fix(verify): judge the current turn against its tool results

[jkyberneees]

Problem

With verify: {enabled: true, mode: "strict"}, accurate answers built from tool output came back verdict: fail and shipped as unverified. Three defects in internal/loop/verify.go:

1. The trace sent to the verifier contained only tool names and arguments clipped to 160 bytes, never a result. Every claim taken from tool output was unsupported by construction, so strict mode failed nearly every evidence-based answer and hint mode triggered needless answer_superseded retries.
2. The task was the first user message of the whole history. On serve sessions, from the second prompt on, the answer was judged against an earlier request.
3. The trace walked every turn with no cap, so the side-call prompt grew with session length.

Fix

• The trace is scoped to the latest real user turn (bg-notice user messages skipped, matching the engine's existing turn convention) and pairs each call with its result by tool_call_id.
• Each result is rendered as a head+tail excerpt (~1.5 KB) with an explicit [… N bytes omitted] marker; the stored TOOL RESULT frame is stripped, secrets are redacted, and every excerpt is wrapped as untrusted content via the engine's wrapper or its own nonce'd default. The whole trace is capped at ~14 KB with omitted calls counted.
• The task is the latest user prompt. Earlier turns reach the verifier only as a bounded block of user prompts and final assistant answers (~1 KB per message, ~4 KB total, most recent kept, redacted, wrapped as untrusted), so a follow-up answer that restates what an earlier turn established is judged against it; earlier turns' tool calls and results never enter the trace.
• The prompt keeps "prefer fail when in doubt" but tells the verifier excerpts may be truncated: a claim consistent with the visible excerpt counts as supported; contradictions, phantom tool runs and unanswered parts fail. Section labels now match what is sent. Result contents are framed as data, never instructions.
• No verifier prose reaches clients (unchanged).

Tests

• Current-turn scoping with results and no leakage from earlier turns; unmatched result handling; per-result and overall-budget truncation markers; untrusted wrapping plus redaction; latest-turn task selection.
• Regression: the real engine over a two-turn history with a stub verifier that passes only when a figure present solely in the tool output reaches its prompt.

Verification

• go test ./internal/loop/, go test -race ./internal/loop/ -run 'Verify|AnswerSuperseded', go test ./cmd/odek/, ODEK_E2E=true go test ./cmd/odek/ -run TestE2E_ all pass.
• Live reproduction with glm-5.3-flash as agent and verifier in strict mode on this branch (4 files, +348/−52): HEAD passed in a single-turn run and on the second turn of a saved session; a control run of the main binary reproduced the bug, failing a correct answer.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	odek	b04c026	Commit Preview URL Branch Preview URL	Oct 08 2026, 07:50 PM