메뉴
Use when accumulated AgentCanon eval evidence is missing, stale, or failing; runs registered eval producers, validates eval family accumulation, and stores evidence through the log archive instead of hand-writing reports.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when agent-side working philosophy, interaction lessons, task retrospectives, repeated routing misses, missed skill invocation, or recurrence-prevention feedback should be logged without mixing them into user preferences.
Use when exploring, refactoring, or choosing an algorithm under proof obligations; builds JIT-canonical IR, lemma dependency graphs, algorithmic blocker frontiers, and algorithm-change guidance before handing terminal proof work to formal-proof-workflow.
Use when designing, implementing, reviewing, or diagnosing numerical optimization, solvers, preconditioners, convergence, gradients, Jacobians, Hessians, KKT conditions, tolerances, or optimization benchmarks; fixes the mathematical and validation contract before code or experiment changes.
Use when a change needs oracle/spec-risk classification or resilient, adversarial static test design, including behavior contracts, oracle choice, property/metamorphic candidates, mutation adequacy, or brittle-test diagnosis.
Use when updating AgentCanon itself, refreshing a vendored vendor/agent-canon submodule pin, repairing AgentCanon root runtime views, applying AgentCanon update TODOs, or routing local AgentCanon source commits through a proper AgentCanon branch and PR before parent pin updates.
Use for code review, doc review, or AI-generated diff review when you need findings-first output focused on bugs, regressions, missing tests, and broken assumptions.
| name | agent-eval-accumulation |
| description | Use when accumulated AgentCanon eval evidence is missing, stale, or failing; runs registered eval producers, validates eval family accumulation, and stores evidence through the log archive instead of hand-writing reports. |
Use the command packet before applying this skill's workflow:
python3 tools/agent_tools/skill_tool_commands.py show --skill agent-eval-accumulation --format text
Execute the required and task-matching conditional commands that the packet prints.
agents/skills/agent-eval-accumulation.md.python3 tools/agent_tools/eval_accumulation_check.py --root . --compact-out reports/agents/<run-id>/eval-accumulation-before.json --format text; use the compact JSON and stdout counters as the first evidence.python3 tools/agent_tools/run_accumulated_agent_evals.py --root . --run-id <run-id> --report-dir reports/agents/<run-id> and pass every used skill with repeated --skill-used <skill>..agent-canon/log-archive/** or agents/evals/results/**; registered producers own accumulated reports.reports/agent-eval-runs/<run-id>/; avoid broad raw archive searches. Treat those stdout / stderr files as transient, summarize the needed lines into the run bundle, then remove them before closeout.eval_accumulation_check.py with --compact-out reports/agents/<run-id>/eval-accumulation-after.json and require EVAL_ACCUMULATION=pass plus EVAL_ACCUMULATION_BLOCKING_FINDINGS=0 before using the accumulated evidence as green closeout evidence.workflow_monitor.py --runtime-feedback "source=eval target=<skill-or-workflow-or-tool> action=prompt_repair reason=<producer-failure>" and repair the target surface before closing the task.eval_accumulation_check.py fails, record
failing_contract, observation_level, cause_classification,
intent_preservation, and evidence before treating the family as green.
Use intent_preservation for the same-intent repair or escalation route.
Do not skip producers to pass, delete intended eval/oracle coverage, weaken
an oracle, downscope validation, or hand-write source-tree substitutes.
Route producer bugs, oracle/spec mismatches, fixture/environment/stale
generated artifacts, unrelated failures, and approved-design/user-request
conflicts to owner repair, residual, or escalation.python3 tools/agent_tools/runtime_log_archive_git.py sync or push so append-only eval evidence is saved on the log archive branch.python3 tools/agent_tools/generated_artifact_guard.py --root . and require GENERATED_ARTIFACT_GUARD=pass; do not leave regeneratable reports/agent-eval-runs/<run-id>/*.stdout.txt or *.stderr.txt files in the source tree.