rlaope/oh-my-hermes

Problem OMH can count tool calls at the session level, but it cannot show which files its native file tools actually read, wrote, or patched. The current activity receipt intentionally drops tool identity and arguments, while the file-operations workflow names an observed file manifest without an implemented session-level producer. A run can therefore finish with a summary and evidence trail that cannot answer the basic question, "Which workspace files did this run touch, and how?" Complete user goal As an operator reviewing an OMH run, I want a bounded, privacy-safe session file-activity record so I can inspect the files read or changed, distinguish attempted from observed operations, reopen relevant artifacts, and carry exact file lineage into review or handoff without collecting file contents or inferring activity from shell commands. Intended surfaces Hermes chat or wrapper Hermes runtime or handoff All executors (executor-neutral) Observed evidence Current OMH main has a metadat

Problem OMH can execute and validate paired agent runs, separate infrastructure state from behavioral results, and produce a scoped Pareto outcome. It does not provide a mechanism-level diagnostic contract for explaining why an intervention changed the result. A comparison can therefore show that one arm performed better while leaving several materially different explanations mixed together: avoiding context overflow, reaching the first useful edit sooner, preserving required state, using fewer summary calls, verifying more completely, or simply terminating earlier. Without matched intervention metadata and explicit termination accounting, maintainers can mistake lower cost caused by early failure for an efficiency improvement or promote a bundled change without knowing which component helped. Complete user goal As an OMH maintainer, I want paired agent evaluations to isolate one declared intervention at a time, hold relevant conditions constant, and report mechanism-specific measureme

Problem OMH makes several consequential policy decisions through deterministic rules and static fixtures: route selection, coding-owner recommendation, task-scale routing, continuation state, context-budget actions, and verification tiering. Those rules can be changed and regression-tested, but current main has no shared way to run a candidate decision policy in shadow mode beside the active policy, retain bounded disagreement evidence, and require reviewed evidence before promotion. Without that lifecycle, maintainers can prove that a proposed rule passes known fixtures but cannot answer whether it disagrees on real eligible decisions, whether those disagreements cluster around one input class, or whether latency, failure, and coverage make the candidate safe to activate. Each subsystem must either invent its own comparison log or promote directly from test evidence. Complete user goal As an OMH maintainer, I want any consequential deterministic decision point to register an active po

Problem OMH's visual QA contract requires motion_interaction_capture/v1 when transitions or interactive states are in scope, but current main provides no typed importer or durable receipt for a host-produced recording and its derived changed-frame contact sheet. A host can produce a video, yet OMH cannot bind that artifact to the exact target revision, route/state/viewport condition, capture interval, and derived stills needed for a visual QA verdict. The result is either a permanent missing-evidence block or an informal file reference that cannot safely support PASS/REVISE decisions. Complete user goal As a maintainer or reviewer validating an interactive web surface, I want to import a bounded host-produced motion capture into OMH and receive a lineage-bound, metadata-only evidence record so visual QA can assess transitions, loading states, hover/focus behavior, and other time-dependent defects without treating an arbitrary video file as verified evidence. Intended surfaces Hermes r

Problem OMH routes stuck, looping, drifting, and repeatedly failing agent runs to agent-debug, but the current workflow only declares report names and high-level guidance. It does not define how to bind a diagnosis to an observed session region, safely inspect large session records, cite each finding, reconcile existing runtime receipts, or produce a privacy-reviewed export. A plausible narrative can therefore look like a diagnosis even when no observed event supports it. Complete user goal A user can ask why a specific agent run repeated work, ignored its goal, lost context, stalled on tools, or consumed unexpected time or tokens and receive a bounded incident report. Every finding is tied to an observed receipt or an exact source location, observed facts are separated from hypotheses and unavailable evidence, competing explanations are considered, and the report ends with the smallest reversible recovery action. If the user explicitly asks to share the case, the workflow can prepare