existence-master/Sentient

About 330 lines across the tests, mostly tests/memory/, tests/channels/, tests/evolution/ and tests/conftest.py, use "Sarthak" as the user's name. Our rule for demo and test data is the fictional persona Maya Rao from desktop/scripts/seed-*.py, never real people. riddhigupta1110 pointed this out after copying the name from existing tests in #125. Fix: replace the real name with "Maya" (and any related real details with Maya Rao's fictional ones) in tests and code. Keep each test's meaning the same; some tests check the name inside generated text, so update those expectations together. Done when: git grep -i sarthak -- tests sentient desktop/src docs finds nothing and the full suite passes.

Follow-up to #202. browser_close (sentient/browser/service.py, the close_tool path) waits for unfinished downloads for the full 120 s download limit while holding the browser lock. With a stuck download, a real test measured 118.9 s. Meanwhile the chat turn stalls, other browser tools are blocked, and a script calling browser_close runs into the 120 s sandbox limit. Fix: wait a short time (5 to 10 s), then cancel what's left and report it in download_errors. Add a sentence to docs/API.md saying browser_close waits briefly for downloads, since it currently says actions never wait. Done when: closing with a stuck download returns within about 10 s and reports the cancelled download, with a test.

docs/API.md section 19 lists hermes_deliver on every job, but skipped jobs in the preview leave it out, so a client reading it raises an error. Fix: always include hermes_deliver (null when unknown). Done when: every job in the preview has the key, with a test. Found in the real end-to-end run of 2026-10-10 (qwen3:8b on an RTX 4060 8 GB).

Found while fixing flaky tests (#228, #235). Windows' clock advances every 15.6 ms by default, so rows saved in the same tick share a timestamp. Message order without a tie-break: sentient/agent/loop.py (around lines 1094 and 1221) and sentient/memory/episodic.py (around line 199) order messages by created_at only. Messages saved in the same tick can come back in the wrong order, which can scramble what the model sees. #235 fixed the same problem in the user model by adding rowid as a tie-break; apply the same here and check the store for other ORDER BY created_at queries. Profile upkeep watermark: upkeep picks new facts with updated_at > since. A fact saved in the same tick as an upkeep run, after its query, is never added to the profile. Use a watermark that can't collide (for example the highest fact rowid seen, or a sequence number). Done when: both are fixed, with tests that reproduce the same-tick case (for example by freezing the clock).

Follow-ups to the memory review inbox (#137, #232): Instruction-shaped memories never auto-save: sentences like "from now on, always ..." or "whenever X, do Y" change how Sentient behaves, so they should be proposed (like rules in #130), never silently saved as facts. Undo on every save: a small "Remembered: ... Undo" toast in the chat each time Sentient saves a memory. Your own words in a chat that read outside content: today they wait for review too. Keep facts that come only from the user's own messages active, and hold only those that depend on outside content, without trusting the model's say-so. Where: sentient/memory/, the chat UI, related to #130. Done when: each works as described, with tests. Related: #137, #232, #130.