Nightly 2026-04-20 — 2 cycles, +15 goal weight, 1 dream finding closed

[boshu2]

Nightly summary

Bounded autonomous nightly pass on nightly/2026-04-20 branched from origin/main @ 6bbc0307.

	Baseline	Final	Delta
Pass / fail	14 / 5	17 / 2	+3 pass
Weighted fitness	80 / 109	95 / 109	+15 weight
Percent	73.4%	87.2%	+13.8 pp

Anchor: previous nightly branch origin/nightly/2026-04-17-v2 @ 4269c665 (no nightly git tags existed yet — this run creates the first).

Per-cycle summary

Cycle 1 — 6103c2e2 fix(provenance): accept structured source frontmatter in learning audit

Dream's INGEST stage flagged .agents/learnings/2026-04-19-orchestrator-compression-anti-pattern.md every iteration with yaml: cannot unmarshal !!map into string, because learningFrontmatter.Source was typed as string while the file uses a map (source: {session, evidence}). Swapped Source for a flexibleString wrapper whose UnmarshalYAML accepts scalar, map, or sequence nodes; structured values decode to a non-empty sentinel so the "missing source" predicate keeps working. Added a regression test (TestProvenanceAudit_StructuredSourceAccepted) covering all three shapes.

Cycle 2 — 8778fa97 fix(codex): convert remaining Skill() invocations to $skill notation

Closed the low-severity council finding under 2026-04-19-rpi-dag-hardening in .agents/rpi/next-work.jsonl. Rewrote all 11 remaining Skill(skill="…", args="…") calls in skills-codex/{design,implement,red-team,validation}/SKILL.md to the canonical $<name> form ($vibe recent, $test coverage --quick, $post-mortem <epic-id>, etc.). Prose examples inside skills-codex/shared/references/strict-delegation-contract.md left as documentation of the Claude-native shape. Regenerated codex hashes (manifest + per-skill markers); bash scripts/audit-codex-parity.sh and bash scripts/lint-codex-native.sh pass.

Findings

Finding	Status
Dream provenance-audit YAML parse error on structured source:	closed (cycle 1)
2026-04-19-rpi-dag-hardening: remaining Skill() notation in codex	closed (cycle 2)
2026-04-19-rpi-dag-hardening: "Decompose skills/crank/SKILL.md under 248 lines"	superseded — tests/skills/lint-skills.sh no longer enforces line count when references/ is present (crank has one)

No new findings opened, no auto-reverts, no quarantined goals.

Goal results

Still failing (both out of nightly scope):

• [8] flywheel-compounding — σρ > δ not met in the measurement window; this is a knowledge-lifecycle metric that needs corpus activity, not a code fix.
• [6] go-complexity-ceiling — gocyclo binary not installed in this runtime environment; no code regression.

Moved from fail → pass:

• [7] flywheel-proof — baseline failure was transient: sum.golang.org 503 DNS cache overflow during proof-run toolchain download.
• [4] compile-freshness — baseline lacked .agents/defrag/latest.json; generated by a dry-run ao defrag during cycle selection (runtime state, not a committed artifact; nightly CI runs ao defrag before measuring, so this also reflects correct CI flow).
• [4] compile-no-oscillation — same root cause as compile-freshness.

Final validation bundle

Check	Result
ao autodev validate --file PROGRAM.md --json	valid: true
go build ./... / go vet ./...	clean
go test ./cmd/ao ./internal/autodev ./internal/provenance ./internal/overnight	all pass
skills/heal-skill/scripts/heal.sh --strict	All clean. No findings.
tests/docs/validate-links.sh	1528 links, 0 broken
scripts/audit-codex-parity.sh --skill {design,validation,implement,red-team}	pass
scripts/lint-codex-native.sh	0 errors, 0 warnings
scripts/pre-push-gate.sh --fast	1 FAIL, 32 skipped — the sole failure is worktree-disposition, which the nightly pipeline documents as a known false positive on a nightly/* branch (the script expects the canonical root on main). All other checks pass.

Commit links

• 6103c2e2
• 8778fa97

Anchor for tomorrow's audit

This branch will be tagged nightly/2026-04-20 after the PR is opened so the next nightly has a concrete fitness anchor to diff against.

---

Generated by Claude Code