Skip to content

improvement(copilot): stop persisting tool-call result outputs in transcripts#4829

Merged
waleedlatif1 merged 1 commit into
stagingfrom
waleedlatif1/copilot-drop-dual-write-v1
Jun 1, 2026
Merged

improvement(copilot): stop persisting tool-call result outputs in transcripts#4829
waleedlatif1 merged 1 commit into
stagingfrom
waleedlatif1/copilot-drop-dual-write-v1

Conversation

@waleedlatif1
Copy link
Copy Markdown
Collaborator

@waleedlatif1 waleedlatif1 commented Jun 1, 2026

Problem

Opening a Mothership task (copilot chat) could take many seconds — sometimes effectively hang. Verified against prod: a single persisted assistant message in copilot_messages.content can reach hundreds of MB, almost entirely inside contentBlocks[].toolCall.result.output (e.g. a get_workflow_logs / run_workflow result). The DB query is ~2 ms; the cost is detoasting that payload, shipping ~200 MB to the browser, and parsing it. 449 prod rows exceed 1 MB, totaling ~2.2 GB.

Why dropping output is safe

These outputs are dead weight on the Sim side:

  • Not rendered — the chat thread shows only tool name/title/status (tool-call-item.tsx); the only component that renders result.output (generic-resource-content.tsx) is unwired.
  • Not replayed to the model — the outbound copilot request (payload.ts) sends only the new message + chatId + context, never the transcript. The upstream copilot service owns conversation memory and retains the full results.

Change

Drop toolCall.result.output before persisting, keeping result.success/error and the tool metadata (id/name/state/params/display).

  • stripToolResultOutput() in persisted-message.ts
  • applied in messages-store.ts toRow — the universal write choke point (assistant turns, update-messages, fork, import, inbox)
  • applied in loadCopilotChatMessages (lifecycle.ts) — pre-backfill rows return a small GET payload, so existing tasks load fast on read

Follow-up (separate, not in this PR)

A one-off backfill strips output from the 449 existing oversized rows to also reclaim the server-side detoast cost. Transform validated via dry-run on the worst row (1316→1316 blocks, 0 residual outputs); pending a write-capable DB credential.

Tests

Unit tests for stripToolResultOutput; write-path strip in messages-store.test.ts; read-path strip in lifecycle.test.ts. Biome + check-api-validation pass.

@vercel
Copy link
Copy Markdown

vercel Bot commented Jun 1, 2026

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment
Project Deployment Actions Updated (UTC)
docs Skipped Skipped Jun 1, 2026 6:14pm

Request Review

@cursor
Copy link
Copy Markdown

cursor Bot commented Jun 1, 2026

PR Summary

Medium Risk
Changes what is stored in chat transcripts (drops tool outputs permanently on write); behavior is intentional and outputs are unused in UI/model replay, but operators lose stored tool payloads until a separate DB backfill.

Overview
Copilot transcripts no longer store bulky toolCall.result.output payloads in copilot_messages.content, which were slowing Mothership task loads when tool results (e.g. workflow logs) reached hundreds of MB.

A new stripToolResultOutput() keeps tool metadata plus result.success / error and drops only output. It runs on write in messages-store (toRow, covering append/replace and related paths) and on read in loadCopilotChatMessages so pre-backfill rows still return small API payloads. Unit tests cover the helper, persist paths, and lifecycle reads.

Reviewed by Cursor Bugbot for commit 0ced63b. Configure here.

@greptile-apps
Copy link
Copy Markdown
Contributor

greptile-apps Bot commented Jun 1, 2026

Greptile Summary

This PR strips toolCall.result.output before persisting copilot messages to the database, fixing a serious performance issue where a single persisted assistant message could contain hundreds of MB of tool-call output that was bloating DB rows, causing slow task loads, and shipping unnecessary payload to clients.

  • Adds stripToolResultOutput() in persisted-message.ts that drops output from each PersistedContentBlock's tool-call result while keeping success and error, using an identity-return optimization when nothing needs stripping.
  • Applies the strip on both the write path (messages-store.ts toRow) and the read path (lifecycle.ts loadCopilotChatMessages) to handle pre-backfill rows already in the DB, with full unit test coverage for both paths.

Confidence Score: 5/5

Safe to merge — the change is narrowly scoped to dropping a field that is confirmed unused by the display layer and never replayed to the model.

The strip logic is null-safe, the identity-return optimization correctly handles the no-op case, and stripping is applied symmetrically on both write and read paths. Test coverage is thorough across unit, write-path, and read-path scenarios.

No files require special attention.

