* feat(vector): add Valkey Search vector database backend
Add a new opt-in VectorDatabase backend backed by the Valkey Search module
(valkey/valkey-bundle), accessed via the official valkey-glide client's native
ft command namespace.
- Implements the full VectorDatabase ABC: VECTOR, FULL_TEXT and HYBRID search,
all 8 metadata filter operators, and pagination with exact totals.
- HYBRID uses filter-then-KNN (no app-side weighted fusion); vector_weight is
accepted for interface parity but NOT honored (docstring + one-time warning +
docs caveat).
- Lazy connect so a down Valkey never blocks boot; mandatory
client_name=langbot_vector_client; optional auth + TLS (never logged).
- Registered via a single elif branch in vector/mgr.py; disabled by default
(vdb.use stays chroma) for toC compatibility.
- Adds valkey-glide>=2.4.1,<3.0.0; no protobuf/pydantic downgrade; no ORM
change so no Alembic migration.
- Unit tests (fast lane, no server) + slow-gated integration tests
(TEST_VALKEY_URL, valkey/valkey-bundle:9.1.0) + integration doc.
* fix(vector): paginate Valkey Search deletes and guard delete_by_filter
Address self-review follow-ups for the Valkey Search VDB backend:
- _search_keys now paginates through the full result set in batches of
_DELETE_SCAN_BATCH instead of capping at a single hard-coded 10000-key
page, so delete_by_file_id / delete_by_filter fully remove files and
filters that match more than one page of chunks (no orphaned vectors).
- Add unit regression tests for the delete_by_filter mass-deletion guard:
a filter referencing only non-indexed fields must skip and return 0
(never fall back to match-all), and a supported filter still deletes
matching keys.
* refactor(vector): harden Valkey Search backend and add adversarial tests
Address the self-review NICE-TO-HAVE items for the Valkey Search VDB backend:
- Guard the username-without-password credential edge (skip auth + warn
instead of building ServerCredentials(password=None, ...), which glide
rejects).
- Add an async close() teardown that closes the glide client and resets
cached state (re-init is safe via the existing None guard).
- Hoist 'import json' to module top (was imported inside three methods).
- Document the FT TAG literal-brace limitation in _escape_tag (fails closed,
never widens).
Tests:
- Add an adversarial-input integration test proving crafted file_id /
query_text cannot break out of or widen a query (fail-closed on braces).
- Add unit tests for close() and the credential-build guard.
Signed-off-by: Daria Korenieva <daric2612@gmail.com>
* fix(vector): make Valkey Search file_id TAG support arbitrary characters
Valkey Search's FT TAG query parser cannot handle '{', '}' or '*' even when
backslash-escaped, so a file_id containing those characters previously
produced an unparseable query (it failed closed / raised). Percent-encode
exactly those FT-unsafe characters (plus '%' for reversibility) in the
file_id TAG value, applied identically at write time and query time, so an
arbitrary file_id round-trips. For normal UUID/hash ids this is a no-op and
the stored value is unchanged; the original file_id is always preserved
verbatim in metadata_json.
Strengthen the adversarial integration test to assert a brace/star-bearing
file_id matches and deletes exactly its own row (no widening, no raise), and
add unit tests for _encode_file_id and the filter encoding.
Signed-off-by: Daria Korenieva <daric2612@gmail.com>
* refactor(vector): address Valkey Search review feedback
- Add configurable request_timeout (default 5000ms; glide default 250ms is
too low for KNN); expose in config.yaml + docs table
- Validate embedding dimension consistency in add_embeddings (fail fast on
mixed lengths to avoid silent KNN corruption)
- Use ft.info (O(1)) instead of ft.list (O(n)) for index existence checks in
the query hot path; also closes the check-then-create TOCTOU window
- Pipeline HSETs via a non-atomic Batch instead of N sequential awaits
- Extract shared _iter_reply_docs to deduplicate reply parsing between
_reply_to_chroma and list_by_filter
- Parenthesize multi-condition pre-filters before the => KNN clause
- Fail closed when a username is configured without a password
- Catch only RequestError on ft.dropindex (let connection/auth errors surface)
- Bound the delete_collection SCAN loop with a safety cap
- Add VectorDatabase.close() (no-op default) + VectorDBManager.shutdown()
- Simplify _MATCH_ALL literal; normalize typing to builtin generics
* fix(vector/valkey_search): address round-2 review feedback
- Serialize lazy client creation with an asyncio.Lock (double-checked) so
concurrent first-use callers don't construct and leak duplicate clients.
