* feat(cli): pull latest LLM models dynamically in the crew wizard
The JSON-crew creation wizard hardcoded a short model list per provider,
which goes stale as vendors ship new models every few weeks. Add a
three-tier resolver that prefers live data and falls back to a curated list.
- New `model_catalog.get_provider_models(provider, fallback)`:
1. Vendor API (openai/anthropic/gemini/groq/cerebras/ollama) when the
provider key is already in the environment — the only reliably-fresh
source (real release dates / display names).
2. Curated hardcoded fallback — hand-verified, used when no key is set.
3. LiteLLM feed — only for providers with no curated list; it lags real
releases, so it must never preempt the curated fallback.
- Rank by date/version parsed from model ids, humanize labels, 6h cache,
short timeouts, silent fallback on any error.
- Wire it into `create_json_crew._select_model()` (picker only).
- Refresh the curated fallback against each vendor's official model docs
(Anthropic Fable 5 / Opus 4.8 / Sonnet 5; OpenAI GPT-5.5(+pro); Gemini
3.5 Flash / 3.1 Pro preview / 3 Flash preview; Groq Llama 4 / GPT-OSS).
- Tests for ranking, chat filtering, caching, and the tier order (17 tests).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): address model_catalog review findings
- Key the Ollama catalog cache by its base URL so a changed OLLAMA_API_BASE /
API_BASE no longer serves the previous host's models for up to the TTL.
- Negatively cache the curated fallback after a failed/empty fetch (short
_NEGATIVE_TTL) so the picker doesn't repeat a timeout-prone vendor/LiteLLM
request on every call — most impactful for a down local Ollama server.
- Guard _read_catalog_cache / _write_catalog_cache against a non-dict cache
root (corrupt JSON array no longer raises AttributeError).
- Replace the two empty `except OSError: pass` blocks with
contextlib.suppress(OSError) plus an explanatory comment (CodeQL empty-except).
- Tests: negative cache, base-keyed Ollama cache, corrupt-cache no-crash (20 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): guard null litellm_provider and paginate Gemini models
- _from_litellm: coerce a present-but-null `litellm_provider` before string
ops so it's skipped instead of raising AttributeError (keeps the documented
"never raises" contract).
- _fetch_gemini: walk models.list pages via nextPageToken (bounded to 10) —
the API is paginated and not guaranteed newest-first, so a single page could
drop models the ranking should consider.
- Tests for both (22 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* ci: ignore nltk PYSEC-2026-597 in pip-audit (no fix, not reachable)
pip-audit newly flags nltk 3.9.4 for PYSEC-2026-597 (CVE-2026-12243), a path
traversal via percent-encoded `..%2f` in nltk.data.load()/find(). It affects
all nltk versions <=3.9.4 with no patched release, so it can't be resolved by a
version bump — same situation as the already-ignored PYSEC-2026-97.
nltk is a transitive dependency (unstructured[local-inference, all-docs] in
crewai-tools) used for text tokenization; we never pass untrusted resource
URLs/paths to nltk.data, so the traversal is not reachable. Add it to the
curated --ignore-vuln list with a justification, matching the existing pattern.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): address second Cursor review round on model_catalog
- Cache ignores new API keys: include API-key presence in the cache key
(`<provider>#key|#nokey`), so a key added after a no-key/negative-cached
lookup triggers a fresh live fetch instead of serving the stale fallback.
- Bad LiteLLM cache crashes picker: `_from_litellm` now requires a dict from
`_load_litellm_data` (a non-mapping JSON root is skipped, not `.items()`'d).
- Stale LiteLLM refetch loop: memoize the feed load once per process
(`_litellm_memo` + `_reset_litellm_memo` test hook) so repeated uncurated-
provider lookups don't each re-attempt a timed download when offline.
- Tests: new-key bypass, corrupt-litellm-cache no-crash, one-fetch-per-process
(25 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): keep partial Gemini results on later-page fetch error
_fetch_gemini paginates; a network/HTTP error on page 2+ previously raised out
through _from_vendor, discarding models already parsed from earlier pages and
forcing the curated fallback. Catch per-page fetch errors and return the
partial set instead (a first-page failure still yields an empty list -> fallback).
Test: test_vendor_gemini_keeps_partial_on_later_page_error (26 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): don't let an invalid fresh LiteLLM cache block download
_fetch_litellm_data treated any truthy JSON root in a fresh provider_cache.json
as the feed and returned it, so a non-mapping root (e.g. a JSON array) was
memoized and the tier never re-downloaded until the file aged out — leaving
uncurated providers with an empty picker despite a recoverable cache. Only
short-circuit on a usable dict; otherwise fall through to the download.