Important Files Changed

Filename Overview
apps/sim/lib/copilot/chat/persisted-message.ts Adds stripToolResultOutput(); logic is correct, null-safe (checks !result before typeof), returns original reference when unchanged.
apps/sim/lib/copilot/chat/messages-store.ts Applies stripToolResultOutput in toRow(), covering both appendCopilotChatMessages and replaceCopilotChatMessages write paths.
apps/sim/lib/copilot/chat/lifecycle.ts Applies stripToolResultOutput on the read path in loadCopilotChatMessages to handle pre-backfill bloated rows.
apps/sim/lib/copilot/chat/persisted-message.test.ts Adds comprehensive unit tests for stripToolResultOutput covering stripping, error preservation, identity return, and multi-block messages.
apps/sim/lib/copilot/chat/messages-store.test.ts Adds write-path strip tests for both append and replace operations, verifying output is absent from persisted rows.
apps/sim/lib/copilot/chat/lifecycle.test.ts Adds read-path strip test for getAccessibleCopilotChatWithMessages, confirming output is absent from returned messages.

Sequence Diagram

sequenceDiagram
    participant C as Copilot Stream
    participant PM as persisted-message.ts
    participant MS as messages-store.ts
    participant DB as copilot_messages (DB)
    participant LC as lifecycle.ts
    participant Client as Browser

    C->>PM: buildPersistedAssistantMessage(result)
    Note over PM: redactSensitiveContent + mapContentBlock<br/>(result.output still present)
    PM-->>MS: PersistedMessage (with output)

    MS->>PM: stripToolResultOutput(message)
    Note over PM: drops output, keeps success/error<br/>returns same ref if nothing to strip
    PM-->>MS: PersistedMessage (output stripped)

    MS->>DB: "INSERT content = stripped message"
    Note over DB: rows written before backfill<br/>still carry output

    Client->>LC: loadCopilotChatMessages(chatId)
    LC->>DB: SELECT content FROM copilot_messages
    DB-->>LC: rows (may have legacy output)
    LC->>PM: stripToolResultOutput(row.content)
    Note over PM: no-op for new rows (identity return)<br/>strips output for legacy rows
    PM-->>LC: PersistedMessage (output stripped)
    LC-->>Client: PersistedMessage[] (lean payload)
Loading

Reviews (2): Last reviewed commit: "improvement(copilot): stop persisting to..." | Re-trigger Greptile

…nscripts

Opening a Mothership task could take many seconds because a single persisted
assistant message in copilot_messages.content can reach hundreds of MB, almost
entirely inside contentBlocks[].toolCall.result.output (e.g. a get_workflow_logs
or run_workflow result). The DB query is ~2ms; the cost is detoasting that
payload, shipping it to the browser, and parsing it.

These outputs are dead weight on the Sim side: they are never rendered (the
thread shows only tool name/title/status) and never replayed to the model (the
upstream copilot service owns conversation memory). So drop result.output before
it is persisted, keeping result.success/error plus the tool metadata.

- add stripToolResultOutput() in persisted-message.ts
- apply it in messages-store toRow (covers every write path) and in
  loadCopilotChatMessages (existing rows render fast on read)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@waleedlatif1 waleedlatif1 force-pushed the waleedlatif1/copilot-drop-dual-write-v1 branch from 7f97352 to 0ced63b Compare June 1, 2026 18:14
@waleedlatif1
Copy link
Copy Markdown
Collaborator Author

@greptile

@waleedlatif1
Copy link
Copy Markdown
Collaborator Author

@cursor review

Copy link
Copy Markdown

@cursor cursor Bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

Reviewed by Cursor Bugbot for commit 0ced63b. Configure here.

@waleedlatif1 waleedlatif1 merged commit 3a2ebd1 into staging Jun 1, 2026
10 checks passed
@waleedlatif1 waleedlatif1 deleted the waleedlatif1/copilot-drop-dual-write-v1 branch June 1, 2026 18:20
waleedlatif1 added a commit that referenced this pull request Jun 1, 2026
* improvement(logs): object storage backed tracespans (#4787)

* improvement(logs): obj storage backed tracespans

* fix storage write context

* fix tests

* address comments

* address comments

* chore(db): remove migration 0219 to regenerate after staging merge

Drops the 0219_robust_shard SQL, its snapshot, and the journal entry so the
trace-spans/cost schema migration can be regenerated on top of the latest
staging migration chain (avoids a number collision with staging's migrations).

Co-authored-by: Cursor <cursoragent@cursor.com>

* improvement(billing): accurate per-member usage via shared ledger helper

Per-member/per-user usage in the org-member routes now adds the usage_log
ledger to the currentPeriodCost baseline (which is no longer incremented),
via a shared getOrgMemberLedgerByUser helper to avoid repeating the
subscription→period→ledger lookup across the admin and member-facing routes.

