* fix(archives): wrap the bare EOFError a truncated tar.gz raises
`tarfile` wraps most decompression failures in `TarError`, but a gzip
stream that ends before its end-of-stream marker escapes as a bare
`EOFError` from the gzip layer. `EOFError` derives from neither
`TarError` nor `OSError`, so it bypassed all three of the tar handlers
added with tar archive support (#3874):
- the format probe in `detect_archive_format`, which caught only
`tarfile.TarError`;
- `tarfile.open` in `safe_extract_tar`;
- member iteration in `safe_extract_tar`.
A truncated `.tar.gz` — an interrupted download, a partially written
file — therefore raised a raw `EOFError` straight through the caller's
`error_type`, so callers catching `ValueError`/`ExtensionError`/
`PresetError` never saw it. In `specify workflow add` the effect is worse
than a traceback: Typer treats a bare `EOFError` as a Ctrl-D abort, so
the command printed only "Aborted." with no diagnostic at all. The ZIP
twin reports "Invalid workflow archive: Invalid ZIP archive: <path>".
Route all three sites through a shared `_TAR_DECOMPRESSION_ERRORS`
tuple so they stay in sync. `zlib.error` is included alongside
`EOFError`: it is likewise neither a `TarError` nor an `OSError` and can
surface from a corrupt deflate block. `OSError` is kept only on the two
`safe_extract_tar` sites, which report genuine I/O failures; adding it
to the probe would silently swallow them instead.
Truncated tar.gz now reports the same clean, domain-typed error as the
ZIP path. Tests cover both the short prefix that fails in
`tarfile.open` and the longer ones that fail during member iteration —
`tarfile` decompresses lazily, so the leak surfaced at different sites
depending on how much of the stream survived.
Assisted-by: Claude Opus 5 (1M context)
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* test(archives): cover the bare zlib.error a corrupt deflate block raises
Review feedback: the `zlib.error` arm of `_TAR_DECOMPRESSION_ERRORS` was
not exercised. Every regression added with the fix truncates a valid
deflate stream, which raises `EOFError`, so `zlib.error` could regress
independently of the EOF handling.
It is genuinely reachable, but only under a narrower condition than the
truncation cases. `tarfile` converts `zlib.error` to `ReadError` while
reading a member *header*, but the forward seek it performs to skip
member *data* (`tarfile.next`) sits outside that conversion, so a corrupt
region past the first header escapes raw. Reaching that seek needs
members larger than the gzip read buffer: with small members the whole
stream is decompressed during the first header read and the error is
wrapped. The new fixture therefore uses two 256 KiB members at
`compresslevel=1` — a ~7 KiB archive — corrupted past the midpoint so
the first header still reads clean.
Adds four tests: the two `safe_extract_tar` sites (plain and with a
caller-supplied `error_type`), the `safe_extract_archive` entry point
with a caller-supplied `error_type`, and a guard asserting the fixture
still reaches the module as a bare `zlib.error` — so if a future Python
wraps it, that fails loudly instead of the coverage silently decaying
into a duplicate of the `EOFError` cases.
Verified test-the-test: the three wrapping tests fail against the
unmodified `_download_security.py` with a raw
`zlib.error: Error -3 while decompressing data: invalid distance code`,
and pass with the fix.
Also corrects the scope claimed for the probe site. Fuzzing 2800 corrupt
archives never produced a bare `zlib.error` from `tarfile.open` alone,
because the only read it performs is the header read that `tarfile`
already converts. The probe's `zlib.error` arm is defensive, not
load-bearing; the tuple comment and a detection test now say so rather
than implying coverage that cannot exist.
Assisted-by: Claude Opus 5 (1M context)
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* test(archives): make the corrupt-deflate fixture zlib-version independent
CI failure on macos-latest/3.13: `test_corrupt_deflate_fixture_raises_bare_zlib_error`
failed with `gzip.BadGzipFile: CRC check failed`. The other five pytest jobs
were fail-fast cancellations, not real failures, and ruff was already green.
The fixture built its corruption by XOR-ing 64 arbitrary bytes mid-stream.
Whether that produces a *structural* deflate error is zlib-version dependent:
on the macOS runner the mangled bytes still decoded, so the stream instead
failed the trailing gzip CRC check and raised `BadGzipFile` -- an `OSError`,
which the pre-fix `(TarError, OSError)` handler already caught. The guard test
exists precisely to catch that degradation, and it did its job.
Replaces the XOR with a deflate block header whose `BTYPE` is the reserved
value `0b11`. Every zlib rejects that identically as "invalid block type", and
it fails during decompression rather than at the CRC check, so no version can
turn it into a `TarError` or `OSError`. The stream is assembled by hand
(`compressobj(-15)` + explicit gzip header/trailer) so the invalid block lands
a controlled 256 KiB into the first member's data -- past the gzip read buffer,
so the first header still reads clean and the failure surfaces from the forward
seek in `tarfile.next`, which is the site the raw `zlib.error` escapes from.
A sweep over clean-prefix sizes confirms a wide margin: with 512 KiB members
every prefix from 160 KiB up yields a bare `zlib.error`, versus the transition
below ~131 KiB where `tarfile` still wraps it as `ReadError`. The hand-built
gzip header also zeroes the mtime field, so the fixture is now byte-identical
across builds instead of embedding a timestamp.
Strengthens the guard to assert what the fix actually depends on -- that the
exception is neither a `TarError` nor an `OSError` -- so the fixture cannot
silently decay into an already-caught type again.
Production code is unchanged from ef49acc; this is test-only.
Verified test-the-test by dropping the `zlib.error` arm from
`_TAR_DECOMPRESSION_ERRORS`: the three wrapping tests fail with the raw
`zlib.error: Error -3 while decompressing data: invalid block type`, and pass
with it restored. `tests/test_download_security.py`: 193 passed.
`ruff check src tests` (the exact CI command): all checks passed.
Assisted-by: Claude Opus 4.8 (1M context)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
`CommandRegistrar.render_frontmatter` calls `yaml.dump()` without `width=`,
so PyYAML applies its default ~80-column wrap and folds any long scalar onto
a continuation line.
A `description` longer than roughly 80 characters is therefore rendered as:
---
name: speckit-implement
description: Execute the implementation plan by processing and executing all tasks
defined in tasks.md
---
The YAML remains valid and round-trips faithfully through `yaml.safe_load`,
so this is not data loss. It is a shape inconsistency with real consequences:
- Hand-written core command templates always keep `description` on one line,
so preset- and extension-rendered commands do not match the files they sit
beside in the same directory.