Test renamed to test_invalid_litellm_cache_falls_through_to_download (asserts
recovery via refetch).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): honor OLLAMA_HOST and treat empty vendor list as authoritative
- Ollama host mismatch: _ollama_base now also reads OLLAMA_HOST (the Ollama
runtime convention) after OLLAMA_API_BASE/API_BASE, normalizing a scheme-less
value (e.g. "127.0.0.1:11434" -> "http://127.0.0.1:11434"), so users who set
only OLLAMA_HOST see models from the server the crew will actually use.
- Empty vendor list: a successful vendor fetch returning no models is now
authoritative instead of collapsing to the curated fallback. A reachable
Ollama with nothing installed yields an empty list (the picker prompts for
manual entry) rather than offering hardcoded models that aren't installed; a
failed fetch still falls back. _from_vendor now returns [] on success-empty
and None only when the tier is unavailable.
- Tests: ollama empty->manual, ollama down->fallback, OLLAMA_HOST resolution
(29 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): Gemini first-page failure falls back instead of showing empty
Interaction from the prior two fixes: _fetch_gemini swallowed a first-page
error and returned [], which _from_vendor reported as a successful-empty result
and get_provider_models treated as authoritative — skipping the curated Gemini
fallback and jumping to manual entry. Now a first-page failure (nothing gathered
yet) re-raises so _from_vendor returns None and the curated list is used; a
later-page failure still keeps the partial results.
Test: test_vendor_gemini_first_page_error_uses_fallback (30 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): Gemini GOOGLE_API_KEY + Ollama recovery not blocked by cache
- Gemini ignores GOOGLE_API_KEY: _PROVIDER_KEY_ENV now maps each provider to a
tuple of accepted env vars; Gemini accepts GEMINI_API_KEY or GOOGLE_API_KEY
(matching crewai's own Gemini provider). A new _provider_api_key() resolver
is used by both _from_vendor and the cache key, so a GOOGLE_API_KEY user gets
the live models API instead of the stale curated fallback.
- Ollama recovery blocked by cache: skip the negative (fallback) cache for
Ollama. It's a local, fast-failing server, so re-probing each call is cheap
and lets the picker pick up real installed models as soon as the server comes
up, instead of serving suggestions for the negative-cache TTL.
- Tests: GOOGLE_API_KEY live fetch, Ollama down->recover (32 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* style(cli): ruff format model_catalog.py
Add the blank line ruff format expects after _provider_api_key; no behavior
change. Fixes the lint-run `ruff format --check lib/` step.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): don't treat 'search' substring as a non-chat model marker
The 'search' entry in _NON_CHAT_MARKERS matched anywhere in a model id, dropping
legitimate completion models like gpt-4o-search-preview and anything containing
'research' (e.g. o3-deep-research, since 'search' is a substring). Remove it;
the remaining markers (embedding/audio/image/moderation/etc.) still filter
genuine non-chat models. Test: test_search_substring_not_treated_as_non_chat (33).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* Revert "ci: ignore nltk PYSEC-2026-597 in pip-audit"
Do not suppress an unpatched security advisory to make CI green. Remove
PYSEC-2026-597 from the pip-audit ignore list; leave the scan failing so it
keeps surfacing the nltk path traversal (CVE-2026-12243). This PR should not be
merged until nltk ships a fix (or the vulnerable transitive dep is otherwise
resolved).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): exclude fine-tuned models and checkpoints from the picker
With a live OPENAI_API_KEY, /v1/models returns the user's fine-tunes and
training checkpoints (ft:..., ...:ckpt-step-N). Their recent `created`
timestamps ranked them above the base models and filled every slot, so the
picker showed a wall of `ft:gpt-4o-mini-...:crewai::...` with mangled labels and
no foundation models at all. Skip fine-tunes/checkpoints in the OpenAI-shaped
fetcher so clean base models surface; a user who wants a fine-tune can still
enter it via the picker's "Other" option. Test:
test_openai_excludes_fine_tunes_and_checkpoints (34 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
* fix(cli): cleaner model labels + filter Ollama non-chat models
Anthropic/Gemini already use vendor display names; OpenAI/Groq/Cerebras/Ollama
fall to _humanize for anything outside the curated map, which produced mediocre
labels ("GPT Oss 120b", "qwen3 32b", "Deepseek r1", "llama3.3:70b").