Co-authored-by: Cursor <cursoragent@cursor.com>

* regen migrations

* update migration

* address comments

* more code cleanup

* incorrect type cast

---------

Co-authored-by: Cursor <cursoragent@cursor.com>

* improvement(providers): harden OpenAI-compatible providers + add tests (#4796)

* improvement(providers): harden OpenAI-compatible providers + add tests

* fix(vllm): let tool-loop errors propagate instead of returning silent partial success

* fix(litellm): force tool_choice 'none' on final structured-output call

The deferred final call used tool_choice 'auto', so the model could emit
another tool_calls round instead of the structured answer, leaving content
stale. Use 'none' (matching vLLM/Fireworks) on both the streaming and
non-streaming final calls so the model must return the structured response.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(providers/ollama): drop tools from post-tool streaming call

Ollama ignores tool_choice (not in its supported fields), so vLLM/Fireworks'
tool_choice:'none' guard is a no-op here. Omit tools from the final streaming
payload instead so the summarization turn can't emit dropped tool calls.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(litellm): spread payload into deferred final call so reasoning_effort carries over

The non-streaming deferred finalPayload hand-picked fields and dropped
reasoning_effort (and any future payload field), diverging from the streaming
path which spreads ...payload. Spread payload here too for consistency.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* chore(providers/ollama): restore enrichment TSDoc block

Keeps parity with sibling Chat Completions providers (cerebras/mistral/xai).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(fireworks): restore TSDoc on utils helpers

Restore the TSDoc blocks on supportsNativeStructuredOutputs,
createReadableStreamFromOpenAIStream, and checkForForcedToolUsage —
TSDoc is the codebase documentation standard and should not have been
stripped.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* chore(litellm): remove inline rationale comments (codebase uses TSDoc)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* chore(providers/ollama): drop orphaned enrichment TSDoc

The block documented a function that now lives in trace-enrichment.ts, so it
documents nothing in this file.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* chore(copilot): deprecate mcp server (#4797)

* chore(copilot): deprecate mcp

* update error codes

* deprecate copilot api v1 route

* feat(integrations): hosted API keys for Findymail, Prospeo, and Wiza (#4777)

* feat(integrations): hosted API keys for Findymail, Prospeo, and Wiza

Add hosted-key support across all credit-consuming Findymail, Prospeo, and Wiza operations so Sim provides the key when a workspace has not brought its own. Register the three BYOK providers, consolidate Wiza's two-step reveal into a single polling wiza_individual_reveal op, and hide the API key field on hosted Sim for hosted operations.

* fix(integrations): harden Wiza reveal polling, soften enrichment getCost guards

Address Greptile + Cursor Bugbot review on #4777: return explicit failures from the Wiza individual_reveal poller instead of throwing (thrown errors were swallowed into a false queued success), short-circuit when the initial reveal is already terminal, tolerate transient 5xx/429 during polling, and return 0 (not throw) from Findymail getCost when the contacts/employees array is absent.

* chore(integrations): biome formatting after wiza merge resolution

* fix(wiza): type isTerminalReveal param structurally for next build typecheck

* feat(enrichments): add Findymail, Prospeo, Wiza to work-email waterfall

* feat(enrichments): add Wiza + Prospeo phone reveal to phone-number waterfall

* feat(enrichments): opportunistic identifiers + LinkedIn URL input across work-email & phone cascades

* fix(tables): reduce column header chevron size and fix sidebar shadow bleed (#4800)

* feat(slack): add install + privacy section to integration landing page (#4799)

* feat(slack): add install + privacy section to integration landing page

Adds a hand-authored, slug-keyed landing-content module (separate from the generated integrations.json so it survives regeneration) and renders an install walkthrough + privacy-policy link on integration pages when present. Also refreshes generated docs (data-enrichment entry, icon mappings, tool mdx).

* fix(landing): render privacy section independently, align CTA analytics label

* docs(landing): clarify the Slack install button is behind sign-in

* refactor(landing): bake integration landing content into generated json via docs-gen

Moves landing content (install walkthrough + privacy) out of a render-time augment and into the generation pipeline: generate-docs reads the pure-data content map and writes landingContent into integrations.json, so the page reads a single source (integration.landingContent). Canonical types live in integrations/data/types.ts.

* improvement(enrichments): align enrichments sidebar with design system (#4801)

* improvement(enrichments): align enrichments sidebar with design system

* fix(enrichments): consistent close button pattern and fix url link hover

* fix(misc): upgrade path change for new better-auth version, billing issue for workflow block agent usage (#4803)

* fix(misc): upgrade path change for new better-auth version, double-billing for workflow block agent usage

* fail loudly if stripe sub id missing

* fix(copilot): seq migration (#4804)

* chore(db): drop redundant idx_webhook_on_workflow_id_block_id index (#4809)

Removed because (workflow_id, block_id) is a left-prefix of idx_webhook_on_workflow_id_block_id_updated_at_desc, which fully covers it. The dropped index was non-unique and enforced no constraint.

