Merge mixed speculative decoding points on agentic curves

[cquil11]

Summary

• Treat speculative decoding as point-level metadata for agentic_traces, so one hardware/framework curve can contain MTP, EAGLE, or standard-decoding points.
• Keep fixed-sequence behavior unchanged: speculative decoding remains part of those curve keys.
• Align agentic availability and workflow-run filtering with the merged series key.
• Preserve spec_method on every point, include it in point-level identity, and label each agentic point as STP, MTP, or its other speculative method.
• Keep each merged agentic series within one workflow run, including latest charts, date comparisons, history, replay, AI charts, and AgentX overview scoring.
• Enable agentic Performance Over Time and replay history without numeric ISL/OSL, while keeping tracked MTP and standard-decoding points distinct.
• Align AgentX overview links and changelog matching with the merged curve identity, and version affected persistent caches.
• Works for both official runs and ?unofficialrun= overlays; verified locally through the shared transform/tooltip regression tests and Cypress overlay coverage, and on the Vercel preview with both MTP and standard-decoding point tooltips.

Root cause

The dashboard used spec_method in the hardware key for every scenario. That assumed an entire curve used one decode method, so mixed AgentX submissions were split into separate legend lines and availability/run filters continued to treat them as distinct series.

Main InferenceX repository assessment

No structural master.yaml change is required in SemiAnalysisAI/InferenceX. Its current agentic schema already defines spec-decoding on each agentic-coding.search-space entry. The matrix generator propagates the value per job, launchers route from the per-job value, and single-node and multi-node artifact filenames already include _spec-${SPEC_DECODING}. A mixed curve can therefore be submitted under one top-level agentic config with different spec-decoding values on individual search-space entries. The spec suffix must remain in job and artifact filenames because it prevents collisions; this PR removes it only from agentic display-series identity.

Reference: Slack discussion

Validation

• bun run lint
• bun run fmt
• bun run typecheck
• bun run test:unit — 3,745 tests passed
• Cypress component smoke — 22 tests passed
• Cypress integration smoke — 63 tests passed, including unofficial-run overlay coverage
• bun run build — passed
• bun run security — reports the repository's existing 20 lockfile advisories; this PR changes no dependencies

---

Note

Medium Risk
Touches core chart identity, deduplication, and history API paths across inference, overview, and DB queries; incorrect merging could mis-attribute agentic points or drop comparison/history data.

Overview
Agentic traces now treat speculative decoding as point-level metadata instead of splitting curves by spec_method. getHardwareKey, availability keys, default precision counting, overview/AgentX grouping, and changelog legend matching all omit spec suffixes for agentic_traces while fixed-sequence behavior stays the same.

Shared point-identity helpers keep scatter, replay, and “track over time” distinct for overlapping STP/MTP points; agentic scatter labels and tooltips show decode mode. New benchmark-run-selection dedupes agentic series by workflow run (latest chart, history, comparisons, replay, AI charts).

The benchmark history API accepts benchmarkType=agentic_traces without ISL/OSL; DB/history queries return workflow_run_id / run_started_at. Cache keys bump to benchmark-history-v2 and workflow-info-v2.

^{Reviewed by Cursor Bugbot for commit ee605cb. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 7, 2026 10:21pm

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4bf754c. Configure here.}