Subagents: restore announce chain + fix nested retry/drop regressions by tyler6204 · Pull Request #22223 · openclaw/openclaw

tyler6204 · 2026-02-20T22:33:37Z

Summary

This PR restores subagent orchestration behavior to the pre-regression model and fixes the remaining nested announce delivery issue observed in live depth tests.

1) Restore subagent announce to agent-only path

Reverts the direct send completion delivery pattern.
Keeps completion delivery on the agent path only (system-message injection into requester session).
Removes direct-channel completion framing behavior so parent LLM always owns user-facing voice.

2) Fix nested/sub-subagent announce completion timing

Defers announce when a finished run still has active descendants.
Detects placeholder parent replies like "waiting for nested result" and waits for the parent's follow-up synthesized reply.
Prevents announcing stale placeholder text.

3) Fix announce drop under deep nesting

Live failures showed top-level runs hitting retry limit before descendants finished, dropping final announces.

In finalizeSubagentCleanup, deferrals caused by active descendants no longer consume retry budget.
Retry budget is still enforced for real repeated failures; expiry guard is preserved.

4) Cron isolated session behavior

Restores fresh-session-per-run behavior for isolated cron/webhook runs.
Keeps valid override-preservation behavior while rolling session IDs.

5) Default max spawn depth

Introduces shared default DEFAULT_SUBAGENT_MAX_SPAWN_DEPTH = 2.
Wires all depth checks/prompt policy fallbacks to the shared default.
Updates config defaults and tests/docs accordingly.

Validation

pnpm exec vitest run src/cron/isolated-agent/session.test.ts src/config/config.agent-concurrency-defaults.test.ts src/config/config.identity-defaults.test.ts src/agents/openclaw-tools.subagents.sessions-spawn-depth-limits.test.ts src/agents/subagent-registry.announce-loop-guard.test.ts src/agents/subagent-registry.nested.test.ts
pnpm exec vitest run --config vitest.e2e.config.ts src/agents/system-prompt.e2e.test.ts src/agents/subagent-announce.format.e2e.test.ts src/agents/openclaw-tools.subagents.sessions-spawn.lifecycle.e2e.test.ts src/agents/pi-tools.policy.e2e.test.ts src/agents/subagent-registry.persistence.e2e.test.ts

Notes

Addresses the reported cases where depth-3 chains completed internally but final announces were missing at the top-level requester session.

Greptile Summary

This PR restores subagent orchestration to a reliable announce-chain model and fixes nested completion delivery regressions observed in deep nesting scenarios.

Key Changes:

Restored subagent completions to agent-only delivery path (system message injection) instead of direct send calls, ensuring parent LLM owns user-facing voice
Fixed nested announce timing by detecting active descendants and deferring announce until descendant runs complete and parent synthesizes final reply
Fixed announce drop regression under deep nesting by separating normal deferrals (active descendants) from retry budget consumption, preventing premature give-up before descendants finish
Introduced shared DEFAULT_SUBAGENT_MAX_SPAWN_DEPTH = 2 constant wired through all depth checks and policy fallbacks for consistency
Restored cron isolated session behavior to always create fresh session per run while preserving per-session override settings
Added placeholder detection logic (isLikelyWaitingForDescendantResult) to defer announcing stale intermediate replies and wait for synthesized follow-ups

Architecture:

The PR addresses a critical timing issue where top-level runs would exhaust retry limits before descendants finished, causing final announces to be dropped. The fix introduces descendant-aware deferral logic in finalizeSubagentCleanup that doesn't consume retry budget when waiting for active descendants, while still enforcing limits for actual repeated failures.