- Make the filter operator chain exhaustive: raise on an unhandled op rather
than silently dropping the condition (which could widen delete_by_filter).
- Cast numeric range (///) values to float, failing closed on
non-numeric input and pre-empting a future NUMERIC-field injection surface.
* refactor(vector): remove shutdown/close from base ABC per maintainer feedback Per maintainer request, interface changes to VectorDatabase ABC and VectorDBManager should be in a separate PR with implementation across all backends. The ValkeySearchVectorDatabase.close() method remains but does not override an ABC method.
Signed-off-by: Daria Korenieva <daric2612@gmail.com>
* docs(test): list valkey_search in vdb coverage exclusions Add valkey_search to the documented vector/vdbs/ coverage-exclusion list, matching the existing chroma/milvus/pgvector/qdrant/seekdb entries. These adapters require a live database instance and are covered by env-gated integration tests instead of unit tests.
Signed-off-by: Daria Korenieva <daric2612@gmail.com>
---------
Signed-off-by: Daria Korenieva <daric2612@gmail.com>
Operators can now set a global default memory limit for all stdio MCP
servers in config.yaml or via environment variable:
config.yaml:
box:
default_memory_mb: 2048 # default: 1536
env:
BOX__DEFAULT_MEMORY_MB=2048
The default is raised from 1024 to 1536 MB — a safer floor for
Node.js V8 + WASM (undici llhttp) under nsjail cgroup limits.
Individual MCP servers can still override via their own box.memory_mb.
Previously the fallback was hardcoded to 1024 MB, causing OOM kills
(return_code=137) on node/npx MCP servers that need more RAM.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
All MCPs share one Box session (mcp-shared). When session memory_mb differed
by command type (512 for python, 1024 for node), the second MCP to call
create_session raised BoxSessionConflictError. Fix: always use 1024 MB for
the shared session so python and node MCPs coexist without conflict.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
Node.js MCP servers (npx/bunx) were being OOM-killed (return_code=137) by the
default 512MB nsjail cgroup_mem_max. Node V8 reserves large virtual address
space and instantiates WebAssembly modules (undici llhttp) on startup, easily
exceeding 512MB resident. This caused every node-based MCP (memory,
sequential-thinking, filesystem, weather, docker, excel) to crash-loop.
Fix: when the stdio command is npx/bunx/pnpm, default memory_mb to 1024 unless
the operator explicitly set a value. Python/uvx servers keep the 512MB default.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
_ColdStartRetry was caught in _lifecycle_loop_with_retry which set
_preserve_managed_process = True — but by then the finally block inside
_lifecycle_loop had already run and called _cleanup_box_stdio_session(),
stopping the live managed process (return_code=143 SIGTERM). The cold-start
retry then restarted a fresh process, eliminating the warm-up advantage.
Fix: add an explicit except _ColdStartRetry in _lifecycle_loop that sets
_preserve_managed_process = True before re-raising. The finally block then
sees the flag and skips stop_managed_process, leaving the live process
untouched for the next handshake attempt.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
The previous _TransferredStack approach broke anyio lexical context:
websocket_client/ClientSession use anyio task groups whose cancel scope is
bound to the frame that entered them. Deferring their aclose via a transferred
exit stack left the underlying memory streams closed once initialize() returned,
so the very next request (refresh -> list_tools) failed with Connection closed.
