feat(providers): chat tool-calling via openclaude + explicit provider mode#106
Conversation
…volution-foundation#52) Vault audit §2.S1 CRITICAL: /api/shares/<token>/view had zero rate limiting. Add flask-limiter (in-memory, single-process MVP) with: - 60 req/min/IP on view_share (Vault §2.S1) - Global default 600 req/min on all other routes (non-blocking baseline) - Referrer-Policy, Cache-Control no-store, Pragma, HSTS, X-Content-Type-Options headers on every public share response (Vault §2.S2) The Limiter singleton lives in rate_limit.py to break the circular-import chain between app.py (which imports route blueprints) and the blueprints that need @limiter.limit() decorators. Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…ortals (evolution-foundation#53) * feat(security): rate-limit public share endpoint + security headers Vault audit §2.S1 CRITICAL: /api/shares/<token>/view had zero rate limiting. Add flask-limiter (in-memory, single-process MVP) with: - 60 req/min/IP on view_share (Vault §2.S1) - Global default 600 req/min on all other routes (non-blocking baseline) - Referrer-Policy, Cache-Control no-store, Pragma, HSTS, X-Content-Type-Options headers on every public share response (Vault §2.S2) The Limiter singleton lives in rate_limit.py to break the circular-import chain between app.py (which imports route blueprints) and the blueprints that need @limiter.limit() decorators. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat(plugins): B2.0 public_pages capability — read-only token-bound portals Add the public_pages plugin capability (B2.0 scope, read-only): plugin_schema.py: - Add Capability.public_pages and Capability.safe_uninstall enum values - Add PluginPublicPageTokenSource + PluginPublicPage Pydantic models (bundle must be under ui/public/, revoked_when disallowed in v1 to prevent SQL injection) - Extend ReadonlyQuery with public_via + bind_token_param fields - Add 4 PluginManifest model validators: capability required, table slug-prefix, unique ids/route_prefixes, readonly_data references valid page routes/plugin_public_pages.py (new): - GET /p/<slug>/<route_prefix>/<token> — serve portal bundle (60 req/min/IP) - GET /p/<slug>/<route_prefix>/<token>/data — serve token-bound readonly query (120/min) - GET /p/<slug>/public-assets/<path> — serve ui/public/ static assets - Token validation via parametric SQL (identifiers validated at install by schema) - Module-level _PLUGIN_PUBLIC_PREFIXES cache for install/uninstall lifecycle - Vault §B2.S2: Referrer-Policy, Cache-Control no-store, HSTS, X-Content-Type-Options on all responses - CSP: default-src 'self' on portal bundles - Rate limiting via rate_limit.limiter (imported from PR evolution-foundation#52) app.py: - Import and register plugin_public_pages_bp - /p/... paths already bypass auth_middleware (non-/api/ paths are passthrough) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…vation, sandboxed hook (evolution-foundation#54) - plugin_schema.py: PluginSafeUninstall, PluginPreUninstallHook, PluginUserConfirmation models + validators (block_uninstall enforcement, preserved_tables slug-prefix check, safe_uninstall_enabled_requires_confirmation, no _orphan_* refs in readonly SQL) - routes/plugins.py: uninstall gate (admin-only, confirmation_phrase, exported_at check, zip_password); sandboxed pre-hook subprocess (no secrets, read-only DB copy, 600s timeout); cascade-DELETE filtering for preserved_host_entities; orphan table rename (_orphan_{slug}_{table}); EVONEXUS_ALLOW_FORCE_UNINSTALL=1 escape hatch with audit; reinstall SHA256 check against plugin_orphans before orphan recovery - app.py: plugin_orphans table migration (id, slug, tablename, orphaned_at, orphaned_by_user_id, original_plugin_version, original_sha256, original_publisher_url, recovered_at, UNIQUE(slug, tablename)) - PluginUninstall.tsx: 3-step wizard (regulatory reason+checkbox → ZIP password → typed phrase); force-uninstall orange banner; integrated in PluginDetail.tsx - docs/plugin-contract.md: full plugin.yaml contract for public_pages + safe_uninstall Vault §B3 mitigations: S1 block_uninstall gate, S2 admin enforcement, S3 sandboxed hook, S4 no _orphan_* SQL refs, S5 AES-256 ZIP password, S6 force-uninstall audit trail. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…er auto-injection (evolution-foundation#55) Two related Wave 2.1.x extensions to the plugin contract that close the last two gaps blocking endpoint-level RBAC for plugin authors (gap inventory in evonexus-plugin-nutri Step 3 RBAC decision). Changes - PluginWritableResource.requires_role: Optional[List[str]] — when set, the POST/PUT/DELETE handler returns 403 if current_user.role is not in the list. 'admin' role always passes (super-user override). Backwards compatible: resources without the field accept any authenticated user (legacy default). Validator enforces kebab-case role names (^[a-z][a-z0-9-]*$). - routes.plugins.writable_data: enforces requires_role at the endpoint, with a 403 message naming the required roles and the actor's current role. - routes.plugins.readonly_data: auto-injects :current_user_id and :current_user_role as bind params on every readonly query. Plugins can reference them directly in SQL for server-enforced scoping without an app-layer wrapper. The two parameter names are reserved — client requests carrying them in the query string get 400 (no identity spoofing). Tests - tests/backend/test_plugins_rbac_and_scoping.py — 13 cases covering Pydantic acceptance/rejection, 403/200 paths for writable, scoping/spoofing for readonly, backwards compat for resources without requires_role and queries without :current_user_id refs. Compat - Existing plugins (PM Essentials) continue to work unchanged — the new field defaults to None and the auto-injected bind params are silently ignored if the SQL doesn't reference them. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ser, raw bundle for clients (evolution-foundation#56) The portal_page handler at /p/{slug}/{route_prefix}/{token} previously served the plugin's JS bundle with mimetype application/javascript regardless of the caller. Browsers navigating to a portal URL saw the JS source instead of a rendered page. Discovered during evonexus-plugin-nutri Step 5. Changes - portal_page: when the Accept header includes text/html and NOT application/javascript, render a minimal HTML shell that loads the bundle as <script type="module" src="/p/{slug}/public-assets/{file}"> and instantiates the plugin's declared custom_element_name. Token reaches the element via data-token attribute (no need for the bundle to re-parse window.location). - _serve_html_shell: new helper. Defensive: validates bundle path is inside ui/public/ and custom_element_name matches alphanum-dash before emitting. Sets X-Content-Type-Options: nosniff + a tight CSP (default-src 'self', frame-ancestors 'none', no external scripts). - Programmatic clients (curl, fetch with Accept: application/javascript) keep getting the raw bundle — backwards compatible. - Bundle is fetched from the existing /p/{slug}/public-assets/{path} route (no token), which is safe because the bundle contains no patient data — data lives behind the token-gated /data endpoint. Tests - tests/backend/test_plugin_public_pages_html_shell.py — 9 cases: HTML accept renders shell, JS accept returns bundle, default Accept (*/*) returns bundle, invalid token returns 404 even with HTML accept, CSP + X-Content-Type-Options present, custom element name appears exactly once, XSS-safe (single <script> tag, token only inside data-token), legacy programmatic fetches unchanged. Compat - Existing public-page consumers using fetch() with Accept: application/json or default see no behaviour change. Plugins that already shipped a bundle start rendering correctly in browsers without any plugin update. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…volution-foundation#57) Bumps version to 0.33.0 so the plugin nutri (which requires this version) can install. Bundles the five plugin-contract PRs (evolution-foundation#52→evolution-foundation#56) merged today into a single release. Plus a UX fix on the install wizard so 409s say why they conflicted. The fix - lib/api.ts buildError now falls back to data.conflicts[0] when the standard error/message fields are absent. The plugin preview endpoint returns {conflicts: string[], manifest, ...