Improve _humanize:
- split on ':' too (Ollama tags: llama3.3:70b -> "Llama3.3 70B")
- uppercase size suffixes (70b -> 70B), acronyms OSS/IT, brand casing
(DeepSeek, ChatGPT, QwQ)
- capitalize the leading letter of fused family+version tokens (qwen3 -> Qwen3)
while preserving OpenAI o-series lowercase (o3, o1-mini)
Also fix _fetch_ollama: /api/tags lists everything installed, so filter
non-chat (embedding) and fine-tune entries the same way the other tiers do.
Tests: expanded test_humanize + test_ollama_excludes_embedding_models (35 total).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBYGqJHC2TMC6fonFziuuh
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Lorenze Jay <63378463+lorenzejay@users.noreply.github.com>
* Implement message setup and feedback handling in AgentExecutor
- Added method to streamline message preparation for agent execution, allowing for integration with human input providers.
- Introduced and methods to manage the state during feedback processing.
- Enhanced and methods to re-run the executor flow using existing feedback messages.
- Updated tests to verify the new message setup and feedback handling functionality, ensuring compatibility with human input providers.
* dont commit runner
* Remove xfail marker from test_crew_train_success as training feedback migration to AgentExecutor is complete.
* fix runtype errors
* fix test
* revert
* mypy fix
* handled reset iterations
- added client_name header to the 4 tavily tools to classify incoming requests as 'crewai' requests.
- This is for internal analysis
Co-authored-by: Lorenze Jay <63378463+lorenzejay@users.noreply.github.com>
Inline agent and crew actions can now use repository-backed agents
without duplicating role, goal, and backstory in each definition.
Examples:
* `agent.with.from_repository: support_specialist`
* `crew.with.agents.researcher.from_repository: researcher`
`PlusAPI.get_agent` now uses the shared synchronous request path so
project loaders can fetch repository agents without nested event loops.
Flow action inputs now support `${...}` inside strings, not only
strings that are fully wrapped in one expression. This lets authored
flows use simple prompt-like values such as:
* `query: "News about ${state.topic}"`
* `input: "Ticket ${text(state, "ticket.id", "unknown")}"`
* `sources: ["${state.primary_source}", "archive-${state.topic}"]`
Whole-expression values still preserve their runtime type, so
`${state.limit}` remains a number and `${state.domains}` remains a list.
Mixed literal and expression strings render as text.
This removes the need to build labeled strings with CEL concatenation,
which was hard to read, easy to quote incorrectly in YAML, and a poor
fit for the Flow authoring skill examples.
### Overview
`BedrockCompletion.acall()` (the async completion path used when a crew is kicked off asynchronously) requires `aiobotocore` to build its async client. The `bedrock` extra, however, only declared `boto3`. Crews configured with an AWS Bedrock model work fine under a synchronous `kickoff()`, since that path only needs `boto3`, but raise `NotImplementedError: Async support for AWS Bedrock requires aiobotocore` as soon as they're kicked off asynchronously, since `aiobotocore` was never installed.
The fix adds `aiobotocore` to the `bedrock` extra, so `crewai[bedrock]` installs both the sync (`boto3`) and async (`aiobotocore`) dependencies the native Bedrock provider needs. The lockfile is regenerated to match. The exception message is also corrected — it previously pointed to a `bedrock-async` extra that never existed in `pyproject.toml`.
### Changes
- `lib/crewai/pyproject.toml`: add `aiobotocore~=3.5.0` to the `bedrock` extra
- `uv.lock`: regenerated to reflect the updated `bedrock` extra
- `lib/crewai/src/crewai/llms/providers/bedrock/completion.py`: fix the install hint in the `NotImplementedError` message to reference the real `bedrock` extra instead of the nonexistent `bedrock-async`
* Document flow agent options
Document and type inline Flow agent options so authored flows can set:
* `llm.model`, `llm.max_tokens`, and `llm.max_completion_tokens`
* `planning_config.max_attempts`
* `allow_delegation`
* `max_iter`
* `max_rpm`
* `max_execution_time` in seconds
Also tell the flow skill to omit optional fields unless needed.
* Potential fix for pull request finding 'Unused import'
Co-authored-by: Copilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
---------
Co-authored-by: Copilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
A method that listens to its own name can re-trigger itself or collide
with router events. Rejecting that definition keeps declarative and
Python-authored flows aligned before kickoff.
