fix(oneshot): honor --skills / --ignore-rules / --ignore-user-config in -z mode by beimingju-dafu · Pull Request #59402 · NousResearch/hermes-agent

beimingju-dafu · 2026-07-06T05:35:47Z

Summary

hermes -z accepts --skills, --ignore-rules, and --ignore-user-config on the CLI but silently drops all three before run_oneshot() — despite the module docstring in hermes_cli/oneshot.py:8 explicitly promising:

Rules / memory / AGENTS.md / preloaded skills = same as a normal chat turn.

This PR connects the three flags to the actual agent construction so -z mode behavior matches the docstring and mirrors what hermes chat already does with -s/--ignore-rules/--ignore-user-config.

Motivation

Downstream evaluation harness (libero) hit this bug hard:

Ran 20 smoke tests with hermes -z "..." --skills libero-decompose --ignore-user-config
Every system_prompt_len came out identical (31311) with all user MEMORY, USER PROFILE, and AGENTS.md content — the flags had no effect
Root-caused via source read: run_oneshot(prompt, model, provider, toolsets) signature never consumed skills/ignore-* args; main.py dispatch sites (lines 12458, 13891) didn't forward them either
After the fix, the model can quote SKILL body content verbatim in -z mode, which was previously impossible

Changes

hermes_cli/oneshot.py (+78 lines)
- Extend run_oneshot() signature with skills: Optional[str], ignore_rules: bool, ignore_user_config: bool
- Set HERMES_IGNORE_USER_CONFIG / HERMES_IGNORE_RULES env vars early (mirrors main.py:cmd_chat lines 2328-2343)
- Forward skip_context_files / skip_memory to AIAgent when --ignore-rules is set
- Inject preloaded skills into agent.ephemeral_system_prompt (mirrors cli.py:15190-15202); use ephemeral_system_prompt — not _cached_system_prompt — because the cached prompt is rebuilt during the conversation loop
- New helper _parse_skills_list() normalizes "a,b,c" / list / None
hermes_cli/main.py (+6 lines)
- Two run_oneshot() dispatch sites (lines 12458, 13891) now forward the three params
tests/cli/test_oneshot_preloaded_skills.py (+154 lines, new file)
- 10 tests covering _parse_skills_list (5), env var setting (3), skill injection (2)
- Verified against existing tests/cli/test_cli_preloaded_skills.py — no regression (3/3 chat mode tests still green)

Verification

# Model can now quote SKILL body content verbatim in -z mode:
hermes -z "quote the first sentence of principle #9 in libero-decompose SKILL" \
  --skills libero-decompose --ignore-rules --ignore-user-config
# → outputs the actual SKILL body text (previously impossible)

# --ignore-rules now removes MEMORY from state.db system_prompt:
# Before this fix: system_prompt_len = 31311 (with MEMORY/USER PROFILE)
# After:           system_prompt_len = 23478

# All tests:
pytest tests/cli/test_oneshot_preloaded_skills.py tests/cli/test_cli_preloaded_skills.py -v
# → 13 passed

Design notes

Why ephemeral_system_prompt and not _cached_system_prompt?
_cached_system_prompt is rebuilt by _restore_or_build_system_prompt() inside the conversation loop, which would overwrite any direct mutation. ephemeral_system_prompt is injected at API-call time (see agent/conversation_loop.py:490 and :815) after the cached prompt is resolved — the same mechanism chat mode uses (via HermesCLI.system_prompt → ephemeral_system_prompt in cli_agent_setup_mixin.py:359).
Why not just set env vars in main.py?
Env vars for --ignore-rules alone are insufficient — AIAgent also needs skip_context_files=True and skip_memory=True passed to its constructor (see agent/agent_init.py:1166-1170). Setting only env vars was tested and did not fully suppress MEMORY injection.
Behavior for unknown skills: prints stderr warning, does not abort (mirrors chat mode's tolerance in cli.py:15195-15197, though chat currently raises — this PR intentionally chooses tolerance for -z because scripted callers can't respond to interactive errors).

