Add an opt-in `include_evidence` to both chat endpoints. The response then
carries the conclusions and messages the agent read while answering, plus the
tools it called.
Evidence is collated from what the agent accessed rather than reported by the
model. That over-reports -- a conclusion appears because the agent saw it, not
as proof the answer used it -- but it is deterministic, costs no model tokens,
and behaves the same at every reasoning level. Asking the model to cite its
sources fails quietly instead: weaker models produce incomplete or invented
citations, and a sparse citation list is indistinguishable from a sparse
answer. The prefetch already makes the point, feeding explicit conclusions
into the prompt without their IDs, so the model could not cite them if asked.
An accumulator is threaded from the router through the agent into ToolContext,
and read handlers hand it the rows they already loaded. Nothing is re-queried
when the response is built, so evidence inherits the scoping of the reads that
produced it and cannot become a way around a session allowlist.
Three details worth knowing:
- Prefetched conclusions never pass through the tool executor, so they are
captured via a new `documents_out` sink on `search_memory`. On a query that
answers without a tool call they are the whole of what was read.
- Conclusions dedupe by ID. `Representation`'s own deduplication keys on
content and timestamp and ignores IDs, which would collapse distinct
conclusions that happen to read alike.
- Conclusion timestamps are re-stamped UTC. `Representation` strips tzinfo so
observations render compactly into prompts, which would otherwise put naive
timestamps in the API beside timezone-aware message ones.
The two chat routes had byte-identical nested SSE formatters; they now share
one helper, so the terminal event carries evidence on both.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(mcp): add stdio host for local clients
* feat(mcp): add Streamable HTTP host and image
Long-lived HTTP entry for Docker and other process hosts, reusing
createServer(). Dedicated mcp/Dockerfile; compose service beside api.
* fix(mcp): stdio launcher cwd/silent and HTTP session bounds
Pin bun --cwd so bunfig loads. Silence bun run. Require Bearer on HTTP.
Idle-expire and cap in-memory MCP sessions.
* fix(mcp): re-check bearer on established HTTP sessions
Session lookup returned early without Authorization, so a missing or
wrong token still 200'd after initialize. Bind each session to the
init key and 401 on mismatch.
* fix: nit cleaning claude command
---------
Co-authored-by: ajspig <dragon@monstercode.com>
* chore: scaffold @honcho-ai/harness-core
* feat(harness-core): resolve shared root config
* feat(harness-core): send client identity headers on SDK requests
* feat(harness-core): drop cloud vs custom api header
* feat(harness-core): migrating v0 config to schema v1 on read
* chore(harness-core): clean up
* feat(config): describe oauth and host overrides in the v1 schema
* chore: rename to harness-plugin-core
* feat(harness-plugin-core): update telemetry headers on a live client.
The workspace agent's prefetch is an orientation overview — scale, active
peers, their cards — not the corpus. `low` is the only reasoning level that
explicitly sets TOOL_CHOICE="auto", so the model was free to skip tools
entirely, and it did: every workspace_chat call in CI run 33662772219 made
zero tool calls. It answered when the overview happened to carry the fact and
otherwise wrote out the search it should have run, then asked the caller which
option to take — at an endpoint with no caller to answer.
Add a `_tool_choice` seam alongside `_select_tools` and override it on
WorkspaceDialecticAgent to require a tool call. `execute_tool_loop` already
relaxes "required"/"any" to "auto" after the first iteration, so this costs one
search round rather than pinning the loop, and the model can still stop and
synthesize. Any value a level configures other than None/"auto" passes through.
The pair agent is unaffected: it prefetches the observations for its query and
can legitimately answer from context alone.
Also tell the workspace prompt it is non-interactive. It had "Do not narrate
tool use" but never said the caller cannot reply, and three of the five traced
responses ended in a menu of lookups.
Unified subset goes 1/5 -> 5/5, and search_memory — the recall path that never
once ran — now fires on 6 of 7 workspace queries. workspace_chat_scope is the
notable one: its two not_contains assertions were passing vacuously because
nothing was ever retrieved, and it now recalls the in-scope fact while still
excluding the out-of-scope vault code.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(docker): make API worker count configurable
Add API_WORKERS with a single-worker default and document database pool sizing.
Refs #1063
* fix(docker): address API worker review feedback
* chore(docs): Add section about harness integrations and deepseek harness to docs
* chore: Add section about harness integrations and deepseek harness to docs
* fix(deriver): eliminate create_documents deadlock and stop silently burning batches on transient errors
Two concurrent work units writing the same (workspace, observer, observed)
collection deadlocked on times_derived reinforcement UPDATEs issued in
batch order (DEV-1975, 682 events in 90 days). The deadlock was swallowed
per-document, the loop cascaded PendingRollbackErrors against the dead
session, the whole batch was lost, and the queue item was marked processed.
- serialize writers per collection with a transaction-scoped advisory lock
(pg_advisory_xact_lock + SET LOCAL lock_timeout), skipped for insert-only
batches; covers all three row-lock sites in one move
- hoist external-vector-store dup-candidate resolution ahead of the first
DB statement so the lock's critical section contains no network calls
- abort the batch on SQLAlchemyError instead of continuing through an
aborted transaction; per-document skip semantics kept for non-DB errors
- classify transient errors (new src/utils/retryable_errors.py) and retry
them via a bounded in-process counter instead of marking items errored
* fix(deriver): replace create_documents advisory lock with id-ordered row locks
Advisory locks are database-scoped and would serialize every writer to a
collection, including across Groudon tenants that share names. Collect
reinforcement and replace ops during the loop, lock target rows with
SELECT ... ORDER BY id FOR UPDATE, then apply. populate_existing reloads
times_derived so a prefetched identity-map row cannot lose a concurrent
increment.