CEL string concatenation currently fails when prompt builders read
missing or null fields. This commit adds `text(root, "path", "default")`
custom CEL helper so prompt text can safely read nested state/output
values.
Point `crewai template list`/`template add` at the crewAIInc-fde GitHub
org so the FDE template_* repos are listed and installed instead of the
crewAIInc ones.
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Lorenze Jay <63378463+lorenzejay@users.noreply.github.com>
* Support inline skill definitions
This commit adds inline skill loading without a need for a file. It also
DRYs the skill loading feature.
* Address code review suggestions
* Type tool and app in CrewDefinition
This commit fixes a bug in the CrewDefinition class where the tool and
app were not being added.
* Type mcps= parameter
* Add generated Flow Definition authoring skill
Generate a portable skill from the Flow Definition schema so agents can
author valid declarative flows with the same reference CrewAI uses to
validate them. New declarative flow projects now write this skill.
```python
from crewai.flow.flow_definition import FlowDefinition
skill = FlowDefinition.skill(skips=(), examples_format="yaml")
```
* `examples_format` accepts `"yaml"` or `"json"`.
* Supported skips: `conversational`, `non_linear_flows`, `each`, `hitl`, `persistence`, `config`, `expression_action`, `script_action`, `tool_action`
The generated skill includes authoring rules, a routed crew example, and
an API reference extracted from the Flow, action, state, agent, crew,
and task Pydantic schemas.
* Fix declarative flow scaffold without framework import
* Fix skipped expression action guidance
* Fix markdown links in skill
* fix: freeze docs version nav from Edge instead of previous release
The docs cut copied every Edge file into the new `docs/v<X.Y.Z>/`
snapshot but built that version's `docs.json` navigation by cloning the
previous frozen release and only rewriting path prefixes. Pages added to
Edge since the last release were therefore copied to disk yet never
linked in the version selector, which is why the v1.15.0 cut shipped
without the Datadog guide. `_build_new_entry` now clones the Edge nav
entry and rewrites `edge/<locale>/` to `v<new>/<locale>/`, so promoting
Edge to Latest carries every current page and nav restructuring.
* docs: link the v1.15.0 Datadog guide dropped during the cut
The v1.15.0 freeze copied `enterprise/guides/datadog` into the snapshot
for every locale but never linked it in `docs.json`, because the cut
cloned the v1.14.7 nav instead of Edge. This backfills the missing nav
reference in the `en`, `pt-BR`, `ko`, and `ar` v1.15.0 blocks so the
already-shipped page is reachable from the version selector. Pairs with
the `_build_new_entry` fix that prevents future cuts from dropping pages.
* docs: link the v1.15.1 Datadog guide dropped during the cut
The v1.15.1 cut ran before the freeze-from-Edge fix landed, so it
inherited the same bug as v1.15.0: `enterprise/guides/datadog` was
copied into the snapshot for every locale but never linked in
`docs.json`. This backfills the missing nav reference in the `en`,
`pt-BR`, `ko`, and `ar` v1.15.1 blocks so the page is reachable from the
version selector.
JSON-formatted stdout is now the only supported log shape in CrewAI
Enterprise — the `CREWAI_LOG_FORMAT=json` opt-in env var is gone and
no longer needs to be configured in AMP. Removes the "Enabling JSON
output" section, the env-var setup step, the troubleshooting check,
and the `legacy text mode` comparison across the four locale copies
(`en`, `ko`, `pt-BR`, `ar`) of `docs/edge/<lang>/enterprise/guides/datadog.mdx`.
* Require explicit CrewAI project definitions
JSON crews and declarative flows now resolve from `[tool.crewai]`
metadata instead of implicit filename discovery. This makes project type
selection deterministic, prevents stray `crew.json(c)` files from changing
CLI behavior, and centralizes definition path validation for run, install,
deploy validation, plotting, and memory reset paths.
`[tool.crewai].definition` must be a project-local file path. Absolute
paths, `~`, missing files, directories, and paths escaping the project root
are rejected so deploy and runtime commands use the same contract.
Breaking changes and migration paths:
* JSON crew projects are no longer discovered from `crew.json` or
`crew.jsonc` alone. Add explicit metadata:
```toml
[tool.crewai]
type = "crew"
definition = "crew.jsonc"
```
* Declarative flow projects must use a valid project-local definition path:
```toml
[tool.crewai]
type = "flow"
definition = "flows/research.yaml"
```
* `Flow.from_definition(definition)` is removed. Use:
```python
Flow.from_declaration(contents=definition)
```
* `FlowDefinition.to_json()` and `FlowDefinition.to_yaml()` are removed.
