fix(update): preserve and rebase committed local patches, and report real rebase conflicts

[mhuttes14]

Problem

hermes update assumes all local work is uncommitted (and therefore stashed). An install whose main is deliberately ahead of origin — carrying committed local patches — hits the ff-only-pull failure path, which unconditionally runs a hard reset to origin/<branch> and silently destroys that work.

What this does

Commit 1 — preserve committed local patches. Adds updates.committed_local_changes (refuse | rebase, default refuse, so behavior is unchanged unless opted in):

• refuse — if local main is ahead, stop with reconcile instructions instead of discarding commits.
• rebase — create a durable backup ref (refs/hermes/update-backups/<branch>/<ts>, pruned to the newest 5 by reflog time), rebase the patch stack onto the remote, and continue the normal pipeline. On any failure it aborts, verifies HEAD is restored to the pre-pull SHA, and keeps the backup ref.

The historical reset is retained verbatim for the no_local_commits case, which is the only situation it was ever correct for.

Commit 2 — report what actually conflicted. When the rebase fails, the operator was shown one line: Rebasing (1/6).

git splits rebase-conflict output across streams: stdout carries CONFLICT (content): Merge conflict in <path>; stderr carries Rebasing (n/m), error: could not apply <sha>... and hint: lines. The failure path read (stderr or stdout) — stderr is non-empty, so stdout was discarded wholesale — and then took .splitlines()[0], which is the progress banner. Every actionable detail was dropped twice.

_rebase_conflict_summary scans both streams for the failing commit and reads unmerged paths from the index before git rebase --abort destroys that state, falling back to scraping stdout CONFLICT lines for rename/delete conflicts that leave no unmerged entries.

Before:

✗ Could not rebase committed local patches automatically.
  Rebasing (1/6)

After:

✗ Could not rebase committed local patches automatically.
  error: could not apply f231e8d... local patch
  Conflicting file(s): conflict.txt
  ✓ Rebase aborted; the original committed state is restored.
    Resolve by hand with:
      git -C <root> rebase --empty=drop --no-keep-empty origin/main

Testing

103 tests pass on this branch against current main (test_update_committed_changes.py, test_update_autostash.py, test_cmd_update.py), including new coverage for the conflict-reporting path and the rename/delete fallback.

Context

Both commits have been running in production on a self-hosted install that carries local patches; the rebase path reconciled a 332-commit upstream jump successfully, and the conflict reporting was what made a real two-patch conflict diagnosable.

🤖 Generated with Claude Code

[teknium1]

Thanks for addressing a real destructive updater path. Current main still hard-resets after a failed fast-forward merge at hermes_cli/update_cmd.py:3431-3440, so the underlying fix remains needed.

Problems

• This branch predates updater extraction commit 927463efcc: the live implementation is now hermes_cli/update_cmd.py:_cmd_update_impl (:3052), while hermes_cli/main.py only re-exports updater symbols (:4991). The reconciliation code must be ported rather than merged in its original location.
• Default config data moved in 1fe06115d1; add the new setting beside updates.non_interactive_local_changes in hermes_cli/config_defaults.py:2674, not the old default block in config.py.
• The cross-referenced fix(update): reconcile local branches during update #22037 local-branch case remains separate: current main switches any non-target branch at update_cmd.py:3246, and this diff reconciles only the target branch after that switch.

Suggested changes

• Port the helpers and failed-merge branch into update_cmd.py, preserving its _m() test/compatibility seam, then run the retained integration tests against the relocated path.
• Clarify the fix(update): reconcile local branches during update #22037 scope or cover it in a focused follow-up.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Three PRs address unsafe handling of committed local patches across two checkout shapes: #4142 and #70953 protect commits on the update branch from the failed-fast-forward hard-reset path, while #22037 reconciles a separate local-only branch that would otherwise be left inactive. #70953 is the stronger successor to #4142, whereas #22037 covers a complementary cause.

Related pull requests

• feat(update): rebase local commits instead of discarding on update #4142 [closed] duplicate — (+57/-16) — n/a: Detects commits ahead on the update branch and rebases instead of hard-resetting them; although closed pending a fresh rebase, it remains relevant as the earlier implementation now superseded by fix(update): preserve and rebase committed local patches, and report real rebase conflicts #70953.
• fix(update): reconcile local branches during update #22037 related — (+283/-1) — n/a: Reconciles a separate local-only branch using patch-id comparison, but the contributor keep_open review remains blocking because git cherry failure is treated as an empty comparison, the diff hard-codes main instead of the resolved update target, and conflict/error recovery lacks tests. Its distinct branch-reactivation logic remains salvageable.
• fix(update): preserve and rebase committed local patches, and report real rebase conflicts #70953 related — (+760/-29) — n/a: Adds fail-closed refusal or backed-up rebase for commits on the update branch, with verified abort recovery and real-Git conflict tests. Consistent with the maintainer-bot keep_open verdict, the implementation must be ported from obsolete hermes_cli/main.py and config.py locations to hermes_cli/update_cmd.py using the _m() seam and hermes_cli/config_defaults.py; it does not cover fix(update): reconcile local branches during update #22037's separate-branch case.

Duplicates

#4142 is substantially duplicated and superseded by #70953 for update-branch divergence; #22037 is complementary because it handles a separate local-only branch.

Suggested consolidation

Author action on #70953: rebase onto current main and port the preserved-commit helpers, failed-merge handling, configuration default, and retained integration tests to the live updater/config locations identified by the maintainer-bot keep_open verdict. Keep #22037 open with a salvage path for separate-branch reconciliation, but require fail-closed git cherry handling, resolved-target support, and comparison-failure and rebase-conflict tests; #4142 can remain closed as a duplicate of #70953.

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
    subgraph Dup4142 ["PRs duplicating each other"]
        P4142["PR #4142 (closed)"]
        P70953["PR #70953 (open)"]
    end
    class P4142 closed
    class P70953 open
    class P70953 target
    click P4142 "https://github.com/NousResearch/hermes-agent/pull/4142"
    click P70953 "https://github.com/NousResearch/hermes-agent/pull/70953"

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 3 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 61 kB of PR diffs, 6 kB of issue/PR text, 3 kB of discussion (3 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.