feat(gateway): expose subagent tool calls and thinking to user (fixes #169) by cutepawss · Pull Request #186 · NousResearch/hermes-agent

cutepawss · 2026-02-28T14:10:29Z

PR: feat(gateway): expose subagent tool calls and thinking to user

Fixes #169

What & Why

When subagents run via delegate_task, the user sits with zero progress indication — just a static spinner for potentially minutes. This PR relays the child agent's tool calls and reasoning to the parent display in real time.

Implementation

Three minimal, isolated changes:

1. `agent/display.py` — `KawaiiSpinner.print_above()`

New method that prints a line above the active spinner without disrupting animation. Uses the captured self._out reference (not sys.stdout) so it works inside the child's redirect_stdout(devnull).

2. `tools/delegate_tool.py` — `_build_child_progress_callback()`

Builds a per-child callback with two display paths:

CLI: prints tree-view lines above the delegation spinner:

  🔀 Delegating: research quantum computing
     ├─ 💭 "I'll search for papers first..."
     ├─ 🔍 web_search  "quantum computing"
     ├─ 📖 read_file  "paper.pdf"
     └─ ⠹ working... (18.2s)

Gateway (Telegram/Discord): batches tool names and sends a summary every 5 calls to avoid message spam. Remaining items are flushed when the subagent completes.

Returns None when no display mechanism exists → zero behavior change for headless/test use.

3. `run_agent.py` — thinking callback

After each API response, fires tool_progress_callback("_thinking", first_line) with the model's first line of text content. <REASONING_SCRATCHPAD>, <think>, and <reasoning> XML tags are stripped before display.

What This Does NOT Change

quiet_mode=True and redirect_stdout(devnull) on child agents — unchanged
Existing spinner behavior for non-delegation tools — unchanged
Existing completion summary (┊ 🔀 delegate "goal" 18.4s) — unchanged
Any agent/API logic — unchanged

How to Test

# Run test suite (20 new tests)
pytest tests/agent/test_subagent_progress.py -v

# Full suite
pytest tests/ -v
# 638 passed, 1 failed (pre-existing, unrelated)

# Manual: run hermes CLI, ask it something that triggers delegation
hermes
> Research the latest advances in quantum computing and summarize
# Should see tree-view lines above the spinner during subagent work

Platform Tested

macOS (Python 3.14)

Checklist

Fixes the reported issue (Feature Request: Expose subagent actions/thoughts to the user #169)
Adds 20 regression tests
All tests pass
No breaking changes
Follows conventional commits (feat(gateway):)
Follows CONTRIBUTING.md PR description format

…ousResearch#169) When subagents run via delegate_task, the user now sees real-time progress instead of silence: CLI: tree-view activity lines print above the delegation spinner 🔀 Delegating: research quantum computing ├─ 💭 "I'll search for papers first..." ├─ 🔍 web_search "quantum computing" ├─ 📖 read_file "paper.pdf" └─ ⠹ working... (18.2s) Gateway (Telegram/Discord): batched progress summaries sent every 5 tool calls to avoid message spam. Remaining tools flushed on subagent completion. Changes: - agent/display.py: add KawaiiSpinner.print_above() to print status lines above an active spinner without disrupting animation. Uses captured stdout (self._out) so it works inside the child's redirect_stdout(devnull). - tools/delegate_tool.py: add _build_child_progress_callback() that creates a per-child callback relaying tool calls and thinking events to the parent's spinner (CLI) or progress queue (gateway). Each child gets its own callback instance, so parallel subagents don't share state. Includes _flush() for gateway batch completion. - run_agent.py: fire tool_progress_callback with '_thinking' event when the model produces text content. Guarded by _delegate_depth > 0 so only subagents fire this (prevents gateway spam from main agent). REASONING_SCRATCHPAD/think/ reasoning XML tags are stripped before display. Tests: 21 new tests covering print_above, callback builder, thinking relay, SCRATCHPAD filtering, batching, flush, thread isolation, delegate_depth guard, and prefix handling.

Two fixes to the subagent progress display from PR #186: 1. Task index prefix: show 1-indexed prefix ([1], [2], ...) for ALL tasks in batch mode (task_count > 1). Single tasks get no prefix. Previously task 0 had no prefix while others did, making batch output confusing. 2. Completion indicator: use spinner.print_above() instead of raw print() for per-task completion lines (✓ [1/2] ...). Raw print collided with the active spinner, mushing the completion text onto the spinner line. Now prints cleanly above. Added task_count parameter to _build_child_progress_callback and _run_single_child. Updated tests accordingly.

…ousResearch#169) (NousResearch#186) When subagents run via delegate_task, the user now sees real-time progress instead of silence: CLI: tree-view activity lines print above the delegation spinner 🔀 Delegating: research quantum computing ├─ 💭 "I'll search for papers first..." ├─ 🔍 web_search "quantum computing" ├─ 📖 read_file "paper.pdf" └─ ⠹ working... (18.2s) Gateway (Telegram/Discord): batched progress summaries sent every 5 tool calls to avoid message spam. Remaining tools flushed on subagent completion. Changes: - agent/display.py: add KawaiiSpinner.print_above() to print status lines above an active spinner without disrupting animation. Uses captured stdout (self._out) so it works inside the child's redirect_stdout(devnull). - tools/delegate_tool.py: add _build_child_progress_callback() that creates a per-child callback relaying tool calls and thinking events to the parent's spinner (CLI) or progress queue (gateway). Each child gets its own callback instance, so parallel subagents don't share state. Includes _flush() for gateway batch completion. - run_agent.py: fire tool_progress_callback with '_thinking' event when the model produces text content. Guarded by _delegate_depth > 0 so only subagents fire this (prevents gateway spam from main agent). REASONING_SCRATCHPAD/think/ reasoning XML tags are stripped before display. Tests: 21 new tests covering print_above, callback builder, thinking relay, SCRATCHPAD filtering, batching, flush, thread isolation, delegate_depth guard, and prefix handling.

Two fixes to the subagent progress display from PR NousResearch#186: 1. Task index prefix: show 1-indexed prefix ([1], [2], ...) for ALL tasks in batch mode (task_count > 1). Single tasks get no prefix. Previously task 0 had no prefix while others did, making batch output confusing. 2. Completion indicator: use spinner.print_above() instead of raw print() for per-task completion lines (✓ [1/2] ...). Raw print collided with the active spinner, mushing the completion text onto the spinner line. Now prints cleanly above. Added task_count parameter to _build_child_progress_callback and _run_single_child. Updated tests accordingly.

…ousResearch#169) (NousResearch#186) When subagents run via delegate_task, the user now sees real-time progress instead of silence: CLI: tree-view activity lines print above the delegation spinner 🔀 Delegating: research quantum computing ├─ 💭 "I'll search for papers first..." ├─ 🔍 web_search "quantum computing" ├─ 📖 read_file "paper.pdf" └─ ⠹ working... (18.2s) Gateway (Telegram/Discord): batched progress summaries sent every 5 tool calls to avoid message spam. Remaining tools flushed on subagent completion. Changes: - agent/display.py: add KawaiiSpinner.print_above() to print status lines above an active spinner without disrupting animation. Uses captured stdout (self._out) so it works inside the child's redirect_stdout(devnull). - tools/delegate_tool.py: add _build_child_progress_callback() that creates a per-child callback relaying tool calls and thinking events to the parent's spinner (CLI) or progress queue (gateway). Each child gets its own callback instance, so parallel subagents don't share state. Includes _flush() for gateway batch completion. - run_agent.py: fire tool_progress_callback with '_thinking' event when the model produces text content. Guarded by _delegate_depth > 0 so only subagents fire this (prevents gateway spam from main agent). REASONING_SCRATCHPAD/think/ reasoning XML tags are stripped before display. Tests: 21 new tests covering print_above, callback builder, thinking relay, SCRATCHPAD filtering, batching, flush, thread isolation, delegate_depth guard, and prefix handling.

Two fixes to the subagent progress display from PR NousResearch#186: 1. Task index prefix: show 1-indexed prefix ([1], [2], ...) for ALL tasks in batch mode (task_count > 1). Single tasks get no prefix. Previously task 0 had no prefix while others did, making batch output confusing. 2. Completion indicator: use spinner.print_above() instead of raw print() for per-task completion lines (✓ [1/2] ...). Raw print collided with the active spinner, mushing the completion text onto the spinner line. Now prints cleanly above. Added task_count parameter to _build_child_progress_callback and _run_single_child. Updated tests accordingly.

… proof, living models The forecasting-ability arc: make the SYSTEM research like a superforecaster, make the OPERATOR one, and measure both honestly. R1 — VOI-directed research + adequacy judge (forecasting/research_audit.py): - build_research_plan turns the question's OWN levers (update_triggers, change_my_mind clauses, outcome_paths) into hunt angles alongside the standard four (primary source, base rate, recent, contrarian) — pure template work, no LLM - audit_research: six deterministic checks (reference class, evidence floor, source independence, disconfirming stance via the existing EvidenceItem stance field, recency, trigger coverage) + an optional fail-open LLM change_my_mind-coverage check; weighted 0-100 score - the autonomous chain re-runs research (max_audit_rounds=2, gap list injected via a new supplemental seam) keyed ONLY on research-stage-controllable gaps (reference classes are a base_rate artifact — correctly excluded), with a no-progress spend guard - research_adequate hook rule (lenient OFF / standard WARN / strict ERROR, participates in the resolved-policy pass); research_plan/research_audit tool actions; protocol teaches the loop R2 — operator practice loop (the goal is the USER becoming the forecaster): - operator_estimates table; resolution scores the human's numbers (fail-open, binary Brier; non-binary marked-skipped honestly) - forecast practice <ref> and forecast drill --n (replay resolved questions with pre-resolution evidence only, instant feedback, polite non-TTY refusal) - operator_calibration_summary: personal curve + trend + vs_system pairing on shared questions; calibration --operator CLI + gateway payload section - estimate-first mode strictly opt-in (forecasting.practice.estimate_first) R3 — market-hidden benchmark arm completed (task NousResearch#186): the withhold core was found already committed (verified, not duplicated); added the missing scoring report market_hidden_pool_report — agent vs withheld-market Brier, win-rate, equal-weight log-odds pool beats-both, and the P1.3 fitted-simplex LOO complementarity (reused, not reimplemented). Hand-verified the leak: the hidden arm exposes neither the price nor the marker to the agent while keeping both for scoring; the open-arm control proves the probe bites. R4 — living models: model_runs scored at resolution (binary Brier / numeric interval coverage + abs error; a refresh run's own proposal can never masquerade as a model observation), on-read model_skill (track_record math, cold-start 1.0, min 10 resolved), skill multipliers applied at the deterministic re-pool with audit-trailed metadata (config default ON, identity on cold start), nightly model refresh default ON. Adversarial review: 4 findings — 1 real (uncached model_skill rescan per refresh; cached) and 3 disproved with documented reasoning. tests/forecasting: 2070 passed (+103). E2E journey harness green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

cutepawss force-pushed the feat/subagent-progress branch from 862ccd7 to 62bc53c Compare February 28, 2026 14:15

teknium1 merged commit dd69f16 into NousResearch:main Mar 1, 2026

teknium1 mentioned this pull request Mar 1, 2026

feat: expose subagent tool activity as persistent CLI lines #187

Closed

This was referenced Aug 9, 2026

fix(security): patch website transitive dependencies dizhaky/hermes-agent#157

Merged

fix(deps): resolve August Dependabot alerts dizhaky/hermes-agent#158

Closed

Navigation Menu

Uh oh!

feat(gateway): expose subagent tool calls and thinking to user (fixes #169) - #186

feat(gateway): expose subagent tool calls and thinking to user (fixes #169)#186
teknium1 merged 1 commit into
NousResearch:mainfrom
cutepawss:feat/subagent-progress

cutepawss commented Feb 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

cutepawss commented Feb 28, 2026

PR: feat(gateway): expose subagent tool calls and thinking to user

What & Why

Implementation

1. agent/display.py — KawaiiSpinner.print_above()

2. tools/delegate_tool.py — _build_child_progress_callback()

3. run_agent.py — thinking callback

What This Does NOT Change

How to Test

Platform Tested

Checklist

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

1. `agent/display.py` — `KawaiiSpinner.print_above()`

2. `tools/delegate_tool.py` — `_build_child_progress_callback()`

3. `run_agent.py` — thinking callback