Distillation

Date: September 9, 2026

• 🔴 [enforced-source-order] (12:30) User stated new binaries and the packages/cli/install script must always check the ordered upgrade sources exactly as: 1. getsentry/toolkit / ghcr.io/getsentry/toolkit, 2. getsentry/cli / ghcr.io/getsentry/cli.
• 🔴 [enforced-fallback-policy] (12:30) User required falling through to the next upgrade source only on HTTP 404; network and fetch failures remain transient errors and must never poison a source.
• 🔴 [enforced-source-parity] (12:30) User required one shared upgrade-source definition for TypeScript and the generated install script, with a check that fails if the two source lists or their ordering diverge.
• 🔴 (12:30) User stated old binaries know only getsentry/cli; every migrated stable CLI release and nightly must therefore be dual-published to Toolkit and legacy CLI endpoints during the compatibility window.
• 🔴 (12:30) User stated getsentry/cli#1244 is a hard prerequisite of getsentry/cli#1248: one stable bridge CLI release must ship from getsentry/cli before any CLI history is imported into getsentry/sentry-mcp; it must work before the Toolkit rename, discover Toolkit releases afterward, and preserve upgrades for older binaries through legacy endpoints.
• 🔴 (12:30) User specified implementation scope: add UPGRADE_SOURCES across src/lib/binary.ts, ghcr.ts, upgrade.ts, and delta-upgrade.ts; mirror ordered fallback in install; add a check: script proving installer/TypeScript source-order parity; preserve sha256-<binary> / from-version delta chains and cleanup/keep-count parity across both GHCR packages; configure post-migration Toolkit stable/nightly releases to publish to both repositories; keep getsentry/cli writable through the compatibility window.
• 🔴 (12:30) User specified the pre-import bridge-release gate in order: 1. cut the next stable release from getsentry/cli using the release system’s computed version, never a manually pinned target; 2. verify a pre-bridge binary upgrades to it through getsentry/cli; 3. verify the bridge binary handles a 404 from not-yet-existing getsentry/toolkit and upgrades through getsentry/cli; 4. verify identical installer behavior; 5. record the bridge version in getsentry/cli#1244; only then begin getsentry/cli#1239.
• 🔴 (12:30) User specified the post-rename compatibility matrix: 1. publish one staging CLI release from Toolkit to both repositories; 2. verify a pre-bridge binary upgrades through getsentry/cli; 3. verify bridge and Toolkit-built binaries upgrade through getsentry/toolkit; 4. simulate Toolkit 404 and verify both new binaries fall back to getsentry/cli; 5. compare release assets, SHA-256 values, GHCR manifests, and delta-chain metadata and require exact matches; 6. repeat the matrix for the install script and one nightly.
• 🔴 (12:30) User required dual-publishing for at least 2 stable CLI releases and roughly 90 days of nightlies; getsentry/cli may be archived only after telemetry and release evidence show the legacy endpoint is unnecessary, and must remain writable until then.
• 🟡 (12:30) Assistant reproduced the installer parity regression: packages/cli/test/lib/install-script.test.ts had 2 tests with 1 failure because packages/cli/install did not define UPGRADE_SOURCES; assistant planned source-selection and deterministic fake-curl coverage without modifying concurrent TypeScript work.
• 🟡 (12:33) packages/cli/install was modified, and packages/cli/test/lib/install-script.test.ts was replaced with expanded deterministic installer coverage using temporary fake curl and gunzip executables, recorded request URLs in SENTRY_TEST_REQUESTS_FILE, recorded setup arguments in SENTRY_TEST_ARGS_FILE, and scenario selection through SENTRY_TEST_SCENARIO.
• 🟡 (12:33) Fake-curl scenarios covered Toolkit stable listing, legacy latest release, exact Toolkit cli@... tags, exact legacy tags, Toolkit/legacy repository probes, GHCR token and nightly manifest/blob requests, redirect downloads, gzip downloads, HTTP 404 fallback, HTTP 403/429 non-fallback, and simulated network exits 6 and 7.
