fix(orchestrator): truncate long completion relays instead of destroying them

[NubsCarson]

fix(orchestrator): truncate long completion relays instead of destroying them

Defect

elideLongBlocks (plugins/plugin-agent-orchestrator/src/services/transcript-sanitizer.ts) hard-replaced ANY post-strip text over 2000 chars with the literal string [output elided — N chars] — no truncation, total data loss. Regression from 3781312 (#11605); before that commit the coordinator relayed record.response verbatim.

Two concrete failure paths on develop:

1. Long pure-prose deliverable destroyed. A dashboard/API-spawned task ("write a detailed migration plan in your final message") whose 2.4KB pure-prose answer contains no tool-output envelopes (strip is a no-op) synthesizes as literally [output elided — 2450 chars]. sessionHasRouterOrigin is false for these sessions so synthesis IS the relay path, and buildTaskResultLine (packages/agent/src/api/server-helpers-swarm.ts) posts completionSummary verbatim to the connector — no LLM pass, no bypass for opencode/codex. The user receives the marker instead of their deliverable.
2. App-verification verdict shadowed then destroyed. runCustomValidatorAndDispatch spreads enrichedData (retaining the raw ACP finalText in response) and adds summary: "App verification passed.". The coordinator read ladder takes response first, so when finalText exceeds 2KB the user sees the elision marker — and the verdict, which exists ONLY on the custom-validator record (the raw task_complete was withheld until validation), never posts at all.

The router's own relay path never had this problem: composeNarration uses only stripToolTranscript, and its 2KB verbatim cap applies to captured tool-output blocks with a summarized fallback. Synthesis-path-only regression.

Fix

• elideLongBlocks now truncates: head + … [output truncated — N chars total] marker, with the whole result bounded to maxChars so the defensive re-sanitize in buildTaskResultLine is a provable no-op (idempotent).
• Coordinator read ladder (swarm-coordinator-service.ts): for custom-validator completions the verdict in record.summary now leads the completion summary, with the sanitized deliverable appended under a budget that keeps the combined text inside the relay cap. response no longer shadows the only record of the validation outcome.

Reproduction (failing tests on unmodified develop)

7 new/updated regression tests fail before the fix, reproducing both scenarios exactly:

× TRUNCATES an over-cap remnant, preserving the head (#11605 destroyed it)
  AssertionError: expected '[output elided — 3000 chars]' not to be '[output elided — 3000 chars]'
× does NOT reduce a long pure-prose deliverable to a bare marker (#11605 regression)
× relays the head of a long pure-prose deliverable instead of destroying it (#11605)
× posts the validated verdict plus the deliverable head when finalText exceeds the relay cap (#11605)
  expected summary.startsWith("App verification passed.") — got '[output elided — 3007 chars]'
...
Tests  7 failed | 56 passed (63)

Verification (after fix)

• bunx vitest run __tests__/unit/transcript-sanitizer.test.ts __tests__/unit/swarm-coordinator-service.test.ts → 63 passed (63)
• Full plugin suite bun run test:unit → 1276 passed (1276), 118 files
• Consumer suite packages/agent/src/api/server-helpers-swarm.test.ts → 13 passed (13)
• bun run --cwd plugins/plugin-agent-orchestrator typecheck → clean
• bunx biome check on the four touched files → clean

Evidence notes

• Real-LLM trajectory: N/A — deterministic string-pipeline defect in the relay sanitizer; the failing-then-passing unit tests drive the exact task_complete event payloads the ACP service emits (response = raw finalText), which is the full real path for this bug (no model in the loop between record.response and the connector post).
• UI screenshots/video: N/A — no UI surface changed; the connector-visible change is the relayed completion text, asserted byte-for-byte in the tests.

Refs 3781312 (#11605), #11578.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: ccf265d5-0272-4768-a357-cf5e8670dca4

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/rh-2

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[lalalune]

Post-merge validation pass from a PR worktree on current develop:

• Generated standard keyword artifacts with node packages/shared/scripts/generate-keywords.mjs --target ts for test imports.
• bunx vitest run plugins/plugin-agent-orchestrator/__tests__/unit/transcript-sanitizer.test.ts plugins/plugin-agent-orchestrator/__tests__/unit/swarm-coordinator-service.test.ts --reporter=dot -> 63 tests passed.
• bun run --cwd plugins/plugin-agent-orchestrator typecheck -> passed.
• bunx biome check on the four touched files -> passed.
• bunx vitest run packages/agent/src/api/server-helpers-swarm.test.ts --reporter=dot -> 13 tests passed.
• bun run --cwd plugins/plugin-agent-orchestrator test:unit -> 118 files / 1279 tests passed.
• git diff --check origin/develop..HEAD -> passed.

Reviewed the relay cap behavior and custom-validator path. The sanitizer now preserves the head of long pure-prose completions and appends a bounded truncation marker, while validated completions lead with the app-verification verdict and append the sanitized deliverable head.

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.