* fix(gateway): gate declared multi-instance deployments and require the redis stream bridge
GATEWAY_WORKERS / WEB_CONCURRENCY only count the uvicorn workers of one
process tree, so a Kubernetes Deployment with replicas > 1 (one worker
per Pod) passed every multi-worker startup check while each Pod's
startup orphan reconciliation wrote the other Pods' lease-less runs off
as crashed on every rolling update.
Add `deployment.multi_instance` as an explicit declaration that runs the
same gate, plus a DEER_FLOW_MULTI_INSTANCE env override so deploy tooling
can set the declaration from its replica count without editing
config.yaml (wiring the Helm chart to export it is a separate change).
Both the worker count and the declaration now also require a Redis
stream bridge and refuse an explicit memory sandbox ownership store,
since the memory bridge and ownership store are process-local. The
agent-storage divergence warning, the inbound webhook dedupe warning and
the WeChat QR-login guard honor the same declaration.
* docs(changelog): link the multi-instance gate entry to #6328
* fix(ci): bump chart config_version to 53 and keep AGENTS.md under the size limits
config.example.yaml moved to config_version 53 for the new deployment
section, so the Helm chart's embedded config and its README example
follow. The multi-instance notes in backend/AGENTS.md and
backend/app/gateway/AGENTS.md shrink to one line each: main already sits
within 55 bytes of the per-file hard limit for the gateway guide and
within 163 bytes of the chain hard limit under backend/, and the full
contract lives in backend/docs/CONFIGURATION.md.
* fix(gateway): name every active knob in the multi-process rollback
When GATEWAY_WORKERS > 1 and a multi-instance declaration are both
active, the refusal offered only "Set GATEWAY_WORKERS=1", so an operator
who followed it bounced through a second refusal on
DEER_FLOW_MULTI_INSTANCE / deployment.multi_instance at the next start.
The rollback now withdraws both knobs, and the browser-tools refusal
takes the same path instead of the worker-count-only message.
multi_instance_declaration() returns a frozen MultiInstanceDeclaration
(source, value) with knob/rollback properties instead of a display
string, so deps.py no longer branches on message prose. Tests pin the
combined rollback text and that following it actually clears the gate.
* docs(gateway): keep the gateway AGENTS.md under the size hard limit after merging main
main grew backend/app/gateway/AGENTS.md by 30 bytes, which left it 25 bytes
under the 49152-byte hard limit; the one-sentence multi-instance note from
this branch pushed the merged file to 49164. Shorten that sentence by 18
bytes so scripts/check_agent_guidance.py passes on the merge result.
* ci: re-run checks after test_stop_cancels_a_hung_driver_poll flaked on shard 4
tests/test_mcp_task_service.py::test_stop_cancels_a_hung_driver_poll timed
out once on backend-unit-tests (4) for 2c16206ac. The MCP task service and
its test are identical to main, main's run for cfe2023df passed, and 15
consecutive local runs pass; the test gives service.stop() a 1s budget
while cancellation drain allows 5s, so a loaded runner can exceed it.
Fork contributors cannot re-run a job, hence this empty commit.
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* perf(summarization): skip the summarizer LLM call for an input already proven no-op
When the exact rendered summarizer prompt (plus candidate models) previously
produced output byte-identical to the current summary_text, reuse that summary
instead of calling the LLM. Compaction proceeds through the normal path, so
messages, hooks and state transitions are unchanged.
- Only proven no-op inputs are cached (bounded LRU, thread-safe); failed
attempts are never cached and any cache error falls back to the LLM call.
- The middleware:summarize event gains changes.llm_call_skipped.
- Applies to both compact_state and acompact_state.
Builds on the per-call telemetry in the P0 branch.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(summarization): keep fallback no-ops retryable
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* docs(scheduler): document tool_enabled, goal notices and goal-evaluation change
Follow-up to #6229:
- CONFIGURATION.md: add scheduler.tool_enabled to the example and notes.
- CONFIGURATION.md, README (en/zh/pt): unmet goal runs send a goal-unmet
notice instead of a completion notice, and three unmet runs add an
auto-pause notice; the docs still said only success/failed notify.
- CHANGELOG (en/zh): the #6229 entry now states the default-on goal
evaluation change for scheduled/webhook/autonomous goal runs and that
evaluator usage counts toward run usage.
* feat(scheduled-tasks): show goal outcomes and end conditions in run history
Follow-up to #6229. The API already returns these fields; the tasks page
did not show them, so a met goal read as plain "Success", an unmet
reason code rendered as a red execution error, and the goal-unmet IM
notice pointed to a workspace showing neither objective nor verdict.
- Task detail shows the per-run goal, max automatic runs and end time
when a conversation-created task has them.
- Goal runs show "Goal met" (or met relying on stated assumptions);
unmet runs show a localized reason plus the host code in neutral
styling. The run whose agent asked to stop the schedule is marked.
- Runs without a goal, and failed runs, render as before.
Read-only and frontend-only: no API change.
* docs(changelog): add #6326 scheduled-task goal outcome entry
* fix(scheduled-tasks): show readable unmet reasons instead of raw codes
Unmet runs now show only the localized reason; the host code moves to the
tooltip so it can still be matched with the IM notice. Codes without a
label are still shown verbatim.
* fix(scheduler): show readable goal reasons instead of internal codes
Task detail now localizes the host-written last_error values from goal and
stop finalization (agent stop, automatic pause, unmet reason codes) under a
matching label, and Chinese copy is tightened. Goal-unmet and auto-pause IM
notices state the reason in words instead of printing the raw code; unknown
codes are still not forwarded.
* docs(changelog): describe readable goal reasons in #6326 entry
* test(scheduler): pin host goal strings with a cross-language contract
Address review on #6326:
- Add contracts/scheduled_goal_notes_contract.json listing the task
last_error strings and unmet reason codes the tasks page and IM notices
translate. A backend test checks finalization constants, blocker and
stand-down codes against it; a frontend test checks every listed value
has a label.
- Move the two last_error strings in finalization.py into constants.
- Frontend reason lookup strips the blocked: prefix like the IM map and
now covers goal_not_met_yet, with en/zh labels.
---------
Co-authored-by: Totoro-qaq <279883115+Totoro-qaq@users.noreply.github.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
Store validated image bytes in the configured content-addressed blob backend and resolve checkpointed refs before local compatibility paths.
Refs: #4189
Signed-off-by: PeaceMaker-best <221849497+PeaceMaker-best@users.noreply.github.com>
Co-authored-by: PeaceMaker-best <221849497+PeaceMaker-best@users.noreply.github.com>
* fix(projects): convert shelf documents outside the SQLite write lock
Lazy conversion ran pymupdf/markitdown inside the BEGIN IMMEDIATE
transaction that serializes the derived/converted.md publish against
trash and purge. On SQLite that lock is database-wide, so every unrelated
writer (run status, thread metadata, the scheduler) waited for the whole
conversion and failed with "database is locked" past busy_timeout (30s).
Conversion now writes into .staging/ outside any transaction. The
document row lock is taken only to revalidate the live row and atomically
rename the output into place (publish_under_live_lock, renamed from
convert_under_live_lock); a row trashed or purged meanwhile discards the
staged output, so nothing is published after purge. A read-only probe
skips converting rows that are already gone, a failed conversion is
still revalidated so a purged row reports content_missing, and the
conversion is drained on cancellation so its staging file is always
removed.
* docs(changelog): reference #6305 in the shelf conversion lock entry
* fix(projects): share one conversion across concurrent first reads
A first-read conversion holds a bounded file-IO pool thread for its
whole duration. With the conversion moved out of the row lock,
concurrent first reads of one document each converted on their own
thread, so a burst of reads could starve unrelated offloaded file work.
