* fix(scan): reduce setup and tool output in Standard scans
* fix(workbench): keep policy paths optional for scan lookups
* fix(sdk): retain the preflight callback before scan dispatch
* fix(workbench): only advertise scoped inventory for Standard scans
* docs(workbench): explain optional scope report failures
* fix(workbench): preserve the Deep Scan registration response
* test(workbench): use the current scan context request in the Deep Scan fixture
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
* docs: add a local import guide for Semgrep hosted MCP findings
* docs: extend the hosted Semgrep import example through Cloud
* docs: clarify the Semgrep Cloud workspace prerequisite
---------
Co-authored-by: Michael D'Angelo <269034524+mldangelo-oai@users.noreply.github.com>
* fix(plugin): scope legacy scan continuations to the attached scan
* fix(plugin): rejoin attached legacy deep scans by target
* fix(plugin): keep attached deep scan results with their conversation
* fix(scan): isolate Git display settings in content snapshots
* fix(scan): pin hunk, file order, and rename rendering in snapshots
* test(scan): write Git order fixtures with LF on every platform
* fix(scan): stabilize submodule, indentation, and rename presentation
* Capture the configured submodule diff before asserting
* Normalize Git driver annotations and binary patch rendering
* fix(windows): preserve long Git metadata in reviewed patch replay
* fix(windows): keep reviewed patch replay within source limits
* fix(windows): preserve metadata-free LF patch replay
* test(workbench): select short replay case only on Windows
---------
Co-authored-by: Codex <codex@users.noreply.github.com>
* fix(scan): verify assumed-unchanged files for committed diffs
* test(workbench): read the index flag before asserting
* fix(workbench): isolate Git settings during private index refresh
* fix(workbench): preserve explicit usage database paths
* test(workbench): close usage fixtures before renaming them
* test(workbench): use supported database names on each host
* fix(runtime): honor literal POSIX executable override paths
* fix(runtime): retain the lookup origin of padded executable paths
* fix(runtime): keep bare executable overrides on their original lookup path
* fix(windows): resume Deep Scans before the first checkpoint
* fix(windows): retain the resume directory handle through registration
* fix(windows): finish Deep output registration after cancellation
* fix(windows): retain registration control across plugin entrypoints
* fix(windows): read Deep resume identities from pinned handles
* fix(windows): preserve Deep Scan checkout identity across Python versions
* fix(windows): verify bound checkouts throughout Deep Scan recovery
* fix(workbench): keep scan history when a legacy checkout moves
* test: include checkout binding in migration expectations
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
Co-authored-by: Codex <codex@users.noreply.github.com>
* fix(plugin): publish new patch files on musl without hard links
* test(plugin): close the rename fixture directory explicitly
* test(plugin): select the raw rename syscall test only on Linux
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
* fix(plugin): fall back to system Python when a Linux UID has no home
* fix(sdk): discover system Python when a Linux UID has no home
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
* fix(scans): archive long POSIX output directory names
* fix(logs): recover bound sessions from short scan archives
* fix(package): include the scan archive path helper in the release contract
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
* fix(windows): resume Deep Scans before the first checkpoint
* fix(windows): retain the resume directory handle through registration
* fix(windows): finish Deep output registration after cancellation
* fix(windows): retain registration control across plugin entrypoints
* fix(windows): read Deep resume identities from pinned handles
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
* fix(sdk): build bundled plugins from long Windows paths
Use the extended Windows path only when creating the temporary build directory. Keep the selected parent spelling for the compiler, publication, and cleanup, and exercise the package build from a long source snapshot.
* fix(plugin): build the selected MCP app on long Windows paths
Run the declared app build with its local dependency binaries when the Windows Node task runner would skip a long package or tool path. Keep the native runner on shorter paths, other platforms, and Bun.
* fix(plugin): run the selected Bun build script in deep Windows checkouts
Use Bun run build for long Windows MCP app paths so the reserved bundler command cannot shadow the package script. Exercise the real Bun process and its selected package lifecycle in the long source-build regression.
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
* fix(plugin): select a SQLite-capable Node runtime on Linux
* fix(plugin): limit Linux SQLite fallback to system installations
* fix(plugin): distinguish missing SQLite by probe exit status
* fix(plugin): follow fixed system Node installation symlinks
* test(plugin): pass the system Node fixture path as data
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
* Invoke the selected Windows policy helper without re-expanding paths
* Keep the reviewed Windows policy workflow available to custom validation
* Keep PowerShell native preferences local to the policy command
---------
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
Deep Scan selects workers from per-run settings, legacy TOML, or the bundled default of four. The former automatic setting also resolves to four. Stop sampling host CPUs in the MCP client, setup inspection, and workbench launch. Preserve validation for callers that still send the legacy argument.
