everyinc__compound-engineering-plugin/docs/solutions/integration-issues/grok-camelcase-structuredoutput-and-nonfinal-peer-position.md

title, date, category, module, problem_type, component, severity, symptoms, root_cause, resolution_type, tags

title	date	category	module	problem_type	component	severity	symptoms	root_cause	resolution_type	tags
Grok's camelCase structuredOutput and schema-valid non-final positions slip past cross-model peer acceptance	2026-08-15	integration-issues	cross-model structured output (skills/ce-pov)	integration_issue	tooling	medium	grok-cli peer route in a ce-pov oracle panel finished with worker exit 0 and a schema-valid artifact whose position was a placeholder ('blocked: gathering subject evidence') instead of a settled verdict parse_structured() only checked snake_case .structured_output, so grok's camelCase .structuredOutput envelope key fell through to the text-scan recovery path grok-4.6 sometimes emits a schema-shaped placeholder JSON on its first turn and, if it stops after that turn, the placeholder becomes the final structuredOutput	wrong_api	code_fix	cross-model structured-output json-schema peer-delegation ce-pov grok

Grok's camelCase structuredOutput and schema-valid non-final positions slip past cross-model peer acceptance

A ce-pov grok-cli peer returned, in ~20s with exit 0, a schema-valid artifact whose position was "blocked: gathering subject evidence". Every mechanical gate passed it; only orchestrator judgment dropped the voice. The Codex peer in the same panel returned a grounded position, so the failure was route-specific.

The shipped fix (PR #1403) is visible in the code: parse_structured() in skills/ce-pov/scripts/cross-model-pov.sh reads .structured_output // .structuredOutput; skills/ce-pov/references/pov-schema.json requires a final boolean; the worker retries a non-final artifact once on the same route inside the remaining hard window (CROSS_MODEL_RETRY_MIN_SECS), then drops it with peer skip evidence: non-final position; candidate selection scores final-and-shaped (2) > shaped (1) > any position-bearing object (0); tests/skills/ce-pov-cross-model-routes.test.ts pins each path. What the code does not say is why two more obvious designs were rejected.

Rejected: "evidence must cite something beyond the payload" as the finality condition

A document-only POV (a spec with no code to inspect) legitimately cites only the payload. Gating on "cites more than the payload" false-positives on every valid document-only POV, not just the placeholder.

Rejected: reading finality out of the prose

An earlier draft matched a phrase list (blocked|pending|gathering|…). Two PR review rounds showed both failure directions: a settled Blocked — insufficient project grounding verdict misclassified as unfinished, and any routine wording variation (Blocked: I am still gathering…) slipping past. Model prose cannot be exhaustively classified; an owned schema field can. So finality is the peer's own declaration through required final, and a shaped artifact that omits it is non-final (fail-closed). The acceptance semantics live once, in skills/ce-pov/references/cross-model-panel.md; the deterministic check lives once, in the worker. Do not duplicate one into the other.

Also not the cause: --json-schema blocking tool use on grok-cli

A direct repro on grok 1.0.4 showed tools do work under --json-schema (num_turns: 2, evidence cited a read file). The model's habit of emitting a schema-shaped placeholder before spending read turns is the cause. In 1 of 4 repro runs the envelope's text field held the placeholder concatenated with the final object while structuredOutput held only the final, which is why selection scores every position-bearing object in the envelope rather than trusting one key.

Why the recovery fallback had masked this

recover_pov_json returned the last dict containing a position key, key-agnostic and finality-blind. Because the snake_case lookup always missed on grok, execution always fell into that scan, and it "worked" until the model's final object was itself a placeholder. Guard-shaped fixes were patched four review rounds in a row; the version that held restated the block as a scoring rule.

When adding a peer CLI route

Verify the envelope field name against that CLI's actual headless JSON output (grok-cli 1.0.4 uses camelCase structuredOutput; the code-review and doc-review workers already handled it, ce-pov's copy had lagged). If the schema's own position description admits an in-progress or blocked value, schema validity does not imply usability; the worker needs an explicit finality field. Every route's stub fixture must carry a new required field, because the gate is fail-closed.

Related Issues

• docs/solutions/skill-design/dispatch-script-failure-degrade-outcome-not-boundary.md — the same-route, boundary-frozen, bounded-retry principle this fix reuses.
• docs/solutions/integration-issues/portable-structured-output-schemas-across-model-clis.md — schema rejection at startup vs. schema acceptance of non-final content here.
• docs/solutions/skill-design/cli-output-buffering-for-progress-detection.md and docs/solutions/skill-design/quiet-interval-floors-for-streaming-peer-routes.md — grok-cli --json-schema buffering and hard-only timeouts for the same route.
• Issue #1270