技能创建器
skill-creator
创建、改进和评估自定义技能,把专业流程转化为可重复使用的 AI 能力。
tokenchronicle-setup【数据与研究】Initialize, diagnose, or open the local TokenChronicle Codex activity archive without importing any packaged user data.
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
Resolve the plugin root first. If a signed bundled executable exists under
bin/<platform>/tokenchronicle/, use tokenchronicle on macOS or tokenchronicle.exe on Windows. Use the
exact executable matching macOS arm64, macOS x86_64, or Windows x86_64; never execute a binary for a
different platform and never download a replacement. A public plugin release may support only a subset
of platforms. If this installed plugin has no executable for the current platform, report that the
platform client has not been published and stop; do not download or substitute an unrelated launcher.
For an unpacked development bundle only,
resolve the plugin root relative to this skill and use its bundled Python package. Use python3 plus
PYTHONPATH on macOS/Linux and py -3 plus PowerShell environment syntax on Windows. Never copy a
Unix environment-assignment command unchanged into Windows.
Start every setup request by running the read-only readiness command. Installation alone is never
proof that TokenChronicle is initialized or protecting activity. Continue from the returned state
instead of restarting completed steps. Do not describe onboarding as complete while the state is
not_initialized, first_archive_required, schedule_choice_required,
first_scheduled_run_pending, daily_protection_off, or attention_required.
Treat the Marketplace default prompt, "initialize TokenChronicle", "set up TokenChronicle", and their
clear Chinese equivalents as requests to start the guided initialization flow. Before explicit privacy
acceptance, only explain the product and run read-only commands such as guide and preflight; do not
create directories, write configuration, start an archive, enable background work, or send network
requests.
After presenting the available choices, summarize the selected Codex source, archive directory, language, viewer port, and whether to run the first full archive. Ask the user to confirm that summary in natural language. Do not require a memorized or verbatim consent sentence. The confirmation must unambiguously accept the current privacy notice and the listed initialization choices. Consent to setup or the first archive never implies consent to scheduling, Codex Automation, feedback transmission, cloud backup, or historical migration.
guide, then explain that TokenChronicle privately preserves searchable Codex inputs, replies,
process evidence, and usage trends for review, audit, idea recovery, and future personal insights.PLUGIN_DATA for plugin operational state when available.preflight command. Report Python support, Codex state availability, selected data
path, current archive size, free disk space, viewer port status, and every failing check.--codex-home;--archive-dir;--port.--language auto, or select Simplified Chinese with zh-CN
or English with en.
Do not ask the user to create these directories manually.
Explain that Documents is visible and portable but may be processed by iCloud, enterprise sync,
search indexing, or backup software. The default application-data location is less visible.
If the selected data directory is outside the current sandbox, show the exact path and request the
required filesystem approval. If the user declines, do not work around the decision; offer another
user-selected writable location and rerun preflight.0/day.
Recommend the bundled local OS scheduler after a successful manual archive; it runs deterministic
commands with zero model tokens and requires --confirm-background-schedule. If the user instead
asks to create a daily Codex Automation, disclose the planning estimate: 0.7M-2.0M
total tokens per run, commonly 1.0M-1.5M, or about 21M-60M over 30 daily runs. State that
cached input is included, the estimate is not a billing quote, and first or broad scans can reach
approximately 1.5M-4.0M.setup --accept-privacy with the selected flags.
Do not enable optional features.doctor and usage-notice, report every failing check, then run the consented first-use
sequence: archive, serve, and optionally memory-daily. After the first archive passes, do not
end the guided task until the user makes one explicit choice:
run-daily, then explicitly consent to schedule enable; orschedule manual --confirm-manual-only after the user accepts that no automatic daily archive
will occur.
Keep any legacy Codex Automation active until the new schedule has a verified run, then avoid
running both schedulers long term.
Treat an unavailable viewer port as either an already-running viewer or a conflict; check before
recommending a different port.readiness. operational means verified automatic daily protection.
manual_only means onboarding is complete but automatic daily protection is off. Report this
distinction prominently. Any other state means onboarding is incomplete and its next_action
must be handled or clearly reported as blocked.