实时 Agent 正在用 TokRepo
最近 14 天 agent 漏斗匿名聚合计数。所有事件由 agent funnel API 记录；不记任务文本、文件内容、tokens 或私有资产数据。
设置 TOKREPO_TELEMETRY=0 可关闭埋点。数据源：/api/v1/tokenboard/agent/funnel。
高价值 Agent 任务的运行模型
这里是机器契约背后的产品面：TokRepo 借鉴了哪些任务经济机制、哪些生命周期 gate 已经接入、哪些 KPI 已经有公开 eval 证据。
Percentage of realistic task prompts that complete with a verified user-outcome oracle.
Count of reference tasks where TokRepo discovery avoids rebuilding an equivalent local asset from scratch.
Median token reduction from using discovered reusable assets on reference tasks.
Ratio of resolved/reused capabilities to newly harvested private drafts over a release window.
Percentage of installable assets covered by verify, install-plan, policy, evidence bundle, and rollback checks.
Percentage of harvested or handoff candidates whose quality_gate passes without unsafe files or unresolved sensitive findings.
Percentage of recurring or delayed agent tasks with owner, schedule, latest evidence, next action, and rollback or handoff state.
Mini harness 证据
每条 case 都记录 Task、Environment、Tools、Trace 和 Grader 输出，TokRepo 不只根据最终回答判断任务经济闭环是否通过。
Decide whether the repository README proves plugin-system support.
Before rebuilding a browser automation harness, resolve reusable TokRepo capability and inspect a safe install plan.
After improving a reusable task harness script, package the change as a private draft and produce a handoff plan without publishing.
生命周期 gate
- 01detect_high_value_task
Classify whether the user task is a repeatable high-value task, a one-off local fix, or a post-task reusable artifact opportunity.
task_value_hypothesisreuse_potentialrisk_profilesuccess_oracle - 02resolve_capability
Call tokrepo_resolve_capability or tokrepo agent-check before rebuilding local tools.
selected_candidate_or_controlled_empty_statenext_mcp_callsfallback_cli_commandslifecycle_contract - 03plan_safe_run
Verify trust and request an install plan before writes, execution, credentials, global config, or recurring work.
policy_decisionpermission_envelopeevidence_bundlerollback_plan - 04execute_or_stage
Use dry-run, stage-only, confirmation-required, or direct execution according to policy and user intent.
changed_files_or_no_writestate_deltapost_verify_commands - 05evaluate_task_outcome
Evaluate the user's real task outcome with a task-specific oracle instead of claiming success from a local command only.
task_outcome_verdictevidenceopen_risks - 06record_memory_and_state
Record install state, task evidence, audit snapshots, and project memory so future agents inherit the context.
TokRepo.lock_or_state_referenceaudit_referenceproject_memory_reference - 07harvest_reusable_work
Run tokrepo_harvest or tokrepo harvest --changed --json when the task creates reusable instructions, scripts, configs, or evals.
private_package_draftsquality_gatesensitive_findingsdedupe_discovery - 08measure_task_economics
Update aggregate evidence for task completion, duplicated rebuilds avoided, token/time saved, safe install coverage, and handoff quality.
task_completion_rateduplicate_rebuilds_avoidedmedian_tokens_saved_pctsafe_install_gate_coverage_pcthandoff_quality_pass_rate
可借鉴机制
- mini_agent_harness
TokRepo now publishes a deterministic mini task harness eval where each case records the task, fixture environment, tool calls with arguments/results, answer, and grader checks.
https://tokrepo.com/evals/agent-task-harness.json - high_value_task_economics
Every agent-facing surface should answer which high-value task it helps complete, what duplicate rebuild it avoids, and how the task result is verified.
https://tokrepo.com/evals/agent-baseline.json - production_harness
TokRepo's harness is the MCP/CLI/install-plan/evidence bundle lifecycle: resolve, discover, verify, plan, dry-run, install or stage, audit, rollback, and handoff.
https://tokrepo.com/.well-known/tokrepo.json - skill_supply_chain
Treat skills, prompts, scripts, MCP configs, workflows, project rules, and evals as typed supply with private-by-default harvest and human-confirmed publishing.
https://tokrepo.com/agent-ecosystem.json#supply_handoff - long_running_clock
Represent recurring and delayed work as task runs with schedule, status, state, evidence, next action, and rollback or handoff plan rather than as disconnected one-shot commands.
https://tokrepo.com/evals/agent-task-ledger.json - outcome_oracle_receipts
TokRepo now publishes deterministic task receipts so long-running work can be audited for freshness and outcome, not only remembered as a scheduled state row.
https://tokrepo.com/evals/agent-task-receipts.json - memory_and_persona_context
Project memory from tokrepo init-agent, .tokrepo/agent.json, installed-state files, and audit records must carry the user's operating rules into future agent sessions.
https://tokrepo.com/schemas/agent-memory.schema.json - multi_agent_reliability
Acceptance gates must include cross-agent project memory, local and hosted MCP parity, npx latest packages, install safety, handoff quality, and production live checks.
https://tokrepo.com/evals/multi-agent-compatibility.json - task_to_supply_flywheel
After a task is verified, agents should run tokrepo_harvest or tokrepo harvest --changed --json, inspect quality_gate, keep drafts private, and only push explicit reviewed files after human confirmation.
https://tokrepo.com/evals/handoff-quality.json
长任务状态
Treat recurring jobs, delayed follow-ups, and automations as first-class high-value task runs with state and evidence.
长任务、周期任务和延迟任务必须跨会话保留 owner、schedule、evidence、next action 以及 rollback 或 handoff 状态。
- taskrun_agent_discovery_smoke_dailyhealthyowner: tokrepo-release-agentdailypass_with_warning
- schedule
- 2026-05-28T00:20:00Z
- latest_evidence
- https://tokrepo.com/.well-known/tokrepo.json
- next_action
- Rerun after the external MCP Registry catches up to live manifest 2.16.1.