- Consumers that read frontmatter line-wise rather than with a YAML parser
see the description truncated at the fold, followed by a stray line. Spec
Kit itself hand-builds SKILL.md frontmatter in the skills path (see #3391),
so this is not a hypothetical class of consumer.
- `speckit.implement`'s own description is 89 characters, so a preset that
overrides it hits this immediately.
`width=float("inf")` disables the line-wrapping only; escaping, quoting and
the handling of genuinely multi-line values are unchanged, since PyYAML
selects the scalar style before applying width.
Adds a regression test that fails without the change.
Verified against the repo's own suite: 6354 passed. Four failures in
tests/integrations/test_integration_subcommand.py are present on a clean
checkout too (ANSI escapes in captured output) and are unrelated.
* fix: skip corrupted run state files in list_runs
* fix: address review comments - add UnicodeDecodeError, dict validation, and regression tests
- Catch UnicodeDecodeError for invalid UTF-8 encoding
- Validate loaded JSON is a dict with required 'run_id' key
- Add 5 regression tests for corrupted state files
Fixes#3814
PresetRegistry._load() catches json.JSONDecodeError and
FileNotFoundError to start fresh on a corrupted or missing registry, but
a registry file with invalid UTF-8 bytes raised UnicodeDecodeError
before JSON parsing began, crashing every preset command.
Catch UnicodeDecodeError in the same clause: undecodable bytes are the
same corruption class as unparseable JSON. OSError stays uncaught on
purpose — the data may be intact on disk, and starting fresh would let a
later _save() wipe it (same fail-closed reasoning as the workflow
catalog cache loader).
Assisted-by: GitHub Copilot (model: claude-fable-5, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
_merge_toml_fragment() and _remove_toml_entries() read the user's
config.toml with bare read_text() calls, so a non-UTF-8 (or otherwise
unreadable) file crashed install_integration_events() and
remove_integration_events() with a raw UnicodeDecodeError — and the
merge path regenerates the file from what it read, so it would have
discarded the user's bytes had it not crashed first. Every JSON
merge/remove path already goes through _load_user_json(), which skips
on an unreadable file to preserve user content (#22).
Abort the merge (returning False so the caller skips tracking, S5) and
skip the teardown cleanup with a warning, leaving the user's bytes
untouched in both directions.
Assisted-by: GitHub Copilot (model: claude-fable-5, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The rescue-retry loop in install_from_directory() reads each staged
backup with bare stat()/read_bytes() calls, so a staged config that
cannot be read crashed the reinstall with a raw OSError. Every sibling
read in this path — the live twin four lines below, the packaged
baseline check, the mode sidecar — already catches OSError.
Treat an unreadable staged file like an uncomparable live config:
add it to the conflict set so both copies are preserved and the retry
aborts with the existing resolution guidance while dest_dir is still
untouched.
Assisted-by: GitHub Copilot (model: claude-fable-5, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
`requires.speckit_version` was presence-checked but never type-checked in
both the extension and preset manifest validators, so an unquoted YAML
`speckit_version: 1.0` (a float) passed validation and reached
`SpecifierSet(required)` in `check_compatibility()`.
That call is guarded by `except InvalidSpecifier` alone, which a
non-string escapes two different ways:
- a float/int/bool/None raises `TypeError: 'float' object is not
iterable` from the `SpecifierSet` constructor;
- a list or dict is an *iterable*, so `SpecifierSet` accepts it and the
failure surfaces much later as `AttributeError: 'str' object has no
attribute 'filter'` from inside `.contains()`.
Neither is a `CompatibilityError`/`PresetCompatibilityError`, so both
bypass the CLI's "Compatibility Error" handler in `_commands.py` and exit
1 with a raw traceback that names no field, leaving the author with no
hint which manifest key is wrong.
Type-check the field in both validators, requiring a non-empty string,
and additionally guard `check_compatibility()` in both managers since
each is public and reachable with a hand-built or mutated manifest. This
mirrors the sibling `IntegrationDescriptor`, which already requires a
non-empty string for the same key, and completes the type-checking pass
started in #3943 for the neighbouring `extension`/`preset` fields.
Adds 33 regression tests across both modules covering every escape path;
26 of them fail without this change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Assisted-by: Claude Code (model: Claude Opus 5, supervised)
The keep-config rescue branch of install_from_directory() reads each
preserved config with bare read_bytes()/stat() calls, so a kept config
that cannot be read (permission or I/O error) crashed the reinstall
with a raw OSError. The sibling symlink guard four lines above already
rejects with a ValidationError and resolution guidance for the same
reason: bytes that cannot be safely rescued must not reach the rmtree
below.
Wrap the read and raise ValidationError with guidance, while dest_dir
is still untouched so the preserved bytes are never lost.
Assisted-by: GitHub Copilot (model: claude-fable-5, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
_script_command() split the configured command with a bare
shlex.split(), so a command string with unbalanced quotes crashed event
dispatch with a raw ValueError. The dispatcher-template twin a few
lines up already wraps the same call in try/except ValueError and
returns None so dispatch falls back cleanly.
Wrap the split the same way and return None, restoring parity between
the two paths.
Assisted-by: GitHub Copilot (model: claude-fable-5, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* feat(events): context injection for opencode and JSON-envelope agent hooks
Adds first-class context injection to agent runtime events:
1. opencode: maps session_start to experimental.chat.system.transform (injects into system prompt) and user_prompt_submit to chat.message (injects synthetic TextPart). TS plugin captures runEvent stdout (stdio pipe, encoding utf-8) and pushes into output objects. Part IDs derive from output.parts[last].id to preserve OpenCode's prt_ brand and prevent session schema crashes.
2. JSON-envelope hook wrapping: adds events_context_envelope to IntegrationBase so agents that require JSON on stdout receive their target envelope via the dispatcher's 5th argument:
- gemini, tabnine, qwen, devin: hookSpecificOutput.additionalContext on session_start/user_prompt_submit; suppress on non-injectable events (prevents systemMessage user-facing noise)
- copilot: top-level additionalContext on session_start
- cursor: top-level additional_context on session_start; suppress elsewhere
- claude, codex: plain stdout passthrough (already injected)
3. Dispatcher template and resolve_and_run_event_command parse the 5th envelope arg and wrap stdout accordingly.
Tests added for opencode TextPart schema, part ID derivation, envelope command generation, and dispatcher output wrapping. All 162 events/integration tests pass.
