Nightly 2026-05-08 — 1 productive cycle, +0 code-driven goals (already 100%), closed ONE-TIME #7 probe registry

[boshu2]

TL;DR

• Code-driven score: 100.0% → 100.0% (19/19 passing, 1 skipped — flywheel-compounding under f-2026-04-30-002 dormancy precondition).
• Runtime-artifact score: 0/2 → 2/2 (compile-freshness, compile-no-oscillation; both flipped because Dream re-ran during this routine and rewrote .agents/overnight/latest/defrag/latest.json, which the goal scripts now fall back to. Per goals.yaml runtime-artifact tag these are excluded from the headline).
• One productive cycle (registry refactor); zero stale-audit cycles; zero auto-reverts; zero transient-flake non-reverts.
• Open-PR blocker matrix: empty. No nightly/triage PRs against main at prelude.
• Inherited-from-main CI failures: 3 (all environment-related: bats not installed in canonical VM image fails pre-push-bats-suite and release-gate-contract-tests with exit 127; cursor-runtime-smoke fails for similar reasons). Verified by running pre-push-gate.sh --fast at HEAD == origin/main before any cycle began. None introduced by this branch.

Cycle summary

#	Type	Target	Commit	Code-driven before/after
1	refactor + tests	cli/cmd/ao/overnight_packets.go: dream-curator probe-shape registry table	3df3689	100.0 → 100.0

The headline is flat because the repo entered the run at code-driven 100%. The cycle closes one of the prompt's named ONE-TIME REPO FIXES (#7: "Promote dream-curator probe shapes to a registry table; emit-side probing must consult it"). PR #252's body explicitly listed this as deferred — it is now landed.

Cycle 1 — dream-curator probe-shape registry

Refactor probeDreamPacketStaleness from a hardcoded chain of if probeX(...); ok calls into a registry walk over dreamProbeShapes []dreamProbeShape{ID, Probe}. Adding a new false-positive title shape used to require two coordinated edits (write a new Go function AND extend the if chain) which created drift; now it is a single registry append.

Registers the four shipped shapes:

• target-files-exist
• repo-ref-tokens
• skill-line-limit-claim
• add-to-skill-claim

Behaviour-preserving: order in the registry matches the prior if chain, so degraded-finding reason strings surface in the same order. New test TestDreamProbeShapes_RegistryIDsAreUniqueAndStable guards the registry contract (every shape has a non-empty unique ID; canonical four are present). All existing probe-shape tests pass verbatim.

End-to-end verification: cd cli && go test ./cmd/ao 30.9s green; ao overnight start --max-iterations 1 --warn-only re-run against this branch synthesizes 0 morning packets against the dormant corpus, same as the baseline, and the curator's probe pipeline still walks the registered shapes (no behavior delta).

Fitness delta (per goal)

Goal	Weight	Tags	Baseline	Final
flywheel-compounding	3	long-cycle, corpus-state	skip	skip
dream-end-user-coverage	3	—	pass	pass
flywheel-proof	7	—	pass	pass
skill-frontmatter	6	—	pass	pass
hook-preflight	6	—	pass	pass
go-cli-builds	8	—	pass	pass
go-cli-tests	8	—	pass	pass
go-vet-clean	5	—	pass	pass
go-complexity-ceiling	6	—	pass	pass
security-gate	6	—	pass	pass
manifest-versions-match	5	—	pass	pass
wiring-closure	7	—	pass	pass
contract-compatibility	5	—	pass	pass
goals-validate	5	—	pass	pass
compile-freshness	4	runtime-artifact	fail	pass
compile-no-oscillation	4	runtime-artifact	fail	pass
competitive-freshness	3	—	pass	pass
codex-parity-drift	5	—	pass	pass
install-smoke	5	—	pass	pass
flywheel-lifecycle	6	—	pass	pass
eval-workbench-verify	6	—	pass	pass
state-path-resolver-coverage	3	warn-only	pass	pass

The two runtime-artifact flips (compile-freshness, compile-no-oscillation) are a Dream side effect (Dream wrote .agents/overnight/latest/defrag/latest.json and the goal scripts read that fallback path); they do not contribute to the headline delta per the prompt's runtime-artifact tabulation rule.

