An agent calling t3_thread_send could invoke a recipient's skill merely by mentioning its name. In the real app, a mid-sentence $unslop loaded full skill instructions on Claude and Codex; a leading /unslop also invoked Claude's native command parser. Codex additionally expanded mentions followed by punctuation.
Neutralize skill syntax in cross-thread sends before choosing queued or steered delivery. Use the recipient provider's existing skill catalog to break known mentions with a word joiner, including Codex's punctuation and linked forms, and break leading slash commands. Preserve shell variables, amounts, paths and surrounding text.
The shared service covers text and attachment sends. Direct user messages, scheduled prompts and initial launch/delegation briefs retain their existing behavior. No wire contracts or persisted fields change.
Fixes https://github.com/pingdotgg/t3code/issues/24667
Triaged bug via the automated issue flow. GitCafe: #24667.
$HOME and $BUILD_DIR beside a known skill.t3_thread_send mode auto, at head 33bc6f1edb5c48792c147caaf2578504ff60cee0.| Runtime probe | Before | After |
|---|---|---|
| Agent → Claude, mid-sentence mention | Native command wrapper and full skill body | Plain text; no injected skill |
| Agent → Claude, leading slash | Native command wrapper and full skill body | Plain text; no injected skill |
| Agent → Codex, mid-sentence mention | Native skill block | Plain text; no injected skill |
| Agent → Codex, attached period | Native skill block | Plain text; no injected skill |
| Agent → Codex, linked mention | Covered by regression test | Plain text; no injected skill |
| Direct user controls: Claude mention/slash, Codex mention | Intentional invocation | All three still invoke the skill |
Before transcript, punctuation reproduction, final after transcript, final direct user controls.
Runtime observations are recipient reports of native prompt expansion, backed by focused service tests, rather than raw provider-protocol captures. No rendering changes; the visual fixture matrix is empty. Other providers were not exercised with real runtime probes.
Implemented with Claude Opus 5.5 through the Claude Agent SDK; triage, proof and review with GPT-6.1-Sol through the Codex harness.
Added 6 commits
Verdict: Approved at 33bc6f1
Macroscope's review found this PR approvable — This is a focused server bug fix that neutralizes recipient skill syntax only for agent-originated cross-thread messages while leaving direct user invocation unchanged. The production change is isolated and covered by targeted regression tests, with no API, persistence, deployment, or default-setting changes.
You can add or adjust custom eligibility rules. Learn more.
✅ Thread transfer remains within every enforced ceiling.
| Provider | Metric | Main baseline | This PR | Impact | PR ceiling | |
|---|---|---|---|---|---|---|
| Codex | Total thread wire | 5.0 KiB | 5.0 KiB | 0 B (0.0%) | 6.8 KiB | ✅ |
| Codex | Thread snapshot wire | 3.8 KiB | 3.8 KiB | 0 B (0.0%) | 4.9 KiB | ✅ |
| Codex | Live turn WebSocket wire | 1.2 KiB | 1.2 KiB | 0 B (0.0%) | 2.0 KiB | ✅ |
| Codex | Live turn WebSocket decoded | 20.9 KiB | 20.9 KiB | 0 B (0.0%) | 29.3 KiB | ✅ |
| Codex | Live turn messages | 2 | 2 | 0 (0.0%) | 8 | ✅ |
| Claude | Total thread wire | 5.0 KiB | 5.0 KiB | 0 B (0.0%) | 6.8 KiB | ✅ |
| Claude | Thread snapshot wire | 3.8 KiB | 3.8 KiB | 0 B (0.0%) | 4.9 KiB | ✅ |
| Claude | Live turn WebSocket wire | 1.2 KiB | 1.2 KiB | 0 B (0.0%) | 2.0 KiB | ✅ |
| Claude | Live turn WebSocket decoded | 21.2 KiB | 21.2 KiB | 0 B (0.0%) | 29.3 KiB | ✅ |
| Claude | Live turn messages | 2 | 2 | 0 (0.0%) | 8 | ✅ |
Baseline: b3d382e · PR result: 33bc6f1 · Source CI: success
10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.
No actionable comments were generated in the recent review.
3a07a7a1-a905-4240-a719-924b36ab76b8Reviewing files that changed from the base of the PR and between b3d382e1009d5430de8c528aef99a39854a7cbc4 and 33bc6f1edb5c48792c147caaf2578504ff60cee0.
apps/server/src/orchestration-v2/ThreadManagementService.test.tsapps/server/src/orchestration-v2/ThreadManagementService.tsIncluded review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 3 remain after this review.
When an agent sends a message to another thread, ThreadManagementService resolves the recipient provider’s skill names and escapes matching skill syntax and leading slash commands. Messages without a sender thread retain their original text. Tests cover these cases in auto and steer modes.
Changes: Agent-to-thread message handling
| Layer / File(s) | Summary |
|---|---|
Resolve recipient skills and escape skill syntax apps/server/src/orchestration-v2/ThreadManagementService.ts | The service resolves skill names from provider and workspace snapshots. It inserts a word joiner into matching skill references and leading slash commands. |
Apply escaping during thread sends and validate behavior apps/server/src/orchestration-v2/ThreadManagementService.ts, apps/server/src/orchestration-v2/ThreadManagementService.test.ts | When a sender thread is present, sendToThread escapes text using the explicitly selected model’s provider instance or the target thread’s provider instance. Tests cover auto and steer modes, skill references, and surrounding text. |
Priority: External image Normal
Estimated code review effort: 2 (Simple) | ~10 minutes
Change: Bug fix · Severity of issue fixed: Medium
Possibly related PRs
Suggested reviewers: t3dotgg, macroscopeapp
33bc6Agent-to-agent messages now arrive as plain text instead of invoking recipient skills. The one known edge case is that a message starting with a single-segment absolute path gets an invisible joiner after the slash, which is low impact. No blocking risk remains.
| Check name | Status | Explanation |
|---|---|---|
| Title check | Passed | The title clearly and concisely describes the main change: preventing agent messages from invoking skills in recipient threads. |
| Description check | Passed | The description covers the problem, implementation, scope, linked issue, focused verification, runtime results, limitations, and agent attribution. It provides sufficient evidence for the reported fix… |
| Linked Issues check | Passed | Issue #24667 requires agent-to-agent messages to remain text when they contain known skill mentions or leading slash commands. ThreadManagementService.sendToThread now applies `breakAgentSkillSyntax… |
| Out of Scope Changes check | Passed | The diff is limited to the server thread-management service and its focused regression tests. The changes directly implement issue #24667 and preserve existing behavior for messages without `senderThr… |
Explanation
The description covers the problem, implementation, scope, linked issue, focused verification, runtime results, limitations, and agent attribution. It provides sufficient evidence for the reported fix.
Explanation
Issue #24667 requires agent-to-agent messages to remain text when they contain known skill mentions or leading slash commands. ThreadManagementService.sendToThread now applies breakAgentSkillSyntax only when senderThreadId is present. The function uses the recipient provider and workspace skill catalogs, inserts a word joiner for matching mentions and leading slash syntax, and preserves other text. The added tests cover queued and steered delivery, Claude and Codex mention forms, links, punctuation, shell variables, amounts, paths, and the user-origin control.
Explanation
The diff is limited to the server thread-management service and its focused regression tests. The changes directly implement issue #24667 and preserve existing behavior for messages without senderThreadId. No unrelated product behavior, wire contract, or persisted field changes are shown.
Comment @coderabbitai help to get the list of available commands.