fix(ci): skip native-host-build when no native-relevant paths changed

[carlos-alm]

Summary

• `native-host-build` (3-OS matrix: ubuntu/macos/windows) ran an unconditional Rust build + `cargo test` on every push to main and every PR, regardless of whether the diff touched anything native. `test`, `parity`, and `pre-publish-benchmark` all depend on its artifacts.
• Added a `detect-changes` job that diffs the PR base (or the previous push SHA) for `crates/**`, `Cargo.toml`, `Cargo.lock`, and the native addon loader (`src/infrastructure/native.ts`)
• `native-host-build` now skips entirely (all 3 OSes) when none of those changed — fail-open: if the detector itself fails or is inconclusive, the build still runs, so a broken detector can't silently hide a real Rust regression
• `test`, `parity`, and `pre-publish-benchmark` keep running on all 3 OSes regardless — they just fall back to the last-published native binary (via the normal `npm install` optional-dependency resolution) instead of a freshly PR-built one when `native-host-build` was skipped. This preserves full cross-platform test/parity coverage for every PR; only the redundant Rust rebuild+`cargo test` is cut for non-native changes.

Why

Part of the broader CI/CD cost-reduction effort (#2108). This is the biggest single lever in that effort — most day-to-day PRs in this repo don't touch `crates/**`, so they were paying for a 6-leg macOS/Windows/Linux native rebuild (plus `cargo test` on each) for no reason. 880 CI runs in the last 30 days were all paying this cost unconditionally.

Closes #2111

Test plan

• YAML parses, `actionlint` shows no new findings (pre-existing shellcheck notes only, unrelated lines)
• A pure-TS/docs PR should show `native-host-build` skipped and `test`/`parity`/`pre-publish-benchmark` still green on all 3 OSes
• A PR touching `crates/**` should show `native-host-build` still running on all 3 OSes as before
• Push to main (no native changes) should show the same skip behavior

[greptile-apps]

Greptile Summary

This PR adds a detect-changes job that diffs the repository against the last semantic-version release tag and gates native-host-build (3-OS Rust matrix) on the result, with a documented fail-open design so a broken detector can never silently mask a Rust regression.

• detect-changes: runs git diff $LAST_TAG HEAD and emits native_changed=true/false; fails open (no tag found, script error) so native-host-build always runs in ambiguous cases.
• native-host-build: skips all three matrix legs only when detect-changes succeeded AND reported no native-relevant paths changed; artifact download/install steps in downstream jobs are individually gated on native-host-build.result == 'success'.
• test / parity / pre-publish-benchmark: updated with always() + != 'failure' / != 'cancelled' guards so they keep running against the published npm binary even when the native build was legitimately skipped.

Confidence Score: 5/5

Safe to merge — the fail-open design ensures correctness in all edge cases, and no downstream job can silently use a stale binary when native code has changed.

The logic is sound and the control flow is well-reasoned: detect-changes can only suppress native-host-build when it positively confirms no native paths changed since the last release tag, and every other outcome keeps the build running. The one limitation — that the skip only fires when no unreleased native commits exist — is intentional, documented, and correct from a safety standpoint.

Only .github/workflows/ci.yml changed. The detect-changes script block (lines 76-90) is the section worth re-reading — specifically the tag-based diff logic and the comment explaining why a PR-base diff was rejected.

Important Files Changed

Filename

Overview

.github/workflows/ci.yml

Adds detect-changes job that diffs against the last release tag to gate native-host-build; all three matrix legs are skipped together when no native paths changed. Downstream jobs use always() + result guards to keep running against the published binary. Logic is correct and fail-open, but the tag-based diff means the skip only triggers when no native commits exist between the last release and HEAD — a condition the code's own comments say doesn't hold today.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A([push / pull_request]) --> DC[detect-changes\ngit diff LAST_TAG HEAD]
    DC -->|script fails / no usable tag| NHB_RUN
    DC -->|native_changed=true| NHB_RUN[native-host-build\n3-OS matrix]
    DC -->|native_changed=false| NHB_SKIP[native-host-build\nSKIPPED]

    NHB_RUN -->|success| DL[Download + install\nPR-built .node artifact]
    NHB_SKIP -->|skipped| PUB[Use published npm\nbinary as fallback]

    DL --> TEST[test / parity / pre-publish-benchmark\nrun on all 3 OSes]
    PUB --> TEST

    NHB_RUN -->|failure| SKIP_DOWN[test / parity / benchmark\nSKIPPED via failure guard]

    TEST --> CP[ci-pipeline\nfail on any failure/cancelled]
    SKIP_DOWN --> CP

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A([push / pull_request]) --> DC[detect-changes\ngit diff LAST_TAG HEAD]
    DC -->|script fails / no usable tag| NHB_RUN
    DC -->|native_changed=true| NHB_RUN[native-host-build\n3-OS matrix]
    DC -->|native_changed=false| NHB_SKIP[native-host-build\nSKIPPED]

    NHB_RUN -->|success| DL[Download + install\nPR-built .node artifact]
    NHB_SKIP -->|skipped| PUB[Use published npm\nbinary as fallback]

    DL --> TEST[test / parity / pre-publish-benchmark\nrun on all 3 OSes]
    PUB --> TEST

    NHB_RUN -->|failure| SKIP_DOWN[test / parity / benchmark\nSKIPPED via failure guard]

    TEST --> CP[ci-pipeline\nfail on any failure/cancelled]
    SKIP_DOWN --> CP

Loading

_{Reviews (2): Last reviewed commit: "fix(ci): diff native-relevant changes ag..." | Re-trigger Greptile}

[carlos-alm]

Thanks for the review — confirming the fail-open design and the tag-based diff are intentional, as noted. CI on 485dd63 now shows native-host-build correctly running on all 3 OSes (previously it was wrongly skipped, which was the root cause of the Test/Engine-parity failures on the prior push — the fallback binary had drifted 62 native commits past the pinned v3.15.0 release). Watching Test/Engine-parity/pre-publish-benchmark now that a fresh binary is available.