frontend-patterns/research/reviews/TruncationDiagnosis.md

Truncation diagnosis

Date: 2026-09-25. Reviewer: codex, self-review. Owner: FEP-WP-0002-T02.

Findings

Inspected llm-connect checkout 37436bb. Its _payload.py explicitly drops reasoning_effort before OpenRouter execution. openrouter.py retains only three aggregate usage counters in its normal LLMResponse, omitting detailed token counts. Per-call /execute?debug=1 already exposes provider-bound settings and raw usage; no deployment, credential retrieval or upstream repository edit was required.

Fresh local-first -04: requested low reasoning effort, 1,000 completion tokens; first response ended length. Live diagnostics confirmed neither reasoning field was sent; the provider reported zero reasoning tokens. Output was 30,595 characters. Reported usage: 525 prompt + 1,000 completion tokens. The numeric provider cost field was retained; it is not an independent billing reconciliation.

Fresh -05: omitted the ineffective requested reasoning override and kept the same 1,000-token limit. The rejected generated text was quarantined under reports/rejected-response.md, separate from workflow outputs. It contains three lines: a claim-table heading and partial header, with 30,544 spaces out of 30,595 characters. This is direct evidence of pathological table padding in this response, not excessive substantive analysis. No raw provider debug body or header is stored.

The two diagnostic attempts each stopped after one service request. This diagnosis does not retroactively establish the content of prior runs whose text was not saved, or prove a general model defect. It identifies an actionable failure mode here.

Intervention

Replaced the claims table with labeled bullet records. All four prompts explicitly forbid padded tables. Preserved source bytes and historical workspaces. Fresh -06 requested 3,000 completion tokens per response, at most four service requests, no client retries, same llm-connect profile; diagnostics enabled. Internal retries are outside the client ceiling. No claim that low reasoning effort was applied.

The client now retains only allowlisted diagnostics, plus normal generated incomplete text as a rejected artifact. Tests cover diagnostic data exclusion and quarantine; a compatibility check documents the upstream adapter's current ignored parameter.

Outcome and editorial review

Local-first -06 completed all four stages with finish reason stop; source hash is unchanged. Receipt. This successful run supports the compact-format workaround for this packet. It does not prove future runs cannot truncate; the client still rejects incomplete output.

Structural screening passes, with admission still not_granted. The model returned a valid terminal JSON candidate ledger without Markdown fences. Screening now accepts one unambiguous terminal JSON document as well as a fenced block; it does not repair syntax or infer a ledger from prose. Regression tests cover this exact live result and reject ambiguous/malformed JSON.

Editorial disposition: retain the bounded negative result; admit no candidates.

• C01 faithfully paraphrases the local-primary-copy fragment.
• C02 substitutes "authoritative version" for "primary copy". Treat that as an interpretation requiring additional evidence, not a second source-established fact. The assisted critique correctly flagged this leap; the editorial review agrees.
• The candidate stage explicitly finds insufficient evidence for a complete pattern; its candidate ledger and relation array are empty. This is a legitimate research outcome, not a pipeline failure or rejection of the existing manual hypothesis.
• Persistence, concurrency, recovery and independent known uses remain unestablished by the packet. VC-02 relevance is a hypothesis, despite the critique's "Pass" label; no scenario walkthrough has been performed by this run.
• Manual review claims C02/C03 from broader source sections are distinct from these workspace-local labels. Do not merge them by ID alone.

The immediate execution obstacle is resolved for this run. T02's next research step is to expand the source packet to located persistence/synchronization/conflict passages and independent implementation evidence. Then T03 can compare restart and reconnection behavior in VC-01/VC-02. No more live calls are needed for this diagnosis.

Completed run reported total tokens: 5660.

Whitespace validation exception: the quarantined -05/reports/rejected-response.md intentionally retains the provider's trailing spaces as diagnostic evidence. All other staged files pass git diff --check; trimming this file would destroy the captured failure and invalidate its receipt hash.