claude-agent-os

Mine past Claude Code session transcripts (JSONL) to extract user corrections, successful patterns, and recurring failure modes. Proposes diffs against lessons.md and rules/*.md for human review. Triggers on /mine-transcripts, "mine my sessions", "what did I learn this week", "session retro". Use weekly or after a fleet completes.

Run the bounded harness caretaker — deterministic state gather (predictions, verifier gain, agent health, sensors, drift), then triage ONLY what's flagged under a hard authority boundary (files work, never ships changes). Triggers on /caretaker, "run the caretaker", "harness health check".

Apply the diff/patch outputs from a /mine-transcripts run to lessons.md and rules/*.md after human review. Use after running /mine-transcripts and reviewing the proposed diffs. Triggers on /apply-mining-patches, "apply the mining patches", "apply mining run", "approve and apply the trajectory diffs".

Show the RL value-function snapshot for all agents — mean reward, recent trend, anomalies, top tuning candidates. Use weekly to monitor agent quality. Triggers on /rl-status, "agent rewards", "rl status", "which agents are degrading".

Propose a Reflexion-style patch to a specific agent's prompt based on evidence of recurring failures. Use when the same agent makes the same class of mistake twice, when /mine-transcripts surfaces an agent in agent-tuning.md, or when /eval flags a regression on an agent. Triggers on /tune-agent, "tune the X agent", "fix the agent prompt".

A/B test an agent's current prompt against a candidate variant from policy/<agent>/candidates/. Runs both on the same scenarios, aggregates rewards, recommends promotion if candidate beats current by >1 stderr with n≥10. Triggers on /ab-test, "compare prompts", "test the candidate", "is the new prompt better".

Run a panel of 5 specialized verifiers (correctness, edge-case, security, regression, style) in parallel against a change, with confidence-weighted voting. Disagreement (40-60% pass share) escalates to lead. Replaces single-critic for high-stakes review. Triggers on /aspect-panel, "panel review", "multi-aspect verify", "review with multiple critics".

Run an agent N times in parallel against the same prompt, then aggregate via one of 5 modes — critic argmax, USC consistency, confidence-weighted, hybrid, or AggAgent synthesis. Use for high-stakes invocations where you'd rather pay Nx to be sure. Triggers on /best-of-n, "best of n", "run multiple", "give me three options", "synthesize across rollouts".

Show prompt cache hit rate trends and flag regressions. Use when user asks about cache performance, asks why sessions feel expensive, or wants to verify cache health. Triggers on /cache-report, "cache hit rate", "cache stats", "is caching working".

Use when multiple PRs are stuck BLOCKED or failing CI. Clusters failures by signature, distinguishes systemic main-line bugs from per-PR content bugs, and dispatches O(root-causes) agents rather than O(PRs). Triggers on /diagnose-ci-queue, "queue is stuck", "PRs blocked", "CI is red", "merge queue jammed".

Generate scenario stubs for /eval against a target agent or skill. Reads recent successful invocations from session transcripts and proposes 5+ scenarios covering happy path, edge cases, and negative cases. Triggers on /eval-author, "author scenarios for X", "create eval scenarios", "generate test cases for the X agent".

Run evaluations against skills, agents, or rules. Use when the user wants to test that a skill triggers correctly, measure agent output quality, compare prompt variants, or detect regressions. Triggers on /eval, "run evals", "test skill", "benchmark agent".

Run an agent N times in parallel against a code-change task, with each rollout in its own sandboxed copy of the codebase, scored by ACTUAL test pass rate (not just critic opinion). Use for SWE-bench-style code changes where the test suite is the ground truth. Triggers on /exec-grounded, "execution-grounded", "verify by running tests", "best-of-N with tests".

Run a complete SCRUM sprint with all ceremonies — Product Owner authors stories, Tech Lead designs, Scrum Master orchestrates implementers, Verifier+Aspect-Panel guard quality, Sprint Auditor adversarially audits, Retro feeds back to /tune-agent. Triggers on /scrum, "run a sprint", "scrum me this", "ship this with full process".

Author a three-file spec (requirements / design / tasks) before non-trivial implementation, then verify the implementation satisfies the spec. Adopts the Kiro spec-driven development pattern. Use for features that span 3+ files, new public APIs, or anything you'd want a human to design-review. Triggers on /spec, "write a spec", "spec this out", "design before build".

Spin up a native Claude Code Teams fleet to tackle a multi-step task — lead orchestrates, specialists work in parallel worktrees, critic + verifier loop guards quality. Use when the task has 3+ independent sub-tasks, when you need different specialists, or when isolation between concerns is required. Triggers on /team, "spin up a team", "fan out", "fleet this".

Prove that work actually behaves correctly by spawning the verifier agent to run the code and capture evidence. Use before claiming a task complete, before commit, when in doubt about whether tests prove the right thing. Triggers on /verify, "prove it works", "verify this", "did the change actually work".

Verify UI changes by capturing before/after screenshots and asking Claude vision to judge whether the change matches intent and didn't break anything else. Mano-verify pattern (arXiv 2509.17336 —