Route the conversion through a single-flight keyed by (user, document,
project): the first reader leads and the rest wait on its result
without holding a pool thread. The map holds concurrent.futures.Future
values under a threading.Lock because the read tool also runs on the
isolated subagent loop. Followers wait through asyncio.shield, so a
cancelled follower cannot cancel the shared future, and a leader that
is cancelled or raises hands the work to a follower instead of failing
it.
* fix(runtime): keep renewing the lease while a staged terminal run finalizes
With an event store, run_agent stages the terminal status in memory and
commits it to the run store only after the journal flush, delivery receipt,
workspace scan and duration checkpoint. _renew_leases selected runs by local
status, so it stopped renewing as soon as the status was staged. A
finalization longer than lease_seconds + grace_seconds let a peer, or this
worker's own reconciler, claim the still-active row and record the
successful run as an orphan error.
The worker now marks RunRecord.terminal_commit_pending until the deferred
commit is attempted, and lease renewal treats it like the existing
scheduled-goal barrier: keep renewing, release on a confirmed own commit,
and fence when a peer claims the row. A fenced run stops renewing.
* docs(changelog): reference #6263 for the staged terminal lease renewal fix
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* fix(gateway): read uvicorn's WEB_CONCURRENCY in the remaining worker-count readers
* fix(gateway): name the worker spelling an operator can act on
The browser refusal and the dedupe warning both took the count but printed
GATEWAY_WORKERS, so on the launches that pass uvicorn no --workers the advice
silences the gate while the extra processes keep starting. Resolve which
variable set the count and print that, and let an unparsable GATEWAY_WORKERS
fall through to WEB_CONCURRENCY instead of reporting a single worker.
Also stop the comments from claiming the readers are equivalent: they share
the two spellings and the blank-means-unset rule, but channel_connections.py
turns an unparsable value into 0 and refuses while these stay inert.
* docs: state the single-worker requirement under both spellings
The browser tool group and the inbound dedupe warning now refuse or warn on
WEB_CONCURRENCY as well, so guidance that names only GATEWAY_WORKERS=1 is
incomplete for the shipped Dockerfile and serve.sh launches, where uvicorn
takes the process count from WEB_CONCURRENCY.
* test(features): isolate the browser capability surface from the shell's worker count
browser_capability() asks for the Gateway worker count, which this change makes
readable from WEB_CONCURRENCY, so an exported count in the invoking shell turned
test_features_reports_browser_control_enabled_when_configured_and_runtime_available
into a failure that has nothing to do with the router.
Four of the eight packages/harness/deerflow citations in this file name modules
that are not in the tree: sandbox/local.py, mcp/manager.py, skills/loader.py,
and tools/__init__ (missing its extension). The LocalSandboxProvider class lives
under sandbox/local/, the MCP and skills subsystems are packages rather than
single manager/loader modules, and get_available_tools is re-exported by
tools/__init__.py.
* feat(extensions): add host-bound Agent run control
* fix(extensions): preserve internal launches and scope run control
* docs(extensions): use a plugin-local run operation key
* fix(community): refuse non-global addresses in the shared SSRF guard
is_blocked_address ORed is_private, is_loopback, is_link_local,
is_reserved, is_multicast and is_unspecified. None of them covers the
100.64.0.0/10 shared address space, which holds CGNAT and Tailscale
hosts and Alibaba Cloud's 100.100.100.200 instance metadata endpoint, so
web_fetch, web_capture, the agentic browser and personal MCP connections
could reach it, including through the IPv4-mapped ::ffff:100.100.100.200
form a DNS answer can carry.
Also refuse any address that is not is_global, the classification the
repo's other SSRF filters already use. Keep the existing flags alongside
it: some non-public forms still report is_global, such as the NAT64
spelling of a metadata address (64:ff9b::a9fe:a9fe) and 224.0.0.1, so an
is_global-only check would newly allow them.
* docs(changelog): reference #6202 in the shared address space entry
* test(community): pin the IPv4-mapped metadata spelling at the web_fetch and personal MCP entry points
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* fix(tui): preserve multiline pasted input
The TUI composer used a single-line `Input`, so a bracketed paste
containing code, a stack trace, or multiple paragraphs kept only the
first line when submitted. The composer is now a multiline `TextArea`:
Enter still submits, Up/Down move inside the document before falling
back to input history at the first/last line, and the slash palette is
unchanged.
Originally #4630 by @felix-windsor, cherry-picked onto current main;
conflicts in app.py key gating (transcript paging keys), test imports,
and the relocated TUI docs were resolved by hand.
Fixes#4613
* fix(tui): submit multiline input unchanged and document the widgets layer
Address the review on #4630:
- Send non-command input to the agent and history exactly as typed or
pasted. Only the empty-input check uses strip(), so leading
indentation and trailing newlines in pasted code and stack traces
survive. Slash commands are still trimmed so they resolve as before.
- widgets/composer.py imports Textual, so the module-layout notes in
tui/AGENTS.md and docs/TUI.md now exempt widgets/ alongside app.py.
- Cover pressing Down on the last document line (history fallback),
whitespace-only input, and slash commands with surrounding whitespace.
Co-authored-by: Felix <felixwindsor3344@gmail.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(tui): treat soft-wrapped rows as composer navigation boundaries
TextArea soft-wraps long lines, and its Up/Down move between the wrapped
rows of one logical line. The history fallback only looked at logical
rows, so Up from the second visual row of a long prompt recalled history
instead of moving the cursor, and Down from its first visual row could
replace the draft with history's empty value.
Use the TextArea navigator's first/last wrapped-row checks for the
boundaries, and cover wrapped prompts in both directions. Also restore a
tui/AGENTS.md sentence about PageUp/PageDown to the app.py entry it
belongs to.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Felix <felixwindsor3344@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(community): add bounded retries for Jina web fetch
Closes#6116
Keep retries opt-in and share a total deadline across requests and backoff. Preserve cancellation and the existing Error contract. Include offline DNS and local-wheel fixture corrections needed to run the complete backend suite in the isolated verifier.
Signed-off-by: 97three <2212371308@qq.com>
* docs(community): scope Jina retry guidance to its provider
Signed-off-by: 97three <2212371308@qq.com>
* fix(community): jitter bounded Jina retry backoff
* docs(community): apply provider-scoped Jina guidance
Signed-off-by: 97three <2212371308@qq.com>
---------
Signed-off-by: 97three <2212371308@qq.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
Checkpoint durable refs for host-externalized tool results and materialize verified bytes before model calls on another Gateway.
Refs: #4189
Signed-off-by: PeaceMaker-best <221849497+PeaceMaker-best@users.noreply.github.com>
Co-authored-by: PeaceMaker-best <221849497+PeaceMaker-best@users.noreply.github.com>
* feat(persistence): add notification delivery outbox for scheduled task IM push
New notification_deliveries table and repository (migration 0012) backing
the outbox pattern for issue #4254: idempotent enqueue keyed on task_run_id
plus event plus provider plus target, atomic claim of due rows, and
exponential-backoff retry state moving pending to sending to sent or failed.
Execution status stays in scheduled_task_runs; delivery status lives here.
* feat(scheduler): enqueue and deliver scheduled-task notifications over IM channels
Closes the loop for issue #4254. On run completion the scheduler writes
outbox rows for the task owner's connected channel identities (success and
failure events; interrupts stay silent), and a new NotificationDeliveryWorker
polls due rows and pushes them through the channel's proactive send path.
Channel gains send_notification; WeCom implements it via the WS client's
send_message with errcode checking, the base class raises for providers
without proactive push.