* fix(deriver): harden create_documents candidate hoist and test isolation
Skip empty embeddings on the external-store path, isolate per-document
resolve failures, and keep replacement times_derived in the in-batch
ledger. Patch get_external_vector_store in the hoist test and cover
in-loop SQLAlchemyError abort.
* fix(deriver): address CodeRabbit findings on create_documents deadlock fix
- Distinguish external resolve failure ([] skip) from pgvector fallback (None)
so _semantic_dup_decision never re-enters external I/O under an open session
- Bound external candidate hoist concurrency with a semaphore
- Map in-loop IntegrityError to ValidationException for a uniform contract
- Persist transient retry attempts on the oldest unprocessed queue item so
every deriver instance shares one MAX_RETRYABLE_ATTEMPTS budget
- Cover resolve-failure skip and multi-manager reclaim of the retry budget
* fix(deriver): harden retry metadata cleanup and stale reinforce fallback
- Strip _retry_attempts from payloads in the same transaction as
mark_queue_items_as_processed / mark_queue_item_as_errored
- Clear shared retry metadata only after a successful terminal mark
- On reinforce, if the locked target is gone or soft-deleted, insert the
incoming document instead of dropping it
- Skip pgvector semantic lookup when embedding is empty so query_documents
cannot embed under an open session
* fix(deriver): address review on deadlock retry and row-lock apply
Strip _retry_attempts before payload validation so non-representation
tasks are not burned as extra_forbidden. Re-raise retryable observer
save errors after telemetry so the queue actually retries. Skip
same-batch reinforce fallbacks after a replace. Revert unordered
FOR UPDATE on mark processed/errored and drop post-commit retry
cleanup from the success path.
* fix: add test and simplify queue query
---------
Co-authored-by: Vineeth Voruganti <13438633+VVoruganti@users.noreply.github.com>
* fix(deriver): chunk scope backfill so large sessions don't OOM the worker
_run_backfill embedded, wrote, and synced every planned copy at once, holding
one Python float list per document. A 14k-document session is ~580MB of
vectors alone, and several backfills run concurrently, which OOM-killed the
deriver at its 1000Mi limit and crash-looped it since the work units never
completed. Phases 2-4 now run per chunk of 500 specs and drop each chunk's
embeddings once synced.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(deriver): hydrate backfill embeddings per chunk
Phase 1 no longer materializes every source embedding into plans.
load_only skips the vector column on the plan queries, and each chunk
reloads only its source embeddings before embed/write/sync.
* fix(deriver): lock scope membership across backfill chunk writes
SELECT ... FOR UPDATE on the active SessionPeer row so a concurrent
leave cannot commit between the membership check and the copy inserts.
Adds a concurrency test that asserts the leave blocks until commit.
* fix: add test for memory bound
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Vineeth Voruganti <13438633+VVoruganti@users.noreply.github.com>
* fix(deriver): strip NUL bytes from model-generated observations
Postgres rejects NUL (0x00) in text columns and in jsonb strings. API
ingress has always stripped it from user-supplied content, but the
deriver's own output did not go through any equivalent: a model can emit
a \u0000 escape in its tool-call arguments, which the JSON parser decodes
into a real NUL byte. Seen in production when models transcribe shell
output (`tr '\x00' '\n'`) or Windows paths (`c:\<NUL>users\amal`).
The NUL reached the exact-content dedup pre-fetch in create_documents as
a bind parameter, so the query raised DataError before any row was
written and the whole batch for that observer was dropped.
Strip in _normalized_observation and _normalized_observation_input --
the points that already normalize text for persistence and embedding --
so the embedded text matches the stored text. premises and sources are
covered too, since they ride along in internal_metadata. The emptiness
check now runs after normalization, because str.strip() does not remove
NUL and all-NUL content would otherwise be stored as an empty string.
DocumentCreate.content gets a mode="before" validator as a backstop for
callers that bypass those paths; running before the length constraint
makes all-NUL content fail min_length rather than silently empty out.
The NUL helpers move out of schemas/api.py into utils/sanitization.py as
a single recursive strip_nul, so ingress and internal paths share one
implementation. It is overloaded to keep str -> str for the callers that
chain .strip(), and passes None through so optional fields need no guard.
Fixes HONCHO-4XZ
* fix: broaden nul strip check
* chore: code simplification
---------
Co-authored-by: Vineeth Voruganti <13438633+VVoruganti@users.noreply.github.com>
* chore(docs): Add Documentation for Scopes
* chore(docs): add scopes to API reference, architecture, and design patterns
- Add the seven /scopes routes and their schemas to openapi.json, plus the
scope/kind fields on chat, representation, session-create, and peer-list
schemas; generate the endpoint pages and register a scopes nav group
- Add a Scopes subsection and diagram node to the architecture data model
- Add scope guidance to design patterns: quick-reference rows, an isolation
boundary comparison (workspace / scope / session allowlist), and common
mistakes (scope-per-reader, scopes-as-access-control)
- Replace the "Underneath the Facade" section in scopes.mdx with behavioral
guardrails and pointers to the implementation source
* chore(docs): tighten scopes doc to decision-level detail
- Drop the recall-resolution diagram (restated the Two Arms table)
- Replace the enumerated Rules table with prose; caps and error shapes
now live in the API reference schema descriptions
- Trim backfill/removal internals to observable behavior and note that a
backfilled scope deepens through subsequent dreams
* chore(docs): reserve "scope" for the scopes feature
Using it as a verb for session design, recall filters, and CLI targeting
collides with the named-session-set feature.