* fix(events): address code review on #3934
- Qwen/Gemini/Tabnine/Devin: include native hookEventName inside
hookSpecificOutput envelope (required by Qwen's hooks spec). Thread
the native event name from the integration's CANONICAL_TO_NATIVE
through _dispatcher_command as a 6th dispatcher argument, through
the dispatcher template's main()/_run_inline()/_emit(), and through
resolve_and_run_event_command()/_emit_event_stdout().
- Copilot: map user_prompt_submit to additionalContext (previously
unmapped, breaking per-prompt context injection despite Copilot CLI
supporting it via userPromptSubmitted).
- OpenCode: guard experimental.chat.system.transform so canonical
session_start handlers only run when input.sessionID is present —
OpenCode fires this hook for non-session operations (e.g. agent
generation) with no sessionID.
Assisted-by: opencode (model: glm-5.2, supervised)
* fix(events): address second Copilot review round on #3934
- Positional arg alignment: always emit default timeout (60s) as the
4th dispatcher argument even when timeout_seconds is omitted, so the
envelope (5th) and native_event (6th) land in the correct argv slots.
Previously, omitting timeout_seconds caused the envelope to be parsed
as an invalid timeout, silently falling back to plain stdout.
- OpenCode session_start caching: cache handler output per sessionID in
the generated TS plugin so non-idempotent handlers (setup, telemetry,
file-mutating scripts) run once per session instead of on every LLM
request. Cache is evicted on session.deleted.
- Updated PR description to reflect Copilot user_prompt_submit now maps
to additionalContext (was documented as plain/unprocessed).
Assisted-by: opencode (model: glm-5.2, supervised)
* fix: cap stdin read at 1 MiB to prevent DoS
Unbounded sys.stdin.read() allowed a malicious caller to exhaust memory
by sending a multi-gigabyte payload. Cap at 1 MiB and raise typer.Exit
if truncated.
* fix: improve stdin payload limit error handling in event.py
- Rename _MAX_PAYLOAD to MAX_STDIN_BYTES (clearer constant naming)
- Improve error message to suggest truncation or smaller payload
- Better code formatting for readability
Assisted-by: GitHub Copilot (model: mimo-v2-free, supervised)
* fix(presets): restore core skills instead of deleting them on preset remove
Skill restoration looked for core command templates under
.specify/templates/commands, a directory specify init never populates
in real projects. Since that lookup always missed, presets overriding
a core command (e.g. speckit.plan) had their skill deleted outright on
removal instead of restored — the actual core templates live in the
bundled core_pack (wheel install) or the repo-root templates/ tree.
Restoration now falls back to that bundled location, gated behind a
restore_from_bundled_core flag so the existing "retire a stale skill
superseded by a command-mode winner" path keeps deleting rather than
resurrecting a duplicate skill.
Fixes#3928
* fix(tests): use explicit utf-8 encoding reading restored skill content
read_text() defaults to the platform locale encoding, which is cp1252
("charmap") on Windows. The bundled specify.md core template contains a
UTF-8 multi-byte emoji whose bytes aren't valid cp1252, so the Windows CI
job failed decoding the restored SKILL.md with UnicodeDecodeError.
* fix(presets): keep extension restore priority over bundled-core fallback
The bundled-core fallback added for #3928 ran before the
extension_restore_index lookup, so a skill an installed extension owns
could be silently replaced by lower-priority bundled core content on
preset removal instead of preserving the extension's winning layer.
* fix(manifests): reject non-string metadata instead of crashing on it
`ExtensionManifest` and `PresetManifest` checked only key PRESENCE for
`id`/`name`/`version`/`description`, then fed the values straight to
`re.match()` and `packaging.Version()`. Both raise a bare `TypeError` on a
non-string, which is neither `ValidationError` nor `PresetValidationError`,
so it escaped every caller that already handles a malformed manifest.
YAML makes this an easy authoring slip rather than a contrived one: an
unquoted `version: 1.0` parses as a float and `id: 2` as an int.
The user-visible symptom is the one the in-tree comment above the section
guards was written to prevent (#3898 for presets, and its extension twin):
`list_installed()` degrades a bad manifest to "⚠️ Corrupted extension" but
catches only the domain error, so a single bad manifest made
`specify extension list` / `specify preset list` exit 1 with a raw
traceback and *no output at all* — hiding every healthy extension/preset
too, not just the broken one.
Also unguarded on the same path:
- extension `provides.commands[].name` → `TypeError` from the command-name
pattern match. The sibling `file` field was already safe, since
`relative_extension_path_violation()` rejects a non-string.
- preset `provides.templates[].name`/`.file` → `TypeError` from `re.match`
and `os.path.normpath` respectively.
The third manifest twin, `IntegrationDescriptor`, is already hardened: it
type-checks the same four fields and catches `TypeError` alongside
`InvalidVersion`. This brings the other two in line with it.
Tests: 68 added across both suites, covering each field against float, int,
None, list, dict, and bool, plus an end-to-end guard per manifest type
asserting a healthy entry still lists while the bad one degrades. All 68
fail with the source change reverted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* fix: narrow exception in invoke separator resolution and add regression test
Narrow 'except Exception' to 'except (ImportError, ValueError, KeyError)'
in register_commands() invoke separator resolution. Add regression test
that verifies TypeError propagates instead of being silently swallowed.
* fix: remove duplicate pass statement in agents.py
Remove redundant second pass statement in the except block
for invoke separator resolution. The narrowed exception handler
now has a single clean pass statement.
Assisted-by: GitHub Copilot (model: mimo-v2-free, supervised)
The step documents the default twice:
class docstring: "Because workflows run unattended, the step defaults to
``--ignore-agent-tools``"
field docs: "Skip checks for the coding agent CLI (defaults to ``true``)"
It implements that with `config.get("ignore_agent_tools", True)`, which
applies the default only when the key is ABSENT. A bare
`ignore_agent_tools:` in YAML parses to None, and `_resolve_bool(None)`
returns False:
key ABSENT -> True flag emitted: YES
bare ignore_agent_tools: -> False flag emitted: NO <-- bug
explicit true -> True flag emitted: YES
explicit false -> False flag emitted: NO
So the flag is dropped, `specify init` re-runs the agent-CLI presence
check, and an unattended run fails with "Agent Detection Error" for any
integration whose CLI is not installed on the runner.