Verified workbench startup and the standalone helper with omitted, zero, one, and sixteen as the CPU count. Project setup, saved settings, coordinator resume, and the MCP store integration suite passed on both revisions. The SDK types and format checks and portable plugin checks also passed.
Use the sealed contract's path grammar to validate a draft's finding locations
and code evidence. The draft retains each original spelling and its existing
error type. The sealed reader still returns the normalized path and allows a
dot only in contexts that previously allowed it. CSV keeps its own rules,
including a different Windows check on drive-prefixed paths.
The two maintained plugin sources lose 19 lines. Regenerating the SDK mirror
adds five lines, for 14 deleted overall. Compared 1,242,746 inputs under each
of Node's POSIX and Windows path semantics, including the Unicode code-point
range. Windows behavior was modeled on Linux.
Replayed 99 cases on the standalone plugin under Node 22.13.0 using examples
from releases 0.1.0 and 0.8.3. Four exported CSVs and an actual sparse CSV
preserved their source locations through scan import. The 316 focused MCP
tests and 263 SDK tests passed, along with types and plugin source checks.
Keep one compressed payload per generated MCP launcher so bundle growth
no longer requires changing a numbered file list. The server and helper
entrypoints retain their paths, module identities and exports. Rebuilding
removes previous parts, and the package checker still reads split archives.
Both new payloads match the concatenated original parts byte for byte.
The packed npm archive grew by 4,377 bytes. All 81 focused tests passed
on each revision, with three platform skips. The installed npm smoke
passed, including MCP initialization, nested worker launch, retry and
detached plugin restart.
The change to inline scan reviews in #1762 taught both the schema preview
and the first-seal publisher to select a mode and construct the compact
review. Make that decision once and use the same fields for both. A sealed
rerun still leaves the stored review alone, and coverageRef bundles still
use their separate coverage document.
Keep the provisional and final manifest validations. The former catches a
malformed reviewState before the publisher drops it and checks report
inputs before optional writeup handling. The latter checks the real
artifact records. Custom schemas can distinguish the two.
Compared finalization against the published 0.1.0 completed scan and the
0.8.3 current and legacy examples. All 14 output files matched byte for
byte, including each original sealed JSON file. Both trees passed the
541 selected finalization and replay tests and 177 subtests (2 skips).
* refactor(mcp): validate delivered scan identity in one place
Give headless Standard and prompt-only starts one definition of a delivered
scan. Its disposition, UUID, directory and workspace ID must agree. Keep
the additional Standard claim, thread, mode, progress and new-scan phase
checks at the Standard entry point. Prompt-only does not grant artifact
access. Both tools still return the original workbench response.
Compared 10,119 synthetic responses through the registered callbacks on
each of Node 22.13, 24.19 and 26. Independent checks observed actual claim
insertion and rejection. In 27 stdio workflows the released 0.8.3 plugin,
main and this change created and rejoined headless Standard, prompt-only
Standard and diff scans, retaining the expected identity and token
handling. Main and the candidate retained the same telemetry fields.
MCP types, formatting, the plugin build, Ruff and nine portable checks
passed. A broader four-file MCP run encountered host process and thread
creation failures. Native Windows and macOS were not part of this run.
No caller migration or generated-source changes are required.
* fix(mcp): drop the unused prompt-only scan directory binding
Keep scan directory validation and return the complete workbench response. The shorter binding reads only the scan identifier used by the prompt-only result.
Select each accepted worker result once for coverage lookup and review
emission. Keep source order when a recovered row has several matching
inputs, and preserve archived evidence when the registered result is
missing.
All 720 recovery and projection tests pass before and after the change,
including failed workers, missing current results, and frozen replay.
Mark the canonical findings unions that expose the historical JSON Schema
type array and property layout. Derive their type names from Zod's resolved
branches during serialization. A member change now needs one edit instead
of a corresponding metadata array edit. Keep unmarked unions and authored
type restrictions untouched. Fail generation on a marked union with an
unresolved, untyped, or duplicate member rather than emitting a broader
contract.
All eleven generated JSON files are byte-identical. Dropping the layout
outright changes published property pointers and validation diagnostics.
A temporary numeric rootCause member verified the new generator accepts
that member while retaining the nested object restrictions. Independent
experiments covered unsupported branches and unmarked restrictions.
Checked generation on Node 22.13, 24.19, and 26.0, MCP typecheck, the three
MCP artifact contract test files, 15 Python schema/checkpoint tests, plugin
source compatibility and its nine tests, and formatting. The two generator
sources add 22 and remove 28 lines, with no generated or consumer changes.