Stale-audit accounting

• Inline-probe rejections this run: 0 (Dream emitted 0 morning packets — see Dream summary below).
• Explicit stale-audit cycles: 0 (cap held at 0 because probe-stale rate is 0/0 — well below the 30% required-trigger threshold).
• Bookkeeping/file-move/comment-only cycles: 0 (the 497dcdd audit-truth bookkeeping commit is a follow-up to cycle 1, not a cycle of its own; cycle accounting follows the productive-cycle definition exactly).

Auto-reverts and transient flakes

• Auto-reverts this run: 0.
• Transient-flake non-reverts this run: 0 (cold-cache go-cli-tests truncation observed in the very first prelude measure as in PR Nightly 2026-05-07 — 3 productive cycles, +0 code-driven goals (already 100%), 2 audit-truth regressions fixed #252; warmed cache via go test -count=1 ./internal/... before snapshotting baseline. The persisted baseline-goals.json shows go-cli-tests PASS).

Quarantined goals (monitored, no fitness contribution)

• None tagged quarantined: true in goals.yaml.
• flywheel-compounding (w=3, long-cycle, corpus-state) remains in SKIP state under the f-2026-04-30-002 precondition — corpus is fully dormant (σ=0 ρ=0, 0 citations, 0 artifacts, 0 learnings created over the last 7 days). Not re-attempted; will not be re-attempted on dormant corpus.

Heavy-goal attempt history

No heavy goal (weight ≥ 5) was failing in code-driven space at the start of this run, so the heavy-goal selection ladder rung was empty. Time-pressure escalation did not fire (no unblocked heavy failing goals exist; per the prompt's rule "the escalation does NOT fire if every heavy goal is in the blocker matrix; in that case drop to ladder step 2" — same outcome with an empty heavy-failing set). No attempts.jsonl entries authored this run.

Dream probe-results summary

• Total morning packets emitted by curator: 0.
• Stale rate: 0/0 (no packets to probe).
• New shape encountered: none (no packets, no shapes).
• Generator candidates aggregated this run: 0 from mine-findings, 0 from external-watchlist. close_loop_promoted=0, harvest_promoted=7 (auto-promotion only, none routed to packets). Corpus dormancy continues to suppress Dream's actionable output, which is the steady state since the f-2026-04-30-002 SKIP precondition.

The probe-shape registry refactor (cycle 1) does not change probe behavior; it only changes how new shapes are added. The next time a recurring stale shape is observed in production, the work to suppress it is a registry append rather than a coordinated function-write + chain-edit.

Open-PR blocker matrix

{"open_prs": 0, "blocked_goals": [], "matrix": []}

No nightly or triage PRs open against main at prelude. Persisted to .agents/nightly/2026-05-08/blocker-matrix.json.

Inherited-from-main CI failures

Verified by running scripts/pre-push-gate.sh --fast at HEAD == origin/main (before any cycle commit). All three failures are environment-related and not introduced by this branch. Persisted to .agents/nightly/2026-05-08/main-ci-baseline.json:

• agentops-core.pre-push-gate-governance / pre-push-bats-suite — exit 127 (bats not installed in canonical VM image)
• agentops-core.release-security-gates / release-gate-contract-tests — exit 127 (bats not installed in canonical VM image)
• agentops-core.runtime-smoke-matrix / cursor-runtime-smoke — exit 1 (cursor runtime smoke harness fails in canonical VM image)

The bats installation is one of the prompt's "ONE-TIME REPO FIXES" (#4 / #4-adjacent — VM image change, not a per-PR fix).

Findings

• Opened: 0 new findings as .agents/findings/f-*.md.
• Closed via implementation: 0 explicitly via the validator-closure path.
• Deferred: none (no findings exist in .agents/findings/; the registry remains empty across nightlies because no run has yet authored to it. This is mechanically correct given corpus dormancy — Dream's mine-findings and external-watchlist generators both report candidate_count=0 this run).

ONE-TIME REPO FIX status (from the routine prompt):

#	Description	Status
1	.gitignore exceptions for persisted JSON paths	DONE (PR #252 cycle 1)
2	f-2026-04-30-002 corpus-active SKIP precondition	DONE (flywheel-compounding returns SKIP)
3	Tag compile-freshness/compile-no-oscillation as runtime-artifact in goals.yaml	DONE (already tagged)
4	Install bd in canonical VM image	open (VM ops; out of per-PR scope)
5	WORKTREE_DISPOSITION_ALLOW=nightly/* env support	DONE (verified working in this run)
6	affects_files: per goal in goals.yaml	DONE (all 22 goals covered in goals-affects-files.yaml)
7	Promote dream-curator probe shapes to a registry; emit-side consult	DONE THIS RUN (cycle 1)
8	Audit .agents/findings/registry.jsonl 7/11 numbered findings	open (registry empty; nothing to audit yet)

bd

bd unavailable in this VM image. Selection ladder step 3 (open ready beads) skipped per spec (single-line note in the prompt's bd-unavailable handling).

Previous-run audit (anchor: 60911ff — Nightly 2026-05-07 #252 merge commit)

The branch origin/nightly/2026-05-07 was deleted post-merge, so the audit anchors on the merge commit instead. Used the persisted, tracked baseline-goals.json from .agents/nightly/2026-05-07/ for the fitness comparison (this is the first nightly post the gitignore force-track fix where a real anchor JSON exists in main).

• Anchor final: code-driven 19/19 passing, 1 skip, runtime-artifact 2/2 passing (Dream side effect during the run; not in main).
• Today's baseline: code-driven 19/19 passing, 1 skip, runtime-artifact 0/2 (Dream had not yet run today's prelude).
• Code-driven delta: 0 (steady at 100%).
• Regressions: none on code-driven goals.
• 11 commits since 60911ff (3 PR merges + 8 chore/fix commits on main) — including CI toil-reduction sweep: registry determinism + codex auto-refresh + local CI dry-run (3 beads) #261 (CI toil-reduction sweep), docs(positioning): wiki-framing sweep across all surface docs (soc-9xn0) #262 (wiki-framing sweep), fix(hooks): reconcile stale .agents/.gitignore deny-all (soc-rv5p) #263 (.agents/.gitignore reconcile), Wave 1A-D: factory-claim-ledger reconciliation (soc-e4ulx) #264 (factory-claim-ledger reconciliation Wave 1A-D).

Full audit: .agents/nightly/2026-05-08/previous-run-audit.md.

Commits

• 3df3689 refactor(dream): promote curator probe shapes to a registry table
• 497dcdd chore(nightly): persist final-goals.json and cycle-history commit SHA

Test plan

• cd cli && go build ./... clean
• cd cli && go vet ./... clean
• cd cli && go test -count=1 ./... green (cmd/ao 29.2s, all internal/ packages cached green)
• cd cli && env -u AGENTOPS_RPI_RUNTIME go run ./cmd/ao autodev validate --file ../PROGRAM.md --json valid
• bash skills/heal-skill/scripts/heal.sh --strict → "All clean. No findings."
• WORKTREE_DISPOSITION_ALLOW='nightly/*' bash scripts/check-worktree-disposition.sh → PASS (verifies ONE-TIME fix: align with Anthropic marketplace standards #5 still works)
• ao goals measure --json → code-driven 100% (19/19 + 1 skip), runtime-artifact 2/2 (final)
• ao overnight start --max-iterations 1 --warn-only re-run against this branch — same 0-morning-packet behavior as the baseline (probe registry is behaviour-preserving)
• CI on this PR (will run on push). Expected red signals: the three inherited-from-main env failures (pre-push-bats-suite, release-gate-contract-tests, cursor-runtime-smoke) — see "Inherited-from-main CI failures" above. No new red signals expected.

https://claude.ai/code/session_01Gq1ohV1YWC55EUdycQEbfX

---

Generated by Claude Code

[boshu2]

Code-only portion cherry-picked to main in 635ca9d. Closing this stale nightly PR rather than merging the repo-root .agents runtime artifacts.