Normalize an explicit null to the default, mirroring the while/do-while
`max_iterations` handling.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat: allow overriding default init integration via SPECKIT_INTEGRATION_DEFAULT
Resolve the non-interactive/init default integration from the
SPECKIT_INTEGRATION_DEFAULT environment variable, fitting the existing
SPECKIT_INTEGRATION_* namespace. Falls back to the hardcoded "copilot"
default when unset, and warns to stderr (rather than silently falling
back) when the value is not a registered integration key.
Wires the resolver into specify init (interactive prompt default and
non-interactive fallback), the init workflow step, and the bundle init
default. Adds unit and CLI tests and documents the variable.
Closes#3939
Assisted-by: GitHub Copilot (model: claude-opus-4.8, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: eecda55f-fa13-42f7-99bf-bfb0bb8565a0
* test: cover env-var default wiring for picker, workflow step, and bundle
Address PR review: add regression tests so each SPECKIT_INTEGRATION_DEFAULT
wiring site cannot silently revert to the hardcoded constant.
- init.py: interactive picker receives the resolved key as default_key.
- workflow init step: no step/workflow default + env var drives output
integration and argv.
- bundle _resolve_init_integration: env-var default applies when unspecified,
while explicit override and manifest-declared integration still win.
Assisted-by: GitHub Copilot (model: claude-opus-4.8, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: eecda55f-fa13-42f7-99bf-bfb0bb8565a0
---------
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: eecda55f-fa13-42f7-99bf-bfb0bb8565a0
* feat(extensions): scaffold config templates on extension add/enable
Deploy an extension's provides.config templates into .specify/ when the
extension is added or enabled. Existing files are never overwritten, so
user customizations are preserved.
Addresses the review on #2000:
- ExtensionManifest.config returns [] unless provides.config is a list of
dicts, so a malformed manifest cannot crash callers.
- scaffold_config returns a consistent (deployed, skipped_existing, failed)
tuple on every path, including a missing manifest.
- Template paths must resolve inside the extension dir and targets inside
.specify/; symlinks and non-regular files are rejected.
- Callers distinguish "already exists (preserved)" from "not scaffolded",
and extension_enable no longer crashes on a corrupt manifest.
- Tests cover traversal, absolute paths, symlinks, directory templates,
malformed provides.config, and the missing-manifest tuple shape.
Ported onto the extensions package introduced by #3014: the manager and
manifest changes land in extensions/__init__.py and the CLI wiring in
extensions/_commands.py.
* fix(extensions): deploy config where it is read, and contain the write
Addresses @Copilot's review.
Config now lands in .specify/extensions/<id>/ rather than the .specify/
root. ConfigManager._get_project_config() reads
.specify/extensions/<id>/<id>-config.yml, and the bundled scripts and
READMEs use the same path, so a scaffolded git-config.yml was being
written somewhere the git extension never looks.
Containment is checked component by component before .specify is used as
the root. Resolving it first and trusting the result let a symlinked
component point outside the project, after which every target satisfied
relative_to and copy2 wrote externally. This matches the project
safe-write path in shared_infra.
mkdir moved inside the OSError handler. A nested target like
foo/config.yml raised out of scaffolding when its parent could not be
created, and on extension add that happened after the extension was
already installed.
The 'Configuration may be required' warning is now conditional. It ran
unconditionally after the scaffolding block, so it contradicted the
success output directly above it and fired for extensions with no
provides.config at all.
Tests cover the corrected location, a symlinked config root, and an
uncreatable nested target.
* fix(extensions): only scaffold config targets that removal preserves
remove(keep_config=True) rmtree's every subdirectory and keeps only top-level
-config.yml / -config.local.yml files; the backup path globs the same
top-level pattern. Scaffolding a nested or differently-named target therefore
handed the user a file that 'extension add --force' silently deleted and
replaced with the template default, losing customization.
Constrain scaffold targets to that convention rather than widening four
removal paths.
---------
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
execute() reads `on_reject = config.get("on_reject", "abort")` and, in the
reject branch, handles only "abort" and "retry" before falling through to
its `# on_reject == "skip"` case. So any other value makes a REJECTED gate
report COMPLETED and the run walks straight past the review the gate
exists to enforce:
on_reject='abort' -> failed "Gate rejected by user at step 'g'"
on_reject='retry' -> paused
on_reject='skip' -> completed (by design)
on_reject='Abort' -> completed <-- rejection silently discarded
on_reject='fail' -> completed <-- same
on_reject='stop' -> completed <-- same
on_reject=None -> completed <-- same
on_reject=5 -> completed <-- same
Reachable by a capitalisation slip, a guessed verb, a non-string, or a
bare `on_reject:` — note `config.get(k, default)` does NOT substitute the
default for an explicit YAML null.
`validate` already rejects anything outside abort/skip/retry, but the
engine does not auto-validate before execute(). Fail loudly instead,
mirroring the `options` and `verdict_input` guards in the same method, and
placed before the non-TTY short-circuit so it surfaces in CI too.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: eliminate TOCTOU race in zip packaging
Open file once and derive both stat info and content from the same file
descriptor to prevent race conditions where the file is modified between
stat() and read_bytes() calls.
* test: add regression test for TOCTOU stat/read consistency in packager
The old implementation called file_path.stat() then file_path.read_bytes()
as separate syscalls. The fix opens the file once and uses os.fstat() +
fh.read() on the same handle. This test verifies the archived bytes and
mode are consistent with the opened file descriptor.
execute() did:
output_config = config.get("output") or {}
if not isinstance(output_config, dict):
output_config = {}
so every non-mapping `output` was silently discarded and the step still
returned COMPLETED — every declared aggregation key vanished, and
downstream `{{ steps.<id>.output.<key> }}` resolved to None and
interpolated as an empty string:
output=[] -> completed, error=None
output=False -> completed, error=None
output=0 -> completed, error=None
output='' -> completed, error=None
output=['a'] -> completed, error=None
output='oops' -> completed, error=None
output=5 -> completed, error=None
`validate` already rejects this and its comment names the flaw exactly:
"execute() silently coerces a non-mapping output to {}, so the author's
declared aggregation keys would vanish with no error." The engine does not
auto-validate before execute(), so on an unvalidated run that is what
happened — and `x or {}` masked the falsy shapes before the isinstance
check even ran.
