fix(oneshot): honor resume state

[SE87H]

What changed

• forward --resume, --continue, and --no-restore-cwd through the top-level -z/--oneshot path
• resolve the durable session tip, restore its conversation history, reopen the same session row, and reuse its session ID
• restore the session working directory by default before constructing the agent
• synchronize TERMINAL_CWD after a successful workspace restore so prompt, file, and terminal resolution use the same resumed workspace
• make -z --continue prefer the current workspace before the historical global-MRU fallback, matching interactive CLI behavior
• add regression coverage for option forwarding, same-session continuation, workspace-scoped selection, unavailable workspaces, explicit CWD opt-out, and stale TERMINAL_CWD
• isolate the workspace-transition tests so their intentional production chdir() calls cannot leak into later tests

Root cause

The CLI parser accepted resume options for one-shot mode, but _run_and_exit_oneshot() and run_oneshot() did not forward or consume that state. A hermes -z --resume <id> invocation therefore created a fresh session instead of continuing the requested durable row.

The first patch also resolved bare --continue directly from the global CLI MRU. The follow-up scopes that lookup to the current Git workspace/CWD first, matching the established interactive resolver and preventing an unrelated project session from winning when the current workspace has history.

A later review identified a second workspace reference: Hermes prompt, file, and terminal resolution can prefer TERMINAL_CWD. Restoring only the process CWD could therefore resume the right history while tools still targeted a stale configured workspace. The current patch updates TERMINAL_CWD only after os.chdir() succeeds and before reopening the session or constructing the agent.

Impact

One-shot callers can now resume a persistent Hermes conversation without forking the session or losing its recorded workspace. Process CWD and TERMINAL_CWD are kept coherent for restored sessions. Failed workspace restoration leaves the prior environment and durable session untouched. --no-restore-cwd remains available for deliberate context changes. Bare --continue now selects the current workspace's latest CLI session when available.

Validation

• earlier focused regression suite: tests/hermes_cli/test_tui_resume_flow.py — 75 passed
• real explicit-resume E2E at that revision — same session row, no durable session-count increase, recorded working directory restored, cache reuse observed
• workspace-context transition tests cover stale TERMINAL_CWD, ordering before session reopen/agent construction, missing workspace, failed chdir, and explicit opt-out
• final test-isolation head: c55ce708ed2889ca3e9ec4529f6da9b1fd48b609
• exact code was validated in disposable fork PR SE87H/hermes-agent#4; the validation branch differed only by contributor-attribution mappings required by fork governance
• full CI run Endless Terminals Environment Integration #24: all 8 Python slices, Python E2E, desktop Playwright E2E, ruff/ty and Windows checks, OSV/supply-chain scans, lockfile and common-history gates passed
• one unrelated asynchronous TTS timing test failed on the first slice-7 attempt; only that job was rerun and passed without code changes
• disposable validation PR closed without merge after evidence capture
• focused security/public-exposure review — no private identifiers, local paths, or real credentials in the published diff

Scope

This PR intentionally remains a focused one-shot resume correction. A separate non-blocking follow-up comment proposes centralizing future workspace transitions behind typed task-local and process-local APIs; that architectural direction is not required for this fix and is not bundled here.

[alt-glitch]

This was generated by AI during triage.

Related: #57859 and #70136 already implement one-shot session hydration. This patch adds explicit CWD restoration but differs in missing-session and first-use behavior, so please consolidate the intended resume contract rather than treating the patches as duplicates.

[SE87H]

Consolidation decision after comparing #57859 and #70136:

• This PR intentionally uses existing-session-only, fail-closed resume semantics. --resume must resolve to a durable session; an unknown ID or unavailable store is an error. Silently minting a new ID or degrading to stateless execution would make a caller believe continuity was preserved when it was not.
• It keeps the useful --continue behavior from fix(cli): honor --resume/--continue in oneshot (-z) — hydrate and chain the session #57859, including latest-session and title resolution.
• It reuses the existing canonical resolve_resume_session_id() path rather than bundling the broader compression-lineage changes from fix: preserve one-shot resume context #70136 into this focused fix.
• It adds the missing workspace contract: restore the recorded CWD by default, with --no-restore-cwd as the explicit opt-out.
• Validation includes a real two-invocation explicit-resume E2E: same requested session ID in usage readback, no increase in durable session count, prior context restored, and recorded CWD restored. The focused regression suite is 73/73 passing.