* chore(docs): clarify the scopes page and document create/status responses
The page now leads with projection rather than partition and points at the
scopes API; OpenAPI declares the 201/409/404 those routes actually return.
* chore(docs): fix broken anchor and core-concepts link
The rebase reintroduced a link to a renamed anchor in scopes.mdx, and
unified-memory-setup pointed at /core-concepts/, which has no index page.
* chore(docs): correct scope arms, listing, and read-surface pointers
The Accepts row mixed named-scope with the allowlist arm, kind=scope on
the peers list does not return facade ids, and chat/context/search never
mentioned scope=.
* chore(docs): drop the 1k-token session batching narrative
Reasoning no longer waits on a per-session token threshold, so product
docs should not tell people to size sessions around that gate.
* chore: minor fix
If consent is withdrawn while array.js downloads, sync() runs before window.posthog exists and the opt-out is skipped. onload now re-checks granted() and resets loaded so a later re-grant retries.
Mintlify swaps pages without a reload, so capture_pageview: 'history_change' records navigation past the landing page.
* fix: stop top_k=0 from reaching Turbopuffer on message search
HONCHO-19Q: dreamer search_messages passed LLM limit=0 through to
Turbopuffer (top_k must be 1..10000). #970 guarded documents; this
closes the message path and floors tool limits at 1.
* fix: preserve pgvector None sentinel on zero top_k
query_external_vector_document_ids must return None when on the
pgvector path before applying the top_k<=0 empty-list guard.
The gate job only runs on `pull_request: labeled`, so it is skipped on
push. A skipped ancestor propagates down the needs chain unless a job
opts out, which `unified-tests` never did — so the suite has been
skipped on every merge to main while still burning a Fly machine.
* feat(cli): fix API key typing
Co-authored-by: Cursor <cursoragent@cursor.com>
* chore: bump honcho-cli to 0.1.4
Ship the masked --setup API key prompt plus the openai-compatible embedding base URL already on main.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(cli): check for newer version
* docs: nit
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(llm): preserve null content on OpenAI tool-call turns
OpenAI-compatible providers can return assistant tool-call messages with
content=null. Coercing that to "" before history replay breaks providers
that bind opaque reasoning state to the exact assistant message shape.
Keep null only when the normalized response has tool calls; tool-less
null still becomes "", and content_override stays authoritative.
Fixes#1061
* test(live_llm): cover OpenAI null content tool-call replay
Add a live multi-turn tool replay that asserts provider content=null stays
null through normalize + OpenAIHistoryAdapter and that the continuation
still answers. Mark gpt_4/gpt_5 families as supports_tool_replay.
* docs(llm): note content_override None sentinel semantics
None means no override, not force-null content. Addresses review on #1064.
author_association on the webhook is CONTRIBUTOR when org membership is
private, so maintainers with write (e.g. ajspig) were labelled
needs-approved-issue. Skip on admin/maintain/write from
getCollaboratorPermissionLevel instead; 404 stays gated.
* docs: make honcho start the documented local path
honcho-cli 0.1.3 can run a personal stack without cloning the repo; point the README, self-hosting, and CLI reference at that, and drop the community installer callouts.
Co-authored-by: Cursor <cursoragent@cursor.com>
* docs: updating with new CLI language
* docs: drop compatibility-guide changes from this PR
Leave that file on main; CLI version cards are updated at release time.
Co-authored-by: Cursor <cursoragent@cursor.com>
* docs: small language changes
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Review findings on #1071. The cookie match now requires the exact
CookieConsent name boundary. A failed array.js request resets the
loaded flag so later consent events retry. And consent events now run
a full sync: withdrawal opts an already running instance out, and a
re-grant opts it back in — same behavior as the landing site's gate.
The runtime stage copies application code but not pyproject.toml, so
src/_version.py cannot find the file it reads the version from. The
image also installs dependencies with --no-install-project, so there is
no honcho distribution for the importlib.metadata fallback to find.
Both lookups fail, so the service falls back to reporting its version as
"unknown" in the OpenAPI schema and in telemetry events.
Copying the file into the runtime stage restores an accurate version.
The file is under 4 KB, so the image size is unchanged.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
DEV-2465 open question 5, option d. Mintlify's built-in integration
loaded PostHog unconditionally on all 249 docs pages — a visitor who
declined on the homepage was tracked one click later in the docs. The
integration key comes out of docs.json; docs/posthog-consent.js loads
PostHog directly instead, only when the CookieConsent cookie grants
Statistics (or holds Cookiebot's -1 marker), and listens for the
consent events so a grant on the docs banner itself loads it too.
Trade recorded on the ticket: this bypasses the ph.mintlify.com proxy,
so ad blockers reduce docs PostHog volume. Verify after deploy that
Mintlify's page CSP allows us-assets.i.posthog.com; if it blocks,
fall back to option c.
DEV-2465 step 1. Mintlify injects gtm.js on all docs pages; the
container is audited to be inert on /docs before this merges, so the
snippet loads and nothing fires. Cookiebot and GA4 arrive later as
container publishes, consent first.
Merging this publishes the docs within minutes, so it stays unmerged
until Marc confirms the container audit.