Use `FlowDefinition.to_dict()` and serialize with the caller's JSON or
YAML library.
* `FlowDefinition.from_dict()` is removed. Use:
```python
FlowDefinition.from_declaration(contents=data)
```
* `FlowDefinition.json_schema()` is removed. Use Pydantic's schema API only
where schema generation is intentionally needed:
```python
FlowDefinition.model_json_schema(by_alias=True)
```
* `crewai_cli.run_crew.find_crew_json_file()` and `_has_json_crew()` are
removed. Use `configured_project_json_crew()` or the shared
`crewai_core.project.configured_project_definition("crew")` helper.
* `crewai reset-memories` now only loads JSON crews declared through
`[tool.crewai].definition`, and invalid declared JSON crew definitions
fail instead of silently falling back to classic crew discovery.
* Address code review comments
* Track conversational flow turn usage in telemetry
* adjusted name to flow:conversation_turn
* only mark on turn completed event
* ensure tui also emits these events
* fix: enforce owner-only permissions on credential files
Credentials stored at rest were left world-readable on multi-user hosts:
- TokenManager._get_secure_storage_path() documented its credential dir as
mode 0o700 but created it via mkdir() with default perms (0o755), leaving
the Fernet secret.key and encrypted tokens.enc in a traversable dir.
- Settings.dump() persisted tool_repository_password (plaintext) to
settings.json via open("w"), producing a 0o644 file, and created the
config dir at 0o755 — despite the sibling token_manager already writing
secrets atomically at 0o600.
Fixes:
- TokenManager: chmod the credential dir to 0o700 after mkdir (robust against
umask and pre-existing dirs).
- Settings: write settings.json atomically at 0o600 (mkstemp + chmod +
os.replace) and chmod the dedicated config dir to 0o700. The /tmp and cwd
fallback parents are deliberately not chmod'd; the 0o600 file mode protects
the credential there.
Adds regression tests asserting 0o600 files and 0o700 dirs, and that shared
fallback dirs are not globally tightened.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Potential fix for pull request finding 'Empty except'
Co-authored-by: Copilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
* Potential fix for pull request finding 'Empty except'
Co-authored-by: Copilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
* Close temp fd on secure settings write failure
* Log secure settings fd close failures
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Copilot Autofix powered by AI <223894421+github-code-quality[bot]@users.noreply.github.com>
`StateProxy` looked like a thread-safety boundary, but it only protected
a small slice of state operations. Some examples of operations that were
not covered:
- `self.state.counter += 1`, `self.state["counter"] += 1` (increments)
- `self.state.user.profile.score += 1` (nested object mutations)
- `self.state.config["limits"]["max"] = 10` (mutation through model fields)
- `self.state.items[0].status = "done"` (list/container mutations)
This commit decided to remove it completely for simplicity and
performance:
- Simpler runtime code
- attr read: 24x faster, attr write: 27x faster, list append: 19x faster (local benchmark)
- Clearer concurrency contract (lifecycle locks remain, but arbitrary
shared state mutation is not presented as thread-safe)
Declarative flows already used `module:qualname` refs for runtime
objects, but crew JSON tools still had their own lookup path. That meant
examples like `project_tools:LookupTool` were treated as named
`crewai_tools` lookups and failed with guidance that only mentioned
`SerperDevTool` or `custom:<name>`. Invalid refs such as
`not_tools:NotATool` also missed the same BaseTool validation used by
flow tool actions.
Move ref resolution into a shared declarative helper, use it from flow
tool actions and crew JSON loading, and require tool refs to resolve to
`BaseTool` classes before instantiation. Validation still checks tool
refs structurally, so validating a crew does not import or execute
project code.
Allow required JSON schema state fields to be supplied by kickoff inputs
instead of requiring every field to exist in state.default before
runtime.
Example: a flow with required lead_name and no state.default can now run
with kickoff inputs={"lead_name": "Ada Lovelace"}.
The page itself already landed on main via #6247. This rebases onto main
and applies the two remaining changes:
- Nest crew-studio + merged-step-card into a collapsible "Crew Studio"
nav group (pencil icon), across edge and v1.14.7 in en, pt-BR, ko, ar.
- Remove the temporary "Rolling out" Note banner (feature ships today).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