- rollback_or_handoff
- handoff · machine-readable endpoint regression or external registry lag persists past the next release window
- taskrun_private_harvest_review_weeklyreadyowner: tokrepo-harvest-agentweeklypass
- schedule
- 2026-06-01T03:00:00Z
- latest_evidence
- https://tokrepo.com/evals/agent-task-harness.json
- next_action
- Run tokrepo_harvest on changed reusable scripts, keep drafts private, and require quality_gate.status=pass before handoff.
- rollback_or_handoff
- rollback · quality gate fails or sensitive_findings is nonzero
- taskrun_delayed_chrome_mcp_recheckneeds_recheckowner: tokrepo-browser-ops-agentdelayedpass_proxy_connected
- schedule
- 2026-05-27T03:30:00Z
- latest_evidence
- local:chrome-mcp-proxy-status
- next_action
- If MCP stdio transport is closed in the current agent session, reconnect the MCP server or start a fresh Codex session before browser-dependent validation.
- rollback_or_handoff
- handoff · mcp__chrome__ list_pages reports Transport closed after proxy health is green
每个长期任务运行都需要最近一次收据，包含新鲜度、验证命令、证据摘要、结果 oracle、判定和用户效果。
- receipt_20260527_agent_discovery_smokefreshness_status: freshpass_with_warning
- last_run_at
- 2026-05-27T04:36:00Z
- next_due_at
- 2026-05-28T04:36:00Z
- outcome_oracle
- machine_readable_surface_contract · All TokRepo agent discovery surfaces return HTTP 200, include the task-economy eval set, and hosted MCP JSON-RPC tools answer without schema regressions.
- evidence_digest
- sha256:443da96fc8c6287ea881348554b3211c8b44fa36dce2cb3dee08d789815a84ef
- user_effect
- Planning agents can discover the live TokRepo MCP, CLI, manifest, harness, ledger, and task-economy surfaces without GUI browsing; the only warning is external registry version lag outside the production site.
- receipt_20260527_private_harvest_reviewfreshness_status: freshpass
- last_run_at
- 2026-05-27T04:38:00Z
- next_due_at
- 2026-06-03T04:38:00Z
- outcome_oracle
- quality_gate_receipt · The private harvest case includes tokrepo_harvest and tokrepo_handoff_plan traces, keeps draft supply private by default, and passes the Grader checks.
- evidence_digest
- sha256:d27a4f80a242278fdd1115bc985fe762895fd742500d6b8b9ebbb92daeb1f389
- user_effect
- Reusable work created during agent sessions can be reviewed as private package drafts with quality gates before any public push.
- receipt_20260527_chrome_mcp_recheckfreshness_status: freshpass_with_handoff
- last_run_at
- 2026-05-27T04:44:00Z
- next_due_at
- 2026-05-27T05:14:00Z
- outcome_oracle
- browser_validation_fallback · Proxy status remains chromeConnected=true and the production browser validation matrix passes with System Chrome when the current MCP stdio transport is closed.
- evidence_digest
- sha256:e6d852212930cd1bf3d70c6dcc35400c70c50bba8513e14d71607510081f0cb1
- user_effect
- Browser-dependent TokRepo production validation remains executable even when the current Codex Chrome MCP stdio transport needs a session-level reconnect.
标记为已量化的数字来自公开 eval 证据。已签约 KPI 会作为 live 义务持续展示，直到运行时 ledger 开始发布这些测量值。
Agent 漏斗
Plan→implementation 边界是 TokRepo 的发力点。每一步把 CLI 与 MCP 两条路合并成一个可读信号，不论用户跑哪个 agent 都能看见。
- 01初始化项目 agent memory（init-agent）init_agent31
- 02Plan 时调 discover / agent-checkmcp_discover78252% 相对上一步
- 03信任、权限、policy 验证verify_asset107137% 相对上一步
- 04生成 install planinstall_plan3836% 相对上一步
- 05执行 install（dry-run + 确认后）install_apply198521% 相对上一步
- 06任务后 supply handoffhandoff_plan3920% 相对上一步
- 07Harvest — 打包私有草稿harvest_plan52133% 相对上一步
- 08Harvest 草稿发布入库harvest_publish00% 相对上一步
- 09发布可复用资产push9
计数不是 per-agent 唯一事件，一次 agent 任务通常触发多个事件。这是 population 视图，不是 session 视图。
每日事件量
全部漏斗阶段的事件按 UTC 日聚合。新增 agent 接入面通常会拉一波。
2026-05-25 → 2026-06-07
Agent 漏斗记录的全部事件
每行是一个 funnel event 类型。Agent 漏斗刻意保持窄口径——只记 gate 决策或为决策提供 evidence 的事件。
| 事件 | 次数 |
|---|---|
| install_apply | 198 |
| install_dry_run | 179 |
| verify_asset | 107 |
| rollback_plan | 105 |
| capability_resolve | 101 |
| mcp_search | 90 |
| agent_check | 78 |
| harvest_plan | 52 |
| install_plan | 38 |
| agent_handoff | 35 |
| audit_asset | 34 |
| init_agent | 31 |
| mcp_discover | 19 |
| find_for_task | 13 |
| push | 9 |
| handoff_plan | 4 |
| mcp_detail | 3 |
想让你的 agent 事件出现在这页？
放进任何使用 Claude Code、Codex、Cursor、Gemini CLI、Copilot、Cline、Windsurf、Roo、OpenHands 或 Aider 的项目即可。init-agent 命令会写入机器可读的 .tokrepo/agent.json 以及 11 个指令接入面。
npx tokrepo init-agent --target all