Backward compatibility

Adds only new optional parameters to run_oneshot() (default: None / False); existing callers unaffected
No changes to argparse — the three flags were already registered upstream
Chat mode untouched (verified by re-running test_cli_preloaded_skills.py)

Downstream impact

libero — a local-first AI time-management coach — depends on hermes -z for its evaluation harness. This bug caused all their smoke tests to run with polluted system_prompt (contaminated with user MEMORY that included decompose-specific preferences), producing a completely misleading 15% drift rate. After this fix + closing pollution, drift dropped to 5%. Full write-up: [link to your smoke report if you want to share]

…in -z mode The three parameters were accepted by argparse but silently dropped by run_oneshot(), contradicting the module docstring promise that 'Rules / memory / AGENTS.md / preloaded skills = same as a normal chat turn'. - run_oneshot() signature: add skills / ignore_rules / ignore_user_config - env vars HERMES_IGNORE_RULES / HERMES_IGNORE_USER_CONFIG set early - skip_context_files / skip_memory forwarded to AIAgent - SKILL body injected via ephemeral_system_prompt (matches chat mode) - main.py two dispatch sites forward the new params Verified: model can quote SKILL body verbatim when --skills is passed, which was previously impossible in -z mode. Carry patch — upstream PR pending.

alt-glitch · 2026-07-06T05:47:42Z

This was generated by AI during triage.

Related: #26633 / #31548 (oneshot drops --ignore-rules / --skills), open PR #26771 (--ignore-rules only) and open PR #51797 (--safe-mode + --ignore-rules isolation). This PR is the most comprehensive of the cluster (all three: --skills, --ignore-rules, --ignore-user-config) but #51797 additionally covers --safe-mode. Competing/overlapping fixes — flagging for a maintainer to pick or merge together.

beimingju-dafu · 2026-07-06T06:29:14Z

This was generated by AI during triage.

Related: #26633 / #31548 (oneshot drops --ignore-rules / --skills), open PR #26771 (--ignore-rules only) and open PR #51797 (--safe-mode + --ignore-rules isolation). This PR is the most comprehensive of the cluster (all three: --skills, --ignore-rules, --ignore-user-config) but #51797 additionally covers --safe-mode. Competing/overlapping fixes — flagging for a maintainer to pick or merge together.

Thanks for the triage @alt-glitch — helpful map.

To align expectations: our downstream (libero) eval harness is currently blocked by the silent system_prompt pollution, so we are looking for the cleanest, fastest path to land this 3-flag fix.

One quick clarification on the overlap with #51797 to help triage:

Where they'll conflict: Both PRs fix the --ignore-rules path in oneshot.py (forwarding skip_context_files / skip_memory to AIAgent). Whichever lands first, the other rebases on those lines.
Unique to fix(cli): preserve one-shot isolation flags #51797: --safe-mode gates plugin/MCP discovery before startup — a completely different lifecycle stage.
Unique to this PR: --skills (via ephemeral_system_prompt) and --ignore-user-config. These are the missing pieces promised by the oneshot.py docstring.

We heavily lean toward keeping this scoped to the 3 flags (Option 1). The discovery-gating in #51797 is a separable architectural stage. Letting #51797 land independently keeps both reviews narrow, and I am happy to rebase on top of it immediately.

Cc @hzhaoy for awareness on the potential rebase coordination. If the maintainers strongly prefer folding the safe-mode axis here instead, let me know.

teknium1

Thanks for tracing the oneshot path. The core premise is verified on current main: hermes_cli/_parser.py:200-232 accepts these flags, while the two run_oneshot() dispatches (hermes_cli/main.py:12675-12681, :14832-14838) still omit them and hermes_cli/oneshot.py:393-417 does not pass the normal CLI skip flags.

Problems

The new test asserts mock_agent.chat.called at tests/cli/test_oneshot_preloaded_skills.py:105, but the implementation calls agent.run_conversation() at PR hermes_cli/oneshot.py:450; the test will fail. Its scalar _run_agent stubs at test lines 57 and 68 also no longer match the (response, result) contract.
The warning at PR hermes_cli/oneshot.py:441 is emitted inside the stderr-to-devnull redirect established at line 226, so unknown skills remain silent in actual -z runs.
Current main added --usage-file support in 7dfd5077ceef4d5a6f7953c050bad1a75e86e215; the PR is conflicting and salvage must preserve that newer argument and reporting behavior.

Suggested changes

Rebase the implementation mechanically onto the current signature and preserve usage_file.
Update tests for run_conversation() and tuple returns, and assert the public forwarding plus the two AIAgent skip kwargs.
Surface all-missing skill failures or route the warning outside the redirect.