• 🟡 (12:34) First expanded Vitest run under v4.1.10 had 10 tests: 8 passed and 2 failed. "selects a nightly source before requesting source-specific GHCR data" did not record "nightly" in setup arguments; "passes setup options through to sentry cli setup" received one malformed string, "clinsetupn--installn--methodncurln--channelnstablen--no-modify-pathn--no-completionsn--no-agent-skillsn", instead of 10 separate arguments.
• 🟡 (12:34) After correcting the fake downloaded binary’s argument recording in packages/cli/test/lib/install-script.test.ts, the focused suite passed all 10 tests in 868ms: transform 285ms, setup 78ms, import 320ms, tests 274ms, environment 0ms.
• 🔴 [enforced-telemetry-isolation] (12:35) User required installer telemetry to be non-blocking: it never blocks installation or propagates failures, and telemetry must never fail the script.
• 🟡 (12:35) Biome formatting initially failed on packages/cli/test/lib/install-script.test.ts:271-275, requiring the runInstaller(["--version", "nightly"], "nightly-fallback") call to be split across multiple lines; the test file was reformatted accordingly.
• 🟡 (12:36) packages/cli/src/lib/upgrade.ts showed stable resolution via fetchLatestFromGitHubWithSource() and resolveUpgradeSource({ getProbeUrl: getGitHubLatestReleaseUrl }), filtering non-draft/non-prerelease tags by source.tagPrefix; nightly resolution used resolveUpgradeSource({ getProbeUrl: getGitHubRepositoryUrl }), then source-specific getAnonymousToken(source) and fetchNightlyManifest(token, signal, source).
• 🟡 (12:36) A subsequent format/check plus focused Vitest run passed all 10 installer tests in 991ms: transform 395ms, setup 65ms, import 440ms, tests 295ms, environment 0ms.
• 🟡 (12:37) A deliberate regression test temporarily removing the non-404 guard caused "stops source selection on a non-404 response" to fail: exit code was correctly 1 and only that test ran, but stderr said "Failed to find a CLI release in getsentry/toolkit" rather than containing "HTTP 403"; 9 tests were skipped.
• 🟡 (12:37) The non-404 guard/error reporting was restored in packages/cli/install; checks then passed and all 10 focused installer tests passed in 1.37s, with transform 506ms, setup 117ms, import 564ms, tests 397ms, and environment 0ms.
• 🟡 (12:38) Installer-specific diff comprised 2 files: packages/cli/install and packages/cli/test/lib/install-script.test.ts, totaling 389 insertions and 54 deletions; install changed by 149 lines and the test changed by 294 lines.
• 🟡 (12:38) Assistant reported implemented getsentry/cli#1244 installer behavior: Toolkit-first ordered resolution; fallback to legacy only on HTTP 404; Toolkit stable-tag filtering with cli@VERSION; exact pinned-release probes; source selection before nightly GHCR requests; gzip-to-raw fallback constrained to the same selected source; non-blocking telemetry; and deterministic request-order/source-affinity tests.
• 🟡 (12:38) Assistant reported final validation: bash -n install passed, ShellCheck passed, Biome passed, focused Vitest passed 10/10 with 0 failures, and git diff --check passed. The deliberate non-404 regression failed as expected before restoration and passed afterward.
• 🟡 (12:38) Working tree also contained concurrent modifications to packages/cli/src/lib/binary.ts, packages/cli/src/lib/delta-upgrade.ts, packages/cli/src/lib/ghcr.ts, packages/cli/src/lib/release-notes.ts, packages/cli/src/lib/upgrade.ts, packages/cli/src/lib/version-check.ts, and their corresponding tests; assistant stated these concurrent TypeScript changes were left untouched.
• 🟡 (12:38) Vitest continued to emit only the existing warning that test.poolOptions was removed in Vitest 4 and former pool options must be top-level.