} on 409 — without this fix the wizard showed only "409 CONFLICT" with the actual reason hidden. - PluginInstallModal: conflicts type was Record<string, unknown>, backend always returned string[]; the JSON.keys() coercion produced index strings. Now typed as string[] and rendered as a list. Tested - Frontend tsc --noEmit clean - Plugin nutri 200 pytest still pass after the 11 `# nosec B603` markers added to subprocess.run calls (false positives from regex security scan — all calls use list args, no shell=True) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… mode Non-Anthropic providers in the agent chat previously bypassed openclaude entirely — a raw /chat/completions fetch with no tool calling, no session history, and a hard API-key requirement that broke Codex OAuth. The terminal was separately blocked by the isCodeModel() name heuristic for agentic models like openrouter/owl-alpha. - chat-bridge: external providers in code mode now go through the Agent SDK with pathToClaudeCodeExecutable pointing at the openclaude binary and a clean whitelisted env (mirrors ClaudeBridge). Restores tool calling, structured streaming, UI tool approval, and session resume. Chat-only models keep the REST path; missing binary falls back with a clear log. - chat-bridge: external providers get full system-prompt replacement for agent personas (append is too weak for GPT models), matching the terminal behavior. - provider-config: new per-provider `mode: code|chat` field overrides the model-name heuristic; the terminal no longer refuses agentic models whose names don't match the code regex. - providers.example.json: mode field on openrouter/omnirouter + new NVIDIA NIM provider entry (OpenAI-compatible endpoint). - tests: provider-config mode override coverage (node --test). Verified end-to-end against OpenRouter: text streaming and Read tool call complete with all UI events (tool_use_start, tool_input_delta, block_stop, result). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![sourcery-ai[bot]](https://avatars.githubusercontent.com/in/48477?s=60&v=4){kind=small}
There was a problem hiding this comment.
Sorry @sistemabritto, you have reached your weekly rate limit of 500000 diff characters.
Please try again later or upgrade to continue using Sourcery
OpenClaude >=0.18 detects integrate.api.nvidia.com and demands the key
in NVIDIA_API_KEY, exiting 1 when only OPENAI_API_KEY is set ("Claude
Code process exited with code 1" in the chat). Derive NVIDIA_API_KEY
from OPENAI_API_KEY in loadProviderConfig so the UI keeps a single key
field, and allowlist the var in both config layers.
Verified: chat session via Agent SDK + openclaude against NVIDIA NIM
(stepfun-ai/step-3.7-flash) completes with result: success.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
| # EvoNexus uses the shared evonexus.db for all plugin tables (no per-plugin DB). | ||
| conn = _get_db() | ||
| try: | ||
| row = conn.execute(sql, (token,)).fetchone() |
There was a problem hiding this comment.
security (python.sqlalchemy.security.sqlalchemy-execute-raw-query): Avoiding SQL string concatenation: untrusted input concatenated with raw SQL query can result in SQL Injection. In order to execute raw query safely, prepared statement should be used. SQLAlchemy provides TextualSQL to easily used prepared statement with named parameters. For complex SQL composition, use SQL Expression Language or Schema Definition Language. In most cases, SQLAlchemy ORM will be a better option.
Source: opengrep
| _recovery_conn.execute( | ||
| f"ALTER TABLE {_orphan_table_name} RENAME TO {_orig_table}" | ||
| ) |
There was a problem hiding this comment.
security (python.sqlalchemy.security.sqlalchemy-execute-raw-query): Avoiding SQL string concatenation: untrusted input concatenated with raw SQL query can result in SQL Injection. In order to execute raw query safely, prepared statement should be used. SQLAlchemy provides TextualSQL to easily used prepared statement with named parameters. For complex SQL composition, use SQL Expression Language or Schema Definition Language. In most cases, SQLAlchemy ORM will be a better option.
Source: opengrep
| } | ||
| if _orphan_table_name in _existing: | ||
| _recovery_conn.execute( | ||
| f"ALTER TABLE {_orphan_table_name} RENAME TO {_orig_table}" |
There was a problem hiding this comment.
security (python.django.security.injection.tainted-sql-string): Detected user input used to manually construct a SQL string. This is usually bad practice because manual construction could accidentally result in a SQL injection. An attacker could use a SQL injection to steal or modify contents of the database. Instead, use a parameterized query which is available by default in most database engines. Alternatively, consider using the Django object-relational mappers (ORM) instead of raw SQL queries.
Source: opengrep
| } | ||
| if _orphan_table_name in _existing: | ||
| _recovery_conn.execute( | ||
| f"ALTER TABLE {_orphan_table_name} RENAME TO {_orig_table}" |
There was a problem hiding this comment.
security (python.flask.security.injection.tainted-sql-string): Detected user input used to manually construct a SQL string. This is usually bad practice because manual construction could accidentally result in a SQL injection. An attacker could use a SQL injection to steal or modify contents of the database. Instead, use a parameterized query which is available by default in most database engines. Alternatively, consider using an object-relational mapper (ORM) such as SQLAlchemy which will protect your queries.
Source: opengrep
| _proc = subprocess.run( | ||
| ["python3", str(_hook_script_path)], | ||
| cwd=str(plugin_dir), | ||
| env=_hook_env, | ||
| capture_output=True, | ||
| text=True, | ||
| timeout=_hook_timeout, | ||
| ) |
There was a problem hiding this comment.
security (python.flask.security.injection.subprocess-injection): Detected user input entering a subprocess call unsafely. This could result in a command injection vulnerability. An attacker could use this vulnerability to execute arbitrary commands on the host, which allows them to download malware, scan sensitive data, or run any command they wish on the server. Do not let users choose the command to run. In general, prefer to use Python API versions of system commands. If you must use subprocess, use a dictionary to allowlist a set of commands.
Source: opengrep
|
|
||
| try: | ||
| _proc = subprocess.run( | ||
| ["python3", str(_hook_script_path)], |
There was a problem hiding this comment.
security (python.lang.security.dangerous-subprocess-use): Detected subprocess function 'uninstall_plugin' with user controlled data. A malicious actor could leverage this to perform command injection. You may consider using 'shlex.escape()'.
Source: opengrep
| # the preservation clause comes from the manifest (validated at install). | ||
| _preserve_clause = _preserved_host_entities[_tbl] | ||
| _where = f"(source_plugin = ?) AND NOT ({_preserve_clause})" | ||
| conn.execute(f"DELETE FROM {_tbl} WHERE {_where}", (slug,)) |
There was a problem hiding this comment.
security (python.sqlalchemy.security.sqlalchemy-execute-raw-query): Avoiding SQL string concatenation: untrusted input concatenated with raw SQL query can result in SQL Injection. In order to execute raw query safely, prepared statement should be used. SQLAlchemy provides TextualSQL to easily used prepared statement with named parameters. For complex SQL composition, use SQL Expression Language or Schema Definition Language. In most cases, SQLAlchemy ORM will be a better option.
Source: opengrep
| # the preservation clause comes from the manifest (validated at install). | ||
| _preserve_clause = _preserved_host_entities[_tbl] | ||
| _where = f"(source_plugin = ?) AND NOT ({_preserve_clause})" | ||
| conn.execute(f"DELETE FROM {_tbl} WHERE {_where}", (slug,)) |
There was a problem hiding this comment.
security (python.flask.security.injection.tainted-sql-string): Detected user input used to manually construct a SQL string. This is usually bad practice because manual construction could accidentally result in a SQL injection. An attacker could use a SQL injection to steal or modify contents of the database. Instead, use a parameterized query which is available by default in most database engines. Alternatively, consider using an object-relational mapper (ORM) such as SQLAlchemy which will protect your queries.
Source: opengrep
| _orphan_name = f"_orphan_{slug}_{_orig_table}" | ||
| try: | ||
| # Rename to orphan name | ||
| _orphan_conn.execute(f"ALTER TABLE {_orig_table} RENAME TO {_orphan_name}") |
There was a problem hiding this comment.
security (python.sqlalchemy.security.sqlalchemy-execute-raw-query): Avoiding SQL string concatenation: untrusted input concatenated with raw SQL query can result in SQL Injection. In order to execute raw query safely, prepared statement should be used. SQLAlchemy provides TextualSQL to easily used prepared statement with named parameters. For complex SQL composition, use SQL Expression Language or Schema Definition Language. In most cases, SQLAlchemy ORM will be a better option.
Source: opengrep
| _orphan_name = f"_orphan_{slug}_{_orig_table}" | ||
| try: | ||
| # Rename to orphan name | ||
| _orphan_conn.execute(f"ALTER TABLE {_orig_table} RENAME TO {_orphan_name}") |
There was a problem hiding this comment.
security (python.flask.security.injection.tainted-sql-string): Detected user input used to manually construct a SQL string. This is usually bad practice because manual construction could accidentally result in a SQL injection. An attacker could use a SQL injection to steal or modify contents of the database. Instead, use a parameterized query which is available by default in most database engines. Alternatively, consider using an object-relational mapper (ORM) such as SQLAlchemy which will protect your queries.
Source: opengrep
Extends the NVIDIA provider beyond chat/code into image generation, and fixes three bugs found while dogfooding the multi-provider flow: - ai-image-creator: NVIDIA NIM provider with FLUX models — flux.2-klein-4b (default, best composition, ~2.5s), flux.1-dev (30 steps), flux.1-schnell. Aspect-ratio mapping constrained to the dimension set and 1.06MP pixel budget the API validates; per-model cfg_scale/steps defaults (schnell requires cfg=0, klein requires cfg>=1); JPEG output auto-converted to PNG via ImageMagick with ffmpeg fallback (temp files beside the output — snap-confined ffmpeg cannot read /tmp). Uses NVIDIA_API_KEY from .env. - backend/providers: saving provider config with a blank key field no longer wipes the stored secret — the UI submits empty password inputs when editing other fields (e.g. model), silently erasing the API key and breaking every session afterwards with auth errors. - claude-bridge: terminal agent personas now resolve from WORKSPACE_ROOT with cwd fallback, matching chat-bridge (cab8966) — sessions started outside the workspace root were silently falling back to a generic "You are the X agent" persona. - Makefile: `make telegram` fails loudly when screen/bun are missing instead of printing a false success message. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…le providers Enforce a minimum interval between chat completion requests and retry 429/503 responses with exponential backoff, honoring Retry-After. Configurable via CHAT_MIN_INTERVAL_MS, CHAT_MAX_RETRIES, CHAT_BASE_DELAY_MS. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
onclose only cleared the keepalive and never reconnected — a dropped socket left the terminal/chat dead until the component remounted (switching tabs). Now both components reconnect with capped exponential backoff and rejoin immediately on visibilitychange, replaying the session buffer on a cleared terminal to avoid duplicate output. If the process died while disconnected, surface it instead of silently restarting the agent. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…r fallback - Pin the CLI conversation to the terminal-server session UUID: first start uses --session-id, restarts use --resume when a persisted conversation file exists — provider crashes and server restarts no longer lose the conversation. - Per-agent model tiers: agents declare model: opus|sonnet|haiku in frontmatter; providers.json maps each tier to a provider model via the new model_tiers field. - Pass --fallback-model from the new fallback_models chain (first entry distinct from the primary). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… totals Older skill versions logged image cost entries without token_usage.total_tokens — the unguarded access in the Image Generation table threw a TypeError and unmounted the whole page. Normalize the image-costs payload defensively, same pattern as normalizeCostData. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…eator New 'openai' provider hitting api.openai.com/v1/images/generations directly, with the 'image2' model keyword (auto-detects provider). Requires a platform API key (AI_IMG_CREATOR_OPENAI_KEY or OPENAI_API_KEY) — ChatGPT Plus/Codex OAuth tokens lack the api.model.images.request scope (verified: 401), so they cannot be used for image generation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ude 0.18 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…e cost estimate Images API bills text input and image output at different per-token rates — use the prompt/completion split when available instead of total_tokens at the input rate. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…reds-bearing test script Top-level business folders (finance/, meetings/, social/, etc.) sit outside the workspace/** rule and were trackable — never commit private operational data, especially in a public PR. Also ignore local backups and scripts/ghost_integration_test.py (has hardcoded Ghost API keys). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…dation - Heartbeats can now run an in-process Python handler instead of spawning a Claude agent: new 'handler' field (max_turns=0), DB migration, schema validation, dispatcher sync, runner execution path. - provider_fallback.py: 429/quota detection + model/provider rotation with cooldown tracking (foundation; not yet wired into the runner). Tests: tests/heartbeats 25/26 pass (the 1 failure is a data-driven seed assertion against the gitignored local heartbeats.yaml, not a regression). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Substituído por #108 — rebaseado limpo no |
2 participants
Summary
This PR makes non-Anthropic providers actually work across EvoNexus — agent chat, terminal, and image generation — and adds NVIDIA NIM as a first-class provider for text/code (LLM) and image (FLUX).