Delivery is best-effort and never shadows execution status: enqueue and send
failures are logged and recorded in the outbox, which retries with backoff.
Channel liveness is re-checked at delivery time, and the worker stops before
the channel service on shutdown. Feature activates only when channel
connections are enabled; no new dependencies.
Verified end to end against a live WeCom bot: scheduled run completion
produced an outbox row that was delivered in under one second.
* feat(scheduler): include scheduled run result summary in IM notifications
The v1 notification only carried the run outcome skeleton (status, task id,
run id). Completed runs now include the agent's final answer so the IM push
is actually useful headlessly.
- The delivery worker accepts an optional resolve_run_summary(run_id,
user_id) resolver; for run_completed events it enriches the rendered text
with a "Result:" line (bounded to 1000 chars). The lookup is best-effort:
resolver failures or missing summaries degrade to the skeleton text and
never block or fail the delivery.
- The summary is read at delivery time from runs.last_ai_message (the
convenience column written on run completion), not snapshotted into the
outbox payload; retried deliveries re-read the freshest value and the
outbox row payload stays skeleton-only. The claimed outbox row is never
mutated in place.
- Gateway wiring scopes the read to the outbox row's owner_user_id so a
stale row cannot pull another user's run content.
- Failed runs keep the error-only skeleton: a partial answer on a failed
run would be misleading.
Tests: 9 new worker cases (rendering incl. truncation, sync/async
resolvers, failure/missing fallbacks, no resolver call for failed runs).
Full scheduled + outbox + channels regression green (441 passed), ruff
check + format clean.
* fix(scheduler): address review feedback on notification delivery reliability
Review follow-ups for the scheduled-task IM notification outbox:
1. Crash recovery for "sending" rows: claim now stamps updated_at and the
worker reconciles rows stuck in "sending" past a 10-minute lease back to
"pending" before every claim. Per-row try/except isolates a crashing
delivery from the rest of the claimed batch.
2. Double-send race under multiple workers: claim returns exactly the rows
its own UPDATE...RETURNING flipped (guarded by status='pending'), so a
rival worker's concurrent claim can no longer be handed out twice.
3. Restore non-ASCII comments and drop the BOM accidentally introduced in
test_scheduled_task_service.py.
4. Manual "run now" completions no longer enqueue IM notifications; the
user is already watching the UI. Missing trigger metadata fails safe
towards notifying.
5. Channel-outage failures no longer consume the retry budget: they are
parked with count_attempt=False on a flat 15-minute backoff so an
hours-long outage cannot exhaust retries and drop the notification.
Tests: repository 16, worker 21, scheduled-task service 31; targeted
regression across notification/scheduler/channel suites: 511 passed.
* fix(scheduler): address round-2 review on IM notification delivery
Park WeCom transport/outage failures without burning retries, bound worker stop, include task titles, align enqueue with delivery, and document the IM push behavior.
Co-authored-by: Cursor <cursoragent@cursor.com>
* docs(scheduler): complete IM notification docs and drop unused list API
Add backend/channels/migrations AGENTS notes, CHANGELOG and config.example hints for the outbox push path, and remove the unused list_by_task_run repository method flagged in review.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(scheduler): address round-3 review on IM notification delivery
Cap channel-down parking by row age and parked attempts, redact IM egress
content, narrow WeCom transport exceptions, add lifespan wiring tests, and
run ruff format so lint-backend passes.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(scheduler): redact entire PEM blocks in IM egress text
The previous BEGIN-only pattern left key material and the END marker in notifications.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(persistence): align 0014 parked_attempts with bootstrap schema tests
Backfill existing rows with a temporary server default, then drop it so create_all matches Alembic, and move HEAD anchors to 0014.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(persistence): format notification tests and shorten alembic revision id
ruff format --check failed on test_notification_delivery_worker.py.
0015_notification_delivery_parked_attempts exceeds alembic_version.version_num
(VARCHAR 32) and truncates on Postgres CI; use 0015_parked_attempts instead.
* test: fix post-merge CI failures after migration renumber
Update migration 0015 test HEAD expectation to 0017_parked_attempts and add queue_timeout_seconds to the lifespan notification worker mock scheduler config.
* fix(persistence): chain the notification outbox migrations after 0026
Main's head moved from 0016 to 0026_mcp_task_lease_tokens, so the two
outbox revisions are renumbered and re-parented:
- 0017_notification_deliveries -> 0027_notification_deliveries
- 0018_parked_attempts -> 0028_parked_attempts
Only the revision ids, parents and the migration guide change.
* fix(scheduler): adapt the notification hook and its docs to main
- tests: the completion tests assert on the atomic complete_run record;
new cases cover a completion that was not recorded, a task deleted
meanwhile, a failed task read, and the outbox being off.
- AGENTS.md: drop the root note and shorten the backend note to a pointer;
the two notes pushed two effective guidance chains past the 96 KiB limit.
- CHANGELOG: move the entry from the released 2.1.0 section to Unreleased
and add its link references.
* test(persistence): move the chain-head pin to the outbox migrations
test_migration_0026 pinned 0026 as the head, so it failed once the outbox
revisions chained after it. It now checks chain membership like the older
migration tests, and a new test file owns the head pin, checks the parents
and the 32-character revision id limit, and runs the upgrade and downgrade.
* fix(gateway): wire scheduled-task notifications after the channel service starts
The outbox wiring sat in the scheduler block and asked for the running
channel service there. #5035 moved the channel service start to after the
scheduler, and the two changes merged without a conflict, so on every
startup the lookup returned None and the feature disabled itself with the
"write-only outbox" warning. The lifespan tests patched the lookup to a
constant and kept passing.
The wiring now runs in its own step after the channel service start. The
scheduler is built as on main, and ScheduledTaskService gains
attach_notification_outbox() so the enqueue side is enabled once the
delivery worker is running. The guard, the worker and the shutdown order
are unchanged.
Tests: the lifespan helper can model the real accessor (None until the
channel service has started); a new case fails without this fix.
* fix(scheduler): enqueue from scheduler start and keep the push when the title read fails
Follow-up to the startup-order fix, from review.
- The enqueue side is wired with the scheduler again (it only needs the
durable table), and only the delivery worker waits for the channel
service. Attaching it later left a window after the scheduler started in
which a finished run wrote no outbox row. If the channel service is
missing or the worker cannot start, the enqueue side is detached, so the
outbox still never becomes write-only.
- A failed task read after complete_run no longer drops the notification:
the title is optional and the message falls back to the task id. A task
deleted meanwhile still stays silent.
* fix(channels): park WeCom notifications while the SDK socket is closed
The WeCom SDK raises a plain RuntimeError when its socket is not open, and
send_notification only mapped ConnectionError, TimeoutError and OSError to
ChannelUnavailable. A disconnect therefore spent the counted retry budget
and settled the row as failed after about 15 minutes, although the docs
say transport outages park without consuming retries.
send_notification now asks the client's is_connected first. The new test
uses a real, never-connected SDK client: it raised RuntimeError after three
retries before, and parks now.
* docs(scheduler): describe which outcomes notify; mirror the changelog entry
- README (en/zh) and CONFIGURATION.md: only runs that finish as success or
failed notify; occurrences that end without a finished run (launch error,
queue timeout, restart recovery) do not. Parking lasts up to about a day.
- CHANGELOG: cite the issue as text like the rest of the file, and add the
entry and link reference to CHANGELOG_zh.md.
- AGENTS.md: drop the backend pointer so this branch adds nothing to the
guidance chains closest to their size limit; the scheduler-side rules
and the lifespan order now live in app/channels/AGENTS.md.