`get_observation_context` resolved scope by fetching every session name the
observer has a membership record in, then expanding that list into
`session_name IN (...)` twice in one statement — once in the CTE and once in
the outer select. That puts psycopg's 65535-bind-parameter ceiling at roughly
32,765 sessions, and the count only ever grows: the loose membership
definition (`active_only=False`) counts sessions the peer has since left, so
leaving a session does not shrink the scope. A workspace with tens of
thousands of sessions for one peer produced a statement the driver could not
serialize at all.
Two new helpers in `crud.message` express the observer half as a correlated
EXISTS over `session_peers`. Scope now costs two bind parameters regardless of
membership size, and the membership query disappears (two round trips become
one). The `session_peers` primary key is `(workspace_name, session_name,
peer_name)`, so the correlated probe is an exact-match index hit.
The caller-supplied allowlist stays an IN clause — it is route-capped at 1000
entries and carries none of the unbounded-growth risk. `resolve_session_scope`
is left in place: three other callers still need the materialized list,
including `_search_messages_external`, which sends session names to the vector
store as a filter payload and cannot take SQL.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
On Redis Cluster a key's slot comes from the substring inside the first
{...}, when one is present. Untagged, one deployment's keys spread over
every slot, so its client opens and holds a connection to every node in
the cluster. Wrapping the namespace in braces puts them all on one slot,
and therefore one node, cutting each deployment's connection count to
the cluster by a factor of the shard count. Namespaces still hash
independently of each other, so keys stay spread across the cluster and
no shard becomes a hotspot.
The tag needs two spellings, because the two ways a key gets built treat
the string differently. cashews runs `prefix=` through format
substitution, so braces have to be doubled there to survive as literals;
keys built by concatenation need them single. A single brace passed to
cashews is read as an empty substitution field and the namespace is
dropped entirely, which would let two deployments collide on one key --
hence two clearly named helpers rather than one string, and a test that
the two paths produce identical bytes.
No key format change for a non-cluster backend, and no migration: the
old keys simply age out by TTL.
* feat(cli): add honcho start/stop/status for a local Docker stack
* feat(cli): fix status command
* feat(cli): improving how we pull docker images and writing a config,toml
* feat(cli): add honcho start --setup wizard for local stack config
* feat(cli): cleaning up unnecessary func, and error throwing
* feat(cli): minor clean up in stack.py
* feat(cli): read setup wizard defaults from the image config.toml
* feat(cli): cleaning up unused commands
* feat(cli): adding ignored docker-compose.yml
* feat(cli): forward host LLM env into honcho start
* feat(cli): share start/stop progress helpers via output.py and cleaning up language
Teach both agents what Honcho, peers, and the harness are instead of comparing them to each other. Render only the tools the request actually offers, and drop the pair prompt's call to a write tool that is not in the loadout.
* fix(crud): preserve joined_at for active session peers
Re-adding an already-active peer no longer advances the membership
window, so peer_perspective search keeps messages from the original
join. Genuine rejoins still start a new window.
* docs: document set_peers membership window and wrap test docstrings
* fix: preserve session observer limit
---------
Co-authored-by: Vineeth Voruganti <13438633+VVoruganti@users.noreply.github.com>
Callers wrapped every ValueError from the embedding client in a
"exceeds maximum token limit" message, so provider and configuration
failures (dimension mismatch, empty response, upstream error) surfaced
to users as though their input were too long.
Add EmbeddingTokenLimitError, raised only by the pre-flight token checks
in embed() and simple_batch_embed(), and narrow the remaps in search.py,
agent_tools.py, document.py and representation.py to catch it. It
subclasses ValueError so existing broad handlers keep working.
Both simple_batch_embed() remap sites pass on_oversize="truncate" and so
could never raise a token-limit error at all; their handlers only ever
mislabelled provider failures.
Fixes#568
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* chore: update issue templates and docs
* chore: slim bug/quality forms and add integration template
Drop high-friction required fields from bug and quality issue forms.
Add an integration-request form (app stores, plugins, frameworks) labeled
integration, and point contributing docs at it.
* chore: route security mail to support@ and polish docs intake
Use support@honcho.dev for private vulnerability email. List the
documentation template in contributing guides, rename Media prove,
and add public-issue redaction/security redirects on the docs form.
* fix: address render issue in templates and add version field
* feat(docs): Initial draft of new contributing policies
* feat(ci): defer issue-gate closes to a scheduled sweeper
Addresses review feedback on #1041.
The gate now reads GitHub's resolved closing references
(closingIssuesReferences) instead of regex-parsing the pull request body,
so an issue linked through the sidebar Development panel counts, and a
bare `#123` mention no longer does.
It also no longer closes on the pull request event. It labels and
explains; pr-sweeper.yml re-checks every six hours and closes only what is
still failing 72 hours after the notice. That re-check is load-bearing:
linking an issue via the sidebar fires no webhook, so an event-only gate
could never observe a contributor complying that way. The sweeper also
closes drafts from outside the org after 30 days.
The shared check lives in .github/scripts/issue-gate.js so both workflows
run identical logic, with a dependency-free self-check wired into static
analysis. Its one regression guard: author_association CONTRIBUTOR stays
gated, since GitHub assigns it to anyone who has previously committed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore(codeowners): drop the third reviewer from most areas
Discussed with @akattelu. Also reassigns SECURITY.md to @Rajat-Ahuja1997
and strips trailing whitespace from the deployment block.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs(v2): port the issue gate policy into the v2 contributing guide
The v2 guide is still published (v2.5.1 in docs.json) but carried no
mention of the issue gate, so a contributor reading it would not learn
that a pull request needs an approved issue until the bot labelled theirs.
