fix: deliver leftover /steer as out-of-band user message (#60543) by astraltrekkin · Pull Request #60555 · NousResearch/hermes-agent

astraltrekkin · 2026-07-07T23:19:58Z

What does this PR do?

Fixes a /steer message being lost / delivered without its trust marker when it arrives after the final assistant turn.

/steer has three delivery paths. Two wrap the user's message in the [OUT-OF-BAND USER MESSAGE] marker so the model treats it as a genuine user instruction rather than tool output / prompt injection (apply_pending_steer_to_tool_results, and the pre-API drain in conversation_loop.py). The third — the "leftover" path, taken when a steer lands after the final assistant turn with no remaining tool batch to drain into — did not. turn_finalizer.py hands the raw steer back as result["pending_steer"], and both consumers delivered it verbatim as the next user turn, so the model saw bare /steer text instead of the marker.

On the gateway path there was a sharper failure: its downstream slash-command discard net drops any pending text starting with /, so a leftover steer whose text begins with / (e.g. steering with /stop the search) was silently discarded. Wrapping it in the marker means it no longer starts with /, so it survives and is delivered.

Approach: add one shared helper next to the existing steer helpers and route both consumers through it, rather than duplicating the wrap at each call site.

Related Issue

Fixes #60543

Type of Change

🐛 Bug fix (non-breaking change that fixes an issue)
✨ New feature (non-breaking change that adds functionality)
🔒 Security fix
📝 Documentation update
✅ Tests (adding or improving test coverage)
♻️ Refactor (no behavior change)
🎯 New skill (bundled or hub)

Changes Made

agent/agent_runtime_helpers.py — add format_leftover_steer_for_delivery() (format_steer_marker(text).strip(); the marker is the whole standalone message here, so the append-oriented leading newlines are stripped) and export it.
cli.py — CLI REPL leftover handler now delivers format_leftover_steer_for_delivery(_leftover_steer) instead of the raw text.
gateway/run.py — gateway leftover handler (TUI / Telegram / Slack / WhatsApp) routed through the same helper.
tests/run_agent/test_steer_leftover_marker.py — new regression suite (6 cases).

How to Test

In the TUI, start a conversation that triggers parallel tool calls.
Send /steer "..." so it lands after the final tool batch (the leftover window).
Confirm it reaches the model wrapped as [OUT-OF-BAND USER MESSAGE], not as raw /steer text — and that a steer beginning with / is delivered rather than silently dropped.

Automated: tests/run_agent/test_steer_leftover_marker.py covers both delivery paths, the marker contract, and the gateway slash-command net. Each assertion was confirmed to fail against the pre-fix raw passthrough (red → green).

Checklist

Code

I've read the Contributing Guide
My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
I searched for existing PRs to make sure this isn't a duplicate
My PR contains only changes related to this fix/feature (no unrelated commits)
I've run pytest tests/ -q and all tests pass — CI Python tests green on Linux (amd64 + arm64); targeted steer/finalizer/gateway suites (52) pass locally on macOS.
I've added tests for my changes (required for bug fixes, strongly encouraged for features)
I've tested on my platform: macOS 26.5 (arm64), Python 3.12

Documentation & Housekeeping

I've updated relevant documentation (README, docs/, docstrings) — new helper is documented via docstring; no user-facing docs affected — or N/A
N/A — no config keys added/changed (cli-config.yaml.example)
N/A — no architecture/workflow changes (CONTRIBUTING.md / AGENTS.md)
I've considered cross-platform impact (Windows, macOS) — change is platform-agnostic (no file I/O, process, or shell surface touched)
N/A — no tool behavior/schema changes

Screenshots / Logs

tests/run_agent/test_steer_leftover_marker.py — 6 passed. Full pytest tests/ -q result noted in the checklist above.

…usResearch#60543) A /steer that lands after the final assistant turn — with no remaining tool batch to drain into — is returned as result["pending_steer"] and delivered as the next user turn. Both delivery sites forwarded it as raw text, so the model received bare /steer text instead of the [OUT-OF-BAND USER MESSAGE] marker the mid-batch and pre-API drain paths use. On the gateway path, a steer whose text starts with "/" could additionally be swallowed by the slash-command discard net. Add format_leftover_steer_for_delivery() beside the existing steer helpers and route both consumers (cli.py, gateway/run.py) through it. Tests: tests/run_agent/test_steer_leftover_marker.py (6 cases) cover both paths, the marker contract, and the gateway slash-command net. 52 passing across the steer, concurrent-interrupt, tui-gateway-queue and turn_finalizer suites.

…pping it A /steer that lands after the final tool batch (e.g. during the final API call when the model returns text with no further tool calls) cannot be injected into any tool result. run_conversation() hands it back in result["pending_steer"]; gateway/run.py and cli.py re-deliver it as the next user turn, but the desktop/web surface (tui_gateway) never read it — the UI accepted the steer (showed 'queued'), then silently lost it. Add _deliver_leftover_steer(): capture pending_steer in _run_prompt_submit's run() closure and dispatch it in the turn tail, after _drain_queued_prompt (an explicit queued/interrupt prompt wins). If a fresh user turn already claimed the session, re-stash the steer via agent.steer() so that turn's own drain picks it up. Upstream context: NousResearch#60543 tracks the leftover- steer mechanism for CLI/gateway (PR NousResearch#60555, unmerged); the tui_gateway consumer gap is unreported upstream as of c552984.

teknium1

Thanks for covering the verified CLI and gateway raw-delivery paths. Current main still has the raw passthroughs at cli.py:12884-12888 and gateway/run.py:20028-20032, and the marker approach preserves the existing trusted-steer contract from agent/prompt_builder.py:595-614.

Problems

The standard Ink TUI remains unfixed. It sends /steer through tui_gateway/server.py:8427-8449, while _run_prompt_submit() calls run_conversation() at tui_gateway/server.py:9103 and processes the result at 9148-9205 without reading pending_steer; the tail only drains queued prompts at 9376-9380. This still drops a leftover steer for the reported TUI flow.
The regression tests duplicate the intended delivery statements (tests/run_agent/test_steer_leftover_marker.py:198-202 and 229-250) instead of invoking cli.py or gateway/run.py, so they do not prove those production call sites are wired.

Suggested changes

Add the tui_gateway consumer after _drain_queued_prompt() so an explicit queued prompt retains priority.
Test the actual CLI, gateway, and TUI delivery seams, including a /stop … leftover through the gateway safety net.

Automated hermes-sweeper review.

teknium1 · 2026-07-15T22:14:37Z

+    pending_input: "queue.Queue[str]" = queue.Queue()
+    _leftover_steer = result.get("pending_steer") if result else None
+    if _leftover_steer:
+        pending_input.put(format_leftover_steer_for_delivery(_leftover_steer))


This mirrors the proposed CLI statement rather than invoking the CLI handler, so it still passes if cli.py forwards raw pending_steer. Please exercise a production delivery seam (or a helper actually called by it) to make the regression test wiring-sensitive.

astraltrekkin · 2026-07-16T14:23:41Z

@teknium1 Adressed both:

Ink TUI gap: The leftover steer was still dropped in the Ink TUI. Fixed by delivering it after queued prompts, with the same OOB marker.
Tests: they weren't wired to the real call sites. Updated them to hit the CLI / gateway / TUI seams, including the /stop leftover case

Cover the Ink TUI leftover path after queued prompts, and exercise the real CLI/gateway/TUI seams so regressions catch raw pending_steer passthrough. Co-authored-by: Cursor <cursoragent@cursor.com>

null-runner · 2026-07-17T12:17:51Z

Verified on a clean origin/main checkout (worktree, patch applied via git apply).

Repro on the tui_gateway/desktop path (independent of the CLI/gateway repro in #60543): a /steer landing in the turn-teardown window (after message.complete is emitted, before running is reset to False) is stashed by agent.steer(); the turn ends without consuming it and turn_finalizer returns it in result["pending_steer"]. On main, gateway/run.py consumes result["pending_steer"] but tui_gateway/server.py has zero references to it (git grep pending_steer origin/main -- tui_gateway/server.py = 0), so on the Ink/desktop surface the leftover steer is silently dropped. _deliver_leftover_steer closes that path.

Semantics match the gateway path: all three seams (cli/gateway/tui) now wrap the leftover via the shared format_leftover_steer_for_delivery (OOB marker); queued prompts keep priority (_deliver_leftover_steer runs after _drain_queued_prompt and re-checks session["running"] under history_lock, mirroring resolve_gateway_leftover_steer yielding when pending/pending_event is set); the turn is claimed atomically to avoid a double-fire.

Tests: tests/run_agent/test_steer_leftover_marker.py = 11/11 pass, including the TUI seam tests. The rest of tests/tui_gateway/ is green except test_subagent_child_mirror::test_prompt_submit_rejected_while_child_run_active and test_projects_rpc::test_discover_repos_from_full_history, which are pre-existing flakes — they fail under the full suite on unpatched origin/main too and pass in isolation, unrelated to this change.

Minimal fix for #60543 on the tui path; #65205 covers the broader queue unification (agent-side TurnQueue).

teknium1 reviewed Jul 15, 2026

View reviewed changes

teknium1 added the sweeper:blast-contained Sweeper blast radius: contained — one narrow path / opt-in / few users label Jul 15, 2026

fix: deliver leftover /steer in TUI and wire delivery-seam tests

045ab05

Cover the Ink TUI leftover path after queued prompts, and exercise the real CLI/gateway/TUI seams so regressions catch raw pending_steer passthrough. Co-authored-by: Cursor <cursoragent@cursor.com>

Xipong mentioned this pull request Jul 17, 2026

fix(agent): preserve /steer through turn-budget enforcement #66401

Closed

GottZ mentioned this pull request Jul 29, 2026

/steer race condition: steer lost when arriving between tool batch drain and next API call #60543

Closed

burak33bb mentioned this pull request Aug 4, 2026

fix(agent): prevent historical OOB steer replay #78386

Closed

teknium1 mentioned this pull request Aug 15, 2026

fix(tui_gateway): stop live-turn replay after redirect and preserve steer-mode message bursts #86593

Merged

13 tasks

alt-glitch mentioned this pull request Aug 28, 2026

[Bug]: /steer - silently discards messages when no turn is running #97371

Open

1 task

burak33bb mentioned this pull request Sep 10, 2026

fix(agent): preserve accepted steers across soft interrupt clears #107234

Merged

12 tasks

Navigation Menu

fix: deliver leftover /steer as out-of-band user message (#60543) - #60555

fix: deliver leftover /steer as out-of-band user message (#60543)#60555
astraltrekkin wants to merge 2 commits into
NousResearch:mainfrom
astraltrekkin:fix/steer-leftover-oob-marker

astraltrekkin commented Jul 7, 2026 •

edited

Loading

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 15, 2026

Uh oh!

astraltrekkin commented Jul 16, 2026

Uh oh!

null-runner commented Jul 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

astraltrekkin commented Jul 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Related Issue

Type of Change

Changes Made

How to Test

Checklist

Code

Documentation & Housekeeping

Screenshots / Logs

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!

astraltrekkin commented Jul 16, 2026

Uh oh!

null-runner commented Jul 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

astraltrekkin commented Jul 7, 2026 •

edited

Loading