benchmark/v2: v2.5 delta — S-02 16k padding + preflight, per-hook sweep, run plans, AWS docs#57
Open
Kaushik985 wants to merge 4 commits into
Open
benchmark/v2: v2.5 delta — S-02 16k padding + preflight, per-hook sweep, run plans, AWS docs#57Kaushik985 wants to merge 4 commits into
Kaushik985 wants to merge 4 commits into
Conversation
…sweep, run plans, AWS docs Incremental v2.5 work on top of the benchmark/v2 harness already on main. Mostly additive (29 new files); the few modified files extend prior versions, no reverts. Harness: - scenarios/s02_long_context.py: dataset token preflight (SystemExit on under-padded data) + mock co-location methodology note + logged VU-halving (BENCH_S02_NO_HALVE). - scenarios/s04_concurrency_sweep.py: BENCH_VU_LEVELS / BENCH_LEVEL_DURATION overrides. - scripts/pad_long_context_dataset.py: token-targeted padding (~16k real cl100k tokens). - scripts/per_hook_sweep.sh, hooks_toggle.sh: per-hook isolation + snapshot-restore toggle. - tests/test_s02_preflight.py: preflight guard unit tests. - datasets/long_context_v2*.json: regenerated to ~16k real tokens. Docs/plans: V25_RUN_PLAN, V25_ABLATION_PLAN, CLAUDE-CODE-V25-* task specs, AWS_RUNBOOK hooks-workflow section, AWS_DEPLOYMENT_PROMPTS, JAMES_* demo artifacts, results + results/invalid quarantine. AWS IPs / instance IDs redacted (public repo). NOTE: the two ai-gateway env flags (NEXUS_TRACE_LATENCY / NEXUS_AUDIT_DISABLED) are NOT in this PR — proxy.go was refactored upstream into a stage pipeline; those flags should be re-applied onto the new structure by the proxy owner. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…re single-read false negatives
The Jul-10 arena session's 'governance ON doesn't take effect' was (in part) a
check-timing ghost: Hub→gateway propagation takes seconds, and the script read
the runtime snapshot ONCE immediately after the PUT — seeing the old state.
(Tiebin's verify_hooks_sync.py already proved the toggle converges when polled.)
Replace the single read with a poll: GET /nodes/{id}/runtime every 0.5s (up to
30s, NEXUS_HOOKS_POLL_MAX overrides) until meta.desired_ver == meta.reported_ver
AND the loaded hook set matches the target (off ⇒ zero response-stage hooks;
on ⇒ every snapshot-restored hook present). Exit semantics are now honest:
0 = verified converged; 1 = did not converge (last-seen dv/rv + hook counts
printed, with the real-bug-vs-propagation-stall discriminator); 3 = gateway
node not found (toggled but UNVERIFIED — callers must not benchmark on it).
Previously both failure shapes exited 0 with a warning.
Poll predicate unit-tested against the CP runtime envelope
({meta:{desired_ver,reported_ver}, snapshot.sources[config.hooks]}, per
node_runtime.go): converged-off OK, version-lag WAIT, on-with-expected-set OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…es box restart Kanishk confirmed in postgres that the arena Nexus boxes lose their governance state on restart: db-init runs 'npm run seed:prod' on every container start, and seedReference upserts HookConfig with update:row — re-asserting the fixture's enabled=false over any hook an operator had toggled ON. So 'governance ON' was silently reset to OFF on the next restart (one of the two root causes behind the Jul-10 'hooks-on shows 0 loaded'; the other was the single-read check-timing artifact fixed in hooks_toggle.sh). Fix: add preserveOnUpdate to REFERENCE_TABLES; upsertRows now strips those columns from the UPDATE payload while keeping them in CREATE. HookConfig marks ['enabled'] — a fresh install still gets the disabled-by-default baseline on create, but a re-seed no longer clobbers the operator's live toggle. Every other column still converges, so reference-definition changes (name/stage/priority) keep propagating. enabled is genuinely per-deployment operational state (on box vs off box) with no single 'correct' fixture value — unlike rule severity, whose fix is a fixture-default correction Tiebin owns (disjoint files; no collision). Tests: upsertRows strips the preserved field from update but not create, and is a no-op with an empty list (existing convergence unchanged); REFERENCE_TABLES asserts HookConfig.preserveOnUpdate==['enabled']; DB-gated integration test flips enabled=true, re-seeds, asserts it survives. 43 seed tests green (31 pass, 12 DB-gated skips) under tsx. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ability + result-integrity validator The earlier hooks-OFF arm was produced by direct DB edits, which do NOT propagate to the running gateway (Category-B config: CP -> Hub -> shadow -> WebSocket -> HookConfigCache.Reload). Those numbers are void. This lands the valid path in the benchmark/v2 harness: - cli.py run-nexus-hooks-off: toggles hooks through the Control Plane by DELEGATING to the proven scripts/hooks_toggle.sh (no CP auth reimplemented in Python — coordinate with Tieben before merge), proves the gateway runtime actually reloaded (desired_ver == reported_ver + zero response- stage hooks), and re-enables hooks in a finally even on failure (writes HOOKS_NOT_RESTORED.json + exits nonzero if restore fails). - engine/hooks_control.py + scripts/nexus_hooks_control.py: structured runtime read (null counts when unreadable — never fabricated 0), poll-until-converged, bash delegation. - Governance proof STAMPED into every result row (engine/metrics.py 'governance' block) — closes the 'runtime proof is exit-code-only' gap flagged in the cross-repo review. - LiteLLM stability (engine/stability.py): run-order forces LiteLLM last, warmup record, docker-stats telemetry (null-not-zero), evidence-gated anomaly classification. - scripts/validate_benchmark.py: result-integrity gate — BLOCKs hooks-off without runtime proof / residual response hooks / NEXUS_AUDIT_DISABLED / BENCH_UNIQUE_PROMPTS!=1 / non-sequential; WARNs LiteLLM caveats. - run/run-nexus-hooks-off gain --vus/--duration/--warmup flags (V25 plan commands now valid as written). - Docs: HOOKS_MODE_METHODOLOGY.md, LITELLM_STABILITY_GUIDE.md, RUNBOOK.md; V25_RUN_PLAN.md reconciled. Built on the poll-fixed base f60a712. Tests: 30/30 stdlib unittest (mocked CP/bash — no live CP, no AWS). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Incremental v2.5 benchmark work on top of the benchmark/v2 harness already on main. ~29 new files + small additive edits (+4651/-27); no reverts of upstream.
Highlights
Not included: the ai-gateway env flags NEXUS_TRACE_LATENCY / NEXUS_AUDIT_DISABLED — proxy.go was refactored upstream into a stage pipeline; those belong on the new structure (proxy owner to apply).
AWS IPs / instance IDs redacted (public repo). 🤖 Generated with Claude Code