* perf(copilot): read chat transcripts from copilot_messages (R+1 cutover) (#4808)

* perf(copilot): read chat transcripts from copilot_messages, not JSONB

Flip user-facing chat reads from the legacy copilot_chats.messages JSONB
array (5.7GB, 99% TOAST) to the normalized copilot_messages table via a
new loadCopilotChatMessages helper ordered by seq NULLS LAST, created_at,
id — the verified canonical order. Both chat-detail getters
(getAccessibleCopilotChat, getAccessibleCopilotChatWithMessages) now drop
the messages column from their metadata select (no more whole-array
detoast on every load) and assemble the transcript from the table after
authorization. This cascades to the copilot + mothership GET endpoints
and to resolveOrCreateChat's conversationHistory (the LLM payload).

The normalize/effective-transcript pipeline is source-agnostic
(copilot_messages.content == a JSONB array element), so transcripts are
byte-identical. Dual-write and the JSONB column stay in place as the
internal-logic source and fallback; removing JSONB writes is a later step.

Prod integrity verified before cutover: 0 messages missing, 0 NULL-seq,
0 dup keys/seq, 0 orphans, order-parity vs JSONB = 0 mismatches.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* test(copilot): cover auth-deny on a found row skips the messages query

Address PR review: exercise the `if (!authorized) return null` contract —
when the chat row exists but authorization fails, the getter returns null
and never issues the copilot_messages read.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* fix(tables): right-align run/stop in embedded toolbar; workflow cells format like normal cells (#4806)

* fix(tables): right-align run/stop in the embedded table toolbar

Add a right-aligned `trailing` slot to ResourceOptionsBar and move the embedded
mothership table's run/stop control into it, so Filter + Sort stay left-aligned
and run/stop sits opposite on the right. No-op for the search-bearing consumers
(logs, resource list), which don't pass `trailing`.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(tables): workflow-output cells format values like normal cells

Workflow-output columns short-circuited in resolveCellRender and rendered their
value as plain text, so a sim-resource URL / external URL / JSON / date produced
by a workflow never got the chip, favicon link, or typed formatting a normal
cell gets. Factor value formatting into a shared `resolveValueKind` helper used
by both the workflow-value branch and the plain-cell branch; the workflow branch
keeps the typewriter reveal for plain streaming text via a `typewriter` flag.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(tables): detect resource/URL links on workflow output regardless of column type

Workflow output columns default to `json` (columnTypeForLeaf), so routing their
values through the type-based formatter (a) gated chip/URL promotion behind
`column.type === 'string'` — a URL produced by a json-typed output never became
a chip — and (b) JSON.stringify'd plain string values, adding quotes and losing
the typewriter reveal. Detect links (sim-resource chip / favicon URL) on the
value string directly for workflow outputs, falling back to the plain `value`
kind; plain cells keep the type-based formatting. Addresses Greptile P2 on #4806.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(icons): repair broken integration icon rendering (#4810)

* fix(icons): repair broken integration icon rendering

Two distinct bugs left integration icons broken on the /integrations page
(visible at 32-40px, hidden at the toolbar's 16px):

1. Corrupted SVG paths (Notion, Greptile, Granola, Calendly, Grafana, Bedrock):
   over-minified data dropped elliptical-arc flag digits (e.g. `A1 1 0 5.9 7`
   instead of `A1 1 0 0 0 5.9 7`); Granola's cubic stream was truncated. Browsers
   abort path parsing at the first invalid arc flag, so each rendered as a fragment
   or blank. Replaced with correct path data from canonical sources, preserving each
   icon's existing fill/gradient and bgColor.

2. Invisible glyph (Bright Data): its icon uses fill='currentColor' but bgColor was
   '#FFFFFF', and every surface forces text-white on the glyph - white-on-white.
   Changed bgColor to Bright Data's brand blue (#3d7ffc) so the white glyph reads,
   matching the white-glyph-on-brand-chip convention.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(icons): restore Calendly dual-tone brand colors

Addresses review feedback: the previous fix replaced the broken Calendly icon
with a monochrome #006BFF path, dropping the cyan #0ae8f0 accent from the
original dual-tone mark. Restored the two-tone logo (blue + cyan) using clean,
valid path data, cropped to a tight square viewBox so it fills the chip.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* improvement(icons): enlarge icons, fix Zoom contrast and Quiver chip

