Agent workflows with evidence and explicit boundaries.
Useful agent practices tend to disappear into personal prompt files and private habits. Overclock turns them into independently installable capabilities: narrow enough to route correctly, explicit about their authority, and tested as the user will actually receive them.
session-memory · learning-loop
critical-thinking · groundwork · project-vocabulary
discipline-gates · debugging-discipline · pr-feedback · agent-bridge
natural-writing · overclock-setup · eval-stack
Install a compatible group or one plugin at a time. The setup advisor recommends; it never changes your machine.
A good instruction can still fail in several ways. The agent may select it for the wrong task, skip it for the right one, interpret a vague permission too broadly, or produce an answer that looks complete without running the checks that make it trustworthy.
Overclock treats the workflow as a small product. It defines where the capability applies, where it does not, what it may read or change, which artifacts it must produce, and how its behavior is evaluated. Every plugin can be installed on its own; teams do not have to adopt one giant instruction bundle to get one useful practice.
Carry verified working state, learned behavior, and solved problems across sessions.
Resume from verified git anchors, retain workflow corrections, and reuse proven fixes when the same symptoms return.
Use the durable lesson ledger on its own when handoffs and solution history are unnecessary.
Challenge the frame, resolve material decisions, and keep domain language precise.
Stress-test consequential reasoning and verify decision-relevant local facts in a bounded, isolated research pass.
Elicit one material decision at a time and stop at a confirmed brief before planning or implementation begins.
Apply one project glossary in conversation and approval-gate every proposed vocabulary write.
Put evidence and explicit authority in front of risky code and collaboration actions.
Leave red evidence before a fix and recover historical intent before weakening defensive code.
Establish the safest useful observation loop before committing to a theory for difficult bugs.
Judge and fix review feedback locally, then publish only a separately approved, digest-locked plan.
Consult or delegate to another installed harness as a bounded leaf collaborator without giving away task ownership.
Improve human-facing work, choose a compatible stack, and evaluate real agent sessions.
Draft and revise long-form prose in the author’s voice while preserving facts, caveats, and deliberate choices.
Inventory the current configuration and return a report-only, conflict-aware installation plan.
Connect real Claude Code and Codex sessions to Ironside, Coeval, human adjudication, and Casefile scanning.
Setup is report-only. Startup hooks are read-only. A workflow that needs to mutate a file declares that scope rather than treating repository access as general permission. Overclock does not commit automatically, and it does not hide consequential actions behind a helpful sounding command.
Evaluation happens at three levels: deterministic repository checks, isolated behavioral fixtures, and routing controls that test both positive and negative selection. The shipped distribution—not only the authoring source—is what the cases exercise.
Overclock can capture real Claude Code and Codex sessions into Ironside, send bounded evaluation work to Coeval, preserve human adjudication, and run Casefile supply-chain inspection. The integration is intentionally contractual: each tool retains its own job instead of becoming a hidden dependency inside one runtime.
Read the Eval Stack integration ↗/plugin marketplace add luka-zivkovic/overclock
/plugin install overclock-setup@overclock
/overclock-setup:setupClaude Code is the primary distribution target. Codex metadata and adapters are included where the capability supports them.