fix(acp): wire stream_delta_callback for real-time ACP text delta streaming by Git-on-my-level · Pull Request #13810 · NousResearch/hermes-agent

Git-on-my-level · 2026-04-22T03:55:22Z

Root Cause

ACP turns emit partial session/update progress (thinking spinners, tool start/complete events, interim commentary), then go silent for the entire duration of run_conversation() before the final output arrives via prompt_return instead of a streamed terminal lifecycle event.

Why: `stream_delta_callback` was never wired in the ACP adapter

The ACP adapter (acp_adapter/server.py) sets several callbacks on AIAgent:

Callback	Set by ACP?	Called by `run_agent.py`?	Result
`thinking_callback`	✅	✅	Works — spinner text
`tool_progress_callback`	✅	✅	Works — ToolCallStart
`step_callback`	✅	✅	Works — ToolCallComplete
`interim_assistant_callback`	✅	✅ (via `_emit_interim_assistant_message`)	Works — commentary
`message_callback`	✅	❌ Never called	Dead code
`stream_delta_callback`	❌ NOT SET	✅ (`_fire_stream_delta`)	Text deltas silently dropped

The critical gap: _fire_stream_delta() (line 5211 in run_agent.py) is the method that fires for every LLM streaming token. It only dispatches to stream_delta_callback and _stream_callback. Since ACP sets neither, every text delta from the LLM is silently discarded.

CAR's completion_source="prompt_return" (not stream_terminal_event) confirms no timely terminal lifecycle signal was received — it only got the final result through the RPC return path, ~30 minutes after the last progress update.

Fix

Add make_stream_delta_cb() factory in acp_adapter/events.py that forwards each text delta to conn.session_update() as an agent_message_chunk
Wire it as agent.stream_delta_callback in the prompt() method
Add unit + integration tests

Files Changed

acp_adapter/events.py — new make_stream_delta_cb() factory (+17 lines)
acp_adapter/server.py — wire callback in prompt() (+4 lines)
tests/acp/test_events.py — TestStreamDeltaCallback class (+28 lines)
tests/acp/test_server.py — integration test test_prompt_wires_stream_delta_callback (+30 lines)

Fixes #9

…t, suppress build stderr Group 1 patches re-applied on upstream main (e94008c): - hermes_cli/update_channel.py: new channel detection (prod vs main) - hermes_cli/main.py: channel-aware fetch/ref in cmd_update, prod pre-sync - hermes_cli/gateway.py: launchd_restart_for_update with healthcheck - hermes_cli/banner.py: channel display in banner - hermes_cli/main.py: capture_output=True on .[all] install fallback - tests updated for new contract

…ompression - telegram.py: _is_nonretryable_auth_error() detects 401/Unauthorized, sets retryable=False to prevent infinite retry loops on bad tokens - run.py: /usage command counts all message roles (not just user/assistant), preventing 2-5x undercount in tool-heavy conversations - run.py: pass full transcript to compressor instead of filtering to user/assistant only, allowing compressor to prune stale tool results

…del runtime - Interrupt handler now checks if queued message is a command (starts with /) and dispatches it via _handle_message instead of treating as plain text - Session-scoped model overrides from /model now flow through to _run_agent's agent creation via _resolve_session_runtime_config() - Fallback model tracking uses session-aware model instead of global default

Add allowed_inbound_targets config support to TelegramAdapter for filtering inbound messages to specific chat/topic combinations. - Parse config from extra.allowed_inbound_targets (dict, string, int forms) - _is_inbound_target_allowed: allowlist check, DMs always pass - _is_inbound_message_allowed: convenience wrapper for PTB Message objects - Guard all four inbound handlers: text, command, location, media - Add tests for parsing, filtering, and handler rejection - Document in telegram.md with YAML examples and behavior notes

…gent patches

Merge-based updates caused compounding conflicts over time and required force-push to recover from anyway. Rebase keeps prod as a clean linear stack of patches on top of upstream main. - Uses --force-with-lease for safer pushes - Warns on large drift (>100 commits, >20 patches) - Pauses on rebase conflicts for manual resolution

_requested_turn is defined inside the run_sync() closure but was referenced from the outer _run_agent() async scope, causing a NameError on every turn where the agent has a model attribute.

Conflict resolutions: - gateway/run.py compression: keep prod model resolution + full history - gateway/run.py /status: keep prod full-history estimation with last_prompt_tokens - gateway/run.py fallback eviction: take upstream _run_failed + _is_intentional_model_switch - gateway/run.py pending handling: take upstream _dequeue_pending_event + command safety net - gateway/run.py queue fallback: take upstream merge_pending_message_event - gateway/run.py _run_agent recurse: pass both session_overrides (prod) + event_message_id (upstream) - hermes_cli/gateway.py: keep prod launchd helper functions + upstream _wait_for_gateway_exit signature

Switch prod pre-sync and update-prod-branch.sh from tracking origin/main to merging only the latest v* release tag. This keeps prod on stable upstream releases rather than pulling in every post-tag commit. - Add _latest_upstream_tag() helper that fetches tags and resolves the highest v* tag via --sort=-version:refname - _attempt_prod_prefetch_sync() now merges the latest tag (not origin/main) and skips sync if fork/prod already contains it - update-prod-branch.sh rebases onto the latest tag, exits early if prod is current, and reports tag-relative patch counts - Update tests and hermes-fork-prod skill docs to match new behavior

Upstream changes absorbed (no prod-only conflict): - run.py: upstream now has session-scoped /model overrides, full-history flush with live memory guard, busy-ack, deprecated cwd warnings, _parse_session_key, channel prompts, try/except in flush agent Prod-only re-applied on upstream base: - telegram.py: inbound-topic allowlist (_parse_allowed_inbound_targets, _is_inbound_target_allowed, _is_inbound_message_allowed) wired into _should_process_message for all message types - telegram.py: non-retryable auth error detection (_is_nonretryable_auth_error) with telegram_invalid_token fatal error (non-retryable) vs connect_error (retryable)

… ACP Wires AIAgent.interim_assistant_callback in the ACP adapter so editor clients (CAR, VS Code, Zed) receive interim assistant commentary messages during tool execution. Commentary is emitted as agent_message_chunk updates with _meta.phase set to commentary, following the ACP schema extension pattern. Fixes the gap where gateway-path commentary works but ACP-path drops it.

…eaming Root cause: ACP turns emitted partial session/update progress (thinking spinners, tool events) then went silent for the entire duration of run_conversation() because stream_delta_callback was never wired in the ACP adapter. The LLM's streamed text deltas were silently discarded since _fire_stream_delta() only dispatches to stream_delta_callback and _stream_callback — neither of which ACP set. Fix: Add make_stream_delta_cb() factory that forwards each text delta to conn.session_update() as an agent_message_chunk, and wire it as agent.stream_delta_callback in prompt(). This gives ACP clients (CAR/PMA/VS Code) real-time token-by-token progress during LLM streaming, eliminating the multi-minute silent gap before final prompt_return delivery. Fixes #9

Git-on-my-level · 2026-04-22T04:27:05Z

Sorry forgot to cherry pick

Git-on-my-level and others added 21 commits April 9, 2026 13:05

chore: glm-5.1 setup, pre-push hooks, doc fixes

0fadeb0

fix(cli): Z.AI coding URL, slash provider syntax, Codex model probing

802e286

fix: repair *** corruption in auth.py, models.py, doctor.py from suba…

37c388e

…gent patches

fix(banner): add missing shutil import

58fdc0d

Merge remote-tracking branch 'origin/main' into HEAD

d8b8bb1

Merge remote-tracking branch 'origin/main' into HEAD

1c2f5a3

fix(gateway): resolve NameError in fallback model tracking

5fd4dad

_requested_turn is defined inside the run_sync() closure but was referenced from the outer _run_agent() async scope, causing a NameError on every turn where the agent has a model attribute.

Merge remote-tracking branch 'origin/main' into HEAD

8fee7dd

Merge remote-tracking branch 'origin/main' into HEAD

967a534

fix: initialize _allowed_inbound_targets in test fixture

54ecc1a

Harden launchd service path generation

75b3fcf

alt-glitch added type/bug Something isn't working P1 High — major feature broken, no workaround comp/acp Agent Communication Protocol adapter comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint labels Apr 22, 2026

Git-on-my-level closed this Apr 22, 2026

Git-on-my-level deleted the fix/acp-stream-delta-bridge branch April 22, 2026 04:26

Git-on-my-level restored the fix/acp-stream-delta-bridge branch April 22, 2026 04:27

Navigation Menu

Uh oh!

fix(acp): wire stream_delta_callback for real-time ACP text delta streaming - #13810

fix(acp): wire stream_delta_callback for real-time ACP text delta streaming#13810
Git-on-my-level wants to merge 21 commits into
NousResearch:mainfrom
Git-on-my-level:fix/acp-stream-delta-bridge

Git-on-my-level commented Apr 22, 2026

Uh oh!

Git-on-my-level commented Apr 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

Git-on-my-level commented Apr 22, 2026

Root Cause

Why: stream_delta_callback was never wired in the ACP adapter

Fix

Files Changed

Uh oh!

Git-on-my-level commented Apr 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Why: `stream_delta_callback` was never wired in the ACP adapter