Why NVIDIA: build.nvidia.com offers a free tier with generous credits across strong open models (DeepSeek, Qwen, Llama, StepFun for text; BFL FLUX.1/FLUX.2 for images) behind an OpenAI-compatible endpoint. For day-to-day operation this gives every EvoNexus user a zero-cost third leg to triangulate between Anthropic and OpenRouter: when one provider rate-limits or runs out of credits mid-workflow, switching providers in the dashboard now actually works — which is the bug this PR fixes.
Problems fixed
1. Agent chat silently bypassed openclaude (root cause of "switching providers does nothing")
chat-bridge.jsrouted every non-Anthropic provider to a rawfetchon/chat/completions:[system, user]provider-config.jsdeletesOPENAI_API_KEYforcodex_authFix: external providers in
codemode now go through@anthropic-ai/claude-agent-sdkwithpathToClaudeCodeExecutablepointing at theopenclaudebinary and a clean whitelisted env (mirrorsClaudeBridge). Tool approval UI, streaming, session resume, and subagents all work unchanged. Chat-only models keep the REST path as fallback.2. Terminal blocked agentic models by name heuristic
isCodeModel()gated the terminal on the model name matching a code-ish regex —openrouter/owl-alpha,stepfun-ai/step-3.7-flashetc. were refused. Fix: explicit per-providermode: "code" | "chat"field inproviders.jsonthat overrides the heuristic.3. openclaude >=0.18 requires NVIDIA_API_KEY for NIM endpoints
With only
OPENAI_API_KEYset the CLI exits 1 ("Claude Code process exited with code 1" in the chat UI). Fix:loadProviderConfig()derivesNVIDIA_API_KEYfromOPENAI_API_KEYwhen the base URL is*.nvidia.com— the UI keeps a single key field.4. Saving provider config wiped stored secrets
The Configure modal submits blank password inputs when the user edits other fields (e.g. model) —
update_provider_configaccepted the empty string and erased the saved key, breaking every subsequent session with auth errors. Fix: empty values no longer overwrite stored*KEY/*SECRET/*TOKENvars.5. Terminal agent personas resolved from the wrong directory
claude-bridge.jsread.claude/agents/{agent}.mdfrom the session cwd; sessions outside the workspace root silently fell back to a generic persona. Fix: resolve fromWORKSPACE_ROOTwith cwd fallback — same fix chat-bridge got in cab8966.6. ai-image-creator: NVIDIA NIM image provider (FLUX)
--provider nvidiawith three models: flux.2-klein-4b (default — best composition, ~2.5s), flux.1-dev (30 steps), flux.1-schnell (fastest). Keywords auto-select the provider.cfg_scale/stepsdefaults — the API schemas differ per model (schnell requirescfg=0, klein requirescfg>=1)./tmp).7.
make telegramfailed silentlyThe target printed a success message even when
screen/bunwere missing. Now it fails loudly with install instructions.Verification (all live, not simulated)
npm testin terminal-server: 7/7 pass (5 new tests for the mode override)stepfun-ai/step-3.7-flash) -> full event stream (tool_use_start,tool_input_delta,thinking_delta,result: success)Read, result returned, all UI events emittedfile; flux.1-dev and flux.1-schnell keywords verifiedNotes for reviewers
providers.example.jsongains themodefield and an NVIDIA NIM entry; userproviders.jsonis untouched/gitignored.🤖 Generated with Claude Code