- Zoom: glyph was blue-on-blue (#0B5CFF on #2D8CFF chip); switched to
  currentColor so it renders as a white glyph on the blue chip.
- Quiver: chip bgColor #000000 -> #FFFFFF to match the icon's near-white box,
  and enlarged the mark slightly (viewBox crop).
- Enlarged (tightened viewBox, verified no clipping): RevenueCat, Prospeo,
  Granola, Firecrawl, Enrich.so, and the AWS icons (RDS, DynamoDB, SQS,
  CloudFormation, Athena, CloudWatch, SES, Bedrock, S3).
- ZoomInfo left unchanged: it is a full red rounded-square logo that already
  fills its frame, so a crop would clip it.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(icons): use Bright Data wordmark on white chip; repair Circleback

- Bright Data: replaced the flame glyph with the official two-tone 'bright data'
  wordmark (provided asset), centered in a symmetric viewBox. Reverted the chip
  bgColor from #3d7ffc to #FFFFFF since the blue wordmark is invisible on a blue
  chip (the wordmark is designed for a light background).
- Circleback: a minifier had rounded the pattern's image scale to scale(0),
  collapsing the embedded logo to zero size (invisible). Restored the correct
  scale (1/280 = 0.00357142857) so the C. mark renders.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(docs): sync Quiver block color card to white chip

Reflects the Quiver bgColor change (#000000 -> #FFFFFF) in the docs block info card.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* improvement(icons): enlarge AWS/Cloudflare/Dagster icons, fully white Zoom

- Enlarged (tighter viewBox, render-verified, no clipping): Cloudflare, Dagster,
  and the red AWS icons AWS IAM, Identity Center, Secrets Manager, SES, STS.
  Identity Center was anomalously small (filled ~32% of its frame); the group is
  now sized consistently (~80% fill).
- Zoom: the camera lens triangle was still #0B5CFF (blue-on-blue); switched it to
  currentColor so the whole camera renders white on the blue chip.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* docs(wiza): consolidate individual reveal into a single operation

Merges the separate Start/Get Individual Reveal operations into one Individual
Reveal operation in the Wiza docs and integrations data (operationCount 5 -> 4).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* improvement(icons): size remaining AWS icons to match the set (~80% fill)

Bring RDS, DynamoDB, SQS, CloudFormation, Athena, CloudWatch and S3 up to the
same ~80% fill as the AWS IAM/Identity Center/Secrets Manager/SES/STS group, so
all AWS icons are visually consistent. Bedrock left as-is (already ~92% fill).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(icons): use Bright Data flame mark, enlarge ZoomInfo

- Bright Data: the full 'bright data' wordmark was illegible at chip size.
  Replaced with just the flame-'i' brand mark (blue #4280f6 on the white chip),
  centered.
- ZoomInfo: cropped the viewBox toward the white 'Zi' so it's larger; the red
  rounded-square background still fills the chip.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* improvement(icons): enlarge CrowdStrike icon

The falcon mark sat small in its chip because the icon used a wide 768x500
viewBox (letterboxed in the square chip). Switched to a square viewBox centered
on the mark so it fills ~80%, consistent with the other icons.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* fix(tables): serialize schema mutations to prevent parallel column clobber (#4812)

* Make workflow description nullable

* fix(tables): serialize schema mutations to prevent parallel column clobber

* fix(tables): load workflow outside schema lock; use DbOrTx for getTableById

* fix(tables): scale idle timeout in updateColumnType to avoid aborting large type changes

* fix(tables): skip stale remap types when workflowId changes concurrently

* fix(tables): scale idle timeout in updateColumnConstraints for large tables

* fix(wait): resume live/draft async waits and preserve cell context on chained waits (#4814)

* Make workflow description nullable

* fix(wait): resume live/draft async waits and preserve cell context on chained waits

* improvement(knowledge): polish tag filter dropdowns

* improvement(knowledge): soften filter section labels

* improvement(knowledge): soften list filter labels

* fix(security): harden SSO domain registration, webhook path isolation, and CSV export (#4813)

* fix(security): harden KB file access, SSO domain registration, webhook path isolation, env secrets, and CSV export

* fix(sso): scope domain conflict query with indexed lower(domain) filter

Address PR review: avoid a full-table scan on every SSO provider
registration by filtering candidate rows in SQL with
lower(domain) = <normalized>, keeping the in-memory ownership check.
Also tighten the normalizeSSODomain TSDoc.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* chore: condense env route security comments

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* icons update

* chore(security): tighten inline comments in CSV export and KB file authorization

Condense verbose comment blocks to concise TSDoc/single-line form; no behavior change.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(security): validate internal serve origin in KB file authorization

Replace the bypassable isInternalFileUrl substring check in resolveInternalKbKey
with an origin allow-list (base URL, internal API base URL, TRUSTED_ORIGINS).
A crafted external host whose path is /api/files/serve/<victim-key> no longer
resolves to the victim key. Relative same-origin URLs are unaffected.

* style(sso): use idiomatic sql lower() comparison for domain conflict query

Match the repo's prevailing `sql`lower(col) = value`` idiom for the
case-insensitive SSO domain conflict lookup.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(security): align workspace env admin gate with hasWorkspaceAdminAccess

Use the same admin check the secrets UI uses (owner, admin permission, or
org-admin) so owners and org-admins are not wrongly denied their own decrypted
workspace secrets, while read-only members remain restricted to names only.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(sso): rely on lower(domain) match for conflict detection, drop dead in-memory recheck

Address PR review: the SQL `lower(domain) = <normalized>` predicate already
excludes rows that the in-memory `normalizeSSODomain(...) === domain` recheck
claimed to catch, making that recheck dead/misleading code. Match on the
canonical lower-cased domain and filter purely by ownership. Malformed legacy
values (wildcards, schemes, ports) never match an email domain at sign-in, so
excluding them is not a gap. Test DB mock now applies the lower() predicate so
the casing-variant case is genuinely exercised.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(security): scope webhook deploy path conflict to active webhooks

findConflictingWebhookPathOwner omitted the isActive filter that the
runtime dispatcher (findAllWebhooksForPath) applies, so an inactive but
non-archived webhook from another workflow (e.g. after undeploy or
failure auto-disable) would permanently block any new deployment on that
path even though it never receives deliveries. Align the guard with the
runtime isActive + archivedAt filter; the earliest-owner runtime check
remains the authoritative cross-tenant protection. Also trims verbose
TSDoc on the webhook path-isolation helpers.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(security): exclude archived workflows from webhook deploy path conflict

findConflictingWebhookPathOwner now joins workflow and filters
isNull(workflow.archivedAt), matching the runtime dispatcher
(findAllWebhooksForPath). A webhook on an archived workflow can never
receive deliveries at runtime, so it must not block legitimate path reuse
with a 409.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(security): anchor KB file ownership to earliest document in any state

A KB file's owner is now the earliest document referencing its key regardless of
state (active/archived/deleted/excluded); access is granted only when that owning
document is still active. Closes the residual where an attacker could plant an
active document to claim a file whose original document was archived or deleted.

* updated greptile icon

* revert(security): drop KB file authorization changes

Reverts the knowledge-base file-access work (origin-pinning / owner-pinning /
origin allow-list in verifyKBFileAccess) and its test. The other hardening fixes
(SSO domain registration, webhook path isolation, workspace env secrets, CSV
export) are unchanged. apps/sim/app/api/files/authorization.ts is restored to its
origin/staging baseline.

* fix(sso): treat caller's own user-scoped provider as owned during conflict check

Self-hosters often register SSO user-scoped via the CLI script (no
SSO_ORGANIZATION_ID). If they later enable organizations and reconfigure the
same domain org-scoped through the UI, the conflict check previously treated
their own user-scoped row as another tenant's and returned a misleading 409.
Recognize the caller's own user-scoped provider as owned so that migration is
allowed, while still blocking another user's or another org's domain.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* revert(security): remove workspace-env admin gate

Defer to a credential-based access model (separate change). Restores
GET /api/workspaces/[id]/environment to main behavior and removes the test.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* refactor(security): consolidate webhook path-collision check into one helper

Extract findConflictingWebhookPathOwner to lib/webhooks/utils.server.ts as
the single source of truth for cross-tenant path-collision detection, used by
both webhook creation paths (deploy sync and the manual /api/webhooks route).

This also repairs two latent issues in the manual route's previous inline
check, which queried with limit(1) and only webhook.archivedAt:
- limit(1) inspected one arbitrary row, so a same-workflow row could mask a
  foreign collision (false negative). The shared helper scans all matching
  rows.
- It omitted isActive/workflow.archivedAt, so inactive or archived-workflow
  webhooks (which never receive deliveries) permanently blocked path reuse.
  The helper mirrors the runtime dispatcher's filter.

Same-workflow webhook reuse for upsert is now a separate, explicit lookup.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* fix(security): block private/reserved IPs for hosted 1Password Connect SSRF (#4818)

* fix(security): block private/reserved IPs for hosted 1Password Connect SSRF

* test(security): use real isPrivateOrReservedIP and cover IPv6 edge cases

* improvement(integrations): validate and expand devin, cursor, and greptile (#4820)

* improvement(integrations): validate and expand devin, cursor, and greptile

- devin: fix missing org_id path segment on all session endpoints, add 7 session sub-resource tools (list messages/attachments, get/append/replace tags, archive, terminate), pagination, and is_archived output
- cursor: add get_api_key_info, list_models, list_repositories tools
- greptile: align block and docs
- normalize array outputs to default [] and tighten types

* refactor(cursor): simplify list_repositories v2 array normalization

Collapse the redundant `?? []` + `Array.isArray` double-guard into a
single Array.isArray check, per PR review feedback.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(devin): scope session-tag mapping to tag ops and normalize array tag inputs

- Only map sessionTags into the tools tags param for append/replace operations,
  preventing stale sessionTags state from clobbering create_session tags
- Fall back to a wired tags value when sessionTags is empty for tag operations
- Normalize tag inputs (string or wired string[]) via normalizeTags so array
  values from other blocks no longer throw on .split

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(cursor): restore base64 file data in legacy download_artifact metadata

The legacy CursorBlock exposes only content + metadata (no v2 file
output), so metadata.data was the only way legacy-block workflows could
access downloaded artifact bytes. Restore the base64 data field and
document it in the outputs/type instead of dropping it.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(devin): coerce terminateArchive to archive flag for boolean-wired input

* docs(integrations): regenerate tool docs for new devin and cursor operations

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* fix(search-replace): don't auto-navigate when content edits invalidate the active match (#4819)

* fix(search-replace): don't auto-navigate when content edits invalidate the active match

* fix(search-replace): clear afterReplaceIndexRef on apply failure and zero matches

* fix(search-replace): remove duplicate setActiveSearchTarget(null) on close

* fix(search-replace): move afterReplaceIndexRef write inside handleApply past the guard

* fix(search-replace): auto-navigate when hydration resolves with no prior active match

* chore(search-replace): remove inline comments

* fix(search-replace): revert !activeMatchId guard that caused immediate re-navigation after deselect

* improvement(enrichments): limit company-info to fields both providers return (#4817)

Hunter's company dataset returns null industry/foundedYear for many large companies (verified against the live API for Microsoft, Amazon, Google), so under the first-non-empty-wins cascade those columns appeared inconsistently across rows. Limit company-info outputs to employee count and description — the fields Hunter and PDL both reliably return — so every row is consistent. employeeCount is a string so Hunter's range bucket and PDL's exact count share the column.

* fix(files): don't reject external URLs containing '..' in file parse validation (#4821)

* fix(files): don't reject external URLs containing '..' in file parse validation

The file block's file_fetch operation rejected any external URL whose path
contained '..' (e.g. Slack files-pri slugs with a literal '...') with
'Access denied: path traversal detected'. Traversal checks only apply to
local paths — external http(s) URLs are fetched with SSRF protection
downstream and are never resolved against the filesystem, so they now
short-circuit as valid. Internal /api/files/serve/ URLs keep full traversal
protection.

* test(files): fix external-URL assertion to handle undefined error

* test(files): assert success explicitly in external-URL traversal test

* fix(files): keep traversal protection for https URLs matching internal serve paths

* feat(google-sheets): add row filtering to read with numeric operators (#4822)

* feat(google-sheets): add row filtering to read with numeric operators

Adds client-side row filtering to the Google Sheets read (v2) operation.
Filter the returned rows by a header column using text operators
(contains, not_contains, exact, not_equals, starts_with, ends_with) and
numeric/ordering operators (gt, gte, lt, lte). Filtering lives in a pure,
unit-tested helper (filterSheetRows) and runs over the fetched read range;
an optional `filter` output reports whether the column was found and how
many rows matched.

Also hardens the surrounding tools:
- trim spreadsheetId in write/update/append URL builders (matches read)
- URL-encode the v1 read default range
- expose valueInputOption for the update operation in the block

Backwards compatible: with no filter requested, read output is byte-
identical and the `filter` field is omitted. The filterMatchType union is
widened additively (4 -> 10 values).

* fix(google-sheets): correct filter metadata for missing column and header-only sheets

- matchedRows is now 0 (not totalRows) when the filter column is not found,
  so it no longer contradicts applied=false / columnFound=false
- columnFound now reflects an actual header lookup for empty/header-only
  sheets instead of being hardcoded true
- add tests covering header-only and empty sheets with present/absent columns

* fix(selectors): fetch all pages for paginated dropdown list routes (#4823)

* fix(selectors): fetch all pages for paginated dropdown list routes

Dropdown selectors fetched only the first page of paginated provider
APIs, silently hiding results past page one. Add bounded server-side
draining to the list routes across Microsoft Graph, Google, Notion,
Atlassian, Linear, AWS CloudWatch, and offset/token REST APIs, plus a
shared client-side drain cap in the selector hook. Response shapes,
stored values, and tool execution are unchanged; CloudWatch list tools
still honor a caller-supplied limit. Also fixes the Word file picker
that was searching for .xlsx files.

* fix(selectors): harden JSM and Monday pagination draining

- JSM service-desk/request-type drains advance `start` by the actual row
  count returned (not the fixed page size) and stop on an empty page, so a
  short non-final page can't skip items.
- Monday boards drain now checks `response.ok` per page, surfacing a
  mid-drain HTTP failure instead of treating it as an empty final page and
  returning a partial 200.

* docs(selectors): clarify JSM drain advances start by actual row count

The offset-advancement fix (advance `start` by the rows returned, not the
fixed page size) landed in 7b19788; update the TSDoc to match so it no
longer reads as advancing by `limit`.

* fix(selectors): drain fetchPage in direct fetchList callers

Making `fetchList` optional left three direct callers (outside the
useSelectorOptions hook) calling it unguarded, which broke the build's
type check. Route them through a shared `loadAllSelectorOptions` helper
that uses `fetchList` when present and otherwise drains `fetchPage`.
This also prevents a regression: `confluence.spaces` / `knowledge.documents`
now paginate via `fetchPage` only, and these callers (search/replace,
value resolution) would otherwise have silently returned no options.

* chore(selectors): rename MAX_PAGE_PAGES to MAX_NOTION_PAGES for readability

* fix(sso): re-check domain conflict before write and reject IP-address domains (#4825)

* improvement(copilot): make copilot_messages the sole transcript store, remove JSONB dual-write (#4826)

Stop writing/reading the legacy copilot_chats.messages JSONB column now that
reads are cut over to copilot_messages. Make appendCopilotChatMessages the
primary write (throws on failure instead of swallowing), repoint peripheral
readers (workspace VFS, chat cleanup, data drains, fork, superuser import) to
copilot_messages, and persist the assistant turn inside finalizeAssistantTurn's
transaction so it commits atomically with the stream-marker clear. The column
itself is dropped in a follow-up migration after this bakes.

* feat(tables): expand filter operators (not-contains, starts/ends-with, not-in, empty) (#4827)

Add does-not-contain ($ncontains), starts-with ($startsWith), ends-with
($endsWith), not-in-array ($nin, previously executed server-side but unexposed
in the UI), and is-empty/is-not-empty ($empty) filter operators end-to-end —
SQL builder, condition types, query-builder converters/constants, the filter
UI, the Table tools/block descriptions, and docs.

Also fix correctness bugs in the filter builder surfaced by the wider operator
set:
- Same-column AND rules (e.g. age > 18 AND age < 65, or name startsWith 'A'
  AND name endsWith 'Z') silently overwrote each other because the AND group
  was keyed by column name. They now merge into one operator object, which
  also makes Filter -> rules -> Filter round-trip losslessly for multi-operator
  columns.
- $nin values were not split into an array like $in, and textual-match values
  like "123" were numeric-coerced (breaking the ILIKE path).
- A non-boolean $empty operand from the raw API silently inverted the check; it
  now coerces 'true'/'false' strings and otherwise returns a 400.

* improvement(copilot): stop persisting tool-call result outputs in transcripts (#4829)

Opening a Mothership task could take many seconds because a single persisted
assistant message in copilot_messages.content can reach hundreds of MB, almost
entirely inside contentBlocks[].toolCall.result.output (e.g. a get_workflow_logs
or run_workflow result). The DB query is ~2ms; the cost is detoasting that
payload, shipping it to the browser, and parsing it.

These outputs are dead weight on the Sim side: they are never rendered (the
thread shows only tool name/title/status) and never replayed to the model (the
upstream copilot service owns conversation memory). So drop result.output before
it is persisted, keeping result.success/error plus the tool metadata.

- add stripToolResultOutput() in persisted-message.ts
- apply it in messages-store toRow (covers every write path) and in
  loadCopilotChatMessages (existing rows render fast on read)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* feat(providers): add Together AI, Baseten, and Ollama Cloud model providers (#4830)

* feat(providers): add Together AI, Baseten, and Ollama Cloud model providers

* fix(providers): guard Ollama streaming fast-path with hasActiveTools

Match Together/Baseten/Fireworks: when tools are supplied but all are
filtered out (usageControl 'none'), take the single streaming call instead
of an extra non-streaming round-trip.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix(providers): filter non-chat model types from Together model list

* refactor(providers): dedupe Ollama Cloud upstream schema

ollamaCloudUpstreamResponseSchema was byte-for-byte identical to
ollamaUpstreamResponseSchema (both /api/tags endpoints return the same
{ models: [{ name }] } shape). Drop the duplicate and reuse the shared schema.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* fix(knowledge): calendar view sync, deduplicate popover animation classes, type-safe filter cast

* cleanup(knowledge): remove TRIGGER_BORDER_CLASS duplication, inline displayLabel, drop enabledFilterParam alias

---------

Co-authored-by: Vikhyath Mondreti <vikhyathvikku@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Waleed <walif6@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Theodore Li <theo@sim.ai>
Co-authored-by: andresdjasso <andresdjasso@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant