v0.9.12: Power BI integration, library articles, chat fork speedups - #8589

* feat(powerbi): add Power BI actions

* fix(powerbi): harden partial results and selector hydration

* fix(powerbi): compact retained failures and reuse rotated fixture tokens

* fix(powerbi): use standard workflow failure handling

* fix(powerbi): preserve nested query diagnostics

* fix(powerbi): generate docs entirely from source

* fix(powerbi): keep generated catalog descriptions complete

* fix(powerbi): validate resource grants in account selector

---------

Co-authored-by: Bill Leoutsakos <billleoutsakos@Bills-MacBook-Pro.local>

…aging (#8578)

- replace the non-null assertions in the Power BI tests with named guards
- drop export from Power BI types and constants used only in their own file
- inline the refresh-attempt output properties at their one use, so the docs generator (which only resolves exported consts) renders them without a dead export

…ity, Deployment, and Evaluation (#8576)

* feat(library): How to Use a Postgres MCP Server With AI Agents: Security, Deployment, and Evaluation

* Pi Babysit: address PR #8576 feedback

---------

Co-authored-by: Sim Pi Agent <pi@sim.ai>

…8577)

* docs(library): update ai-agent-examples-by-department-and-industry

* Pi Babysit: address PR #8577 feedback

---------

Co-authored-by: Sim Pi Agent <pi@sim.ai>

…release review (#8571)

* fix(audits): close guardrail detector gaps and correct guidance from release review

* fix(audits): detect partialize and .with by AST instead of regex

* fix(audits): exempt only the imported OpenTelemetry context from the .with rule; unwrap cast partialize

* fix(audits): drop the OTel exemption for rebound names, catch optional .with calls and defaulted partialize params

* fix(audits): treat logical and conditional partialize returns of the whole state as leaks

* improvement(audits): let tsc guard ES2023 methods, tighten ratchet hints, fix && partialize false positive

- check:utils drops the ES2023 call-site detector; every tsconfig keeps lib at ES2022 so tsc rejects toSorted/with by type, and the check now fails a tsconfig that raises lib past it (the #5340 cause)
- check:zustand-v5 treats only the right side of && as a returned value
- check:explicit-any counts group-level and bare lint biome-ignore suppressions, and no longer pairs a deleted file with an unrelated new one as a rename
- check:file-names drops its rule-matched rename hint, which pointed new violations at old baseline entries
- check:comment-hygiene skips the parse for files with no possible hit and scans untracked files
- inline the single-consumer source-kind helper and two leftover ref aliases
- CLAUDE.md naming states the check:file-names scope; /ship runs type-check

* chore(guidance): state the tsconfig lib invariant as at or below ES2022

* fix(audits): keep matching toSorted/toReversed/toSpliced in source for any receivers

* chore(guidance): scope the ES2023 tsc guarantee to typed receivers

* fix(knowledge): scope list counts to selected knowledge bases

* fix(knowledge): avoid count-query fan-out for unpaged lists

…ropdown (#8582)

* fix(integrations): list tools for integrations without an operation dropdown

The catalog generator only read operations from a block's own 'operation'
dropdown, so single-tool blocks and provider pickers (DuckDuckGo, SMTP,
Twilio SMS, Embeddings, ...) emitted zero operations, and Notion v2 lost its
dropdown because it spreads NotionBlock.subBlocks instead of the whole config.

Resolve operations through same-file subBlocks spreads, and fall back to the
block's tools.access ids named from the client-safe tool metadata.

* fix(docs): drop the any cast and only follow plain subBlocks spreads in the catalog generator

* fix(sdk): default both SDKs to the www.sim.ai host

The apex https://sim.ai answers every request with a 301 to
https://www.sim.ai, and fetch and requests replay a 301 POST as a GET with
no body, so a default-configured execute call never reaches the API. Both
SDKs now default to https://www.sim.ai. Patch-bump both to 0.2.1 so the
publish workflows release the new default on merge to main, and note in the
API reference that earlier versions still default to the apex.