Fail loudly with validate()'s own message, mirroring the `wait_for` guard
in the same method. An explicit `output:` (YAML null) stays valid.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Add --extension flag to specify init for installing extensions at init time
Adds a repeatable --extension flag to `specify init` so users can opt into
extensions (bundled name, local path, or HTTPS URL) during initialization,
without a separate `specify extension add` step.
- New `_install_extension_during_init` helper in commands/init.py that
auto-detects source type (URL / local path / bundled name / catalog) and
installs via ExtensionManager. Failures are non-fatal and recorded in the
tracker without aborting init.
- Extension tracker steps are pre-registered before the Live context and run
after preset install, before finalize.
- Five new tests in TestExtensionFlag covering bundled name, multiple
extensions, local absolute path, unknown extension (graceful error), and
combination with --preset.
Rebased onto upstream/main and adapted to the refactored init command
(moved to src/specify_cli/commands/init.py) from stale PR #2396.
Assisted-by: GitHub Copilot (model: Claude Opus 4.8, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
* Address review: reuse hardened downloader, refresh events, escape labels, fix bundler call
Responds to review feedback on #3914 and fixes CI (pytest bundler failure).
- Extract shared `install_extension_from_url` helper in extensions/_commands.py
that reuses the authenticated, redirect-guarded, bounded (50 MiB) download
and TOCTOU-safe transient archive used by `extension add --from`. Both
`extension add --from` and `specify init --extension <url>` now go through
this single downloader instead of a second raw urlopen path.
- Refresh native event configuration once after successful extension installs
during init (mirrors `_refresh_events_and_warn` in the add path) so an
extension declaring `events:` has its hooks activated.
- Escape user-controlled extension specs and error text before interpolating
them into StepTracker labels (Rich markup injection).
- Pass `extensions=None` from bundler's `_run_init` so the init callback no
longer receives the typer OptionInfo sentinel ('OptionInfo' object is not
iterable), which broke `test_install_initializes_uninitialized_project`.
- Add init URL coverage in TestExtensionFlag: non-HTTPS rejection and a
successful HTTPS ZIP install with download-cache cleanup assertion.
Assisted-by: GitHub Copilot (model: Claude Opus 4.8, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
* Add default-deny trust confirmation for URL extension installs at init
URL-based --extension installs now require explicit trust, matching the
`extension add --from` posture. Interactive sessions show an "Untrusted
Source" panel and prompt (default no); non-interactive sessions deny by
default unless --trust-extension-urls is passed. Trust is resolved before
the Live display since the prompt can't be answered under the spinner.
- Add --trust-extension-urls option and _ext_spec_is_url /
_confirm_extension_url_trust helpers
- Skip (not abort) unconfirmed URL extensions, consistent with other
non-fatal extension failures
- Pass trust_extension_urls=False from the bundler init callback
- Add tests for deny-by-default, interactive confirm, and trusted install
Assisted-by: GitHub Copilot (model: Claude Opus 4.8, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8bc6802d-81b8-48f4-8f60-cba3aebc3bb3
---------
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8bc6802d-81b8-48f4-8f60-cba3aebc3bb3
A gate with `on_reject: retry` consumes a bound reject verdict before
pausing by resetting the named input to `""` (documented behaviour, so a
later resume prompts again). Every `resume()` re-resolves the persisted
inputs through `_coerce_input`.
Those two rules collide when the bound input declares an `enum` that does
not list `""`. The reset writes a value the input's own enum forbids, and
the run wedges:
inputs:
spec_verdict:
type: string
enum: [approve, reject]
steps:
- id: review
type: gate
options: [approve, reject]
on_reject: retry
verdict_input: spec_verdict
$ specify workflow run wf --input spec_verdict=reject
Status: paused
$ specify workflow resume <run_id> --input note=b
Error: Input 'spec_verdict' value '' not in allowed values:
['approve', 'reject'].
The workflow validates clean and the first run looks fine, so the failure
only appears at the second resume. It is also unrecoverable in practice:
`_resolve_inputs` re-coerces the whole persisted map, so *any* resume that
supplies an input dies on the stored `""`. Only a resume with no inputs at
all still works -- and that is precisely the call that cannot deliver a new
verdict, which is the one thing the retry cycle exists to allow.
Extend the existing `verdict_input` cross-check (which already confirms the
name is declared) to also require that a retry-bound input's `enum` admits
the reset sentinel, and report it with a fix hint. To do that, thread the
input *definitions* through `_validate_steps` instead of just their names.
Rejected the alternative of popping the key instead of writing `""`: that
lets the input's `default` flow back in on the next resume, so a gate the
user just rejected would silently auto-approve.
Docs: note the `enum` requirement next to the reset behaviour it follows
from.
Adds 4 validation tests for the new guard plus a characterization test that
drives the engine directly to pin the wedge it prevents.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Assisted-by: Claude Opus 5 (1M context)
`workflow resolve` printed two lines through `console.print`, which has
Rich markup enabled, without escaping:
1. The layer tier was wrapped in literal brackets:
`f" • [{layer.tier}] {layer.source} ..."`. Rich parsed `[base]`
and `[project-overlay]` as style tags, so the tier label was swallowed
on *every* invocation -- no untrusted input required. The column has
never rendered.
2. Step attribution interpolated `composed.step_id` raw. Step IDs come
from base-workflow / overlay YAML and are only validated against `:`
(see `_parse_edit`), so brackets pass validation. A balanced
`[stuff]` is silently swallowed; an unbalanced `[/red]` raises
`rich.errors.MarkupError`, producing an uncaught traceback and exit 1
-- the workflow cannot be inspected at all.
Route the interpolated fields through `rich.markup.escape` and escape
the literal tier bracket as `\[`, matching the existing pattern in
`workflow info`'s step graph and `workflow_list`'s `\[disabled]`.
Only display is affected; the returned payload was already unescaped and
is unchanged.
Adds 3 regression tests, all of which fail without the fix: the tier
label renders, and a step ID survives both the swallowing and the
crashing markup cases.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Assisted-by: Claude Opus 5 (1M context)
* fix(catalogs): validate the port in the shared catalog-URL validator
`CatalogStackBase._validate_catalog_url()` reads `parsed.hostname` inside
its `try/except ValueError` but never reads `parsed.port`. `urlparse()` and
`.hostname` do not perform port validation — only `.port` does — so a
catalog URL with a non-numeric or out-of-range port passes validation.