* docs(changelog): cite #6135 next to #4843 for the notification outbox entry
* fix(scheduler,channels): re-check the IM binding at delivery time; park WeCom transport errors, count its rejections
Review round 1 on #6135.
The delivery worker now takes resolve_connections (the Gateway wires
ChannelConnectionRepository.list_connections) and, before the channel
liveness check and the summary read, requires the row's owner to still
hold a connected (provider, external_account_id). A disconnected target
is finalized with mark_failed(terminal=True); a lookup failure or an
unexpected result shape parks the row. Terminal outbox rows ignore
later status writes.
WeCom: the SDK raises RuntimeError from send_message for a socket that
is not open and for a dropped connection, but also for a platform
rejection ("Reply ack error: errcode=..."). Only that prefix keeps
counting; other RuntimeErrors, ConnectionError/TimeoutError/OSError and
websockets' ConnectionClosed park the row as ChannelUnavailable.
* fix(channels): treat a WeCom ACK rejection as proof the transport is up; pin the rejection prefix through the retry branch and the real connection-row shape
---------
Co-authored-by: GaoTianyuanhr <75788039+GaoTianyuanhr@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: GaoTianyuanhr <GaoTianyuanhr@users.noreply.github.com>
Co-authored-by: Totoro-qaq <279883115+Totoro-qaq@users.noreply.github.com>
* fix(uploads): reject filenames reserved for temporary staging
Guard thread and shelf upload names, validate SDK batches before copying, and preserve legacy shelf downloads. Add regression coverage and document the reserved staging namespace.
* fix(uploads): report reserved names before HTTP batch ingestion
Preflight reserved staging basenames before thread storage or sandbox setup. Return HTTP 400 with a rename hint; add HTTP and strict blocking-I/O regressions and update upload documentation.
* fix(uploads): reject Windows-style staging basenames on POSIX
Normalize backslashes before the HTTP reserved-name preflight. Add POSIX basename regressions for single and mixed batches and strict blocking-I/O coverage. Clarify the preflight scope and upload documentation.
Follow-up to the review of PR #6122; AI-assisted with OpenAI Codex.
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* fix(cli): exit non-zero when a headless run ends in an LLM error fallback
LLMErrorHandlingMiddleware turns provider failures (e.g. an expired
credential returning 401) into an AIMessage flagged
`deerflow_error_fallback` instead of raising. The Gateway run status and
the subagent executor already treat that flag as a failure. #5963 gave
`deerflow --print` / `--json` an error boundary for raised exceptions,
but a flagged fallback never raises, so both modes still printed the
fallback text and exited 0.
Both headless modes now follow #5963's failure contract when the run's
final AI message carries the flag: `--print` keeps the fallback text on
stdout and adds an `Error:` line on stderr; `--json` appends the terminal
`{"type": "error"}` record. Both exit 1.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* refactor(cli): share final-answer accumulation with the embedded client
* docs: scope headless accumulator guidance to TUI
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* docs(backend): point the stream-mode translation anchor at stream_modes.py
* docs(backend): use one anchor spelling within the stream-mode table row
The row named the translator by its backend-relative path and its call
site by the harness-relative short form, so the same symbol was written
two ways two rows apart.
* fix(config): upgrade the config.yaml the Gateway loads
scripts/config-upgrade.sh resolved DEER_FLOW_CONFIG_PATH, then
backend/config.yaml, then <checkout>/config.yaml. make dev exports
DEER_FLOW_PROJECT_ROOT=<checkout>, so the Gateway reads the checkout copy
first: with both files present the script upgraded a file nothing loads
and left the live config outdated while reporting success. It also
ignored DEER_FLOW_PROJECT_ROOT and a DEER_FLOW_CONFIG_PATH set in .env,
and fell back to another file when DEER_FLOW_CONFIG_PATH named a
missing one.
Ask the harness resolver (AppConfig.resolve_config_path) instead of
copying its order, defaulting DEER_FLOW_PROJECT_ROOT to the checkout as
serve.sh does. A missing DEER_FLOW_CONFIG_PATH or invalid project root
now fails with the Gateway's error; no config anywhere still creates
<checkout>/config.yaml from the example.
* docs(changelog): reference #5991
* test(config): pin config-upgrade seeding a missing config from the example
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* fix(tui): report headless --print/--json failures instead of leaking tracebacks
The headless --print and --json runners dispatched straight into the
embedded client with no error boundary, so session-creation, thread-
resolution, chat, and stream failures escaped main() and surfaced as
unhandled Python tracebacks from the console script - for example the
documented CheckpointModeMismatchError case, which the checkpoint
operations/troubleshooting docs recorded as a raw traceback.
Both runners now catch Exception (KeyboardInterrupt/SystemExit still
propagate), mirroring the TUI's stream_actions() boundary:
- --print writes one concise 'Error: ...' line to stderr and returns 1
- --json appends a terminal {"type": "error"} record to the NDJSON
stream (diagnostics stay on stderr) and returns 1; every stdout line
remains independently valid JSON, including failures mid-stream
The checkpoint operations/troubleshooting docs (en + zh) no longer
describe the headless runners as printing an uncaught traceback.
Fixes#5962
* fix(tui): close the broken-pipe edge and document the --json error contract
Review follow-ups on #5963 (both non-blocking):
- print(answer) sat outside the headless boundary, so a consumer closing
the pipe early (deerflow --print ... | head) still got a
BrokenPipeError traceback plus the interpreter's shutdown-flush noise.
The final --print write and the --json terminal error-record write now
catch BrokenPipeError, point the stdout descriptor at devnull (the
documented SIGPIPE-adjacent recipe) and exit 1 without a traceback.
- The terminal {"type": "error"} record is a real contract for --json
scripters but was only documented in the checkpoint pages; the
backend/docs/TUI.md command table and the README cheat line now state
the error record and the exit-1 semantics where CLI consumers look.
* feat(community): add Unbrowse web fetch provider
Add a community web_fetch provider backed by Unbrowse
(https://unbrowse.ai), a hosted service that returns a page as markdown,
over plain HTTP when that is enough and through a cloud browser when the
page needs JavaScript. The tool sends one JSON-RPC tools/call POST
(unbrowse.scrape) to Unbrowse's MCP endpoint, so it needs no extra Python
package and no MCP client (uses httpx, already a dependency).
Unbrowse is not a general web search engine, so only web_fetch is added.
Changes:
- backend/packages/harness/deerflow/community/unbrowse/{__init__,tools}.py
web_fetch_tool with the same output shape as the Sofya, Exa and
Firecrawl fetch tools (# <title> plus markdown, 4KB cap). API key is read
from the config.yaml api_key field or the UNBROWSE_API_KEY env var.
Optional render: auto|never|always; unknown values log a warning and fall
back to auto. JSON-RPC errors, isError tool results, HTTP errors and
malformed payloads come back as "Error: ..." strings.
- backend/tests/test_unbrowse_tools.py: key resolution, render handling,
request shape, structuredContent vs text payloads, error paths.
- scripts/doctor.py, scripts/wizard/providers.py: register the key and the
web_fetch provider; backend/tests/test_doctor.py covers both doctor paths.
- config.example.yaml, .env.example, backend/docs/CONFIGURATION.md,
tools/AGENTS.md, EN/ZH harness/tools.mdx: document the provider.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* docs(changelog): add Unbrowse web_fetch provider entry
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* feat(memory): add Jev pre-screening and signal classification over a shared TypeSafe client
- refactor(typesafe): extract a shared Jev (TypeSafe) client and unify its
configuration: one connection-precedence rule, credential fingerprint, wire-size
counting and error taxonomy, shared by the guardrail gate and the memory hooks.