Automated hermes-sweeper review.

teknium1 · 2026-07-15T19:33:13Z

+
+        monkeypatch.delenv("HERMES_IGNORE_RULES", raising=False)
+
+        with patch("hermes_cli.oneshot._run_agent", return_value="mock response"):


_run_agent() returns (final_response, result) (also declared by this PR at hermes_cli/oneshot.py:300), but this scalar stub makes run_oneshot() hit its unpacking-error path. Return a tuple and assert _run_agent received the expected flag-derived behavior.

teknium1 · 2026-07-15T19:33:13Z

+            _run_agent("test prompt", skills="libero-decompose")
+
+        assert "SKILL_BODY libero-decompose" in mock_agent.ephemeral_system_prompt
+        assert mock_agent.chat.called


_run_agent() calls agent.run_conversation() at hermes_cli/oneshot.py:450, not agent.chat(). Set mock_agent.run_conversation.return_value to a result dict and assert that method instead; this assertion otherwise fails.

teknium1 · 2026-07-15T19:33:13Z

+            task_id=getattr(agent, "session_id", None),
+        )
+        if missing_skills:
+            sys.stderr.write(


This write occurs under redirect_stderr(devnull) from run_oneshot() line 226, so the advertised warning is never visible in a real -z invocation. Surface it through the saved stderr outside the redirect, or fail when every requested skill is missing as normal CLI does.

alt-glitch added type/bug Something isn't working P2 Medium — degraded but workaround exists comp/cli CLI entry point, hermes_cli/, setup wizard tool/skills Skills system (list, view, manage) area/config Config system, migrations, profiles labels Jul 6, 2026

teknium1 mentioned this pull request Jul 7, 2026

feat(cli): --safe-mode troubleshooting flag (Claude Code v2.1.169-inspired) #45300

Merged

This was referenced Jul 8, 2026

fix(oneshot): honor --max-turns / agent.max_turns in hermes -z #61152

Open

feat(cli): accept --max-turns on the oneshot/-z path #61566

Open

teknium1 reviewed Jul 15, 2026

View reviewed changes

teknium1 added sweeper:risk-compatibility Sweeper risk: may break existing users, config, migrations, defaults, or upgrades sweeper:blast-moderate Sweeper blast radius: moderate — a subsystem or single platform labels Jul 15, 2026

alt-glitch mentioned this pull request Jul 15, 2026

fix: make -z/--oneshot respect -s/--skills flag (#65119) #65249

Open

hey-ynot mentioned this pull request Jul 23, 2026

fix(cli): forward system prompt overlay in oneshot (-z) mode (supersedes #34863) #69957

Open

GottZ mentioned this pull request Jul 25, 2026

hermes -z (oneshot) discards --skills silently #31548

Open

This was referenced Jul 26, 2026

Global --skills flag does not inject skill content in one-shot -z runs #71759

Open

oneshot (-z) has no way to skip built-in memory injection — --ignore-rules is silently ignored #72064

Open

YuYigeng mentioned this pull request Jul 26, 2026

fix(cli): honor --ignore-rules in oneshot mode #72075

Open

GottZ mentioned this pull request Aug 3, 2026

fix(cli): honor -s/--skills in oneshot (-z) mode #63814

Open

webtecnica mentioned this pull request Aug 14, 2026

fix(memory): keep built-in memory store alive under --ignore-rules #86328

Open

This was referenced Aug 17, 2026

Scope proposal: unify agent isolation (--ignore-rules / --safe-mode) across TUI, messaging gateway, and API server #88214

Open

feat: unify agent isolation (--ignore-rules / --safe-mode) across TUI, messaging gateway, and API server #88360

Draft

Navigation Menu

Uh oh!

fix(oneshot): honor --skills / --ignore-rules / --ignore-user-config in -z mode - #59402

fix(oneshot): honor --skills / --ignore-rules / --ignore-user-config in -z mode#59402
beimingju-dafu wants to merge 1 commit into
NousResearch:mainfrom
beimingju-dafu:fix/oneshot-honor-skills-ignore-flags

beimingju-dafu commented Jul 6, 2026

Uh oh!

alt-glitch commented Jul 6, 2026

Uh oh!

beimingju-dafu commented Jul 6, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 15, 2026

Uh oh!

teknium1 Jul 15, 2026

Uh oh!

teknium1 Jul 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants


		monkeypatch.delenv("HERMES_IGNORE_RULES", raising=False)

		with patch("hermes_cli.oneshot._run_agent", return_value="mock response"):

Uh oh!

Conversation

beimingju-dafu commented Jul 6, 2026

Summary

Motivation

Changes

Verification

Design notes

Backward compatibility

Downstream impact

Uh oh!

alt-glitch commented Jul 6, 2026

Uh oh!

beimingju-dafu commented Jul 6, 2026

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants