Add a Hermes Agent adapter (10th agent) + fix kimi sync crash

[KennyUMN]

What

Adds a Hermes Agent adapter (the 10th agent) — both directions:

• reads Hermes chats into every other agent, and
• writes mirrors into Hermes' own store, so they appear in Hermes' native /resume / picker and hermes -c.

Hermes stores everything in one SQLite database, $HERMES_HOME/state.db (default ~/.hermes/state.db), in sessions + messages. Chats are source='cli'; sub-agent/compaction children are filtered out via parent_session_id and the _delegate_from marker in model_config.

Mirrors are marked with origin_json = {"imported_from":{"tool":"samethread"}}, which read() turns into is_hop — so hop never re-adopts its own mirror. The messages table already carries AFTER-INSERT FTS/display triggers, so a plain INSERT keeps the search index correct; nothing FTS-specific is written.

Two things worth a reviewer's eye

1. model is left NULL on purpose. hermes --resume reuses the session's stored model, and a placeholder like hop-import is not a real provider model — Hermes rejects it (HTTP 400: hop-import is not a valid model ID) and the mirror becomes unresumable. Claude/Codex ignore that field; Hermes does not.
2. connect() falls back to mode=rw when mode=ro fails. A WAL-mode db whose -wal/-shm sidecars are gone (Hermes exited cleanly) cannot be opened read-only — and hop sync usually runs precisely when Hermes is not running. sqlite3.connect is lazy, so the open is forced with a pragma inside the try.

Also fixes two pre-existing bugs found while wiring this up

• kimi: updated was the raw ISO string from state.json, so sync.adopt_new's sorted(..., key=...['updated']) crashed with TypeError: '<' not supported between instances of 'str' and 'int'. hop sync died outright on any machine with a Kimi session. Now converted with core.iso_ms.
• kimi: added the missing iso_ms import.

Verified

• Full suite green: 20 tests, on macOS py3.9 (and the repo's own CI command pip install -e . && python -m unittest discover -s tests -v). Suite also stays isolated when HERMES_HOME is set (CI has it unset, so this only bit real users).
• Real end-to-end: mirror written into a copy of a live state.db (395 sessions), then hermes --resume <mirror-id> -z "what codename…" → the model answered the codename from the mirrored history. Writes succeed against the real schema with its triggers.
• hop sync --dry-run against a real machine now completes (871 mirrors planned across Claude/OpenCode/Hermes/Codex/Kimi) — it previously crashed on the kimi bug.
• The machine's real ~/.hermes/state.db was never modified (all probes on copies).

Notes

• No install() hook: Hermes shell hooks need an allowlist entry plus a config.yaml edit that a third-party installer shouldn't make. Hermes mirrors land on hop sync or any other agent's sync, same as MiniMax. Easy to add later if a safe path exists.
• ~/.hermes is the default; $HERMES_HOME is honoured for profiles.

[KennyUMN]

CI is queued as action_required (first-time contributor — needs a maintainer to approve the run). Meanwhile I ran the exact CI steps locally on both matrix Python versions against this branch:

# python 3.12
pip install -e . && hop --version && python -m unittest discover -s tests -v
  samethread 0.2.0
  Ran 20 tests in 0.012s — OK

# python 3.9
pip install -e . && hop --version && python -m unittest discover -s tests
  samethread 0.2.0
  Ran 20 tests in 0.014s — OK

End-to-end receipt (mirror into a copy of a live 395-session state.db, then resume it through Hermes itself):

$ hop write ...  -> 20260923_164505_84dc07  (source=cli, model=NULL, title='PR handoff (Agy)')
$ hermes --resume 20260923_164505_84dc07 -z 'What codename? Reply with just the codename.'
CODENAME-PAPAYA-99

And the kimi crash this also fixes, before/after:

$ hop sync --dry-run
TypeError: '<' not supported between instances of 'str' and 'int'   # before
done: create 871, new 194                                            # after

[Copilot]

Copilot review overview

🟡 Changes recommended

Hermes detection can schedule writes to a missing database, and documentation omits Hermes from resume references.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1 · 1

Open (2)

• Avoid detecting Hermes when its state database is missing · New
• Document Hermes in resume command and flowchart · New

What changed in this PR

Adds a Hermes Agent adapter for native SQLite session synchronization and fixes Kimi timestamp handling.

Changes:

• Adds and registers Hermes read/write support.
• Normalizes Kimi timestamps.
• Updates CLI/docs and adds regression coverage.

File	Summary	Final review notes
tests/​test_core.py	Adds Hermes and Kimi coverage.	No final review comments.
src/​samethread/​cli.py	Updates resume help.	No final review comments.
src/​samethread/​agents/​kimi.py	Converts ISO timestamps to epoch milliseconds.	No final review comments.
src/​samethread/​agents/​hermes.py	Implements Hermes SQLite discovery, reading, writing, and mirroring.	Moderate (3 votes): missing databases can be detected as writable, causing sync failures.
src/​samethread/​agents/​__init__.py	Registers Hermes.	No final review comments.
README.md	Documents Hermes support.	Nit (2 votes): resume references and the flowchart omit Hermes.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[KennyUMN]

All three reproduced failures are fixed in 519f823, plus the nits. Each verified against a copy of a real 395-session state.db (its triggers present):

1. Missing store → HopError, not OperationalError

FIX1  HopError: run Hermes once so its state.db exists

write() raises before touching anything; no more per-turn retry queue.

2. Title collision keeps a title

FIX2  1st='Greeting (OpenCode)'  2nd='Greeting (OpenCode) · f548ad'

Suffix is · <last 6 of session id> — same shape the existing OpenCode collision produces.

3. delete() / retitle() with ~/.hermes gone

FIX3  delete with no store: no crash
FIX3  list with no store -> {}

Moved the existence check into connect(), so every rw caller treats a missing store as a no-op.

Nits done

• started_at is now the first message's time (first<last verified).
• The delete from messages moved inside the with c: block.
• list() uses one grouped query for message counts (was 2 per session), like OpenCode.
• Tests: per-test store reset so no test depends on another having run (the WAL test was order-dependent on real1); the WAL test now seeds its own session; HOP → HERMES_HOME_DIR.
• pyproject description + keywords and the README 'Tested on' line mention Hermes. Hero image left to you (make_hero.py).

New tests (22 total, green on py3.9 and py3.12 with the exact CI commands): missing store → HopError + delete no-op, title collision keeps a suffixed title, WAL store without sidecars.

End-to-end still holds on a real store — a collision-suffixed mirror resumed through Hermes and answered from the imported history:

$ hermes --resume 20260923_171221_1cf3d1 -z 'what was this conversation about?'
Ngebandingin 5 ide hackathon IFEST dari Google Doc pakai prompt di tab PROMPT, terus nulis
hasilnya ke tab Hasil AI Kenny, termasuk bikin revisi KONEK v3 (ganti WhatsApp API ke
WebRTC + Telegram).

One note on failed 15 in a sync run: those are all OpenCode, from a malformed opencode.json in one workspace — pre-existing and unrelated.

[KennyUMN]

Superseded by #2, which now contains the complete Hermes adapter, Agy writable-store, and Agy chat-picker changes.