New design:
- Attach on the owner exit stack (same task as the serve loop, lexically intact)
- A cold-starting process makes initialize() fail; signal _ColdStartRetry up to
the outer retry loop, which reuses the live process without consuming retry budget
- _lifecycle_loop_with_retry handles _ColdStartRetry like _TransportReconnect:
preserves process, no fatal budget, backs off 2s and retries
- Two new unit tests: cold-start raises _ColdStartRetry (not fatal) when process
is alive; raises fatal error when process has actually exited
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
A node/npx stdio MCP server (e.g. firecrawl-mcp via npx -y) failed on first
connect with Connection closed / Failed after 4 attempts, even though the
process was fine. An npx cold start downloads+installs the package before the
server can answer the MCP handshake (measured ~27s for a simple official
server; longer for heavier ones). The old code attached the WS and called
session.initialize() the instant the process was started, so the handshake ran
before the process could answer and failed; the outer lifecycle retry then
rebuilt the process, churning it in a loop.
Verified decisively: attaching + initialize() against a mid-cold-start process
times out on attempt 1 (process still installing) but SUCCEEDS at t+0.6s on
attempt 2 once the process is ready. So the fix is to retry the handshake in
place, not to rebuild the process.
Changes (mcp_stdio.initialize):
- Start the managed process ONCE, then loop attach WS -> ClientSession ->
initialize() within the startup_timeout budget, tearing down each failed
attempt cleanly, until the handshake succeeds or the budget elapses. A
successful transport/session is transferred into the owner exit stack via a
small _TransferredStack adapter.
- Bound each attempt with asyncio.wait_for(initialize, _HANDSHAKE_ATTEMPT_TIMEOUT_SEC=10s)
so a cold-starting process fails fast and retries instead of hanging until
the transport drops.
- Stop retrying ONLY when the process has DEFINITIVELY exited: new
_managed_process_has_exited() (checks EXITED status) replaces the previous
not-_managed_process_is_running() test, which false-negatived on a
just-spawned process that had not yet reported RUNNING and made the loop bail
to the outer rebuild path (relay then rejected the early re-attach with HTTP 400).
Adds a unit test that fails the first two handshakes with the process alive and
asserts the loop retries to success while starting the process exactly once.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
* refactor(mcp): make MCP test reuse the shared Box session instead of a per-test session
Testing an MCP server (config-page "test" button) previously spun up a fresh
isolated mcp-test-<uuid> Box session every time: cold-start the container, run
the dependency bootstrap, probe, then tear the whole session down. That is slow
(tens of seconds) and, on an already-hosted server, wasteful — the server is
already running in the shared session.
Change the test to reuse the shared session / live process:
- _build_box_session_id: transient tests now use mcp-shared, the same Box
session as live servers, so a test reuses the running container (and, for an
existing server, its live managed process) instead of a cold per-test session.
- cleanup_session: a transient test no longer deletes the whole session (which
under the shared model would kill every other MCP server in the container). It
stops only its own process_id, exactly like a live server. Isolation is now at
the process level (distinct process_id per server/test), not the session level.
- test_mcp_server (persisted server): reuse the live connection with a real
list_tools refresh/probe; only fall back to a full start() when there is no
live connection to probe or the refresh fails, instead of an ERROR->start()
rebuild.
Trade-off: a failing test now shares the container with live servers rather than
a throwaway session. Accepted deliberately in favour of near-instant tests;
process-level isolation keeps a test from stopping another server's process.
* chore(deps): pin langbot-plugin 0.4.9 for the nsjail RLIMIT_AS node/npx MCP fix
---------
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
* fix(mcp): survive transient WS transport drops for Box stdio MCP servers
A Box-backed stdio MCP server (e.g. pab1it0/prometheus) would periodically
error on the frontend with Box managed process exited unexpectedly /
Failed after 4 attempts once the session had been alive for a while.
