Problem Statement

Every workflow run re-invokes the model for every node, even when the node's inputs, agent, model and skills are identical to a previous run. There is no result cache in the workflow scheduler, so re-running an unchanged workflow (manually, on a schedule, or after a downstream node fails) re-pays the full token/latency cost and re-reads every skill file from disk.

Evidence (v0.7.36 build; same logic on main)

• No cache/memoization exists in the workflow service: packages/server/src/modules/studio/services/workflow/* has zero hits for cache/memo/reuse/digest apart from the import-preview digest in portability.ts.
• Skill files are re-resolved from disk twice per execution pass. The run preflight calls assertWorkflowNodeSkillDependencies (services/workflow/manager.ts:388-397, invoked at :467 and :538), which calls resolveWorkflowSkillContent (services/workflow/skill-resolver.ts:113-138 — readdir/realpath-walks every candidate root and reads SKILL.md), and buildNodeUserMessage re-reads the same skill file again for every node execution (manager.ts:2476-2485).
• The only result reuse that exists today is intra-run: rerunFromNode seeds initialOutputs from completed node sessions of the same run (manager.ts:2396-2404). A fresh run starts from zero.

Proposed Solution

Add a content-addressed node-result cache at the scheduling layer:

1. Before executing a node, compute a cache key over everything that can change the result: profile, agent, agentMode, provider/model/apiMode/reasoningEffort, the node's resolved input prompt, the resolved upstream outputs, and the selected skills plus a content hash (or mtime+size) of each resolved SKILL.md.
2. On a hit: skip the run and feed the stored structured node result to the scheduler, so edges/orchestration (success/failure/always routing, loop feedback, join) continue exactly as if the node had executed.
3. On a miss: load the skills as today (assertWorkflowNodeSkillDependencies → resolveWorkflowSkillContent), execute the node, then store the result under the same digest.
4. Invalidation: skill content change (hash), node/upstream change (digest), explicit purge, TTL. Never cache nodes marked approvalRequired, and make it opt-out per node for agents that mutate the workspace (coding agents) or otherwise have side effects.

Alternatives Considered

• "Rerun from node" / run snapshots only avoid upstream work inside one run; they do not help a fresh run or a scheduled run.
• Provider-side prompt caching is orthogonal: it lowers token cost but still invokes the model and still re-reads skills.

Use Cases

1. Iterating on the last node of a long workflow — every upstream node re-runs and re-bills today.
2. Scheduled runs where the upstream inputs did not change.
3. Recovering a run after a downstream failure.

Additional Context

Cache hits must stay observable: the node status/run evidence should state that a node result was reused rather than executed, otherwise a cached run is indistinguishable from a real one in the run view.

Environment

• Ekko Studio v0.7.36 (latest release, 2026-10-15)
• Windows
• Agent runtime: Hermes Agent 0.22.2 / DSH

Priority

High - really need this