Confidence Score: 4/5

This PR is safe to merge with low risk given comprehensive test coverage and targeted fixes
Score reflects well-structured logic changes with clear intent, comprehensive test updates, and no obvious security concerns. The PR addresses real production issues with depth-3 chains. Deducted one point because the retry/deferral logic in finalizeSubagentCleanup is complex with multiple conditional paths that could benefit from additional integration testing under high-concurrency scenarios.
Pay close attention to src/agents/subagent-registry.ts lines 324-377 (finalizeSubagentCleanup retry logic) and src/agents/subagent-announce.ts lines 518-564 (descendant detection and placeholder handling)

_{Last reviewed commit: 241b86e}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 241b86e8cd

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-02-20T22:39:15Z

src/agents/subagent-announce.ts

+    if (!requesterIsSubagent && !directOrigin) {
      const { entry } = loadRequesterSessionEntry(targetRequesterSessionKey);
-      directOrigin = resolveAnnounceOrigin(entry, targetRequesterOrigin);
+      directOrigin = deliveryContextFromSession(entry);


Merge direct announce origin with stored delivery context

This branch now uses targetRequesterOrigin as-is whenever it is present, so if a run was spawned from a non-deliverable/internal channel hint (for example webchat) the direct announce call keeps that channel and never falls back to the persisted lastChannel/lastTo route. In that case agent delivery is forced to non-delivery and the completion update is not sent to the user even though the requester session has a valid outbound target; this used to be avoided by always resolving origin against the session entry.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-02-20T22:39:15Z

src/agents/subagent-announce.ts

  // actually on and must take priority over the session entry, which may carry
  // stale lastChannel / lastTo values from a previous channel interaction.
-  return mergeDeliveryContext(normalizedRequester, normalizedEntry);
+  return mergeDeliveryContext(requesterOrigin, deliveryContextFromSession(entry));


Ignore internal channel hints when resolving queued announce origin

resolveAnnounceOrigin now directly prioritizes requesterOrigin without filtering non-deliverable channels, so a queued announce can inherit webchat/internal channel hints and overwrite a valid persisted delivery route from the requester session. When that queued item is sent, the gateway treats the channel as internal and does not deliver outward, causing queued subagent completions to be dropped while the requester run is active.

Useful? React with 👍 / 👎.

)

tyler6204 · 2026-02-20T23:46:22Z

Merged via squash.

Merge commit: fe57bea

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204)

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204)

@tyler6204

…#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (#22223) (thanks @tyler6204)

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204)

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204)

…2223 Commit f555835 ("Channels: add thread-aware model overrides") inadvertently reverted the `resolveAnnounceOrigin` channel guard introduced in fe57bea (openclaw#22223). The revert changed `isInternalMessageChannel` back to `!isDeliverableMessageChannel`, which strips non-standard channel hints (anything not in the built-in deliverable list) from the requester origin. When the stripped origin falls back to the session entry's persisted `lastChannel`/`lastTo`, stale values from a previous interaction (e.g. WhatsApp with no valid E.164 target) cause delivery failures: Error: Delivering to WhatsApp requires target <E.164|group JID> [warn] Subagent announce give up (retry-limit) The fix narrows the guard back to `isInternalMessageChannel` so only the webchat channel is stripped — matching the intent of openclaw#22223. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Two issues caused subagent completion announces to silently fail for plugin channels like Gmail: 1. `resolveAnnounceOrigin` used `!isDeliverableMessageChannel()` to strip non-standard channels, inadvertently reverting the narrower `isInternalMessageChannel()` guard from openclaw#22223 (fe57bea). Restored the original guard so only webchat is stripped. 2. `sendSubagentAnnounceDirectly` and `sendAnnounce` passed the origin channel/to verbatim to `callGateway()`, even when the channel was not in the deliverable list. Plugin channels whose adapter ID differs from their registered plugin ID (e.g. "gmail" vs "openclaw-gmail") were unroutable, causing the gateway to time out (15s x 3 retries) and then fall back to stale session routes (typically WhatsApp with no valid E.164 target). Fix: check `isDeliverableMessageChannel()` before passing channel hints. When the channel is unroutable, omit channel/to/accountId so the gateway resolves delivery from connected clients or session state. Symptoms observed: Subagent completion direct announce failed: gateway timeout after 15000ms [warn] Subagent announce give up (retry-limit) retries=3 Error: Delivering to WhatsApp requires target <E.164|group JID> Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…unceOrigin Restores the narrower internal-channel guard from PR openclaw#22223 (fe57bea) that was inadvertently reverted by f555835. The original !isDeliverableMessageChannel() check strips the requester's channel whenever it is not in the registered deliverable set. This causes delivery failures for plugin channels whose adapter ID differs from their plugin ID (e.g. "gmail" vs "openclaw-gmail"): the requester origin is discarded and the announce falls back to stale session routes — typically WhatsApp — resulting in a timeout followed by an E.164 format error. Replacing with isInternalMessageChannel() limits stripping to explicitly internal channels (webchat), preserving the requester origin for all external channels regardless of whether they are currently in the deliverable list. Fixes: openclaw#22223 regression introduced in f555835 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204)

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204)

…unceOrigin Restores the narrower internal-channel guard from PR #22223 (fe57bea) that was inadvertently reverted by f555835. The original !isDeliverableMessageChannel() check strips the requester's channel whenever it is not in the registered deliverable set. This causes delivery failures for plugin channels whose adapter ID differs from their plugin ID (e.g. "gmail" vs "openclaw-gmail"): the requester origin is discarded and the announce falls back to stale session routes — typically WhatsApp — resulting in a timeout followed by an E.164 format error. Replacing with isInternalMessageChannel() limits stripping to explicitly internal channels (webchat), preserving the requester origin for all external channels regardless of whether they are currently in the deliverable list. Fixes: #22223 regression introduced in f555835 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…unceOrigin Restores the narrower internal-channel guard from PR openclaw#22223 (fe57bea) that was inadvertently reverted by f555835. The original !isDeliverableMessageChannel() check strips the requester's channel whenever it is not in the registered deliverable set. This causes delivery failures for plugin channels whose adapter ID differs from their plugin ID (e.g. "gmail" vs "openclaw-gmail"): the requester origin is discarded and the announce falls back to stale session routes — typically WhatsApp — resulting in a timeout followed by an E.164 format error. Replacing with isInternalMessageChannel() limits stripping to explicitly internal channels (webchat), preserving the requester origin for all external channels regardless of whether they are currently in the deliverable list. Fixes: openclaw#22223 regression introduced in f555835 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204) (cherry picked from commit fe57bea) # Conflicts: # CHANGELOG.md # docs/concepts/session-tool.md # docs/tools/subagents.md # src/agents/openclaw-tools.subagents.sessions-spawn-depth-limits.test.ts # src/agents/openclaw-tools.subagents.sessions-spawn.lifecycle.e2e.test.ts # src/agents/pi-tools.policy.ts # src/agents/subagent-announce.format.e2e.test.ts # src/agents/subagent-announce.ts # src/agents/subagent-registry.announce-loop-guard.test.ts # src/agents/subagent-registry.ts # src/agents/system-prompt.test.ts # src/agents/tools/subagents-tool.ts # src/config/config.identity-defaults.test.ts # src/config/types.agent-defaults.ts # src/config/zod-schema.agent-defaults.ts # src/cron/isolated-agent/session.test.ts # src/cron/isolated-agent/session.ts

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204) (cherry picked from commit fe57bea) # Conflicts: # CHANGELOG.md # docs/concepts/session-tool.md # docs/tools/subagents.md # src/agents/openclaw-tools.subagents.sessions-spawn-depth-limits.test.ts # src/agents/openclaw-tools.subagents.sessions-spawn.lifecycle.e2e.test.ts # src/agents/pi-tools.policy.ts # src/agents/subagent-announce.ts # src/agents/subagent-registry.announce-loop-guard.test.ts # src/agents/subagent-registry.ts # src/agents/system-prompt.test.ts # src/agents/tools/subagents-tool.ts # src/config/config.identity-defaults.test.ts # src/config/defaults.ts # src/config/types.agent-defaults.ts # src/config/zod-schema.agent-defaults.ts # src/cron/isolated-agent/session.test.ts # src/cron/isolated-agent/session.ts

@tyler6204

…openclaw#22223) * Subagents: restore announce flow and fix nested delivery retries * fix: prep subagent announce + docs alignment (openclaw#22223) (thanks @tyler6204)

…unceOrigin Restores the narrower internal-channel guard from PR openclaw#22223 (fe57bea) that was inadvertently reverted by f555835. The original !isDeliverableMessageChannel() check strips the requester's channel whenever it is not in the registered deliverable set. This causes delivery failures for plugin channels whose adapter ID differs from their plugin ID (e.g. "gmail" vs "openclaw-gmail"): the requester origin is discarded and the announce falls back to stale session routes — typically WhatsApp — resulting in a timeout followed by an E.164 format error. Replacing with isInternalMessageChannel() limits stripping to explicitly internal channels (webchat), preserving the requester origin for all external channels regardless of whether they are currently in the deliverable list. Fixes: openclaw#22223 regression introduced in f555835

Subagents: restore announce flow and fix nested delivery retries

241b86e

openclaw-barnacle bot added agents Agent runtime and tooling size: XL maintainer Maintainer-authored PR labels Feb 20, 2026

tyler6204 self-assigned this Feb 20, 2026

chatgpt-codex-connector bot reviewed Feb 20, 2026

View reviewed changes

fix: prep subagent announce + docs alignment (#22223) (thanks @tyler6204

72dc5da

)

openclaw-barnacle bot added the docs Improvements or additions to documentation label Feb 20, 2026

tyler6204 added a commit that referenced this pull request Feb 20, 2026

fix: prep subagent announce + docs alignment (#22223) (thanks @tyler6204

34470ad

)

tyler6204 merged commit fe57bea into main Feb 20, 2026
22 of 25 checks passed

tyler6204 deleted the fix/cron-subagent-regressions branch February 20, 2026 23:39

tyler6204 mentioned this pull request Feb 20, 2026

Fix merge-test regressions around media path roots and status emojis #22274

Merged

github-actions bot mentioned this pull request Feb 21, 2026

📡 Upstream Digest — 2026-02-21 01:16 UTC curtismercier/openclaw-mods#81

Open

dashed mentioned this pull request Feb 21, 2026

fix(cron): prevent stale session state from leaking into cron announce delivery dashed/openclaw#3

Open

8 tasks

mcinteerj mentioned this pull request Feb 21, 2026

fix(subagents): restore isInternalMessageChannel guard in resolveAnnounceOrigin #23030

Merged

2 tasks

tyler6204 added a commit that referenced this pull request Feb 22, 2026

fix(agents): restore exact announce logic from fe57bea (PR #22223)

93798d3

tyler6204 added a commit that referenced this pull request Feb 22, 2026

fix(agents): restore exact announce logic from fe57bea (PR #22223)

cb9b0a5

tyler6204 mentioned this pull request Feb 22, 2026

fix(agents): restore subagent announce chain from #22223 #23166

Closed

QuietAnima mentioned this pull request Feb 22, 2026

[Bug] sessions_spawn announce not delivering to Signal requester session #23617

Closed

steipete mentioned this pull request Feb 23, 2026

sessions_spawn notify_done fires on spawn ack instead of run completion #17483

Closed

github-actions bot mentioned this pull request Mar 1, 2026

cherry-pick: upstream bugfix commits (2026-03-01-0443) hughdidit/DAISy-Agency#140

Closed

6 tasks

tyler6204 mentioned this pull request Mar 5, 2026

fix(subagents): deterministic announce delivery with descendant gating #35080

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Subagents: restore announce chain + fix nested retry/drop regressions#22223

Subagents: restore announce chain + fix nested retry/drop regressions#22223
tyler6204 merged 2 commits intomainfrom
fix/cron-subagent-regressions

tyler6204 commented Feb 20, 2026 •

edited by greptile-apps bot

Loading

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Feb 20, 2026

Uh oh!

chatgpt-codex-connector bot Feb 20, 2026

Uh oh!

Uh oh!

tyler6204 commented Feb 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

tyler6204 commented Feb 20, 2026 • edited by greptile-apps bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

1) Restore subagent announce to agent-only path

2) Fix nested/sub-subagent announce completion timing

3) Fix announce drop under deep nesting

4) Cron isolated session behavior

5) Default max spawn depth

Validation

Notes

Greptile Summary

Confidence Score: 4/5

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Feb 20, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot Feb 20, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tyler6204 commented Feb 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

tyler6204 commented Feb 20, 2026 •

edited by greptile-apps bot

Loading