Codex review: needs real behavior proof before merge. Reviewed July 16, 2026, 8:42 AM ET / 12:42 UTC.

Summary
Detects two Claude CLI <synthetic> placeholder messages, propagates that state through CLI run metadata, and classifies placeholder-only turns as fallback-worthy empty_result outcomes.

PR surface: Source +97, Tests +383. Total +480 across 7 files.

Reproducibility: no. live current-main reproduction was run here, but the parser path, captured JSONL fixture, and repeated timestamped logs in #99131 provide a high-confidence source reproduction of both the placeholder and the late-answer hazard.

Review metrics: 2 noteworthy metrics.

• Fallback classifications: 1 added. A new dependency-specific terminal outcome now enters the existing model/provider fallback pipeline.
• Terminal metadata kinds: 1 added. The new run-result state crosses parser, runner, and fallback-classifier boundaries and therefore needs one consistent lifecycle meaning.

Stored data model
Persistent data-model change detected: serialized state: src/agents/cli-output.test.ts. Confirm migration or upgrade compatibility proof before merge.

Root-cause cluster
Relationship: fixed_by_candidate
Canonical: #90789
Summary: This PR is the direct candidate for the silent-placeholder report; the later reused-session report shares the placeholder but adds a distinct lifecycle consequence that the current fallback policy must account for.

Members:

• canonical: [Bug] claude-cli backend: synthetic "No response requested." placeholder leaves Telegram turn fully silent with no observability #90789 - The PR explicitly proposes to fix this report's synthetic-placeholder message-loss behavior.
• same_root_cause: [Bug]: Claude CLI reused session emits synthetic empty result before real response, causing OpenClaw session discard and background task cascade #99131 - It records the same synthetic placeholder but shows that it can precede a later real result and damage reused-session state.
• partial_overlap: fix(claude-cli): preserve reused session on empty response #99139 - It mitigates session clearing after empty responses but does not settle placeholder terminality or fallback timing.
• adjacent_distinct: fix(agents): preserve streamed assistant text when Claude CLI result event is empty #90450 - It fixes loss of already-streamed text from an empty terminal result, while this PR handles placeholder-only metadata and fallback policy.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Merge readiness
Overall: 🧂 unranked krab
Proof: 🧂 unranked krab
Patch quality: 🦐 gold shrimp
Result: blocked until real behavior proof from a real setup is added.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

• [P2] Gate fallback on proven terminal lifecycle state and add regressions for a late real result after the synthetic marker.
• [P1] Provide redacted after-fix proof from an actual reusable Claude CLI session showing delivery and session behavior.

Proof guidance:

• [P1] Needs real behavior proof before merge: The TSX fixture harness and mocked runner test do not exercise an actual Claude CLI reusable session or user delivery; add redacted live terminal/log proof covering placeholder-only and placeholder-then-late-real sequences, update the PR body, and request @clawsweeper re-review if automatic review does not run.

Risk before merge

• [P2] A reused Claude CLI session can emit this placeholder before its real answer; starting fallback immediately can race the original process, duplicate work, or discard the eventual response.
• [P1] The recognized strings and their lifecycle meaning are based on captured behavior rather than a verified upstream Claude CLI contract, so future or context-dependent uses may not share one recovery policy.
• [P2] No after-fix proof exercises an actual Claude CLI reusable session, session binding, fallback decision, and resulting user delivery together.

Maintainer options:

1. Prove terminality before fallback (recommended)
   Revise the lifecycle so fallback starts only when the Claude turn cannot still produce a real answer, then cover both terminal and late-result sequences.
2. Own the immediate-fallback tradeoff
   Merge only with explicit owner approval and real reusable-session proof showing that late output, duplicate delivery, and session cleanup remain safe.
3. Pause for the session-lifecycle fix
   Keep this PR open or narrow it to metadata until the overlapping lifecycle behavior in [Bug]: Claude CLI reused session emits synthetic empty result before real response, causing OpenClaw session discard and background task cascade #99131 has one canonical recovery design.

Next step before merge

• [P1] An owner must choose the terminal-versus-continuation policy and require live reusable-session proof before the branch can safely merge; this is not a mechanical repair decision.

Maintainer decision needed

• Question: Should a recognized Claude CLI synthetic placeholder immediately terminate the OpenClaw attempt and start model fallback, or should reusable sessions continue waiting for a later real result from the same process?
• Rationale: The same dependency-emitted marker is observed in both terminal no-output cases and nonterminal reused-session cases, so code inspection cannot safely choose one recovery policy without owner intent and live lifecycle proof.
• Likely owner: steipete — The adjacent merged parser rewrite is the strongest current-main policy precedent for empty Claude CLI results.
• Options:
  • Gate on terminal state (recommended): Keep the marker but classify it as empty_result only after process/session evidence proves no later result can arrive.
  • Accept immediate fallback: Approve the current semantics and require live Claude CLI proof that fallback cannot race or discard a later answer in reusable sessions.
  • Split metadata from recovery: Land only placeholder detection and defer fallback/session behavior to a follow-up that resolves [Bug]: Claude CLI reused session emits synthetic empty result before real response, causing OpenClaw session discard and background task cascade #99131.

