fix(server): preserve agent origin in cross-thread provider messages - #24725

Approvability

Verdict: Approved at 138bdc7

Macroscope's review found this PR approvable — Provider prompts now distinguish cross-thread agent coordination from human messages, including steering and context handoffs. The change is limited to message presentation and preserves existing command handling.

^{You can add or adjust custom eligibility rules. Learn more.}

No actionable comments were generated in the recent review.

Merge Risk: Low · previous review fdd06

Agent-to-agent messages that start with a custom slash command will no longer run that command on Pi, because the provenance notice now comes first. The PR documents this as out of scope and it affects a narrow path. The owner should accept the regression or follow up before relying on Pi custom commands across threads.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title names the main change: cross-thread agent messages sent to providers now keep their agent origin. It follows the conventional commit format and matches the changeset, which adds an origin no…
Description check		The description follows the template. It gives the problem, the change, the scope link to issue #24711, and detailed verification with commands, platforms, and observed results. It also states what wa…
Linked Issues check		[#24711] At head fdd0618, withProviderMessageOrigin in apps/server/src/orchestration-v2/providerMessageOrigin.ts returns text unchanged unless createdBy is agent and senderThreadId is set…
Out of Scope Changes check		The changes connect to [#24711] or preserve existing native behavior. The formatter and its call sites in the start, steer, and run-execution paths implement the issue. withInlineContext heading sup…

Full details: Title check

Explanation

The title names the main change: cross-thread agent messages sent to providers now keep their agent origin. It follows the conventional commit format and matches the changeset, which adds an origin notice with sender thread ID and an agent-authored marker to provider-bound text.

Full details: Description check

Explanation

The description follows the template. It gives the problem, the change, the scope link to issue #24711, and detailed verification with commands, platforms, and observed results. It also states what was not checked, such as other providers and clients. The Scope section explains the bug fix and links the issue, and the description notes that the work is a replacement for PR #24483 requested in the issue. The Verification section is thorough. The description is long and includes some follow-up history, but the required sections are all present and substantive.

Full details: Linked Issues check

Explanation

[#24711] At head fdd0618, withProviderMessageOrigin in apps/server/src/orchestration-v2/providerMessageOrigin.ts returns text unchanged unless createdBy is agent and senderThreadId is set. For cross-thread agent messages it prepends a heading that contains the JSON-escaped sender thread ID and states that the message is agent-authored, not from the human user. ProviderTurnControlService.ts applies this formatter to the projected composer-context text passed to steerTurn, so steers carry the marker. ProviderTurnStartService.ts and RunExecutionService.ts apply it to new turns, and withInlineContext in ContextHandoffDelivery.ts accepts an "Agent message:" heading for these cases. The stored message, attachments, and provenance metadata are not rewritten. Human and sender-less messages pass through unchanged. The changed /goal and Pi /compact branches keep the command token first. The current head changes that are specific to goal routing were checked only at the function level, not traced through every branch. That residual uncertainty does not affect the core start and steer requirement. The PR description reports live Linux Codex verification on an earlier head (187051d663); this assessment relies on source evidence and the reported test results, not on a live provider run.

Full details: Out of Scope Changes check

Explanation

The changes connect to [#24711] or preserve existing native behavior. The formatter and its call sites in the start, steer, and run-execution paths implement the issue. withInlineContext heading support and the "Agent message:" label serve the same provenance purpose. The Pi /compact branch and the native /goal routing (acceptsNativeGoalCommands) keep provider-native commands parseable when a cross-thread message starts with /compact or /goal. These are narrow consequences of applying the marker to agent-authored text. The test and fixture updates (ProjectionControlReads.test.ts, RunExecutionService.test.ts, ProviderSwitch.integration.test.ts, codex_transcript.ndjson) exercise the changed prompt text. No unrelated implementation was identified. The goal-routing capability predicate is a broader change than the core fix. Its connection to provenance is stated in the PR and the commit message, so it is not classified as unrelated. Not every changed test or fixture line was read.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: 0bc993a
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: f554cb69-f316-4416-844d-aec35346260e

Commits

Reviewing files that changed from the base of the PR and between 3ed09f8 and fdd0618.

Files selected for processing: 8

• apps/server/src/orchestration-v2/ProjectionControlReads.test.ts
• apps/server/src/orchestration-v2/ProviderTurnControlService.ts
• apps/server/src/orchestration-v2/ProviderTurnStartService.ts
• apps/server/src/orchestration-v2/RunExecutionService.test.ts
• apps/server/src/orchestration-v2/RunExecutionService.ts
• apps/server/src/orchestration-v2/providerMessageOrigin.ts
• apps/server/src/orchestration-v2/testkit/ProviderSwitch.integration.test.ts
• apps/server/src/orchestration-v2/testkit/fixtures/delegated_task_status/codex_transcript.ndjson

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 7 remain after this review.