Every implementation that documents itself as mirroring this function
already reads `.port` inside the same try: workflows/catalog.py (4 sites),
bundler/services/adapters.py (2), bundler/commands_impl/catalog_config.py,
and commands/bundle/__init__.py. The shared base — inherited by
ExtensionCatalog and IntegrationCatalog — is the only one without it.
The accepted URL then escapes as a raw `http.client.InvalidURL`, which is
neither `urllib.error.URLError` nor `json.JSONDecodeError` (the only two
the fetcher converts), so it surfaces as an unhandled traceback rather
than the validator's normal error.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs(catalogs): describe both bad-port failure modes accurately
The comment attributed both malformed-port cases to
http.client.InvalidURL. Only a non-numeric port raises that (when the
connection object is built); an out-of-range port constructs fine and
fails later in the socket layer. Measured:
example.invalid:notaport -> http.client.InvalidURL: nonnumeric port
example.invalid:65536 -> HTTPSConnection() OK, connect() fails
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(extensions): harden URL download cache
Assisted-by: GitHub Copilot (model: GPT-5.6 Sol, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* fix(extensions): retain secure archive descriptor
Assisted-by: GitHub Copilot (model: GPT-5.6 Sol, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Harden extension URL cache anchor opens
Assisted-by: GitHub Copilot (model: MAI-Code-1-Flash, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Use descriptor-safe mkdir for cache components
Assisted-by: GitHub Copilot (model: MAI-Code-1-Flash, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Harden extension URL download cache
Assisted-by: GitHub Copilot (model: MAI-Code-1-Flash, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Align extension manifest regression expectation
Assisted-by: GitHub Copilot (model: MAI-Code-1-Flash, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Make download cache leaf anonymous to remove cleanup TOCTOU
Address review: the best-effort cleanup walk re-derived the downloads
directory by path, so a cache ancestor swapped after the archive was
opened could redirect os.unlink to a replacement leaf, and it silently
no-op'd (failing open) on platforms without descriptor-relative unlink.
_safe_open_download_zip now unlinks the exclusively-created leaf
immediately via the same directory descriptor, returning an fd backed by
an anonymous inode. Installation already consumes that descriptor through
archive_file, so the on-disk pathname is never reopened and no cleanup
walk is needed. The capability gate additionally requires os.unlink in
os.supports_dir_fd, so unsupported platforms fail closed. Removed the
now-unused _safe_unlink_download_zip helper and its cleanup finally, and
updated the tests accordingly.
Assisted-by: GitHub Copilot (model: MAI-Code-1-Flash, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Fix Windows test matrix for cache hardening tests
The hardened cache primitives fail closed on platforms without dir_fd/
O_NOFOLLOW, so on the windows-latest matrix several tests errored instead
of exercising POSIX behavior:
- test_symlinked_cache_ancestor_is_refused and
test_cache_ancestor_resolving_outside_project_is_refused called
_validate_safe_cache_dir directly and expected typer.Exit, but on
Windows it raises NotImplementedError first. Guard both with
_require_secure_dir_fd() so they skip where the primitive is unavailable.
- test_safe_open_fails_closed_without_atomic_platform_support built its
download dir via _validate_safe_cache_dir, which itself fails closed on
Windows; construct the directory directly so the assertion targets
_safe_open_download_zip's platform gate in isolation.
- The _open_test_download_zip stand-in unlinked a still-open file, which
raises PermissionError on Windows. Use O_TEMPORARY there (auto-delete on
close) and keep immediate unlink on POSIX.
Assisted-by: GitHub Copilot (model: MAI-Code-1-Flash, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Support Windows in extension URL download-cache hardening
Replace the fail-closed NotImplementedError on platforms lacking dir_fd
with a portable, still-hardened download path so `specify extension add
--from <url>` works on Windows instead of rejecting the install.
- `_validate_safe_cache_dir` now dispatches to a POSIX dir_fd + O_NOFOLLOW
walk when available, and otherwise a portable path-wise walk that rejects
symlink/junction components before and after each mkdir and requires every
component to resolve back under the project root.
- `_safe_open_download_zip` keeps the POSIX anonymous-inode create/unlink and
adds a portable leaf create using O_EXCL + O_TEMPORARY (auto-delete on
close) plus a post-open fstat/lstat inode-identity check to detect a leaf
swapped underneath us. Installation still consumes only the open
descriptor, so the cache pathname is never reopened.
- Detect the symlink-refusal case via errno (ELOOP/ENOTDIR/EMLINK) instead of
FileExistsError, and add O_CLOEXEC to the descriptor-walk opens.
- Drop the now-unreachable NotImplementedError handling in the --from branch.
- Tests: cover the portable path (success, symlinked-leaf refusal, symlinked
ancestor refusal, full --from install) and keep the POSIX-only cases guarded.
Assisted-by: GitHub Copilot (model: Claude Opus 4.8, autonomous)
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
---------
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 8f71e02a-bc64-4593-b305-2554debe96f6
* Fix bundle-update-force-mislead: add refresh() to DefaultPrimitiveInstaller
Apply the remediation from the bug assessment on issue #3424.
DefaultPrimitiveInstaller lacked a refresh() method, causing
_refresh_component() to fall back to install(), which calls
ExtensionManager.install_from_directory() with force=False. This
raised ExtensionError with a leaked --force hint that bundle update
does not support, leaving users with no valid recovery path.
Fix: add refresh() to each kind manager (ExtensionKindManager and
PresetKindManager delegate to _do_install(force=True); WorkflowKindManager
and StepKindManager delegate to install() as their callables are
idempotent). DefaultPrimitiveInstaller.refresh() dispatches to the kind
manager's refresh(). PresetManager.install_from_directory() and
install_from_zip() gain a force parameter that removes the existing
preset before reinstalling, mirroring ExtensionManager's force semantics.