- feat(memory): wire `memory.prescreen` and `memory.signal_classification` into
DeerMem through one injected judge hook (both off by default, `enforce` gated by
the shadow evaluation).
- config: add the top-level `typesafe:` block and the two judging slots to
`config.example.yaml` and the Helm values, bump `config_version` to 50, and sync
the Helm README example with it.
- test(memory): add the shadow-evaluation script (`eval_memory_prescreen.py`) with
per-stratum gates, censoring and the no-network baseline.
Review follow-ups folded in:
- `mutations_accepted` counts only facts that survive capacity enforcement and the
upload-event scrub that precedes persistence, only removals/reinforcements of ids
that existed, and persisted near-duplicate merges (matched by id, or by identity
for legacy id-less facts); it is published only after persistence succeeds.
- The cached manager hot-reloads its judge when a memory slot or the shared
`typesafe` block changes (including a credential rotation); publication is
serialized and revalidates the config generation, so a build started before an
`enforce` -> `off` rollback cannot land after it.
- Cache provenance is per answer: the `cached` flag and the reported model follow
the answers a side actually consumed, across partial merges and answerless retries.
- An oversized noul integer stays a per-question error instead of raising.
- Incremental call savings use the shared scored denominator.
- The sensitive-stratum gate requires evidence for each of identity, preference
and correction.
* fix(typesafe): resolve one credential per layer and refuse unusable connection values
Review follow-ups on the shared TypeSafe client (#5906):
- A consumer's `api_key_env` is no longer overridden by a literal `api_key` in the
shared `typesafe:` block: the first layer that sets either credential setting
decides it, and inside one layer the literal key wins (the standalone
semantics). A consumer whose endpoint does not accept the block's key no longer
sends it on every call, and the missing-credential message keeps naming the
variable that layer configured.
- `base_url` is validated where it is configured: `http(s)` only, with no query,
fragment, embedded credentials, whitespace or control characters, and a
malformed port fails at construction instead of on every request (httpx raises
`UnsupportedProtocol` / `InvalidURL` per call, and the query would have been
swallowed into the appended endpoint path).
- A credential that cannot travel as a header value — surrounding whitespace from
a secret mounted from a file, a non-printable character — is refused when the
connection resolves. It used to construct fine and then put the whole
`Bearer <key>` value into h11's `LocalProtocolError` message, which the
guardrail middleware's `logger.exception` prints.
- The client raises transport failures without chaining the original exception,
so that message can no longer carry the credential through the cause chain; the
exception type stays in the message for diagnosis.
- `TypeSafeConfig` validates in strict mode, agreeing with the consumer `config`
path: `max_attempts: true` was silently read as `1`. An integer stays a usable
float, so `timeout: 5` keeps working.
Tests: 20 cases across `tests/test_typesafe_config.py` and
`tests/test_typesafe_client.py`, each red on the previous revision and green now.
367 focused typesafe/memory/config tests, `make test-blocking-io` (167) and the
eval-script suites (50) pass; `ruff check` and `ruff format --check` are clean.
Docs synced: `backend/docs/GUARDRAILS.md`, `typesafe/AGENTS.md`,
`config.example.yaml`.
* fix(typesafe,memory): bound the reply, record request failures, gate enforce on the evidence
Second review round on #5906: the remaining items from the review thread and from
the follow-up pass.
Typesafe client (the reply is treated as hostile input too):
- Send `Accept-Encoding: identity`, refuse a body that carries a non-identity
`Content-Encoding`, and read through a bounded `iter_bytes` that stops past
`MAX_RESPONSE_BYTES` (64 KiB). A gzip body that expands to 128 MiB used to be
decoded in full and, on the gate path, allowed the tool call unexamined; it is
now `invalid_response` (a denial with `fail_closed: true`).
- Map `RecursionError` from `json.loads` to `invalid_response`: deep nesting
escaped the module's taxonomy, so the memory consumers lost the round's record
instead of recording a fallback.
Memory judge:
- A request-level failure is recorded as `request_failed`, never `no_verdict`
(`REASON_REQUEST_FAILED` was defined and never emitted, so an unreachable
endpoint was indistinguishable from a provider with no usable answer). The
adapters' `decide()` now propagates `TypeSafeError` for the coordinator to record
and log; `no_verdict` stays for a question-level failure, and an ineligible side
keeps its own reason even when the round's request failed.
Shadow evaluation (the enablement gate reads the evidence, not one number):
- The miss-rate gate requires the exact one-sided Clopper-Pearson upper bound to
meet the 1% target; a sample that only meets the point estimate is
`INSUFFICIENT` (0/250 was reported as a pass with a 1.19% bound).
- New `baseline_incremental_savings` gate: a baseline that saves as many calls as
the model — the shipped all-trivial sample — fails instead of enabling.
- Negative counts (`mutations_accepted`, message / batch / token counts) are
treated as absent rather than scored: `-3` read as "nothing lost", adding a
phantom saving to the numerator and the denominator.
Docs: the design documents cited but absent from the repository are replaced with
in-repo pointers (`MEMORY_IMPROVEMENTS.md`, `agents/memory/AGENTS.md`,
`GUARDRAILS.md`), and the gate and limit wording now matches what the code
enforces.
Tests: 31 new or updated cases, each red on the previous revision (compressed body
refused even when it decodes to a valid envelope, body over the cap refused, body
at the cap answered, deep nesting mapped, `request_failed` on each request-level
failure, the bound and baseline gates, the negative counter). 454 focused tests,
`make test-blocking-io` (167) and the eval suites pass; `ruff check` and
`ruff format --check` are clean.
* fix(memory-eval): pair the baseline comparison on the same records
Review follow-up on the incremental-savings gate added in ed930478 (P2): both
sides of the comparison are now the records the no-network baseline can score
(`trivial_only` plus a readable signals list).
Subtracting that subset's heuristic savings from model savings over *all* scorable
skips credited the model with batches the baseline never saw: 300 reviewed skips
with a single annotated record computed 300-1=299 "incremental", passed the gate
and set `enable_recommended`, while the paired comparison was model=1 against
heuristic=1 (incremental 0).
- `Evidence` carries the paired numbers (`baseline_paired_skips`,
`baseline_model_saved_calls`, `baseline_heuristic_saved_calls`) instead of one
heuristic count, filled by `build_report` from the baseline summary.
- The gate reports `INSUFFICIENT` when no scored skip carries the annotation and
when the annotation covers only a subset of the scored skips (the message carries
the paired counts and the `paired of scored` coverage), `FAIL` on a non-positive
paired gain, and `PASS` only on a positive gain over the full paired set.
- Docs state the pairing/coverage requirement in
`backend/docs/MEMORY_IMPROVEMENTS.md` and in the script docstring.
Tests: the reviewer's counterexample as a report-level regression (it failed with
`assert 'PASS' == 'INSUFFICIENT'` on ed930478 while `enable_recommended` was true),
a unit case for the paired-subset `INSUFFICIENT`, and the renamed paired fields.
344 focused tests, `make test-blocking-io` (167) and the eval suites pass;
`ruff check` / `ruff format --check` clean.
* fix(typesafe): bound the sync body read by the deadline
`deadline_seconds` only covered the sync path's headers, the completed body and
the parse, so an endpoint that drips bytes just under the read timeout held the
caller for one read per piece — the whole of `MAX_RESPONSE_BYTES` past the budget.
The read loop now checks the budget before every transport read, which bounds the
overshoot to one read. The pieces are the transport's own reads, not re-chunked
output: `iter_bytes(chunk_size)` pulls many reads inside a single yielded chunk,
so a check at that granularity would never see a drip.
