fix: preserve one-shot resume context - #70136
Open
karthik-panchap wants to merge 1 commit into
Open
Conversation
alt-glitch
added
type/bug
teknium1
reviewed
Jul 30, 2026
teknium1
left a comment
teknium1
left a comment
Contributor
There was a problem hiding this comment.
Thanks for identifying the one-shot resume gap; current main still drops that state at hermes_cli/main.py:12365-12371.
Problems
- PR
hermes_cli/oneshot.py:327callsresolve_session_id(), which only accepts an exact or uniquely prefixed ID (hermes_state.py:4547-4572). The parser documents--resumeas accepting an ID or title (hermes_cli/_parser.py:157-162), and normal CLI resolves titles athermes_cli/main.py:2526-2533. - PR
hermes_cli/main.py:13377-13378and15537-15538forward only explicit resume/session-ID state.--continueand--no-restore-cwdremain accepted by the parser but bypassed in one-shot mode. - The compression child is now published atomically via
agent/conversation_compression.py:2147-2160; the PR's direct child-creation hunk needs to be relocated during salvage.
Suggested changes
- Use the normal ID-or-title/continue resolution contract, including recorded-CWD restoration and its opt-out.
- Rework the lineage marker at the current atomic publication boundary and validate it against current resume tests.
Automated hermes-sweeper review.
| return None, [] | ||
| if session_db is None: | ||
| raise RuntimeError("Session database is unavailable; cannot resume safely") | ||
| resolved = session_db.resolve_session_id(requested) |
Contributor
There was a problem hiding this comment.
resolve_session_id() only resolves an exact or uniquely prefixed ID. The top-level parser documents --resume as accepting an ID or title, and normal CLI startup resolves titles before resume; add the equivalent title fallback here so one-shot does not reject valid titled sessions.
| provider=getattr(args, "provider", None), | ||
| toolsets=getattr(args, "toolsets", None), | ||
| usage_file=getattr(args, "usage_file", None), | ||
| resume=getattr(args, "resume", None), |
Contributor
There was a problem hiding this comment.
This forwards only explicit --resume. One-shot still bypasses parsed --continue and --no-restore-cwd, while normal cmd_chat() implements both; carry those options through this dispatch and the Termux fast path so the resume contract is consistent.
teknium1
added
the
sweeper:blast-moderate
blairjordan
added a commit
to blairjordan/hermes-agent
that referenced
this pull request
Aug 7, 2026
Three gaps in the oneshot resume path, found reviewing NousResearch#80799 against CONTRIBUTING.md and the review findings on sibling PR NousResearch#70136. 1. `--resume` bypassed the ID-or-title resolution contract. `_parser.py` documents `--resume SESSION` as "by ID or title" and `cmd_chat` resolves titles via `_resolve_session_by_name_or_id`. `_resolve_oneshot_resume` passed the value through verbatim, so `hermes --resume "my project" -z "..."` silently minted an empty session literally named "my project" instead of resuming — the same silently-drops-a-flag class as the bug being fixed, and the exact finding teknium1 gave NousResearch#70136. Now resolves first; an unresolved value still falls through to create-on-first-use, so the design intent (scripted callers minting stable ids) is preserved. The asymmetry was visible inside the PR itself: `--continue "name"` already resolved. 2. Create-on-first-use was an unguarded entry boundary for caller-supplied session ids. A session id becomes a filename downstream — `SessionDB._remove_session_files` builds `sessions_dir / f"{id}.json"` and globs `request_dump_{id}_*.json` without sanitizing — so `hermes --resume ../../x -z "hi"` minted a row that a later `sessions delete`/`prune` unlinks outside the sessions dir (verified: the victim file is removed). `gateway/session.py` already rejects exactly this at its own entry boundary (`_is_path_unsafe`, CWE-22); the new CLI boundary now does too. Path-unsafe values that *resolve* to a real session are still fine — the guard only applies to values about to become new ids, so titles containing `/` keep working. 3. The integration test asserted a dedup property it never exercised. `_RecordingFakeAgent` appended only its own two new messages, so "4 rows, not 8" held by construction regardless of whether the flush's identity-based seeded-history dedup worked. It now drives the real `AIAgent._flush_messages_to_session_db` with the real `messages = conversation_history + this turn` shape; break the identity contract and the assertion fails (verified by mutation). Also adds the coverage teknium1's review would still have flagged: `_resolve_oneshot_resume` had zero tests (title, exact id, unknown id, path-unsafe, `--continue` by name / bare / unmatched), a third-turn accumulation case, and the compression-chain redirect — a caller passing the pre-compression parent id must land on the continuation child, with this turn's writes going there and not into the dead parent. Docs: `-z` help and the `hermes_cli/oneshot` module docstring now state ID-or-title, create-on-first-use, and that an unmatched `--continue` errors while an unmatched `--resume` does not. tests/hermes_cli/test_oneshot_resume.py: 26 passed (was 9).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Prevents one-shot resume calls from silently creating or binding to the wrong durable session.
The top-level
--oneshotshortcut previously dropped--resumeand--pass-session-id; the one-shot runner therefore started a fresh agent without hydrating the requested transcript. Callers that tried to recover identity by selecting the newest same-directory session could then bind to a nested, delegated, tool, or otherwise unrelated session.Changes made
--resumeand--pass-session-idthrough the top-level one-shot path.session_metarecords before replay._compression_from.How to test
uv run --with pytest python -m pytest \ tests/hermes_state/test_resolve_resume_session_id.py \ tests/hermes_cli/test_oneshot_resume.py \ tests/hermes_cli/test_tui_resume_flow.py \ tests/hermes_cli/test_oneshot_usage_file.py \ -o addopts='' -qResult: 97 passed.
Type of change