So the intended contract is: exact durable identity + canonical tip + history + recorded CWD, or fail visibly; never claim a resume after silently creating/forking/statelessly continuing.

@teknium1, you reviewed the directly related #57859 path; could you review this consolidated, current-main variant?

[SE87H]

Follow-up pushed in 9c5274e9b: bare one-shot --continue now prefers the current Git workspace/CWD before the established global-MRU fallback, matching interactive CLI behavior. Added focused coverage for both workspace preference and fallback; the focused suite is now 75/75 passing. Maintainer write access to the fork branch has also been disabled; review remains possible through the PR normally.

[SE87H]

Follow-up hardening pushed in 591ca367c and 284e6bafd:

• CWD restoration now fails visibly when a resumed session records a workspace that no longer exists or cannot be entered, instead of silently continuing in the caller's current directory.
• The durable session is reopened only after workspace restoration succeeds, so a failed resume does not mutate the session state first.
• --no-restore-cwd remains the explicit opt-out and continues to resume the exact session/history without changing directories.
• Added focused regression coverage for both the unavailable-workspace failure and the explicit opt-out.

The two follow-up commits touch only hermes_cli/oneshot.py and the new focused test file. I could inspect the resulting diff here, but did not have a runnable repository checkout in this environment to rerun the pytest suite.

[teknium1]

Thanks for consolidating one-shot resume around the canonical session tip and adding fail-closed workspace handling. The current-main premise is valid: hermes_cli/oneshot.py:245-251 forwards no resume state, and hermes_cli/oneshot.py:410-443 creates a fresh session.

Problems

• hermes_cli/oneshot.py:389 changes only the process CWD. It does not retarget TERMINAL_CWD. The established interactive helper deliberately sets both (cli.py:7134-7142, cli.py:7178-7179), because agent/runtime_cwd.py:67-73 and tools/file_tools.py:314-321 prefer TERMINAL_CWD. With an inherited/configured value, this resumes the history but still gives prompt construction and file/terminal tools the prior workspace.

Suggested changes

• After successful os.chdir(saved_cwd), set os.environ["TERMINAL_CWD"] = saved_cwd.
• Add coverage with a distinct pre-set TERMINAL_CWD and assert it is changed before agent construction.

This is an automated hermes-sweeper review.

[teknium1]

Please also set os.environ["TERMINAL_CWD"] = saved_cwd after this succeeds. Interactive resume does this in cli.py:7178-7179; prompt, terminal, and file resolution prefer that variable, so os.chdir() alone can leave a resumed one-shot operating on the prior configured workspace.

[SE87H]

Addressed in e9db5835a: after os.chdir(saved_cwd) succeeds, one-shot resume now sets os.environ["TERMINAL_CWD"] = saved_cwd before the durable session is reopened and before AIAgent construction. Added focused coverage in b5123936c for stale-value replacement, missing/unenterable workspaces leaving the prior environment and session untouched, --no-restore-cwd preservation, and the value observed during agent construction. The implementation commit diff is limited to this synchronization. I reproduced the transition logic in an isolated environment (4/4); the full repository pytest run still requires a runnable checkout.

[SE87H]

Full validation is now complete. The exact PR code head b5123936cb13005ba446835ee46b5da5d24564f7 was tested against the same main base in a disposable internal PR; its only additional changes were contributor-attribution mappings required by the fork's governance check. CI run #19 passed all 8 Python slices (including the focused one-shot tests), Python E2E, desktop Playwright E2E, ruff/ty and Windows checks, OSV/supply-chain scans, and lockfile/history gates. The validation PR was closed without merge. Leaving this thread for reviewer confirmation rather than self-resolving it.

[SE87H]

Final verification on head c55ce708ed2889ca3e9ec4529f6da9b1fd48b609: the focused workspace tests now also restore the process CWD after each test, preventing their intentional production chdir() calls from leaking into later tests. Disposable validation PR SE87H/hermes-agent#4 completed full CI run #24 successfully; one unrelated asynchronous TTS timing flake passed when only its job was rerun, with no code change. Leaving the thread unresolved for reviewer confirmation.

[SE87H]

Current-main integration postflight:

• exact PR head b5123936cb13005ba446835ee46b5da5d24564f7 has completed the full repository CI successfully in the disposable validation described above
• upstream main has since advanced to 98105f31f46d3de58a8f69a2a439cee3f7a5e389; the PR branch is now 347 commits behind and GitHub reports real merge conflicts
• current main still lacks one-shot resume/continue state in run_oneshot, so the underlying defect remains present
• a non-destructive merge gate in the fork confirmed that the validated branch cannot be merged mechanically without resolving actively changed entry-point files

I am not force-updating or replacing current-main files from the older branch, because that would risk erasing newer upstream work. The next safe operation is a line-by-line rebase/cherry-pick in a real Git checkout, followed by the same full CI validation. The validated semantics and focused tests are ready; only current-main conflict resolution remains.

[SE87H]

Non-blocking architectural follow-up — not proposing to broaden this PR:

Hermes already has a centralized read side in agent/runtime_cwd.py, but workspace transitions are still written through several carriers: durable sessions.cwd, process CWD, TERMINAL_CWD, and task-local _SESSION_CWD.

A useful follow-up could be a small typed write-side API with two deliberately separate modes:

• task-local bind for concurrent gateway/cron/ACP hosts (ContextVar only; never mutate process CWD or process environment)
• process-local activate for finite CLI/one-shot execution (validate and enter the workspace, then publish the canonical TERMINAL_CWD, with fail-closed rollback semantics)

Resume resolution could also return one resolved context — canonical durable session ID, filtered history, recorded workspace — and let each surface apply the appropriate workspace mode before reopening the session or constructing the agent.

That would turn this split-brain class into one contract and test seam instead of requiring each surface to remember every carrier independently. The current PR should remain focused; this is only a possible follow-up design direction.

[SE87H]

Final third-eye validation pass:

• isolated the focused workspace tests so their production chdir() calls cannot leak process state into later tests (c55ce708e)
• revalidated the exact fix in disposable fork PR SE87H/hermes-agent#4
• all eight Python slices passed after rerunning only one unrelated flaky asynchronous TTS timing test; the one-shot resume/CWD suites were green on the original attempt
• Python E2E, desktop Playwright, ruff/ty, Windows checks, supply-chain/OSV, locks, attribution, and common-history gates all passed
• validation PR was closed without merge

No additional product behavior was added in this pass. The separate architectural comment is explicitly non-blocking and should not broaden this PR.

[GottZ]

This was generated by AI during triage.

Summary

Seven PRs cover two one-shot flag-loss causes: #26669, #26771, and #72075 propagate rule isolation into AIAgent, while #40333, #49204, #57859, and #74397 either hydrate resume state or reject unsupported resume combinations. The current implementations differ materially in Termux coverage, session creation semantics, workspace restoration, and whether --resume/--continue is honored or rejected.

