fix(cli): honor --resume/--continue in oneshot (-z) — hydrate and chain the session

[nathansmithopenclaw-alt]

What

hermes --resume <id> -z "<prompt>" (and --continue [name]) now actually chains: the session's prior transcript is loaded as conversation history and the turn is appended to the SAME session id. Ids that don't exist yet are created on first use, so scripted callers can mint a stable session key per conversation up front and pass it on every turn — no output parsing, no discovery step.

Why

The oneshot (-z) dispatch in hermes_cli/main.py called run_oneshot() and sys.exit()ed before the --resume/--continue → chat routing, and run_oneshot() never hydrated anything — the flag was accepted (exit 0, no warning) and silently dropped. Every -z turn was stateless and wrote a new throwaway session.

This is #49195. It bites any frontend driving oneshot with a stable per-conversation id — our concrete case is a gateway that runs a Hermes chat seat over -z one-shots; before this fix the agent was amnesiac on every turn.

Relationship to existing PRs

• fix(oneshot): honor --resume/--continue so -z can be session-aware #40333 implemented essentially this hydration and was closed by its author without review — this PR carries that approach forward (same core mechanism, independently derived) and adds create-on-first-use ids plus best-effort degradation.

• fix(cli): reject --resume/--continue in oneshot instead of dropping it #49204 takes the inverse, minimal path (reject the flags in oneshot) and explicitly defers "full resume hydration" to a later change, citing exact-output replay and compression-lineage concerns. This PR is that later change, and both concerns are addressed:

  • Compression lineage: the resume id is projected to the live tip via the existing SessionDB.resolve_resume_session_id before loading, exactly like interactive resume.
  • Replay duplication: history is seeded via run_conversation(conversation_history=...); _flush_messages_to_session_db skips seeded history dicts by identity, so resumed transcripts are never re-written to the store (same mechanism the gateway relies on).

  If maintainers prefer this hydration path, fix(cli): reject --resume/--continue in oneshot instead of dropping it #49204's guard becomes unnecessary; if you'd rather land the guard first, this rebases trivially on top of it (the guard's two dispatch sites are the ones wired here).

Change

• hermes_cli/oneshot.py — run_oneshot(resume=...) → _run_agent(resume=...) → new _load_resume_history(session_db, resume): resolve the compression chain, load messages in conversation format (dropping session_meta rows), best-effort reopen_session, and pin AIAgent(session_id=...). Unknown ids return as-is with no history (create-on-first-use; the agent's existing create_session upsert makes the row on first persist). A broken session store degrades to a stateless turn — -z never gets less reliable than before.
• hermes_cli/main.py — both oneshot dispatch sites pass resume=_resolve_oneshot_resume(args). --resume is passed through verbatim (no existence check — that's what enables caller-minted ids); --continue resolves by name / most-recent exactly like interactive chat, and an unresolvable --continue errors (nothing sensible to chain onto).
• hermes_cli/_parser.py — -z help documents the chaining semantics.
• With no --resume/--continue, behavior is byte-for-byte unchanged (session id stays agent-generated).

Tests

• New tests/hermes_cli/test_oneshot_resume.py (7 tests): history load + session_meta filtering + lineage resolution, create-on-first-use, broken-store degradation, and wiring assertions that AIAgent gets session_id and run_conversation gets conversation_history (and that the no-resume path is untouched).
• tests/hermes_cli/test_tui_resume_flow.py — two kwarg-capture assertions extended with the new resume key.
• tests/hermes_cli full run: failure set byte-identical to a clean-tree baseline run (pre-existing failures only; +7 passing).

Manual verification (live CLI):

SID="test_$(date +%Y%m%d_%H%M%S)_zeb"          # id Hermes has never seen
hermes --resume "$SID" -z "Remember the word ZEBRA. Reply with exactly OK."
# -> OK
hermes --resume "$SID" -z "What one word did I ask you to remember? Reply with only that word."
# -> ZEBRA

Closes #49195

🤖 Generated with Claude Code

[alt-glitch]

This was generated by AI during triage.

Related: fixes #49195. Carries forward the hydration approach of closed #40333 (independently derived) and is the deferred-hydration change #49204 explicitly punted on. Competing with open #49204, which takes the inverse minimal path (reject --resume/--continue in oneshot rather than honor them). Not a duplicate of either — flagging the cluster (#57859 hydrate vs #49204 reject vs closed #40333) so a maintainer can pick the direction.

[LavyaTandel]

Closing — created in error while comparing Option A/B for prompt-caching issue #57845. The actual Option A/B pair is #57876 + follow-up branch fix/prompt-caching-markers.

[teknium1]

Thanks for carrying forward the session-hydration approach. The current-main premise is confirmed: both oneshot dispatches call run_oneshot() before resume routing and pass no resume value (hermes_cli/main.py:12670-12685, hermes_cli/main.py:14826-14843). The implementation direction matches the existing interactive lineage and history flow (hermes_cli/cli_agent_setup_mixin.py:272-292) and the flusher’s identity-based seeded-history deduplication (run_agent.py:1885-1911).

Problems

• The new tests are mock-only (tests/hermes_cli/test_oneshot_resume.py:63-81), so they do not verify real SQLite create-on-first-use, a second-process reload, or no duplicate transcript rows. The contribution rubric requires an E2E path for session-resolution and file-I/O changes.

Suggested changes

• Add a temp-HERMES_HOME integration regression covering two caller-minted resumed oneshot turns and durable transcript assertions.
• Salvage onto current main while preserving usage_file, added by 7dfd5077ce to hermes_cli/oneshot.py:174 and both dispatch calls.

Automated hermes-sweeper review.

[teknium1]

This harness mocks SessionDB and AIAgent, so it cannot prove create-on-first-use or a second independent -z --resume invocation reloads and appends the real SQLite transcript. Please add a temp-HERMES_HOME integration test for that two-invocation path.

[nathansmithopenclaw-alt]

Closing in favor of #80799, which carries this fix rebased onto current main with real-SQLite coverage.