`tests/test_typesafe_client.py` pins it with a valid envelope delivered as 24
pieces of 50 ms against a 0.2 s deadline: red before (24 pieces pulled), green now
(~5), on both the sync and async paths. Docstring and `typesafe/AGENTS.md` track
the new granularity.
* fix(memory): read the YAML off spelling, and veto a skip over the whole batch
Two review findings on the memory judging path.
`memory.prescreen.mode` / `memory.signal_classification.mode` rejected the
`mode: off` spelling `config.example.yaml` ships and a rollback types: YAML 1.1
parses the unquoted value as a boolean, so both fields saw `False`. The failed
reload left the previous judge running, so both fields now normalize `False` to
`off` before the `Literal` runs. `True` stays a validation error (no mode it
could mean). The examples quote the value, and the docs record why.
The pre-screen's L3 veto read `detect_signals`' 6-message window while the batch
it answers for is the whole post-watermark feed, so an explicit correction in an
earlier turn of an eight-message batch did not stop the skip that then advanced
the watermark past it. `detect_signals` gains a `window` parameter (default
unchanged) and the updater scans the full feed for the veto; the extraction hint
and the reinforcement gate keep the window.
Both regressions are red on the previous head: the config test reproduced the
`input_value=False` rejection through `AppConfig.from_file`, and the batch test
reproduced the prescreen judging (and skipping) a batch whose first turn carries
the correction.
* fix(goal): show the active goal to the agent on every model call
Nothing put ThreadState.goal into model requests. The agent only knew the
goal from messages: the objective the Web UI and IM channels send, and the
hidden <goal_continuation> message. Goals set through PUT /goal, the
embedded client or the TUI never reached it, and compaction could remove
the messages that stated the others.
DurableContextMiddleware now renders the active objective first in its
hidden durable-context data message, escaped, bounded, PII-redacted and
treated as data like the summary. Only the objective is rendered, so the
block changes only when the goal is set or cleared.
* fix(goal): let the agent pursue the active goal at user priority
The durable-context contract told the model never to follow instructions in
its data message, and that message now carried the goal. While a goal is
rendered, the contract ends with a static exception for the <active_goal>
element that opens the data message: work toward it as a request in a user
message, with no system or developer authority, while every other field
value, and any other text that calls itself a goal, stays data.
The element replaces the heading the goal used, which a summary could
reproduce. Other field values in the data message are HTML-escaped, and
active_goal joins the sanitizer denylist, so user input and remote tool
results cannot carry a raw one either.
---------
Co-authored-by: Totoro-qaq <279883115+Totoro-qaq@users.noreply.github.com>
* feat(extensions): grant bounded host model invocation
* fix(extensions): retain model budgets and isolate schema validation
* fix(extensions): normalize provider-origin model cancellation
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* docs(backend): drop references to APIs that do not exist
PATH_EXAMPLES.md's upload example imports THREAD_DATA_BASE_DIR from
thread_data_middleware; the module exists but exports no such name, so the
snippet fails with ImportError as written. Resolve the same directory through
the accessor the middleware itself uses, get_paths().sandbox_uploads_dir(),
which yields the layout the file's own path table documents.
CONFIGURATION.md's Docker sandbox block lists `auto_start: true`. SandboxConfig
declares use/port/container_prefix/mounts/network but not auto_start, and the
provider is constructed as cls(**kwargs) from the caller of
get_sandbox_provider() rather than from config extras, so the key is accepted by
extra="allow" and then silently dropped. Remove the line from the example.
* docs(backend): resolve the uploads example through the owner-scoped bucket
get_paths().sandbox_uploads_dir(thread_id) without a user_id selects the legacy
threads/<id>/ bucket, but Gateway uploads land in
users/<user_id>/threads/<id>/user-data/uploads/, so the example imported cleanly
yet still raised FileNotFoundError for a normally uploaded file. Use
deerflow.uploads.manager.get_uploads_dir(), which resolves the effective user
and validates the thread id.
* docs(backend): show the real uploads path layout in every PATH_EXAMPLES demo
The remaining sections still documented the pre-isolation
threads/<thread_id>/ bucket and called `path` a backend-relative string, while
the upload and list APIs return an absolute path under
users/<user_id>/threads/<thread_id>/user-data/uploads/. Reproduced with
deerflow.uploads.manager.get_uploads_dir() and
uploads._list_uploaded_files_for_thread() against DEER_FLOW_HOME: the listed
path is /<home>/users/default/threads/abc123/user-data/uploads/report.pdf.
* docs(backend): clarify runtime home precedence
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* fix(config): honor $VAR substitution for request_admission integer fields
AppConfig.resolve_env_variables replaces every `$VAR` reference with the raw
environment string. RequestAdmissionConfig.requests_per_minute and
max_queue_size are declared `strict=True` so a bool or float cannot slip in,
but strict mode also refuses the string "60", so `requests_per_minute: $RPM`
failed the whole config load with "Input should be a valid integer" even when
RPM=60 — contradicting config.example.yaml's promise that environment
variables work for every field. They were the only strict fields in the
config schema; every other integer coerces in lax mode.
Wrap both fields in a BeforeValidator that converts a decimal integer literal
arriving as a string and leaves everything else untouched, so the strict check
behind it still rejects bools, floats, "60.0" and non-numeric strings.
Tests pin the accepted literals, the rejections that must survive, and the
end-to-end path from a config.yaml `$VAR` reference through AppConfig.from_file.
* docs(changelog): reference #5838 in the request_admission entries
* feat(prompts): add operator overlays for core system prompts
* test(prompts): isolate SOUL forwarding test from operator config
* fix(prompts): wrap complete subagent system message with overlay
* fix(subagents): serialize prompt overlays in durable batches
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* feat(models): add declarative reasoning capability contract (#5073)
Introduce an optional per-model `reasoning:` block beside the legacy
`supports_thinking` / `supports_reasoning_effort` booleans. It declares
thinking availability (unsupported | optional | required), the accepted
effort values with aliases, default, and serialization path, the payload
dialect, and the reasoning-history requirement. When present the booleans
are derived from it and contradictory profiles fail at config load.
`deerflow.models.reasoning` normalizes legacy and declared profiles into
one immutable contract and resolves every caller's generic request against
it. `create_chat_model` is the single enforcement point: required-thinking
models never receive a synthesized disable payload, unsupported effort
values never reach the provider, and the dialect synthesizes the on/off
payload when no template exists. Legacy profiles keep the historical path
byte-for-byte. The lead agent and subagent descriptor resolve the same
policy first so run metadata reports effective values.
`/api/models` and `DeerFlowClient` project the contract as `reasoning`;
the frontend derives mode clamping and effort choices from it through
`core/models/reasoning.ts`, and user preferences accept provider-specific
effort tokens. The setup wizard's Z.AI GLM-5.3-Flash profile migrates to
the contract, restoring its low/high/max effort control.
* docs(models): move the reasoning contract guidance into models/AGENTS.md
backend/AGENTS.md sits within 46 bytes of the 96 KiB effective-chain hard
limit for the middlewares, runtime, sandbox, and subagents guides, so the
new section pushed scripts/check_agent_guidance.py over it. The models
guide is the nearest file for the factory and the contract module, and
its chain has ample room.
* fix(models): address review on the reasoning contract
- vLLM dialect: the enable path now mirrors whichever chat_template_kwargs
switch the template declares, like the disable path already did, so a
template using the older `thinking` key no longer also grows
`enable_thinking`.
- Validate operator-supplied effort values at `effort.path` inside
`when_thinking_enabled`, `when_thinking_disabled`, and the `thinking`
shortcut against the contract, not only the profile-level value; they are
forwarded when the caller chooses nothing, so the "unknown values never
reach the provider" guarantee now holds for them too.