Security
Cleared: The diff changes internal CLI parsing, metadata, and fallback classification without adding dependencies, downloaded code, broader permissions, secret access, or package-resolution changes.

Review findings

• [P1] Do not fallback before the Claude turn is terminal — src/agents/embedded-agent-runner/result-fallback-classifier.ts:256-266

@clawsweeper re-review

Updated since prior review (which timed out before completing):

1. Reachability fix: executeCliAttempt previously threw FailoverError(empty_response) on every empty-text turn, swallowing the syntheticPlaceholder marker before buildCliRunResult could attach terminalReplyKind. The classifier hook was therefore unreachable from runPreparedCliAgent — the very path Telegram/auto-reply dispatch take. Fixed in 8d6f1a7 by skipping the throw when output.syntheticPlaceholder === true.
2. Integration test: new src/agents/cli-runner.synthetic-placeholder.test.ts drives runPreparedCliAgent end-to-end with a synthetic-placeholder CLI output and asserts both metadata propagation (terminalReplyKind) and the classifier verdict (empty_result). Addresses the prior P1 reachability feedback.
3. Real behavior proof: PR body now contains a populated ## Real behavior proof section with side-by-side baseline (upstream/main) vs candidate (PR head) terminal output captured from pnpm exec tsx running against the real source tree, plus the integration test live output. The Mantis 'agentic native Telegram proof' was capture-infrastructure-blocked (mock provider does not yet emit <synthetic> stream-json); that is documented as an out-of-scope follow-up.

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

• State: Complete
• Detail: The targeted re-review finished, the durable review comment was updated, and the synced verdict was routed.
• Run: https://github.com/openclaw/clawsweeper/actions/runs/27049299546
• Updated: 2026-06-06T02:01:09.953Z

Ready for maintainer review

ClawSweeper verdict: 🐚 platinum hermit (overall + patch quality) / 🦞 diamond lobster (proof) — "ready for maintainer review", no rank-up moves outstanding.

The current head cd1b3211 applies the patch that classifies Claude CLI <synthetic> no-output JSONL records as empty_result so the model-fallback path catches them instead of silently dropping the turn.

Remaining gates are both maintainer-side:

[P2] Merging changes delivery recovery semantics for a narrow Claude CLI no-output case: known <synthetic> placeholders become fallback-worthy empty results instead of generic empty CLI failures.
[P1] Native Telegram proof remains unavailable until qa-lab can emit this Claude stream-json shape; the supplied proof validates the runtime parser/classifier path rather than a full transport recording.

ClawSweeper's recommended landing option is "accept that exact Claude CLI synthetic placeholders with no visible assistant text should route through existing model fallback as empty results", which matches the PR's narrow scope. The qa-lab transport-fixture work is a separate proof gap that this PR can't unblock.

@jalehman / @steipete — pinging since you've recently reviewed and merged adjacent cli-runner.ts work. Happy to add any narrower regression case you'd like before merge.

Friendly bump — this has been sitting since 2026-06-06 with all gates green and no maintainer review yet.

State at head cd1b3211:

• mergeStateStatus: CLEAN, mergeable: MERGEABLE, no conflicts
• All required CI checks passing (Real behavior proof ✅)
• ClawSweeper: 🐚 platinum hermit (overall + patch quality), 🦞 diamond lobster (proof) — "ready for maintainer review", no rank-up moves outstanding
• Scope is narrow: +83 source / +303 tests across 7 files, only the Claude CLI <synthetic> no-output JSONL path

@vincentkoc / @joshavant — pinging since you've recently merged adjacent src/agents/cli-runner.ts work (#90912, #88161). Happy to scope down further or add any regression case you'd like before merge.

@shakkernerd could I ask you to take a look at this, or point me to the right owner?

It fixes a small Claude Code CLI runtime edge case: on timeout/wedge the runtime can emit a <synthetic> "No response requested" placeholder that OpenClaw doesn't recognize, so the agent turn ends silently with no reply to the user.

Asking you specifically because I noticed you recently worked on the adjacent CLI empty-response diagnostics path — same area as the main review surface here: cli-runner.ts / cli-output.ts.

CI is green and tests are included. The diff looks bigger than the bug because it bundles test coverage + some cleanup; the functional surface is really just those two files. Happy to rebase whenever it's useful for review/merge.

Rebased onto latest main and force-pushed a cleaned branch (8706888).

An earlier merge resolution had accidentally reverted an upstream repo-wide import-ordering autofix across ~100 files, which is what had turned this PR conflicting and inflated the diff to 107 files. That churn is gone — the PR is back to the intended 7-file diff (+387/−4) and MERGEABLE again.