Related pull requests

• fix(oneshot): honor --ignore-rules / HERMES_IGNORE_RULES on -z path (#26633) #26669 [closed] partial — (+266/-1) — n/a: This closed reference implementation forwards ignore_rules through both one-shot dispatches and maps it to skip_context_files and skip_memory; it remains relevant as the original tested fix, now superseded by the current-main salvage in fix(cli): honor --ignore-rules in oneshot mode #72075.
• fix(cli): honor --ignore-rules in oneshot mode #26771 partial — (+32/-3) — n/a: This wires the same isolation flags into AIAgent, but its diff covers only one dispatcher and uses the obsolete string return contract. Despite the keep-open review, fix(cli): honor --ignore-rules in oneshot mode #72075 preserves this PR's authorship while adding the missing Termux dispatch and current tuple/usage/cleanup behavior, so fix(cli): honor --ignore-rules in oneshot mode #26771 is superseded by fix(cli): honor --ignore-rules in oneshot mode #72075.
• fix(oneshot): honor --resume/--continue so -z can be session-aware #40333 [closed] fixes — (+270/-1) — n/a: This closed reference implementation hydrates prior history, resolves the live session tip, reopens the durable session, and reuses its ID; it remains relevant as the predecessor to the open hydration approaches, although its diff also includes unrelated MCP discovery.
• fix(cli): reject --resume/--continue in oneshot instead of dropping it #49204 best fix — (+58/-0) — n/a: This recorded best fix for hermes -z (oneshot) silently ignores --resume / --continue (session never hydrated) #49195 takes the minimal alternative of rejecting --resume/--continue at both one-shot dispatches instead of hydrating history. Its keep-open review identifies required author action: move the normal-path guard before startup and add normal-plus-Termux dispatch tests proving startup and run_oneshot are not called.
• fix(cli): honor --resume/--continue in oneshot (-z) — hydrate and chain the session #57859 best fix — (+198/-2) — n/a: This recorded best fix for hermes -z (oneshot) silently ignores --resume / --continue (session never hydrated) #49195 implements hydration, compression-tip resolution, same-session appends, and caller-minted create-on-first-use IDs. Its keep-open review requires a temp-HERMES_HOME two-process SQLite regression for durable reload and duplicate-row prevention, plus preservation of current usage_file plumbing.
• fix(cli): honor --ignore-rules in oneshot mode #72075 best fix — (+97/-7) — n/a: This recorded best fix for [Bug] hermes -z silently ignores --ignore-rules (oneshot path never reads HERMES_IGNORE_RULES) #26633 and oneshot (-z) has no way to skip built-in memory injection — --ignore-rules is silently ignored #72064 is the current-main salvage of fix(cli): honor --ignore-rules in oneshot mode #26771, covering both normal and Termux dispatches, the current (response, result) contract, safe mode, and both context and memory skip flags. The keep-open review found no blocking implementation defects.
• fix(oneshot): honor resume state #74397 best fix — (+655/-1) — n/a: This recorded best fix for hermes -z (oneshot) silently ignores --resume / --continue (session never hydrated) #49195 hydrates an existing canonical session tip and additionally restores the recorded workspace, synchronizes TERMINAL_CWD, scopes bare --continue to the current workspace, and fails closed for unavailable workspaces. The contributor's blocking TERMINAL_CWD concern and the keep-open review are addressed in the shown diff, but the documented current-main conflicts require a line-by-line rebase and renewed validation.

Duplicates

#26669, #26771, and #72075 implement substantially the same --ignore-rules propagation; #72075 is the explicit current-main salvage, so #26669 is the closed predecessor and #26771 can close as duplicate/superseded by #72075. #40333, #57859, and #74397 overlap on resume hydration but are not exact duplicates: #57859 adds create-on-first-use IDs, while #74397 requires an existing session and adds workspace restoration; #49204 is the competing fail-loud design rather than a duplicate.

Suggested consolidation

Keep #72075 open as the consolidated salvage path for #26633/#72064 and close #26771 as duplicate of #72075 despite its keep-open review, because #72075's diff supplies the review's missing Termux forwarding and current tuple contract. Author action: rebase #74397 onto current main, preserve its validated session-tip/CWD/TERMINAL_CWD behavior, and explicitly reconcile #57859's create-on-first-use contract; keep #57859 and #49204 open meanwhile because both are recorded best fixes with distinct supported designs, requiring respectively the requested SQLite E2E coverage and the requested pre-startup guard/dispatch tests before maintainers choose hydration versus fail-loud behavior.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I49195(["issue #49195 (open)"])
    P74397["PR #74397 (open)"]
    P74397 -->|best fix| I49195
    class I49195 open
    class P74397 open
    class P74397 best
    class P74397 target
    click I49195 "https://github.com/NousResearch/hermes-agent/issues/49195"
    click P74397 "https://github.com/NousResearch/hermes-agent/pull/74397"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 7 pull requests and 3 issues in this complex. Each diff was read against this issue; Assessment working set: 88 kB of PR diffs, 33 kB of issue/PR text, 18 kB of discussion (27 comments), 18 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[SE87H]

M6R fresh-main rebase complete.

• Rebased onto frozen main 4371712; main later advanced to fdc342c without touching concerned files.
• Resolved the semantic conflict by preserving current-main test topology and transplanting resume invariants into tests/hermes_cli/test_oneshot_resume_flow.py.
• 6 dedicated tests and 31 current resume/session/CWD tests passed; compile, Ruff, and diff-check passed.
• New head: 5931f54.