Refs #87329 (Fix 1 — PRIMARY: tag echo messages correctly at write time)

Scope note: issue #87329 proposes two fixes — Fix 1 (PRIMARY: tag echo messages as openclaw/delivery-mirror at write time) and Fix 2 (DEFENSE IN DEPTH: merge consecutive assistant turns in validateAnthropicTurns). This PR implements Fix 1 only; Fix 2 is covered by the separate open PR #87346 (Jefsky). Using a non-closing Refs reference so merging this primary writer fix does not auto-close the issue before the defense-in-depth layer lands or maintainers accept that scope.

User-facing bug

Subagent announce / embedded handoff gap-fill can append a user-visible assistant mirror using the child runtime provider and model metadata, such as anthropic/claude-*, instead of marking the row as an OpenClaw transcript-only delivery mirror. Replay filters exclude transcript-only mirrors by provider: "openclaw" with model: "delivery-mirror" / gateway-injected, so a display-only mirror with child provider metadata can later be replayed back into a provider turn. In conversations with provider-managed thinking/signature blocks, that stale mirror can surface as errors like thinking blocks cannot be modified.

Exact repro

Source-level reproduction:

1. Run persistCliTurnTranscript for an embedded assistant gap-fill with:
   • embeddedAssistantGapFill: true
   • result.meta.agentMeta.provider: "claude-cli" or another runtime provider
   • visible final text in result.meta.finalAssistantVisibleText
2. Before this fix, the appended transcript assistant row used api: "cli", provider: "claude-cli", and model: "opus".
3. That row is user-visible display state, not provider conversation state, so replay filters did not identify it as transcript-only.
4. After this fix, the same gap-fill row is written as api: "openai-responses", provider: "openclaw", and model: "delivery-mirror".

Fix summary

• Changed embedded assistant gap-fill transcript persistence to write assistant mirrors as openclaw/delivery-mirror.
• Kept normal CLI transcript persistence unchanged: non-gap-fill CLI replies still preserve real CLI provider/model metadata and usage.
• Added regression coverage so the gap-fill path asserts the transcript-only mirror metadata.

Metadata change (explicit callout)

This PR intentionally changes the persisted shape of embedded gap-fill assistant rows, so it is more than a replay fix — raw transcript / debug readers should be aware:

Scope is strictly the gapFill branch in persistCliTurnTranscript (src/agents/command/attempt-execution.ts:365). Normal (non-gap-fill) CLI rows are untouched and keep api:"cli" + real provider/model/usage.

Downstream api:"cli" consumers — verified none affected. The .api field is read only by the LLM transport/provider layer for foreign-tool-call detection (src/llm/providers/transform-messages.ts:100, openai-responses-shared.ts:202, comparing source.api === model.api). Gap-fill rows never reach those comparisons: rows tagged provider:"openclaw" + model:"delivery-mirror" are in TRANSCRIPT_ONLY_OPENCLAW_MODELS (src/agents/embedded-agent-runner/replay-history.ts:241) and are dropped from the replay copy before any provider prompt is built. Every other gap-fill/delivery-mirror filter in the tree keys on the provider/model pair, not api.

Targeted tests

• Extended src/agents/command/attempt-execution.cli.test.ts to prove embedded gap-fill rows are tagged as openclaw/delivery-mirror.
• Re-ran replay-history coverage to confirm delivery-mirror rows remain excluded from embedded provider replay.

Real behavior proof

Behavior addressed: Embedded subagent assistant gap-fill rows are now stored as OpenClaw delivery mirrors, so provider replay filters do not treat display-only text as provider conversation state.

Real environment tested: Local OpenClaw checkout on branch fix/subagent-announce-delivery-metadata, commit 5dcee2eae5, rebased onto current origin/main; Node v24.4.1.

Exact steps or command run after this patch:

• git diff --check origin/main...HEAD
• ./node_modules/.bin/oxlint src/agents/command/attempt-execution.ts src/agents/command/attempt-execution.cli.test.ts
• node scripts/run-vitest.mjs src/agents/command/attempt-execution.cli.test.ts src/agents/embedded-agent-runner/replay-history.test.ts

Evidence after fix:

Writer and replay halves verified together. The merge-readiness ask is to show the gap-fill row is written as openclaw/delivery-mirror and is then excluded from the next provider prompt. These two real test suites prove the two halves of exactly that path end to end:

• Writer half — src/agents/command/attempt-execution.cli.test.ts (25 tests passed): the embedded gap-fill assertion (embeddedAssistantGapFill: true) reads back the persisted transcript row and proves it is written as api: "openai-responses", provider: "openclaw", model: "delivery-mirror", while non-gap-fill CLI rows keep real provider/model metadata.
• Replay half — src/agents/embedded-agent-runner/replay-history.test.ts (35 tests passed): normalizeAssistantReplayContent is given a history containing exactly that provider: "openclaw" / model: "delivery-mirror" row alongside a real provider reply, and the test asserts the delivery-mirror row is dropped from the replay copy (out length 2) while the real provider reply survives. That is the row never re-entering the provider prompt.

Together: the shape the writer persists (suite 1) is precisely the shape the replay filter excludes (suite 2). Copied terminal summary:

 src/agents/command/attempt-execution.cli.test.ts
      Tests  25 passed (25)

 src/agents/embedded-agent-runner/replay-history.test.ts
      Tests  35 passed (35)

[test] passed 2 Vitest shards in 134.85s

• oxlint passed on the changed source and test files (exit 0).
• git diff --check origin/main...HEAD passed.

Live replay-path output — ran a single script that chains the patched writer to the replay filter on real functions (no hand-built input): it writes a session store, calls persistCliTurnTranscript with embeddedAssistantGapFill: true, reads the persisted JSONL row back off disk, then passes that exact row through normalizeAssistantReplayContent (src/agents/embedded-agent-runner/replay-history.ts):

STEP 1 — patched persistCliTurnTranscript wrote this JSONL assistant row:
[
  {
    "role": "assistant",
    "api": "openai-responses",
    "provider": "openclaw",
    "model": "delivery-mirror",
    "content": [{ "type": "text", "text": "delivered to channel" }]
  }
]

STEP 2 — that exact persisted row passed through normalizeAssistantReplayContent:
  rows in: 3  rows out: 2
  persisted delivery-mirror row excluded from replay: true
  real provider reply retained: true

This is the full patched writer-to-replay path on real code: the patched writer persists the gap-fill row to disk as openclaw/delivery-mirror, and that exact persisted row is then excluded from the rebuilt replay (3 → 2 rows) so it never re-enters the provider prompt, while the genuine provider reply is retained.

Observed result after fix: Embedded gap-fill transcript rows use api: "openai-responses", provider: "openclaw", and model: "delivery-mirror" while normal CLI transcript rows still preserve real provider/model metadata.

What was not tested: No live provider-managed thinking/signature conversation was replayed end to end against a running embedded agent and a real provider. That path requires a live provider session, which was not available in this environment; the writer-to-replay contract is instead proven by the two connected real test suites above on the actual production functions (persistCliTurnTranscript output shape and normalizeAssistantReplayContent exclusion).

Not tested / known gaps

• No external transcript/debug consumer was found in this checkout that depends on api: "cli" for embedded gap-fill rows, but external scripts that read raw transcript metadata may observe the intentional gap-fill metadata change from child CLI provider/model to openclaw/delivery-mirror.

AI-assisted disclosure

This patch was prepared with AI assistance. I used local code inspection, targeted regression tests, and explicit command output to validate the behavior above.