Root cause: the managed MCP process lives in the Box runtime and SURVIVES a
WebSocket transport drop, but _lifecycle_loop treated any monitor completion
as a fatal process death. It then ran the finally-block cleanup — which STOPS
the still-healthy managed process — and did a full 4-attempt exponential
backoff rebuild. Under an occasionally-stalled single-worker event loop the
mcp websocket client misses a ping/pong, the transport drops, and this
self-inflicted teardown loop is what the user sees.
Fixes:
- _lifecycle_loop: when the health monitor completes, re-check the real
managed-process state. If the process is still running, the transport
merely dropped: raise an internal _TransportReconnect signal instead of
Box managed process exited unexpectedly.
- _lifecycle_loop_with_retry: handle _TransportReconnect as a free, uncounted
reconnect (does not consume the fatal retry budget), so a long-lived session
survives arbitrarily many transient drops.
- finally-block: gate managed-process teardown on a _preserve_managed_process
flag so a transport-only reconnect closes just the WS, not the process.
- BoxStdioSessionRuntime.initialize: reuse an already-running managed process
instead of stopping+rebuilding it (which also re-ran the slow dependency
bootstrap); only (re)start when none is running. Adds
_managed_process_is_running() helper.
Pairs with langbot-plugin-sdk fix adding a server-driven WS heartbeat to the
managed-process relay, which prevents most drops in the first place.
* style(mcp): ruff format
* chore(deps): pin langbot-plugin 0.4.8 for the managed-process WS heartbeat fix
---------
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
The /change-password and /bind-space endpoints already refuse when
system.allow_modify_login_info is false, but /set-password did not,
leaving a path to alter login credentials on locked-down deployments
(e.g. public demo instances). Apply the same guard.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
* docs(platform): add HTTP Bot adapter design (RFC)
Standalone server-to-server HTTP adapter for driving a pipeline from external
systems (LangBot Space ticketing et al). Inbound via the existing unified
webhook route; outbound via signed callback POSTs. Preserves pipeline-native
N->1 aggregation and 1->M multi-reply without a long-lived WebSocket.
No core changes required (router/aggregator/pipeline untouched).
* feat(platform): add standalone HTTP Bot adapter
A first-class, vendor-neutral message-platform adapter (http_bot) for
server-to-server integrations (LangBot Space ticketing et al). Drives a
pipeline over plain HTTP with no long-lived connection:
- Inbound: signed POST to the existing unified webhook route /bots/<uuid>,
carrying a caller-defined session_id mapped to the LangBot launcher id via
get_launcher_id -> per-session isolation. Preserves pipeline-native N->1
aggregation for free.
- Outbound: each reply_message / reply_message_chunk becomes one signed
callback POST to the config-only callback_url, delivered in per-session
sequence order with retry/backoff -> 1->M multi-reply.
- Sub-paths: /reset (drop a session) and /sync (block for the collapsed reply).
- Auth: symmetric HMAC-SHA256 both directions (timestamp + replay window),
no JWT/Turnstile, no socket.
Decisions: callback URL is config-only (SSRF closed); reset + sync shipped;
Python + TS reference clients shipped (signing verified byte-identical 3-way).
No core changes: the unified webhook router, aggregator, query pool and
pipeline are untouched. Adapter is auto-discovered from platform/sources/.
Adds:
src/langbot/pkg/platform/sources/http_bot.{py,yaml,svg}
src/langbot/pkg/platform/sources/http_bot_signing.py
docs/platforms/http-bot.md, docs/http-bot-openapi.json
examples/http-bot/{client.py,client.ts,README.md}
Updates docs/HTTP_BOT_ADAPTER_DESIGN.md (status: implemented).
* docs(examples): add interactive HTTP Bot playground (browser debug console)
A single-file aiohttp web app (examples/http-bot/playground.py) that lets you
chat with a RUNNING http_bot bot from the browser and watch the protocol live:
signed inbound POST -> 202 ack -> 1->M signed callbacks streamed back via SSE,
with a debug panel showing the signature, HTTP status, and per-callback
sequence/verification. Light LangBot-styled UI.
