Cross-lineage scrum (Opus 4.7 / Kimi K2.6 / Qwen3-coder via chatd's
/v1/chat) on the harness's first 4 commits surfaced 5 real bugs;
this commit lands the 4 inside the LLM/validator stack. B5 (scanner
skip-list semantics) ships separately as it changes scan behavior
on every target repo.
B1 (Kimi BLOCK + Opus WARN convergent) — internal/validators:
evidencePresent had two flaws: (1) cursor advanced on match in the
trim-line fallback, breaking same-line repeated matches AND skipping
not-yet-considered lines so out-of-order evidence spuriously failed;
(2) strings.Contains on a single `}` trim-matched any closing brace
in the file, defeating the "evidence quotes real text" contract.
Fix: trivial-evidence guard FIRST (reject anything <4 non-whitespace
chars) + per-line search no longer advances a cursor. New regression
test TestEvidencePresent_RejectsTrivialMatches covers `}`, `{`, `)`,
empty, and out-of-order multi-line evidence (which now passes —
order isn't part of the contract).
B2 (Kimi WARN + Opus WARN convergent) — internal/pipeline:
WriteJSON error for rejected-findings.json was swallowed with
`if err == nil`, so a write failure left the validation phase
reporting status="ok" while the audit trail vanished. Mirror the
validated-findings branch: surface the error in
validatePhase.Errors + bump status to degraded + ExitCode=66.
B3 (Kimi BLOCK + Opus BLOCK convergent) — internal/llm/ollama.go:
HealthCheck.basic_prompt_ok was set to true on ANY non-empty
response, so a model emitting `<think>...` traces or apologies
passed silently. Now requires the response to contain "OK"
(uppercase, substring). Substring rather than equality lets minor
whitespace/punctuation variations through (some models add a
trailing period). Errors now record what the model actually said
when it fails the check.
B4 (Opus BLOCK only — same class as today's chatd Anthropic-temp
fix) — internal/llm/ollama.go: chatBody had `if opts.Temperature != 0`
which silently dropped Temperature=0 from the request, so HealthCheck
+ Reviewer (both pass Temperature=0 expecting determinism) actually
ran at Ollama's ~0.8 default. Always forward Temperature now. The
two callers always set explicit values, so "0 means 0" is correct;
if a future caller wants Ollama's default they'll switch
CompleteOptions.Temperature to *float64 like chatd did this morning.
Verified end-to-end: insecure-repo + --enable-llm still produces 25
confirmed findings (16 static + 9 LLM), 0 rejected. Validator unit
tests: 11 pass (added TestEvidencePresent_RejectsTrivialMatches).
Same-day-as-shipping scrum, same-day-as-shipping fixes. The
convergent-≥2 gate caught 3 of these; the 4th was Opus-only but
verified by reading the code (same idiom as today's chatd bug).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Implements PROMPT.md / docs/REVIEW_PIPELINE.md Phase 2:
- internal/llm/ollama.go — real Ollama provider:
- HealthCheck probes /api/tags + a 1-token completion + a JSON-mode
probe ({"ok": true} round-trip), populating the model-doctor.json
schema documented in docs/LOCAL_MODEL_SETUP.md
- Complete + CompleteJSON via /api/chat with stream=false
- think=false set for ALL completions (qwen3.5:latest is reasoning-
capable but the inner-loop hot path wants direct answers, not
reasoning traces consuming the token budget — same finding as
the Lakehouse-Go chatd 2026-04-30 wave)
- internal/llm/review.go — Reviewer wrapper:
- 2-attempt flow: prompt → parse → repair-prompt → parse
- Strict JSON shape enforced; markdown fences stripped before parse
- Severity normalized to enum; out-of-range confidence clamped
- Per-file chunking (file-level for v0; function-level Phase D+)
- Bounded by review-profile max_file_bytes + max_llm_chunk_chars
- pipeline.go — Phase 2 wired between static scan + report gen:
- --enable-llm flag opts in (off by default — static-only is
cheaper and faster)
- Raw output ALWAYS saved to llm-findings.raw.json (forensics)
- Normalized findings → llm-findings.normalized.json
- LLM findings merged into the report findings list (sourced
"llm" so consumers can filter)
- Receipts honestly mark phase status: "ok" | "degraded" | "skipped"
- cli model doctor — real probes replace the Phase A stub.
Verified:
- model doctor: status="ok" with qwen3.5:latest + qwen3:latest both
loaded, basic_prompt_ok=true, json_mode_ok=true
- insecure-repo with --enable-llm: 9 LLM findings; qwen3.5 correctly
flagged SQLi, RCE, hardcoded credentials as critical with verbatim
evidence; 27s wall for 3 chunks
- clean-repo with --enable-llm: 0 LLM findings, 4 parsed chunks, 2.8s
- self-review with --enable-llm: 77 LLM findings + 83 static; 3 of
~30 chunks needed retry (PROMPT.md, REPORT_SCHEMA.md,
SCRUM_TEST_TEMPLATE.md — all eventually parsed); 5min wall
go vet + go test -short clean. Fixture stray.go now `package fixture`
so go-tooling doesn't choke on the orphan.
Phase D (validator cross-check) + Phase E (memory + diff/rules
subcommands) remain.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Claude (review-harness setup)