The real review surface is unchanged:

• src/agents/cli-output.ts — detect the Claude CLI <synthetic> "No response requested." placeholder, flag it instead of swallowing the turn
• src/agents/cli-runner.ts — route flagged turns through the fallback classifier instead of throwing the generic empty-response failover
• src/agents/embedded-agent-runner/result-fallback-classifier.ts + types.ts — classify it as empty_result
• plus targeted parser/runner/classifier tests

Local verification on the rebuilt branch: tsgo:core clean, and the three affected test files pass (cli-output.test.ts, cli-runner.synthetic-placeholder.test.ts, result-fallback-classifier.test.ts — 70 tests). CI is re-running.

Updated this PR to address ClawSweeper's P2 ordering concern.

What changed at head eb2ab0a2:

• Synthetic-placeholder fallback now only classifies empty_result after stronger terminal semantics are respected: committed delivery, hook_block, generic external failure, visible user-facing output, and fallback-safe incomplete_turn all keep their existing precedence.
• Added regressions for:
  • synthetic-placeholder marker + non-empty finalAssistantVisibleText → no fallback
  • synthetic-placeholder marker + visible payload → no fallback
  • synthetic-placeholder marker + hook_block → no fallback
  • synthetic-placeholder marker + fallback-safe incomplete_turn → incomplete_result with preservation metadata

Local verification:

• node scripts/run-vitest.mjs src/agents/embedded-agent-runner/result-fallback-classifier.test.ts src/agents/cli-runner.synthetic-placeholder.test.ts src/agents/cli-output.test.ts — 74 tests passed
• pnpm exec oxlint src/agents/embedded-agent-runner/result-fallback-classifier.ts src/agents/embedded-agent-runner/result-fallback-classifier.test.ts src/agents/cli-runner.ts src/agents/cli-runner.synthetic-placeholder.test.ts src/agents/cli-output.ts src/agents/cli-output.test.ts src/agents/embedded-agent-runner/types.ts — 0 warnings/errors
• pnpm exec oxfmt --check src/agents/embedded-agent-runner/result-fallback-classifier.ts src/agents/embedded-agent-runner/result-fallback-classifier.test.ts src/agents/cli-runner.ts src/agents/cli-runner.synthetic-placeholder.test.ts src/agents/cli-output.ts src/agents/cli-output.test.ts src/agents/embedded-agent-runner/types.ts — clean
• git diff --check — clean
• Independent Oracle review of the final patch: OK, no blocking issues, safe to commit/push.

Heads up on CI: the only failing check (CI / checks-node-compact-small-whole-2) is inherited main breakage, not from this PR.

• It fails in test/scripts/render-maturity-docs.test.ts (maturity scorecard HTML): the renderer now emits maturity-summary-grid / 68%, while the test still asserts the old <span className="maturity-summary-value">67%</span>.
• The same test fails with the identical assertion on main's own push CI — run 7fa26e088, job checks-node-core-tooling.
• This PR's diff only touches src/agents/embedded-agent-runner/result-fallback-classifier.ts and its test, which cannot affect the maturity-docs renderer/test.

Head eb2ab0a2 is green on the relevant surface:

• node scripts/run-vitest.mjs src/agents/embedded-agent-runner/result-fallback-classifier.test.ts src/agents/cli-runner.synthetic-placeholder.test.ts src/agents/cli-output.test.ts — 74 tests passed
• oxlint / oxfmt --check on the 7 touched-area files — clean
• git diff --check — clean

The Real behavior proof / auto-response CANCELLED entries are superseded older runs. I'll rerun / rebase once main is green again on the tooling shard.

Merged using squash.

What changed after review:

• Replaced global fallback classification with bounded lifecycle handling in the Claude live-session owner.
• Limited continuation grace to newly resumed processes and fixed no-output watchdog ordering.
• Added focused lifecycle regression coverage and corrected the PR to close [Bug]: Claude CLI reused session emits synthetic empty result before real response, causing OpenClaw session discard and background task cascade #99131 while keeping [Bug] claude-cli backend: synthetic "No response requested." placeholder leaves Telegram turn fully silent with no observability #90789 as related context.

Verification:

• Rebased/refreshed onto current origin/main before pushing.
• Blacksmith Testbox tbx_01kxpnrhyje1er0j8afmjpnvwt: focused 14-test suite and pnpm check:changed passed.
• Blacksmith Testbox tbx_01kxppc7c4d4mfdrek1wxexm1y: live Claude subscription Docker resume pipeline passed with Claude Code 2.1.211.
• Auto-review: clean after accepted lifecycle and timer-race findings were resolved.
• GitHub checks: full CI green on the patch-identical pre-rebase head; exact-head real behavior proof green; no required checks configured.

Landed on main:

• Final landed SHA: 2972db5; range d71c1fe...2972db5.

Thanks @wangwllu!