On startup it reads the API key + http_bot bot from data/langbot.db and points
the bot's callback_url + secrets back at itself via the LangBot API (live
reload, no restart). README updated with a playground section.
* docs(examples): add Chinese README for http-bot reference clients
* style(platform): use </> code icon for http_bot adapter logo
* docs(examples): point http-bot guide links to docs.langbot.app
* style(platform): make http_bot icon a transparent monochrome </> so WebUI tints it like other adapters
* Revert to colorful </> badge for http_bot icon (WebUI renders it as-is)
CI follow-up to the local/remote MCP work:
- Apply ruff format to provider/tools/loaders/mcp.py and the 0006
normalize-remote-mode migration (Lint job failed on formatting).
- test_migrations.py hardcoded the head revision as 0005_*, which broke
once 0006 landed. Resolve the actual head from the Alembic
ScriptDirectory so future migrations don't require editing the test.
Replace the three-way transport choice (stdio / sse / httpstream) for
connecting LangBot to external MCP servers with two modes: local (stdio)
and remote. Remote servers only require a URL; the runtime auto-detects
the transport (tries Streamable HTTP, falls back to SSE).
- provider/tools/loaders/mcp.py: add _init_remote_server() with
Streamable-HTTP-then-SSE probing; dispatch 'remote' lifecycle, keep
legacy sse/http branches for back-compat
- plugin/connector.py: normalize legacy http/sse marketplace modes to
'remote' on Space install, preserving connection params
- entity/persistence/mcp.py: document mode as stdio, remote (legacy: sse, http)
- alembic 0006: idempotent data migration mapping existing sse/http rows
to remote (downgrade maps back to http)
- api/http/service/mcp.py: stash runtime_info (status + tool list) into
test task metadata before tearing down the temp session
- web: collapse mode dropdown to local/remote, remote renders URL+timeout
only, edit auto-maps legacy sse/http to remote; show tools after test in
create mode from task metadata; remove dead plugins/mcp-server/ tree
- i18n: local/remote labels + mode/url hints across 8 locales
* feat(api): support global API key from config.yaml (api.global_api_key)
Accept a config-defined global API key anywhere a web-UI key is accepted
(X-API-Key / Bearer), with no login session and no DB record. Useful for
automated deployments and AI agents (HTTP API + MCP). Defaults to empty
(disabled); does not require the lbk_ prefix.
- templates/config.yaml: add api.global_api_key with security notes
- service/apikey.py: verify_api_key checks global key first (constant-time)
- docs/API_KEY_AUTH.md: document the global key + security guidance
- tests: cover global-key match, prefix-free, fallback-to-db, disabled
* feat(mcp): expose LangBot management as an MCP server at /mcp
Add an MCP (Model Context Protocol) server so external AI agents can manage a
LangBot instance. Reuses the same API-key auth as the HTTP API (including the
config.yaml global API key).
- pkg/api/mcp/server.py: FastMCP server wrapping the service layer; 21 curated
tools across system/bots/pipelines/models/knowledge/mcp-servers/skills
- pkg/api/mcp/mount.py: ASGI dispatcher fronting Quart; authenticates /mcp
requests with an API key, runs the streamable-HTTP session manager lifespan
- controller/main.py: serve the wrapped ASGI app via hypercorn (was run_task)
- web: new 'MCP' tab in the API integration dialog showing endpoint, auth, and
client config; i18n for 8 locales
- tests/manual/mcp_smoke.py: e2e check (401 unauth, list tools, call tools)
Tool surface is intentionally curated (not all ~25 route groups) to keep the
agent surface small, safe, and maintainable. Extend deliberately.
* feat(skills): add in-repo skills/ as the single source of truth
Migrate the agent skills + QA/e2e test harness from the (now archived)
langbot-app/langbot-skills repo into LangBot/skills/, and add four new skills.