Ports the policy, both linking routes, and the gate's place among the
automated checks, keeping the v2 guide's own structure and unwrapped
prose rather than importing the v3 rewrite wholesale.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(ci): count only bot-authored gate notices, share the exemption list
Two review findings on the issue gate, with a common root cause.
MARKER is an invisible HTML comment, so anyone who can comment on a
public repository can paste it. findNotices accepted any comment
containing it, so a third party could post one on someone else's pull
request: runGate posts a notice only when none exists, so the author
would never be told, and runSweep would then measure the 72-hour grace
window from the stranger's timestamp and close them unwarned. Notices now
require bot authorship.
The stale-draft sweep re-listed the gate's exemptions and had lost the
bot case, so a bot's long-lived draft was closable despite checkGate
exempting bots. Both callers now share one exemptReason(pr) rather than
keeping parallel lists that drift.
Not changed: closingIssuesReferences(first: 20) truncation. It needs a
pull request with 21+ closing references where only a later one carries
the label, and the outcome would be a label plus the grace window, not a
close.
Coverage goes 11 -> 20 cases, including the stale-draft close path, which
had none. Both fixes were confirmed to fail their tests when reverted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Aakash Kattelu <aakash@plasticlabs.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore: update issue templates and docs
* chore: slim bug/quality forms and add integration template
Drop high-friction required fields from bug and quality issue forms.
Add an integration-request form (app stores, plugins, frameworks) labeled
integration, and point contributing docs at it.
* chore: route security mail to support@ and polish docs intake
Use support@honcho.dev for private vulnerability email. List the
documentation template in contributing guides, rename Media prove,
and add public-issue redaction/security redirects on the docs form.
* fix: address render issue in templates and add version field
---------
Co-authored-by: Vineeth Voruganti <13438633+VVoruganti@users.noreply.github.com>
* Add workspace-level chat (DEV-1326)
POST /v3/workspaces/{workspace_id}/chat: agentic dialectic over the whole
workspace instead of a single (observer, observed) pair. Salvaged from
plastic-labs/honcho#373 and re-grown on today's DialecticAgent:
- WorkspaceDialecticAgent subclasses DialecticAgent via four new seams
(_get_tools, _create_tool_executor, _prefetch_intro, _trace_name) instead
of a base-class extraction; observer/observed use empty-string sentinels.
- Routing-accelerated prefetch: workspace stats + top-5 active peers with
their self peer-cards (pure DB, ~7ms measured) so routing-obvious queries
resolve without a discovery tool round.
- Observation search stays pair-scoped (matches per-pair vector namespaces;
avoids workspace-flat top-k dilution): search_memory/get_peer_card take
observer/observed as tool arguments, with pair attribution in results.
- workspace_chat / workspace_chat_stream orchestrators, WorkspaceChatOptions
schema (scope param seam left for the #897 scopes facade), SSE streaming,
structured output via response_format.
- crud: get_workspace_stats, get_active_peers; format_documents_with_attribution.
- SDKs: Python Honcho.chat/chat_stream + HonchoAio mirrors; TypeScript
honcho.chat/chatStream.
- 46 tests (route, orchestrator preflight, tool handlers, executor routing,
attribution formatting) + unified test cases + docs.
Co-Authored-By: doria <93405247+dr-frmr@users.noreply.github.com>
Co-Authored-By: Benjamin McCormick <docterformer@protonmail.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: type SSE stream wrapper as AsyncIterator (basedpyright)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: silence unused db_session fixture warnings (basedpyright failOnWarnings)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: drop docs changes from this PR (defer to follow-up)
Restores docs/v3/documentation/features/chat.mdx to main's version. This
also puts back the peer-chat Structured Outputs section (#896) that the
workspace-chat commit removed as a rebase artifact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: workspace message tools deny-all under rebased session scoping
The #882 rebase changed the unscoped-observer contract from falsy to
'observer is None': resolve_session_scope looked up the workspace
executor's observer='' sentinel as a real peer with no session
memberships and denied every workspace-flat message read (search, grep,
date-range, temporal, observation context) whenever no session was
pinned — the primary workspace-chat shape. Normalize the sentinel to
None at the five read-handler crud boundaries and add regression tests
that run the tools unpinned (verified to fail without the fix).
Also from review:
- wrap the workspace prefetch in the same degrade-to-None protection
the base agent has (an overview query error no longer 500s the
request or kills the SSE stream after headers)
- thread session_allowlist through create_workspace_tool_executor so
the agent-level allowlist seam is honored end to end when scopes
(#897) wire it up; allowlisted grep is covered by a test
- deterministic name tie-break in get_active_peers ordering
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* review: SDK response_format parity, shared query sanitizer, annotations
- TS SDK: WorkspaceChatParams gains response_format; _workspaceChat/
_workspaceChatStream consume the shared interface instead of inline
duplicates; chat/chatStream expose responseFormat.
- Consolidate the three identical sanitize_query validators into one
NulStripped annotation.
- workspace_chat_stream: return annotation + full docstring.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* review: fold active peers into workspace stats; trace + query bounds
- Merge get_active_peers into get_workspace_stats (one discovery round
instead of two); minimal loadout keeps a discovery tool via the merged
stats tool. Fixed top-10 by recent activity; deeper discovery routes
through search_messages.
- get_active_peers CRUD now aggregates over a trailing 90-day window so
the chat-path prefetch never scans a workspace's full message history.
- Workspace agent inherits the "dialectic_chat" trace name; scope stays
distinguished by agent_type/track_name (workspace name was already in
telemetry context).