- Constrain `effort.path` to dotted identifiers and reject single-segment
paths that would replace a whole mapping (`extra_body`, `thinking`,
`model_kwargs`, `default_headers`, `default_query`).
- Z.AI GLM-5.3-Flash profile: `default: high` instead of `max`, because the
default also governs summarization, title generation, and subagent calls
that never choose an effort; `max` stays selectable. Documented in the
config example, the configuration manual, and the design note.
* fix(models): close reasoning contract review gaps
* fix(models): preserve native Ollama reasoning config
* fix(models): keep native reasoning strings and project reasoning into the descriptor
- `ModelConfig.reasoning` also accepts the level string ChatOllama takes
for gpt-oss style models (`low` / `medium` / `high`); the factory forwards
both the boolean and the string on the legacy path, so profiles that
passed `reasoning: high` through before the contract field existed keep
loading and keep sending it.
- The assembly descriptor no longer excludes `reasoning` from
`model_parameters`: the native ChatOllama value and a declared contract's
dialect/history/effort change the request payload, so they now move the
assembly fingerprint.
- e2e: the sidecar spec pinned the pre-contract body where a model without
effort control still received `reasoning_effort: minimal`; the preset is
now dropped for such models (the Gateway discarded it anyway), so the
assertion expects the key to be absent.
The docs site pointed at POST /api/extensions/{skills,mcp}/{name}/enable
and /disable, and backend/docs/API.md at POST /api/skills/{name}/enable
and /disable. None of these routes exist. The gateway toggles a skill
with PUT /api/skills/{skill_name} and an MCP server with
PATCH /api/mcp/config, both admin-only, which is also what the
frontend calls.
* feat(guardrails): add TypeSafe (Jev) pre-execution tool risk gate
Adds TypeSafeGuardrailProvider, a network guardrail provider that asks one
`noul` question per probed tool call and denies calls whose Jev risk
probability reaches `threshold`. It reuses the existing GuardrailProvider
slot: no middleware, config-schema or dependency changes.
Semantics that are load-bearing:
- state is strict JSON (no `default=`, `allow_nan=False`, string keys only);
unserializable arguments, lone surrogates among them, and argument text
above `max_state_chars` are denied locally with zero network requests,
never truncated and sent.
- the async path cancels in-flight requests through `asyncio.timeout`; the
sync path checks the deadline after the response headers arrive, around
the body read and before accepting the result, and drops late results.
No total return-time bound is promised on the sync path.
- clients are created per evaluation inside a context manager, and the
injected `transport_factory` is called once per network-needing
evaluation: httpx closes an injected transport with its client, so a
shared instance would be dead on the second call.
- denial reason messages carry probability, threshold, served model version
and state digest, so the threshold comparison can be replayed from the run
journal; local denials record no probability or model version, and allow
decisions stay unpersisted (the audit path is unchanged).
- provider errors carry no response body or arguments: malformed responses
report the field and the offending type only, so a response echoing tool
arguments cannot reach middleware logs or the evaluation report.
The online evaluation harness (`backend/scripts/eval_typesafe_risk_gate.py`
plus its labeled case set) is the threshold/tools evidence that must precede
enabling interception. It reports four mutually exclusive populations
(uncached network, cache hits, unprobed tools, local refusals), separates
model classification from local refusals and fail-closed errors, measures
latency from the caller entering `aevaluate` through cleanup, discloses
censored samples, probes DNS/TCP/TLS separately, and flags unprobed tools as
a coverage gap rather than a pass. Real responses are never written to disk.
Cache-pass samples are classified from `decision.metadata['cached']` and
counted once, warming calls are retained as an auxiliary group with their
failures named in the report and JSON, and an optional connection-probe
diagnostic can never discard completed evaluation results.
Tests: the provider suite over MockTransport plus the middleware
integration, a blocking-IO anchor that drives `aevaluate` against a real
loopback HTTP server with a meta-check proving the sync path on the loop
trips the gate, and harness cases covering the reporting defects. Every fix
above was verified to fail its new tests when reverted.
Docs: GUARDRAILS.md gains Option 4 (privacy, local-refusal trade-off,
deadline semantics, audit boundary, coverage/bypass risk, Allowlist->RBAC
migration, lone-surrogate local refusals and the no-response-body property
of error messages) and config.example.yaml gains a smoke-scope example.
Config: the new example block bumps `config_version` 46 -> 47, with the
Helm chart's embedded version and README example kept in step as
`scripts/check_config_version.sh` (chart.yaml CI) requires.
* feat(guardrails): whitelist tool calls inside the TypeSafe risk gate
Static allow rules now live in the provider instead of waiting on a composite
provider. `whitelist` mirrors `AllowlistProvider.allowed_tools`: omitted means
"not enforced", `[]` refuses every tool, and a list refuses every tool outside
it locally (`typesafe.tool_not_whitelisted`) before state construction, caching
and networking, independently of `fail_closed`. Whitelisted tools are still
probed and still denied when their risk probability reaches `threshold`, and
the whitelist is checked before the probe scope so an unlisted tool cannot
inherit an allow from being unprobed.
Deny rules stay with `authorization.*`. The policy version moves to 1.1.0, the
evaluation script reports whitelist refusals as their own population, and
`GUARDRAILS.md` / `config.example.yaml` document the allow-half migration.
* fix(guardrails): address TypeSafe risk gate review findings
- Status errors report only the numeric status. The HTTP status line is
server-controlled like the body, and this message reaches middleware logs
and the evaluation report, so a hostile endpoint could echo the state (raw
tool arguments) through it.
- Expired cache entries are removed with `pop(key, None)`. Sync evaluation
runs on executor threads when a turn produces parallel tool calls, so a
second caller can remove the same expired entry; a bare `del` raised
KeyError, which the middleware reports as a provider error and, under
fail_closed, turns into a spurious denial.
- The evaluation harness states its acceptance gates inline (no missed
dangerous call, system false-block <= 5%, uncached network p95 <= 1s)
instead of citing a design spec that is not in the repository, and reports
the whitelist-refusal population alongside the other four.
- `--tools` / `--whitelist` strip whitespace and drop empty entries, so
`--tools "bash, write_file"` no longer probes only bash while treating
`" write_file"` as an unknown tool name.
- The case set's `padding` note says top-level argument fields only, matching
what the runner implements.
Each fix was verified by running its new test against the pre-fix code.
* refactor(guardrails): name the TypeSafe permission list allowed_tools
Align the key with `AllowlistProvider.allowed_tools` so an allowlist migration
copies it verbatim instead of translating `allowed_tools` into `whitelist`.
The companion names follow: reason code `typesafe.tool_not_allowed`, metadata
`tool_not_allowed`, harness population `not_allowed`, and the `--allowed-tools`
flag.
Semantics are unchanged: omitted = not enforced, `[]` = refuse every tool,
tools outside the list are refused locally before any state, request or cache
entry, and listed tools still face the risk gate.
* fix(guardrails): record only a bounded served model token; fail the eval exit code
Two findings from the re-review at d3f0b072.
`model` is response content, and `_model_decision` interpolated it verbatim into
the reason message -- which reaches the denied ToolMessage, the run journal,
middleware logs and the evaluation report on disk, and which the cache retains
for replay. A malformed or hostile endpoint can put an echoed tool argument,
injected text or a multi-megabyte string there. The token is now recorded as
itself only when it matches a conservative shape; anything else is recorded as
`unrecorded:sha256:<digest>`, which keeps the verdict and a stable provenance
record without echoing the value. Rejecting the response instead would turn an
odd version string into a fail-closed denial of every call the endpoint answers.