Migrated:
- langbot-plugin-dev, langbot-testing (e2e), langbot-env-setup,
langbot-skills-maintenance, langbot-eba-adapter-dev
- the bin/lbs CLI (src/, test/, scripts/, schemas/, qa-agent-docs/)
New:
- langbot-dev core backend + web development
- langbot-deploy Docker/K8s deployment + config.yaml + global API key
- langbot-mcp-ops operating the LangBot MCP server (/mcp)
- langbot-space-ops operating the Space marketplace MCP server
- src/cli.ts repoRoot(): recognize the skills assets root (skills.index.json +
bin/lbs) so the CLI works when nested inside the LangBot repo
- README.md: unified skill catalog; skills.index.json regenerated
Parity with source verified: bin/lbs validate + node test suite match the
source repo (only the uncommitted .lbpkg build-artifact fixture differs).
* docs(agents): document agent-facing surfaces + API/MCP/skills sync rule
* docs(readme): add 'Built for AI Agents' section across all locales
Highlight MCP server, in-repo skills (single source of truth), AGENTS.md
sync rule, and llms.txt. Cross-link LangBot Space MCP marketplace.
* style(mcp): fix ruff format + prettier lint in MCP server and API panel
* style(web): prettier format MCP i18n locale entries
* docs(skills): note MCP instance control in dev/testing skills
All development-guidance skills now point to the LangBot instance MCP
server (/mcp) and the Space marketplace MCP server, reusing API keys.
Update _normalize_stream_tool_calls to preserve provider_specific_fields
(including thought_signature) from streaming tool call chunks. Also preserve
provider_specific_fields from delta in invoke_llm_stream.
This ensures Gemini's thought_signature is round-tripped correctly:
1. LiteLLM extracts thought_signature from Gemini response
2. It's preserved in Message/ToolCall entities (via SDK changes)
3. _convert_messages includes it in the next request
Also add unit tests for provider_specific_fields round-tripping.
Fixes: langbot-app/LangBot#1899
Adding a Kimi/Moonshot provider failed model scanning out of the box for
CN-region API keys: the single preset defaulted its base URL to the
global endpoint `https://api.moonshot.ai/v1`, but CN-issued keys are only
valid against `https://api.moonshot.cn/v1`, so scanning returned
`401 Invalid Authentication`. Flipping the default would just move the
breakage to international keys, since the base_url field is plain
free-text and either region is equally common.
Instead, offer two clearly labelled presets, mirroring how the Lark
adapter exposes feishu.cn vs larksuite.com:
- `moonshot-chat-completions` -> "Moonshot / Kimi (Global · api.moonshot.ai)"
- `moonshot-cn-chat-completions` -> "Moonshot / Kimi (China · api.moonshot.cn)"
The existing component name is kept unchanged so provider rows already in
the DB keep resolving; only its display label is clarified. Both presets
keep base_url as a free-text field, so users behind a proxy / one-api
gateway can still enter a custom endpoint. Both carry the same `kimi`
search aliases so either shows up when searching.
Fixes#2232
Ollama's OpenAI-compatible streaming endpoint emits a tool-call delta
carrying an `index` and a `function` payload but never an OpenAI-style
`id`. `_normalize_stream_tool_calls` dropped any tool call without an
`id`, so a tool-only turn yielded neither content nor a tool call: the
stream "completed" with 0 chars, the tool never ran, and the chat
appeared stuck. Models on standard OpenAI APIs (e.g. SiliconFlow) were
unaffected because they always send a `call_...` id.
Synthesize a stable per-index id (`call_<index>`) when the provider
omits one but a function name is present. Providers that do send ids
keep theirs, and parallel id-less calls keep distinct ids.
Adds regression tests for the single and multi id-less tool-call cases.
Fixes#2261
Outbound attachment collection (pipeline wrapper) runs on every turn
regardless of inbound files, but the agent was only told the per-query
outbox path inside the inbound-attachment note in LocalAgentRunner. So on
pure-generation turns (e.g. "generate a QR code"/chart/mermaid where the
user sent no file), the agent never learned the outbox path or the
query_id, wrote the generated file nowhere deliverable, and it was
silently dropped.
