summary

Adds a narrow agent-device debug symbols workflow for crash symbolication in the #694 native diagnostics rollout.

• Supports Apple .ips, .crash, and log-style crash artifacts by matching UUIDs against local .dSYM bundles via dwarfdump --uuid, then running atos.
• Keeps debug scoped to symbols/crash artifact space; logs, network, perf, record/trace, and React Native internals stay routed to existing commands.
• Exposes CLI, typed client, and MCP command metadata with compact path/summary output.
• Documents Android Java/R8 and native symbolication as unsupported/deferred with recovery hints.

Refs #699
Refs #694

Touched files: 21.

Validation

Worker validation passed:

• Focused vitest: 5 files, 127 tests.
• pnpm format.
• pnpm check:quick.
• pnpm check:unit: 249 unit files, 2333 tests, 8 smoke tests.
• pnpm build.
• git diff --check.

Fixture-backed CLI evidence:

• Apple fixture wrote a symbolicated artifact and printed only the artifact path plus a short summary/proof line.
• Android Java crash fixture returned the explicit unsupported/deferred hint for mapping.txt/ndk-stack/addr2line external workflows.

SkillGym case debug-symbols-apple-crash was added but not runnable in the sandbox because external runner approval was denied.

Residual risk: no real simulator crash/dSYM artifact was generated in this sandbox; coverage is fixture-backed with command construction and artifact proof.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://callstack.github.io/agent-device/pr-preview/pr-756/
Built to branch gh-pages at 2026-06-11 07:14 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.2 MB	1.2 MB	+11.3 kB
JS gzip	390.4 kB	394.3 kB	+3.9 kB
npm tarball	501.7 kB	505.6 kB	+4.0 kB
npm unpacked	1.7 MB	1.7 MB	+12.0 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	29.0 ms	27.2 ms	-1.9 ms
CLI --help	44.6 ms	43.6 ms	-1.1 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/9542.js	+8.4 kB	+3.0 kB
dist/src/args.js	+890 B	+305 B
dist/src/1352.js	+1.1 kB	+295 B
dist/src/8173.js	+493 B	+169 B
dist/src/8699.js	+284 B	+101 B

[thymikee]

Code review

Verdict: significant issues — the headline .ips path will fail on real device-produced crash files; the rest of the design is sound.

Findings

1. Major (arguably blocker) — src/debug-symbols.ts:136: real .ips files written by iOS/macOS since iOS 15 are two JSON documents — a single-line JSON header followed by the JSON payload. JSON.parse(text) on the whole file throws, the parser falls through to the text-crash reader, finds no Binary Images: section, and dies with UNSUPPORTED_OPERATION — i.e. the primary advertised input format fails for every real device artifact. The fixture in src/__tests__/debug-symbols.test.ts:91 is a single headerless JSON object, which is exactly why tests don't catch it. Fix: split on the first newline, parse the payload, preserve the header when rewriting output.

2. Minor — src/debug-symbols.ts:243: the Binary Images: regex requires an arch token, but legacy macOS .crash files use +com.example.App (1.0 - 1) <UUID> /path with a version instead — those parse to zero images and error as unsupported; arm64_32 (watchOS) is also missing from the alternation.

3. Minor — src/debug-symbols.ts:161,182,211,254: readNumber accepts any finite number, then BigInt(base)/BigInt(imageOffset) is called on it — a malformed artifact with a fractional value throws an uncaught RangeError instead of a normalized AppError.

4. Minor — src/debug-symbols.ts:307: findDsymBundles calls fs.stat(root) unguarded; a nonexistent --search-path surfaces as a raw ENOENT rather than an INVALID_ARGS AppError with a hint.

5. Minor — src/debug-symbols.ts:268-274: text-crash frames are matched to images by name only; when two binary images share a name (app + extension — common in real crashes), the first image's base is used for all frames with that name, producing wrong -l load addresses and silently wrong symbols.

6. Minor — src/debug-symbols.ts:429-434: unsymbolicated detection relies on atos echoing the address byte-for-byte; variants that zero-pad or print 0x… (in App) get treated as successful symbols. Also, atos output lines are zipped with input addresses after filtering blank lines — any stray blank line shifts the mapping and mis-attributes symbols across frames.

7. Minor — src/debug-symbols.ts:378-386: on no-match, artifactUuids dumps every image UUID into the error details; a real .ips has 300+ usedImages, so the payload is effectively unbounded relative to the "compact summary" goal.

8. Minor — src/debug-symbols.ts:482: readTextFile reads the whole artifact with no size cap; a multi-hundred-MB log is loaded fully into memory (twice, via split/join).

9. Minor (tests) — all five tests mock xcrun/dwarfdump/atos with idealized outputs; no two-document .ips fixture (masking finding 1), no multi-arch/multi-dSYM case, no garbage-input case. At least one fixture captured from a real device .ips should be checked in.

10. Informational — feat: add Apple xctrace perf profiling #755, feat: add Android native perf profiling #757, and feat: add perf memory diagnostics #759 touch the same shared files (cli-grammar/observability.ts, client-command-metadata.ts, cli-flags.ts, cli-command-overrides.ts, cli-help.ts, client-types.ts, skillgym smoke suite, docs); whichever lands second conflicts, though the additions here are append-style and should rebase mechanically.

Verified clean

No command injection (spawn with shell: false argv arrays); the atos -l load address is computed correctly (image base as -l, base + imageOffset as absolute addresses — the right convention); UUID normalization symmetric; dSYM search bounded; output is paths + summary only.

Overall

Good architecture — clean parse/match/symbolicate separation, bounded search, well-normalized errors. But the flagship .ips format can't parse real device files at all; that must be fixed and validated against a real crash before merge. The rest is robustness hardening, worth doing but not individually blocking.

---

Generated by Claude Code