Refs #3424
Assisted-by: GitHub Copilot (model: claude-sonnet-4.6, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address review feedback on primitives.py and test_bundler_primitives.py
- Replace ... with pass in _KindManager Protocol method stubs
- Conditionally pass force= keyword only when force=True in _PresetKindManager
- Fix _StepKindManager.refresh() to remove step before re-installing
- Rename test to reflect actual assertion (refresh succeeds + force=True)
- Remove duplicate install_bundle import
Assisted-by: GitHub Copilot (model: claude-sonnet-4.5, autonomous)
* fix: add missing role/effective_integration to InstallPlan in _plan() and remove redundant import
- Remove duplicate `DefaultPrimitiveInstaller` import inside test body
(already imported at module scope on line 15)
- Add required `role` and `effective_integration` fields to `InstallPlan`
constructor in `_plan()` helper to prevent TypeError at runtime
Assisted-by: GitHub Copilot (model: claude-sonnet-4.6, autonomous)
* fix: address latest PR review comments
Assisted-by: GitHub Copilot (model: gpt-5.6-terra, autonomous)
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Replace unbounded f.read() with chunked iteration to prevent excessive
memory allocation on large or corrupted manifest files. Matches the
pattern used in integrations/manifest.py _sha256().
* feat: bind gate verdict to workflow input via verdict_input
Add an optional `verdict_input` field to gate steps that lets an
external system supply a verdict through a declared workflow input
instead of an interactive TTY prompt.
When the referenced input carries a non-empty string value that
matches one of the gate's `options` (case-insensitive), the gate
auto-decides, records the matched spelling in `output.choice`, and
applies the existing `on_reject` / abort / skip / retry semantics.
If the value is present but does not match an option, or is a
non-string, the gate fails immediately with a clear error message.
When the input is absent, null, or empty, the gate falls back to
today's TTY-prompt-or-pause behaviour unchanged.
The engine now persists `result.error` alongside each step's status
and output so that failed-step error messages survive across runs.
The CLI (`workflow run` and `workflow resume`) surfaces these
persisted errors after a failed or aborted run. `validate_workflow`
cross-references `verdict_input` against the workflow's declared
inputs block and reports an error for undeclared names, consistent
with the existing `wait_for` id cross-check.
Closes discussion: https://github.com/github/spec-kit/discussions/3717
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Assisted-by: GitHub Copilot (model: claude-sonnet-4.6, autonomous)
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Fix workflow JSON error payloads
Include persisted step errors in _workflow_run_payload so workflow run/resume/status --json all surface failure reasons consistently. Add JSON-path tests for failed and successful runs.
Assisted-by: GitHub Copilot (model: gpt-5.3-codex, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* fix(workflows): reject verdict inputs in fan-out
Fan-out items share workflow inputs and cannot safely consume a bound gate verdict. Reject verdict_input bindings during validation and at runtime while preserving unbound gates.
Assisted-by: GitHub Copilot (model: GPT-5.6 Sol, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix: update workflow command handling
Assisted-by: GitHub Copilot (model: gpt-5.3-codex, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Markus <markus@example.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
PR #3847 hardened the prompt step's `timeout` guard against a huge-int
value, but its twin in the shell step — the step the prompt one was
mirrored from — still has the hole.
`math.isfinite(10**400)` raises `OverflowError: int too large to convert
to float`. A 400-digit YAML scalar is an `int` and is not a `bool`, so it
clears every clause before `isfinite()` and raises there, escaping
`_timeout_error()` as exactly the uncaught crash that helper exists to
prevent:
steps:
- id: qa
type: shell
run: echo hi
timeout: 1000...0 # 400 digits
$ specify workflow run wf.yml
Traceback (most recent call last):
...
File "src/specify_cli/workflows/engine.py", line 361, in _validate_steps
step_errors = step_impl.validate(step_config)
File "src/specify_cli/workflows/steps/shell/__init__.py", line 127
or not math.isfinite(timeout)
OverflowError: int too large to convert to float
`workflow_run` calls `engine.validate()` before executing any step, so
the OverflowError propagates out of `validate_workflow` and kills the
command with a bare traceback that names neither the step nor the field,
instead of the "Workflow validation failed" report. `execute()` shares
the same helper, so an unvalidated run raises there too — and the engine
re-raises anything a step throws, aborting the whole workflow after
earlier steps have already run their side effects. The value is
genuinely invalid rather than merely unrepresentable in the check:
`subprocess.run(timeout=10**400)` raises the same OverflowError.
Unlike the prompt step, the shell step checks `isfinite()` *before*
`timeout <= 0`, so a negative huge int (`-(10**400)`) crashes as well
rather than being caught by the sign check.
Wrapped the condition in `try/except OverflowError` and treated the
value as invalid, mirroring the prompt step's guard so both steps reject
the same values with the same message. Now:
Workflow validation failed:
- Shell step 'qa': 'timeout' must be a positive number of seconds,
got 1000...0.
Valid int/float timeouts, non-finite floats, bools, strings and
non-positive values are unaffected — the existing clauses are unchanged.
Regression tests in `TestShellStep`: `validate()` rejects both signs of
the huge int, `validate_workflow()` reports it end to end (pinning the
path the CLI actually takes, not just the helper), and `execute()` fails
only that step with `subprocess.run` patched to assert it is never
reached. Test-the-test: reverting the source change fails all three with
`OverflowError` and leaves the rest of `TestShellStep` passing.
Assisted-by: Claude Code (model: claude-opus-5, under direct human supervision)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(workflows): validate prompt step 'timeout' like the shell step
PR #3768 added a `timeout` to the prompt step and passed it straight into
`subprocess.run(timeout=...)`. Neither `validate()` nor `execute()` checks
it, so a bad value from a user-authored `workflow.yml` escapes as a raw
exception:
steps:
- id: first
type: shell
run: echo side-effect
- id: ask
type: prompt
prompt: do it
timeout: abc
$ specify workflow run wf.yml
> [first] shell ...
Workflow failed: unsupported operand type(s) for +: 'float' and 'str'
The engine re-raises anything a step throws, so this takes down the whole
run — after `first` has already run its side effect — with a message that
names neither the step nor the field. `timeout: .nan` raises `ValueError:
cannot convert float NaN to integer` the same way, and a non-positive
`timeout` (`0`, `-5`) makes `subprocess.run` report an immediate
TimeoutExpired for a command that never got the time to run. `timeout:
true` silently becomes a 1-second limit, since bool is an int subclass.