Move the outbox instruction into BoxService.get_system_guidance(query_id),
which is injected as a system message on every turn the exec tool is
available. The inbound note keeps its own (now redundant but harmless)
outbox line. Add unit tests asserting the outbox path is present with a
query_id and absent without one.
The agent attachment outbox is written by the sandbox container as root over
the bind-mount, so the LangBot host process (non-root) cannot rmtree those
files — the host-side delete failed silently and stale files were re-collected
on a later turn that reused the same query_id (the query_id counter resets to 0
on every restart).
- BoxService.initialize now purges leftover inbox/outbox after the runtime is
available: host rmtree first, then an in-sandbox 'rm -rf' via exec for any
root-owned survivors.
- _clear_outbox now falls back to exec when the host delete leaves root-owned
files behind, instead of silently failing.
- collect_outbound_attachments clears the outbox unconditionally (even on an
empty collection) so a reused query_id never inherits stale files.
- Tests: startup purge (host-owned + root-owned exec fallback + no-workspace
noop) and empty-collection-still-clears.
* feat(box): bidirectional attachment transfer for sandbox
Materialize inbound attachments into the sandbox workspace so agents can
process user-sent files, and collect agent-produced files from the outbox
to attach them back to the reply.
- box(service): add materialize_inbound_attachments / collect_outbound
attachments. Prefer direct host-filesystem read/write on the bind-mounted
workspace (no size limit), falling back to chunked exec only for
non-shared backends (e2b/remote). Clear per-query inbox/outbox dirs at
turn start to avoid query_id-reuse collisions.
- provider(localagent): inject inbound attachment descriptors into the
sandbox and append a system note telling the agent the inbox/outbox paths.
- pipeline(wrapper): collect outbox files on the final stream chunk and
append them as attachment components to the response chain.
- web(debug-dialog): render File components with a download link when
base64/url is present; add base64/path fields to the File entity.
- tests: cover inbound/outbound, large-file transfer without truncation,
and stale-dir clearing (86 passing).
* feat(box): support voice/file attachment round-trip end-to-end
Extends the bidirectional attachment transfer to audio and arbitrary files
through the real webchat UI, and fixes the model-payload errors that
non-image attachments triggered.
- platform(websocket_adapter): resolve Voice/File component storage keys to
base64 (previously only Image), so audio/documents reach the sandbox inbox.
- web(debug-dialog): accept audio/* and any file in the uploader (was
image-only), classify by mimetype, upload Voice/File via the documents
endpoint, and render non-image staged attachments as a chip.
- provider(litellmchat): drop non-image file parts (file_base64 / file_url)
when building the OpenAI/LiteLLM payload. These come from Voice/File
attachments — including ones replayed from conversation history — and the
agent reads their bytes from the sandbox, not the model. Without this the
provider rejects the request: 'invalid content type=file_base64'.
- provider(localagent): also strip those parts from the current user message
alongside the sandbox-path note (model-facing clarity; the requester is the
real safety net for history).
- tests: cover the requester strip/keep behavior (file dropped, image kept and
reshaped to image_url, mixed history, plain-string content).
* test(box): cover inbound/outbound attachment helpers; fix ruff format
- ruff format localagent.py (CI ruff format --check was failing)
- add unit tests for ResponseWrapper outbound-attachment helpers (wrapper.py 78%->98%)
- add unit tests for LocalAgentRunner._inject_inbound_attachments
- add unit tests for WebSocketAdapter._process_image_components (0%->covered)
Lifts PR patch coverage from 68.97% to ~88% (>75% target).
Load the instance creation timestamp from data/labels/instance_id.json
(backfilling+persisting it for instances created before the field existed),
expose it as constants.instance_create_ts, and include it in the heartbeat
payload so Space can anchor Time-To-Value / onboarding analytics on real
install time rather than first-heartbeat.
Verified: py_compile, ruff, pytest tests/unit_tests/telemetry/ (37 passed).
Hyu