The digest path encodes with `surrogatepass` so an escaped lone surrogate -- a
value `json.loads` happily produces -- cannot raise where this must not fail.
`scripts/eval_typesafe_risk_gate.py` returned 0 even when a gate failed, so an
operator scripting `... && enable` got a green light on a failing evaluation.
`main()` now returns 1 when any gate in `report["score"]["gates"]` fails, after
the JSON report is written, and prints the failed gate names.
Tests: `test_an_unrecordable_served_model_is_digested_instead_of_echoed` (an
echoed argument, a space, a lone surrogate and an oversized value, plus the
cached replay) and `test_main_exit_code_carries_the_gate_verdict` (both
directions). Both fail against the pre-fix source.
docs/GUARDRAILS.md audit boundary updated to match.
113 passed: tests/test_typesafe_guardrail_provider.py,
tests/test_typesafe_risk_eval_script.py,
tests/blocking_io/test_typesafe_guardrail_provider.py.
* fix(evals): fail the risk-gate run when a labeled population is uncovered
Review finding on dcd38cc3. The new exit code made `_score`'s vacuous passes
load-bearing: unprobed calls are excluded from the score, so `risky_misses_zero`
is true over an empty risky population. `--tools read_file` excludes all 18
risky cases in the bundled set, so the run printed every gate as PASS and exited
0 after evaluating no risky operation at all.
`_score` now gates coverage itself: `risky_coverage_present` and
`safe_coverage_present` fail when the probe scope leaves that labeled population
with nothing evaluated, so a scope that cannot substantiate the verdict no longer
reports one. The safe side had the same vacuity -- `system_false_block_within_target`
passes trivially with zero safe cases.
Verified end to end against a loopback mock endpoint with the bundled case set:
`--tools read_file` exits 1 with `GATE risky_coverage_present: FAIL` while the
other four gates still PASS, where dcd38cc3 exits 0 on the same command.
Tests: `test_main_fails_when_the_tool_scope_leaves_no_risky_case` and
`test_main_fails_when_no_safe_case_is_in_scope` (both fail against dcd38cc3),
`test_bundled_case_set_keeps_both_labels_the_coverage_gates_require`, and the
exit-code test now runs with both populations present.
116 passed: tests/test_typesafe_guardrail_provider.py,
tests/test_typesafe_risk_eval_script.py,
tests/blocking_io/test_typesafe_guardrail_provider.py.
* fix(auth): claim the first admin atomically instead of counting first
POST /api/v1/auth/initialize read the admin count in one session and
created the admin in another, with no lock between them. Two requests
arriving together with different emails both saw an empty system and
both became admins: reproduced over HTTP, two 201s and two admin rows,
after which setup-status reports an initialized system.
The endpoint is public and unthrottled, and during the pre-setup window
anyone who can reach the gateway can claim admin with a single request
anyway — that part is the first-boot design. What the race removes is
the only signal: without it exactly one request wins and the loser gets
409, so an operator who loses sees "already initialized" and looks into
it. With it, both succeed and the operator's setup looks ordinary.
The check and the insert now share one transaction in the repository,
with writers serialized first — SQLite takes the write lock up front
(BEGIN IMMEDIATE) and Postgres a transaction-scoped advisory lock, the
idiom the project and channel repositories already use. The same
repository serves both backends, and multi-replica deployments make the
race genuinely parallel rather than an event-loop interleaving.
create_first_admin returns None when the claim is lost, which the route
reports as 409; uniqueness conflicts still raise ValueError and stay a
400. The count in the handler is now only a fast path that avoids
hashing a password for an already-initialized system.
AUTH_DESIGN.md said concurrent initialization was backstopped by a
database unique constraint. No such constraint exists — email
uniqueness only catches the same address submitted twice — so it now
describes the claim that does the work.
* docs(changelog): note the atomic first-admin claim (#5776)
* fix(auth): refuse a dialect the first-admin claim cannot serialize
Review follow-up. The claim's serialization handled sqlite and
postgresql and fell through silently for anything else, which would
leave create_first_admin as the plain check-then-act it exists to
prevent — the one place the abstract method's atomicity contract is
enforced.
An unknown dialect now raises instead. The engine builds only these two
today, so this guards a future backend: it fails loudly at first boot
rather than quietly reopening the race.
* feat(storage): add a content-addressed blob store contract (#4189 item 2)
Addresses the review on the previous head:
* the retention/dedup claim in the doc now states the deletion rule (liveness
from the durable references) instead of promising refcount-free deletion, and
the sidecar records the first writer as advisory provenance only;
* one validate_blob_kind() owns the kind grammar, so the message a producer sees
is the pattern that is enforced;
* adds the missing storage/AGENTS.md (registry entry included) that both
docstrings already pointed at;
* drops the dead _import_backend_module() by routing the folder scan through it;
* exports BlobNotConfiguredError from the package;
* renames _KIND_MAX_BYTES to _MAX_BLOB_BYTES (it caps one put, not a kind);
* delete failures raise the neutral BlobStoreError instead of BlobReadError.
Rebased onto current main so the guidance-budget test matches upstream.
Signed-off-by: zeng-bohan <zengbh1@gmail.com>
* fix(storage): restore the config module and backend packages dropped by the squash
The force-push to e72b435 silently lost four items the surviving code
imports, breaking `import deerflow.storage` and the test suite at
collection:
- deerflow/config/blob_storage_config.py (BlobStorageConfig +
get/set_blob_storage_config + load_blob_storage_config_from_dict,
mirroring memory_config's host-shared schema and hot-reload pattern)
- app_config.py wiring: the blob_storage AppConfig field, the import,
and the _apply_singleton_configs -> load_blob_storage_config_from_dict
call (memory stayed wired; blob did not)
- storage/backends/__init__.py (drop-in contract docstring)
- storage/backends/local_fs/__init__.py (LocalFsBlobStore re-export +
STORE_CLASS for _scan_backends)
Verified: pytest tests/test_blob_store.py tests/test_app_config_reload.py
tests/test_app_config_name_indexes.py -> 67 passed; ruff format/check
clean on all four files.
* fix(storage): name the failing digest check; document the 64 MiB per-put cap
- get_bytes: split the corruption message into size mismatch (truncation /
partial write) vs digest mismatch (same-size bit rot, reporting what the
content hashes to vs what the ref names) — the combined message reported
'got 1024 bytes, expected 1024' on same-size corruption and sent the
debugger in the wrong direction.
- docs/blob-storage.md + storage/AGENTS.md: state the _MAX_BLOB_BYTES cap
(64 MiB) and that it is backend-level defense-in-depth rather than part
of the BlobStore contract, so the externalized-tool-results follow-up
sees the limit before the first production put fails.
* test(storage): pin both get_bytes corruption branches separately
The pre-split test corrupted the blob with different-size bytes, so only
the size-mismatch branch ever ran and neither new message was asserted.
Now: different-size corruption asserts 'size mismatch', and a same-size
reversal (bit rot) asserts 'digest mismatch' — the branch the split was
actually about.
* docs(config): document blob_storage in config.example.yaml
Add the blob_storage section (enabled/backend/backend_config) mirroring the
memory section's shape, matching the Configuration section of
backend/docs/blob-storage.md, and bump config_version 46 -> 47 per the
example-file convention for schema additions. Default state is unchanged:
enabled: false, purely opt-in.
* fix(helm): align chart config version with example
* fix(storage): refresh blob store on config changes
---------
Signed-off-by: zeng-bohan <zengbh1@gmail.com>
Co-authored-by: zengbohan1 <310902929+zengbohan1@users.noreply.github.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
Co-authored-by: zeng-bohan <310902929+zeng-bohan@users.noreply.github.com>