The sibling shell step already rejects exactly these values via a
`_timeout_error()` helper shared by `execute()` and `validate()`, so the
same workflow failed validation cleanly as a shell step and crashed as a
prompt one. Mirrored that helper onto PromptStep: `validate()` reports the
contract error, and `execute()` re-checks it so an unvalidated run fails
just that step instead of aborting. Now:
Workflow validation failed:
- Prompt step 'ask': 'timeout' must be a positive number of seconds,
got 'abc'.
caught before the first step runs. Positive int/float timeouts and an
absent `timeout` are unaffected.
Regression tests in `TestPromptStep` mirror the shell step's: validate
rejects "30"/True/inf/nan/0/-5/list/None, validate accepts 300/5/0.5 and
an absent field, and execute fails cleanly with `subprocess.run` patched
to assert it is never reached. With the source fix reverted, all 9
rejection tests fail.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Assisted-by: Claude Code (model: claude-opus-5, under direct human supervision)
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* test(workflows): cover the huge-int timeout OverflowError guard
The autofix commit wrapped the prompt step's `_timeout_error()` check in
`try/except OverflowError` but added no test, so nothing pins the
behaviour it introduced.
`math.isfinite(10**400)` raises `OverflowError: int too large to convert
to float` — the value is an `int`, is `> 0`, and is not a `bool`, so it
clears every other clause of the guard and reaches `isfinite()`. Without
the `except`, validating
```yaml
- id: ask
type: prompt
prompt: do it
timeout: 1000...0 # 400 digits
```
raises that `OverflowError` out of `validate()`/`execute()` — exactly the
uncaught-crash failure mode this guard was added to prevent. The same
value raises `OverflowError` from `subprocess.run(timeout=...)`.
Add `10**400` to both parametrized rejection lists (`validate()` and the
`execute()` fails-cleanly loop). Test-the-test: reverting the `try/except`
fails both new cases with `OverflowError` and leaves the rest passing.
Assisted-by: Claude Opus 5 (model: claude-opus-5, autonomous)
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
`preset catalog add` and `preset catalog remove` interpolate the raw
`--name` and URL into `console.print()`, so Rich parses them as markup.
Two failure modes:
* Silent misreporting — a name like `[bold red]pwned[/]` is printed as
`pwned`, so the confirmed name is not the persisted name and a later
`remove` with the reported name fails.
* Unhandled MarkupError — an unbalanced closing tag raises, and because
the crash happens *after* preset-catalogs.yml is written, the user gets
a traceback for a catalog that was in fact added.
This file already imports `_escape_markup` and escapes name/description/
url in `preset catalog list` (whose invariant `test_catalog_list_escapes_
rich_markup` already pins); `add`/`remove` were the remaining gaps.
Only rendering changes: the raw values are still what get persisted and
what the duplicate-name comparison uses.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
github-actions[bot]GitHubgithub-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>CopilotManfred RiemCopilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
* Fix upgrade-overwrites-copilot-skills: pass force=True to extension skill re-registration after upgrade
Apply the remediation from the bug assessment on issue #3849.
_register_extension_skills() had a skip guard that refused to overwrite
existing SKILL.md files (protecting user customizations). In the upgrade
path, setup() regenerates all core-template SKILL.md files first, then
calls register_enabled_extensions_for_agent(). The guard then sees those
freshly-written core files as 'existing' and skips every extension, leaving
only core template content on disk.
Fix: add force: bool = False to _register_extension_skills() and thread it
through register_enabled_extensions_for_agent() and
_register_extensions_for_agent(). In integration_upgrade(), pass force=True
so extension content layers on top of the just-regenerated core files.
The force flag is off-by-default so plain extension add still protects
user-modified skill files.
Refs #3849
Assisted-by: GitHub Copilot (model: claude-sonnet-4.6, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Potential fix for pull request finding 'Unused local variable'
Co-authored-by: Copilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
* test: add end-to-end regression guard for upgrade-overwrites-copilot-skills (#3849)
The existing regression tests in TestRegisterExtensionSkillsForceFlag exercise
the new force parameter at the helper level, so without the fix they fail only
with a TypeError (unknown kwarg) rather than on the user-facing behaviour.
Add a command-level test that runs 'specify integration upgrade copilot --skills
--force' end-to-end and asserts the installed git extension's SKILL.md is
restored (with its extension content, not a bare core-template stub) when the
skill directory already exists — the exact skill_dir_preexists path the bug
depends on. The test fails on pre-fix source (the skill is never recreated) and
passes with the fix, so it is a genuine behavioural regression guard rather than
an API-surface check.
Refs #3849
Assisted-by: GitHub Copilot (model: claude-opus-4.8, autonomous)
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Manfred Riem <15701806+mnriem@users.noreply.github.com>
Co-authored-by: Copilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
`IntegrationManifest.uninstall()` guards every tracked-file `path.unlink()`
with `except OSError: skipped.append(path)`, but the manifest's own
`manifest.unlink()` is bare. The manifest is deleted *last*, so an
undeletable manifest (read-only file, a directory left at the path, a
Windows lock) raises after the tracked files are already gone.
The caller loses the `(removed, skipped)` result and never runs its
post-uninstall bookkeeping — reassigning the default integration,
rewriting/removing `integration.json`, clearing init options — leaving a
removed integration still recorded as installed.
Report it in `skipped` like any other file we could not remove, mirroring
the `path.unlink()` guard above and the same `except OSError:
skipped.append(...)` pattern in kimi's legacy-directory cleanup.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
CommandRegistrar.render_toml_command passes the raw frontmatter `description`
straight into `_render_basic_toml_string`, which iterates the value and calls
ord() on each character. Frontmatter comes from yaml.safe_load, so description
can be any YAML type:
description='ok string' -> description = "ok string"
description=None -> TypeError: 'NoneType' object is not iterable
description=42 -> TypeError: 'int' object is not iterable
description=True -> TypeError: 'bool' object is not iterable
description=['a','b'] -> description = "ab" <- silently WRONG value
This is a format-branch asymmetry: it is the only renderer reached from
register_commands' format branches that does not normalise description.
render_yaml_command (same class, ~70 lines below) already does exactly
`if not isinstance(description, str): description = str(description) if
description is not None else ""`, render_markdown_command goes through
yaml.dump which handles any type, and TomlIntegration._extract_description
returns "" for a non-str. So only extension/preset commands rendered for the two
TOML agents were affected.
Apply the same coercion the sibling uses. After: None -> "", 42 -> "42",
True -> "True", ['a','b'] -> "['a', 'b']", each still valid parseable TOML.
String descriptions are untouched.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