- Prefetch failure logs carry workspace + traceback; prompt no longer
contrasts against a peer-level agent the model has no concept of;
drop ticket identifiers from comments.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: add `scope` to workspace chat and exclude scope peers from stats
Workspace chat is peer-unanchored, so `scope` is always a session-union
allowlist (single name or list), fail-closed when empty. Stats and
active-peer prefetch drop scope-kind peers and honor the same allowlist.
* test: teach the unified runner `workspace_chat` and parse every case
QueryAction now accepts target=workspace_chat (SDK path, including
scope). A pytest over tests/unified/test_cases/*.json keeps the four
existing workspace-chat cases — and a new scoped one — from rotting
against the schema again.
* docs: tighten workspace-chat scope docs and judge prompt
Scoped workspace_chat uses the SDK, not raw HTTP. The scope fixture's
judge now requires the in-scope tea fact, not merely the absence of the
leak. format_sse_stream matches the peer-chat one-liner.
* fix(dialectic): restore the empty-memory fallback for workspace chat
`search_memory` auto-searches messages when a pair has no observations,
but the gate only admitted `agent_type == "dialectic"`. The workspace
executor passes `workspace_dialectic`, so workspace chat got a bare
"No observations found" and answered that it knew nothing rather than
falling through to message search.
Also fixes the two unified cases that never ran: `deriver` is not a
field on `WorkspaceConfiguration`, so both aborted at load with
`extra_forbidden`. `workspace_chat_scope` additionally enables reasoning,
since it asserts scope isolation and has no reason to depend on the
fallback path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(tests/unified): fail CI when unified tests fail
`runner.run()` tallied failures into `failed_count` and printed them, but
returned nothing, and both entrypoints ignored the result. The workflow
invokes `python -m tests.unified.run` bare, so the job has gone green on
failing and unrunnable cases since it was wired up in #291.
Return the count and exit non-zero on it. `INVALID SCHEMA` already counts
toward the tally, so a malformed case now fails the job instead of being
skipped silently.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(unified): assert scope peers stay out of workspace chat answers
Scope peers are real peer rows, so a regression in the `scope_peer_clause`
exclusion would surface `scope.therapy` through workspace stats or the
routing prefetch. Nothing asserted against that.
Adds the check to the existing scoped query and a new unscoped one, since
the two exercise different `get_active_peers` branches. Verified by
removing the exclusion, which fails the unscoped query.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix: Remove dead code references
---------
Co-authored-by: doria <93405247+dr-frmr@users.noreply.github.com>
Co-authored-by: Benjamin McCormick <docterformer@protonmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Aakash Kattelu <aakash@plasticlabs.ai>
Co-authored-by: Vineeth Voruganti <13438633+VVoruganti@users.noreply.github.com>
Bug (verified vs SQLAlchemy 2.0.49): on GC-cleanup of an abandoned async connection, _finalize_fairy routes through fairy.detach(), which nulls the record's dbapi_connection so NullPool's close is a no-op (the `close` event never fires) then emits `detach` with the record. Listening only to close/invalidate left the marker unpopped, so db_connections_open leaked upward and never reset until restart. Listen to `detach` too — it carries the ConnectionRecord and the marker dedupes, so exactly one decrement occurs.
Also: strip a bare PR-number provenance tag from the test docstring (plastic-labs comment-reconciliation rule); disambiguate db_connections_open from the existing db_pool_connections; note in initialize_bounded_metrics that DB-instrumentation metrics zero-init in their registrar.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add db_connections_open (gauge) and db_connections_established (counter), driven by SQLAlchemy connection-lifecycle events so they report real connections under every pool class — including NullPool, where the pool-object collector (db_pool_connections) reads zero. Under NullPool the establishment rate approximates request rate.
DBConnectionTracker mirrors DBQueryInflightTracker: a ConnectionRecord.info marker makes each physical connection increment once and decrement at most once (no leak, no negative). Registered per-process in the API and deriver; zero-init via the pre-resolved labeled children, matching db_queries_in_flight_gauge.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(filter): make ne on jsonb metadata keys null-safe
* test(filter): cover null-safe ne on nested metadata keys
* test(filter): count actual rows for nested-metadata ne null-safety
Compile-only checks lock the operator map entry but don't catch wrong
row sets under three-valued logic. Adds a live messages/list case with
a message missing the key and one with empty metadata, following the
scalar-column pattern in test_negation_includes_conclusions_with_no_session.
* test(filter): type message_configs with a TypedDict
basedpyright couldn't narrow the heterogeneous metadata dict literals,
so indexing message_configs["content"] came back partially unknown and
broke the sorted() calls under type checking.
* fix(deriver): truncate oversize observations so one cannot drop the batch
simple_batch_embed raised ValueError when any input exceeded the per-input
token cap, which failed the entire deriver save when a single observation
was over-length. Add on_oversize="truncate": oversize inputs are embedded
from a token-capped prefix (re-encoded until it fits, with a warning),
preserving one vector per input. Default stays "raise" so existing callers
are unchanged. RepresentationManager opts into truncate.
Also add a live embedding test that fails on main (raise / missing kwarg)
and passes once a mixed short+oversize batch survives.
Refs #569
* fix(deriver): surface failure when all observer saves fail
When every observer's save_representation failed (e.g. embedding retries
exhausted under a sustained 429), the deriver logged the error and returned
normally, so the queue marked the work unit processed with zero documents
saved. Collect per-observer errors and, after telemetry is emitted, raise
RepresentationSaveError when no observer succeeded. Partial failures stay
processed (saved observers must not be discarded) and are recorded via an
additive failed_observer_count on RepresentationCompletedEvent.
