fix(adapter): silence handling — sentinel, empty-reply config, diagnostic logging (#836)

[howie]

Closes #836.

What

Three fixes for silence/empty-reply handling. All land in AdapterRouter::stream_prompt_blocks (src/adapter.rs) — the single funnel for every adapter — so Discord, Slack, and all gateway-relayed adapters (Feishu, Telegram, LINE, Teams, WeChat) inherit them automatically. No per-adapter changes.

Fix 1 — <silent /> sentinel filter

Steering can instruct agents to output exactly <silent /> when staying silent. The gateway detects this post-loop (not per-chunk, to avoid partial-buffer false positives), deletes any streaming placeholder, and returns without posting.

if text_buf.trim() == "<silent />" {
    info!(platform = %adapter.platform(), "agent emitted <silent /> sentinel -- suppressing reply");
    if let Some(msg) = placeholder_msg.as_ref() {
        let _ = adapter.delete_message(msg).await;
    }
    return Ok(());
}

Addresses Bug 1: the agent now has a positive "output X" instruction instead of a negative "don't output Y", which LLMs follow far more reliably.

Fix 2 — emptyReplyPlaceholder config

New [reactions] empty_reply_placeholder field (default true, fully backward-compatible). When false, empty replies with no error are silently suppressed instead of posting _(no response)_.

# per-agent config.toml (or values.yaml -> reactions.emptyReplyPlaceholder: false)
[reactions]
empty_reply_placeholder = false

Helm chart: configmap.yaml template and values.yaml comment updated.

Fix 3 — diagnostic logging

Three tracing::warn! lines added at the three response_error = Some(...) sites:

Site	Log message
Agent process died	agent process died mid-prompt
Hard timeout exceeded	agent hard timeout exceeded + elapsed_s field
JSON-RPC error	agent JSON-RPC error + code + message fields

Each log carries platform field for cross-adapter filtering. Zero behaviour change. Resolves the "two agents returned _(no response)_ with no server-side evidence" case from 2026-05-17.

Adapter scope

Adapter	Benefits?
Discord native (discord.rs)	Yes
Slack native (slack.rs)	Yes
Gateway-relayed: Feishu, Telegram, LINE, Teams, WeChat	Yes — all re-enter via AdapterRouter::handle_message
Future adapters	Yes — automatically

No changes to discord.rs, slack.rs, gateway.rs, or any gateway/src/adapters/*.rs. Adapters remain pure transport.

Also adds

docs/steering/silence.md — agent steering template explaining <silent /> usage and the multi-mention sub-case (Bug 1B: agent must read the full mention list before deciding silence, not pattern-match on "other bot present").

Tests added

• silent_sentinel_passes_through_directive_parser — <silent /> is NOT consumed by the [[key:value]] directive parser
• silent_sentinel_with_whitespace_passes_through — whitespace/newline variants survive intact
• empty_reply_placeholder_defaults_to_true — backward compat regression guard
• empty_reply_placeholder_can_be_set_to_false — opt-in suppression round-trips through TOML

🤖 Generated with Claude Code

[github-actions]

⚠️ This PR is missing a Discord Discussion URL in the body.

All PRs must reference a prior Discord discussion to ensure community alignment before implementation.

Please edit the PR description to include a link like:

Discord Discussion URL: https://discord.com/channels/...

This PR will be automatically closed in 3 days if the link is not added.

[shaun-agent]

OpenAB PR Screening

This is auto-generated by the OpenAB project-screening flow for context collection and reviewer handoff.
Click 👍 if you find this useful. Human review will be done within 24 hours. We appreciate your support and contribution 🙏

• Title: fix(adapter): silence handling — sentinel, empty-reply config, diagnostic logging (silence handling: agent meta-replies, (no response) placeholder, and missing diagnostic logging #836)
• Source: fix(adapter): silence handling — sentinel, empty-reply config, diagnostic logging (#836) #842
• Status: moved to PR-Screening
• Generated at: 2026-05-18T02:53:23.165Z
• Discord thread: not available

Screening report

screened PR #842, posted the marker comment, and moved project item `PVTI_lADOEFbZWM4BUUALzgtAjzk` from `Incoming` to `PR-Screening`.

GitHub comment: #842 (comment)
Project action: https://github.com/orgs/openabdev/projects/1

Intent

This PR tries to make agent silence explicit and operationally debuggable. The visible problem is that agents can currently produce confusing empty replies like _(no response)_, or fail silently enough that Discord, Slack, and gateway operators have little evidence about whether the agent chose silence, died, timed out, or hit JSON-RPC failure.

Feat

Fix work. It adds a shared adapter-level <silent /> sentinel suppression path, a backwards-compatible empty_reply_placeholder reactions config option, and warning logs for agent process death, hard timeout, and JSON-RPC errors. Because the changes land in AdapterRouter::stream_prompt_blocks, the behavior should apply to Discord, Slack, and gateway-relayed adapters without transport-specific patches.

Who It Serves

Primary beneficiaries are Discord users and Slack users who should stop seeing unwanted empty placeholders when silence is intentional. Secondary beneficiaries are deployers and agent runtime operators who need logs that distinguish intended silence from broken agent execution.

Rewritten Prompt

Implement shared silence handling in the OpenAB adapter router. When a complete streamed response, after trimming, is exactly <silent />, delete any streaming placeholder and return without posting a reply. Add a [reactions] empty_reply_placeholder config field that defaults to true; when set to false, suppress empty successful replies instead of posting _(no response)_. Add structured warning logs with platform for agent process death, hard timeout, and JSON-RPC error paths. Update Helm values/templates and add focused tests for sentinel parsing and config default/override behavior. Keep transport adapters unchanged unless a shared-router integration issue requires otherwise.

Merge Pitch

This is a narrow, high-leverage adapter fix: one shared funnel improves all current and future adapters. The risk profile is moderate-low because the default empty-reply behavior remains backwards-compatible and the sentinel requires an exact trimmed match. The likely reviewer concern is whether <silent /> can accidentally suppress legitimate user-facing output, and whether deleting placeholders is reliable across adapter implementations.

Best-Practice Comparison

OpenClaw relevant practices are explicit delivery routing, run logs, and isolated execution evidence. This PR moves in that direction by making silence an explicit output contract and adding failure logs, but it does not add durable run logs or retry/backoff.

Hermes Agent relevant practices are fresh scheduled runs, atomic persisted state, and self-contained scheduled prompts. Those mostly do not apply because this PR is about live adapter response handling, not scheduling. The self-contained prompt idea applies lightly: docs/steering/silence.md makes the silence contract clearer for agent steering.

Implementation Options

Conservative: merge the shared sentinel and config path as-is after focused review of placeholder deletion behavior and tests. This preserves the current architecture and keeps the blast radius small.

Balanced: merge this PR, then follow with a small observability PR that attaches a stable request/session id to the new warning logs and documents expected log queries for Discord, Slack, and gateway deployments.

Ambitious: replace ad hoc empty response handling with a typed response outcome model such as Reply, Silent, Empty, and Error, then route delivery, logging, placeholder cleanup, and metrics through that model. This would be cleaner long-term but larger than the current bug fix.

Comparison Table

Option	Speed	Complexity	Reliability	Maintainability	User Impact	Fit for OpenAB now
Conservative	Fast	Low	Good if exact sentinel tests pass	Good, shared router only	Immediate reduction in unwanted replies	Strong
Balanced	Medium	Medium	Better operator diagnosis with request correlation	Good, adds observability without redesign	Same user fix plus easier incident triage	Best
Ambitious	Slow	High	Best if completed well	Best long-term model, but more invasive	Cleaner semantics across adapters	Premature for this PR

Recommendation

Use the balanced path. Advance this PR through review as the behavior fix, with reviewer attention on exact sentinel matching, placeholder deletion, and the default-preserving config path. Split request/session correlation or richer run-log work into a follow-up so this fix can land without turning a targeted silence bug into a broader runtime redesign.

[howie]

Discord Discussion URL: https://discord.com/channels/1491295327620169908/1491969620754567270/1505591332457152724