* test(ts-sdk): cover the default www host

* test(ts-sdk): use a typed Response in the default-host test

* feat(search): browse Lucid folders and Notion pages

* fix(search): preserve discovery query boundaries

#8583)

* fix(core): treat EHOSTUNREACH as a retryable network failure

An unreachable host is a connection that never opened, so the request never
reached its destination and one more attempt is as safe as after
ECONNREFUSED or ENETUNREACH.

* fix(mothership): pass typed fork refusals through and stop overlapping retries

The worker refuses a fork it cannot make with 404 (the chat or message is
gone), 409 (the response has not finished) or 413 (the cut is above its
ceiling), but every non-2xx became a generic error and a 500. Those are now
classified and passed through by the fork route with a message the person can
act on.

The retry loop's own catch swallowed every first-attempt failure, so a 400, a
500, a malformed receipt and a timed-out attempt were all sent again. A
timed-out attempt may still be copying, so only a recognized socket failure or
a 502/503/504 gets one more attempt now.

The per-attempt timeout is sized from measured copy latency at the worker's
ceiling with headroom, so a legitimate fork finishes inside one attempt and a
timed-out one is abandoned (the worker rolls back a copy whose caller left).

* fix(mothership): show a fork refusal's own message

A 409 or 413 from the fork route tells the person what to do (wait for the
response to finish, or fork from an earlier message), so the toast shows that
message instead of a generic failure.

* fix(mothership): discard the worker copy of a fork that is not published

When anything after the worker copy failed, most often the final transaction
that publishes the chat, the worker kept a conversation Sim had no chat for.
The fork now asks the worker to clean that chat up, through the same cleanup
endpoint chat deletion uses. It is best effort and a no-op when the worker
never committed.

* fix(mothership): point every reference a fork copies at what the fork holds

- A file whose blob copy failed is never published, yet the fork's messages
  and the worker's history were rewritten to its id and key. References to it
  now stay on the source file, and its resource tab is dropped as before.
- In-app /workspace/<id>/files/<fileId> links were left on the source file;
  the fork stays in the same workspace, so only the file id moves.
- Tool-call arguments and display titles kept the source file's id or key.
- A Sources tab for a response past the cut was copied, pointing at a message
  the fork does not have.

* fix(cleanup): purge a chat's inline images and only its own attachments

- Chat images are stored under the chat's id but were never deleted with the
  chat. Their keys are rebuilt from the assistant messages that published
  them, the same way a fork finds them to copy.
- Message attachments were deleted as copilot storage whatever their key.
  Attachments in Chat are workspace-bucket files owned by their workspace_files
  row, which a fork can share with its source (a failed copy, a deleted file),
  and the copilot bucket falls back to the workspace bucket on GCS and can be
  configured to it on S3. Only copilot keys are deleted from messages now.

* fix(cleanup): delete an organization attachment once no chat references it

Organization Chat attachments (assistant/ keys) have no workspace_files row and
a fork carries the same key, so the purge now deletes one under its own
storage context only after confirming no remaining chat of that organization
still references it.

* fix(mothership): order fork publication against a purge of its source

A fork can share keys with its source (organization attachments, files whose
copy failed), and chat cleanup deletes a shared key once no remaining chat
references it, checking after it deletes the source row. The fork's publish
transaction now holds the source row with FOR KEY SHARE: a purge that already
removed it refuses the fork (404, worker copy discarded), and one that has not
waits for the commit and then sees the fork's references.

* chore(mothership): quote the final copy latency behind the fork attempt timeout

* feat(youcom): add You.com integration

* fix(youcom): unique subblock ids, nullable outputs, reject non-array contents

* fix(youcom): share field ids across operations so agent tool rows expose them

* fix(youcom): validate per-endpoint values in tools, live-verified fields

* fix(youcom): normalize country and language codes

…l, and author slugs (#8587)

* fix(landing): return 404 for unknown library, blog, integration, model, and author slugs

* fix(landing): pre-render reseller model pages so they keep resolving

…8588)

* fix(files): read offset 0 as the first line instead of rejecting it

* test(files): cover the agent CLI forwarding --offset 0

…ider catalog (#8593)