Updated in 972d23a by merging main b424c6e while preserving commit history. The merge keeps main's deferred history for native /goal commands and the Agent message heading for ordinary cross-thread handoffs.

The new main path exposed a compatibility issue: adding the provenance prefix to a native goal command prevents the adapter from parsing it. The regression first failed on the provider-visible /goal text. Codex/Claude goal commands now retain their original command text; attachment-bearing goal text remains an ordinary prompt with provenance. Existing /compact dispatch is unchanged.

Validation: execution, provider-switch and steering integration suites passed 105 tests. check:changed then passed format, lint, both knip checks, GraphQL consistency, affected-package typechecks and 124 adjacent tests. The protocol fixture uses production services with controlled provider runtimes; no real provider request was sent.

merged current main 17a6a01d84 into the existing head without rebasing (5038f0c530). the only conflict was the provider-switch handoff table: main added two cancelled rows and this PR added an agent column, so the merge keeps both (cancelled rows get agent: false). no code changes beyond the merge.

verified on 5038f0c530 through blacksmith: vp run check:changed 17a6a01d84 passed format, lint, both knip checks, graphql, three package typechecks, and 169 tests in six files. the 8 hands off provider-switch cases, including main's new cancelled ones, pass.

unverified on this merged head: a fresh real-provider session. the earlier before/after evidence above still applies; no visual changes.

@coderabbitai resume

Verified and fixed the new Linked Issues warning in cd17a5f after merging main e455200 without rebasing. Native Codex/Claude Agent /goal objectives now carry the sender and agent-origin heading inside the objective, preserving the /goal command token. Goal show/edit/clear/pause/resume controls and human-authored input remain unchanged. Pi /compact custom instructions likewise retain provenance after the command token, while a bare compact remains a native control operation.

Extended existing start-routing and durable-steer tests. The new Pi start/steer cases failed before the fix; the updated matrix covers Codex and Claude Agent goals, goal controls, input whitespace, attachments and ordinary prompts. check:changed passed: 158 tests in six files, affected-package typechecks, formatting, lint, both knip passes and GraphQL validation. Earlier real-provider evidence remains linked in the description; this follow-up used the existing adapter-boundary and persistence tests rather than launching a real provider.

Model: GPT-6. Harness: Codex desktop.

The automated pipeline stopped on this pull request and will not merge it:

the reviewers still had findings after 3 rounds (the work is on main...codex/fix-24711-agent-message-origin):

1. [major] apps/server/src/orchestration-v2/RunExecutionService.ts:1919: Worse than main: provenance is prepended before OpenCode parses registered slash commands, turning working command invocations into ordinary model prompts. Only native-capable /goal commands are exempted. OpenCode's submit path subsequently calls commandOf on the prefixed text, gets undefined, and uses session.prompt instead of session.command. Preserve command dispatch before decorating provider-visible prompt content.
   Fails with: An agent sends an attachment-free '/git/hello WORLD' to an idle OpenCode thread whose workspace registers git/hello, with no pending handoff. Main dispatches session.command({name:'git/hello',text:'WORLD'}); this head sends the marked literal text through session.prompt. Older supported OpenCode sessions registering the goal command also regress for '/goal clear' when supportsNativeGoals is false. Confirmed against origin/main's dispatch code and by running both revisions' actual command parser with original versus decorated input.
2. [major] apps/server/src/orchestration-v2/providerMessageOrigin.ts:16: Worse than main: only Pi gets slash-led cross-thread agent text unchanged. For OpenCode, the T3 adapter (packages/provider-opencode/src/server/v2/adapter.ts:4381-4393, commandOf = /^/(\S+).../ on the trimmed text) runs a bare slash message as a registered command via session.command. With the origin note in front, the text no longer starts with '/', so the command is sent to the model as literal prose. The same applies to steers (ProviderTurnControlService.ts:416). Claude Agent and ACP providers (Antigravity, etc.) also accept slash commands at the start of the prompt and are probably affected the same way; that is from reading the code, not from a run.
   Fails with: A parent agent calls delegate_task or t3_thread_send with the text '/review the diff' to a thread on OpenCode where 'review' is a registered command. On origin/main, OpenCode runs the command through session.command. On this head the provider receives "[Message from another thread's agent (thread "...")...]\n\n/review the diff" as an ordinary prompt and the command never runs. This hits any orchestrator agent that delegates slash commands, skills or prompt templates to OpenCode. Fix: apply the Pi exemption to every slash-dispatching driver (opencode, claudeAgent, ACP drivers), or to all bare slash-led text, and add an OpenCode '/review' row to the RunExecutionService routing table.

It stays open for a person to finish or close.