Refs #728
* fix(embedding): guarantee truncation progress and truncate on re-embed
The retry slice in _truncate_to_token_limit always recomputed the same
keep count, so a slice whose re-encode grew past the cap could oscillate.
Decrement keep after each unsuccessful retry.
Document re-embed in the reconciler used the default on_oversize="raise",
so one oversize document failed every other document in the batch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* chore: drop ticket ids and shrink comments to one sentence
Comments and docstrings describe current behavior, not the PR that
introduced them. Ticket numbers stay in the commit/PR.
* chore: annotate RepresentationSaveError and assert truncate on re-embed
* fix(embedding): truncate on conclusion create paths and document BPE loop
Storage callers in create_observations (API + agent tools) now pass
on_oversize="truncate" so a single oversize item cannot drop the batch.
Docstring on _truncate_to_token_limit notes why decode/re-encode is load-bearing.
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* telemetry: materialize dropped-event counter children at 0
A labeled Prometheus counter exports no series until its first labels()
call, so telemetry_events_dropped stayed invisible until an event was
actually dropped — impossible to alert on or graph, and "no drops" was
indistinguishable from "metric missing / scrape broken".
Pre-create the (namespace, reason) children at 0 on emitter start, for
each reason the emitter can emit, so the metric is always present.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* telemetry: generalize counter zero-init to all bounded-label counters
Extends #927 (which zero-inited telemetry_events_dropped) to every counter
whose label domain is bounded and known at startup, so metrics are present in
Prometheus before their first event — a missing series then signals a broken
scrape rather than "nothing happened yet".
- add initialize_bounded_metrics(instance_type) on PrometheusMetrics; call it
per-process from main.py (api) and deriver/__main__.py (deriver).
- extract a shared _touch() helper; refactor initialize_telemetry_dropped_metrics
onto it (that one stays per-emitter in start() — it's prefix-dependent).
- explicit ALL_EVENT_TYPES / HIGH_VOLUME_EVENT_TYPES registry in telemetry.events,
drift-guarded by tests that walk BaseEvent subclasses.
- only VALID (task_type, token_type, component) tuples for deriver_tokens (the
cartesian product would fabricate impossible always-0 series); only high-volume
event types for sampled_out; high-cardinality labels (endpoint, workspace_name)
left open.
- gauges: zero-init embed_now_tasks_in_flight + telemetry_buffer_size; add a new
message_embeddings_pending backlog gauge, set each reconciliation cycle and
zero-inited at deriver startup (Rajat's pending/in-flight ask).
- backfills the tests #927 shipped without.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* review: task-aware deriver combos + fail-soft gauge zero-init
I1: _DERIVER_TOKEN_COMBOS was factored task-independently, materializing the
impossible (ingestion, input, previous_summary) series — previous_summary is
summary-only. Make combos task-aware (_DERIVER_TOKEN_COMBOS_BY_TASK) so no
always-0 impossible series is fabricated, matching the PR's own goal. Tests
tightened to assert the ingestion/previous_summary series is absent.
I2: the three gauge .set(0) zero-inits were bare while the counter inits go
through the fail-soft _touch. Add _set_gauge_zero() so a gauge init can't
propagate an exception into process startup either.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(telemetry): isolate zero-init namespaces, add deriver-to-api guard
Global-REGISTRY assertions used a fixed "test" namespace, which several
other suites also pin, so another test's materialized children could
satisfy a presence assertion or break an absence one. Each test now runs
under a unique namespace resolved from settings at read time.
Adds the inverse per-process isolation test: deriver-only init must not
materialize API-only series (dialectic tokens, embed_now).
Co-Authored-By: Claude <noreply@anthropic.com>
* review: per-replica backlog gauge, drop duplicated constants and .meta refs
Addresses Vineeth's review on #927.
Blocking:
- message_embeddings_pending is a DB-global count, so drive it from
ReconcilerScheduler._scheduler_loop (runs on every replica, every
interval) instead of run_vector_reconciliation_cycle (runs off the
queue behind work-unit dedup, so one replica per cycle). Combined with
the zero-init, the old placement made every replica that never won the
work unit export a confident permanent 0. Help string now names the
owner so dashboards don't reach for sum().
- guard initialize_telemetry_dropped_metrics on METRICS.ENABLED,
matching its sibling initializer.
- drop the duplicate REASONING_LEVELS; import the one in src/config.
Non-blocking:
- walk BaseSpecialist recursively via a shared utils.types.walk_subclasses
(replaces the direct-children-only __subclasses__() and the test's
private copy of the same helper).
- derive the specialist assertion from the subclasses instead of
hardcoding two names — the hardcoded pair kept passing after
CardRefreshSpecialist landed, leaving it uncovered.
- inline the zero-init rationale and the multi-instance bucket taxonomy;
removes both pointers to a .meta design doc that is not in the repo.
Tests: new tests/reconciler/test_pending_backlog_gauge.py pins both
halves of the relocation (verified it fails when reverted).
* review: fix inert test guard, stale comments, and the REASONING_LEVELS drift claim
Second review pass on the branch. Findings, most severe first:
- tests/reconciler/test_pending_backlog_gauge.py: the _try_enqueue_task stub
was patched onto the class but declared without `self`, so calling it
raised TypeError — which _scheduler_loop swallows. The guard was inert and
the test passed for the wrong reason. Fixed the arity.
- metrics.py still commented that the backlog gauge is "set live each
reconciliation cycle". That is the exact claim the previous commit
overturned; it now contradicted the help string, the bucket-3 docstring
and sync_vectors.py.
- metrics.py claimed REASONING_LEVELS is "derived from the config Literal so
it never drifts", but config.py hand-listed it, so the earlier dedup had
quietly traded away the guarantee the original get_args() call provided.
Made it true instead: config.REASONING_LEVELS = list(get_args(...)), which
keeps the dedup and restores the invariant.
- dropped _set_gauge_zero: all three gauges it zeroed already have identical
fail-soft setters, so it was a second way to do one thing. Using the
setters also makes _handle_metric_error name the actual gauge.
- record_pending_embeddings_backlog's docstring oversold the covering index
as making the COUNT "negligible". The index makes cost proportional to the
pending backlog, not to the table — which is worst precisely when the
backlog matters. Stated honestly.
- _scheduler_loop's docstring said it only enqueues; it also refreshes the
gauge, at a cadence set by the shortest task interval.
- comment reconciliation: stripped #927 / "the generalization" temporal
anchoring, a CardRefreshSpecialist change-narration clause, and
reviewer-directed phrasing from the test file; disambiguated the
src/utils/summarizer.py path.
- CLAUDE.md had no Prometheus section at all, so the new "add a BaseEvent
subclass -> update ALL_EVENT_TYPES" obligation and the never-sum() rule
for non-additive gauges were undiscoverable from the architecture doc.
Verified: ruff + basedpyright clean (0 errors), tests/telemetry + reconciler
+ dialectic + llm 497 passed, full suite 1768 passed with only the 4
pre-existing test_document failures (OpenAI key required, reproduced on
clean origin/main). Re-confirmed the relocation guard fails when reverted.
* fix: silence the two basedpyright warnings inherited from main
CI runs `uv run basedpyright` bare, and basedpyright exits non-zero on any
warning — so these two have been failing the staticanalysis job on every
branch cut from current main, not just this one:
- src/vector_store/__init__.py:209 implicit string concatenation (#496)
- tests/test_cache_redaction.py:5 private import (#869)
Both predate this branch and are unrelated to the telemetry work; fixed
here only because they block this PR from going green. Verified: clean
origin/main also reports "0 errors, 2 warnings" and exits 1.
basedpyright now 0 errors, 0 warnings, exit 0.
* docs(telemetry): make the bucket-3 aggregation rule precise
The multi-instance taxonomy said a service-scoped non-additive metric has
"no aggregation correct once they disagree", then immediately mandated that
every instance refresh on its own timer. Those undercut each other: staggered
timers ALWAYS disagree slightly, so as written the rule reads as "ensure they
don't", which is unachievable, and it leaves the reader unsure whether max()
and avg() survived the fix.
The actual rule is bounded disagreement plus a scale-preserving aggregator.
Instances are N witnesses to one fact, not N parts of one whole, so sum() can
never be correct (it scales with replica count) while max()/avg()/quantiles
are correct precisely because the per-instance timer bounds the spread.
Wording only; no behavior change. The gauge help string already said
"max() or avg(), never sum()" — this makes the normative docstring agree
with it. Surfaced walking Vineeth's comment 3668208059 for comprehension.
* refactor(bench): import REASONING_LEVELS from config instead of re-listing
Third copy of the constant, missed when ee781c0/694e07f deduped the other
two. This one re-declared the ReasoningLevel Literal as well as the list,
so the type alias could diverge from config's with nothing to catch it —
and the list was hand-written, the variant that typechecks clean while
missing a member.
No import barrier justified it: this module already imports from src, as do
seven of its siblings in tests/bench. Concrete effect of the drift was that
a newly added sixth reasoning level would be rejected by the bench CLI's
argparse choices=.
src.config.REASONING_LEVELS is now the single definition repo-wide.
* test(telemetry): pin the METRICS.ENABLED guard on the per-emitter initializer
initialize_telemetry_dropped_metrics gained a METRICS.ENABLED guard in
ee781c0, addressing Vineeth's asymmetry comment, but nothing asserted it —
it had only the enabled half of the pair its sibling has. Deleting the guard
left the suite green, so the fix closed the asymmetry in the guards and
reproduced it one level up in the tests.
Mirrors test_init_noop_when_metrics_disabled. Verified live rather than
assumed: deleting the two guard lines turns this test red.
Uses a unique namespace, without which the absence assertion would be
satisfied by the enabled test's children rather than by the guard.
* docs(telemetry): fold zero-init why-prose behind # region ai markers
Comment/docstring-only pass over the changed files, per the groudon
comment-marker standard: the terse human-facing "what" stays visible, and
load-bearing "why" (the zero-init / absent-series-means-broken-scrape
rationale, gotchas, receipts) folds into # region ai / # ai: blocks.
Behavior-preserving: AST-identical modulo docstrings/comments vs the
pre-pass merge; ruff, ruff format --check, and basedpyright all clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: honor DERIVER_DEDUPLICATE in create_observations
The agent-tool path hardcoded deduplicate=True, so DERIVER_DEDUPLICATE=false
could not disable dedup for observations created through this path. Pass
settings.DERIVER.DEDUPLICATE, matching crud/representation.py.
* test: cover deduplicate setting is forwarded in create_observations
docker/entrypoint.sh runs under dash inside the container image. A Windows checkout with core.autocrlf=true rewrites it to CRLF, and dash aborts at startup because the carriage return becomes part of the -e flag argument.
Eugene Eisenstein
Aakash Kattelu
ajspig
steven-ji
Ulysse Pence
Vineeth Voruganti
Eri Barrett
Rajat Ahuja
Ken Weiner
Phil
adavyas
Serhii Zghama
Daniel Peng
Phil
ZiYun Liu