Datasets:
Changelog — Nomad
All notable changes to the Nomad skill are documented here.
Versions describe the specification's capability, not internal edit history.
The reference files ship in two content-equivalent versions (<name>.md /
<name>.zh-CN.md); every change below lands in both.
[0.5] — 2026-07-23 · Governance artifacts land in-repo
- Hooks layer completed in the same release:
scripts/frontmatter-touch.js(portable change-detection via mtime-vs-frontmatter reconciliation — works in zip-extracted, non-git team folders; six safety guards; selftest grown 102→153) + a charter-pointer injection segment inscripts/session-start.js, with per-platform config templates and enable guides inscripts/README.md§3/§8.
The v0.4 governance layers gain ready-to-use artifacts, upstreamed from an integration pack that assembles team workspaces from this skeleton. That pack now pulls these files from this repo's published HEAD instead of carrying private copies — one source of truth, no drift. Rationale: an integration pack is composed entirely of published building blocks; any capability that isn't pack-specific belongs to the upstream repo it conceptually extends, so users can view, install, and receive updates for it independently.
Added
references/governance/— governance document templates, bilingual pairs:ARCHITECTURE.md/ARCHITECTURE.zh-CN.md— team-structure charter template (the closed-form constitution: fixed slots + an application channel; place at the team root). The in-file §8 registers the sentinel as its machine-guarantee layer.metadata-conventions.md/metadata-conventions.zh-CN.md— the frontmatter + linking operational manual (place under the team's index slot, e.g.00-index/).
scripts/structure-sentinel.py/scripts/结构哨兵.py— dependency-free four-check structure sentinel (must-read existence / memory-index orphans / frontmatter schema in mandatory zones / broken relative links). A guardrail, not a gate: reports, never blocks.scripts/append-team-log.js/scripts/追加团队日志.js— assembled-team variant of the concurrency-safe log append tool (self-locating inside a team tree, enforces the charter's entry discipline). Division of labor with the existing genericappend-log.jsis documented inscripts/README.md§7.
Changed
- Migrated files carry exactly one class of edit: self-descriptions generalized from pack-specific wording to nomad-skeleton wording (6 spots across 4 files). All functional content is byte-identical to what the integration pack shipped before the move.
- Charter template §8.3/§9.2 wording updated (both languages) now that the
hooks layer ships alongside the governance artifacts: boot injection goes
from "this pack ships no hooks" to "shipped, enable per platform" (the
persona mandatory path stays the primary enforcement), and §8.3 notes
frontmatter-touch.jsas the machine half of hook-less compensation.
[0.4] — 2026-07-15 · Governance layers for teams that grow large
The skeleton spec gains three optional governance layers, plus one design lesson that explains why they matter more for an AI team than for a human's notes. All additions are optional — the minimum viable team (§8) is unchanged, and a small team never has to touch any of this.
Added
- Closed-set memory types & slot-application channel (
team-skeleton§9). Closes the memory-type set by type, not by filename (a{domain}placeholder absorbs per-agent variation); a register-on-create anti-orphan rule; an asynchronous application channel for growing the set; and a "design graveyard" — a recorded list of rejected proposals. The graveyard is an AI-specific discipline: a cold-started agent has no fuzzy "we tried that" memory, so without an explicit record a fleet re-proposes dead ideas forever. - File lifecycle (
team-skeleton§10). Workspace close-out (distill-up / archive / recycle, one of three per task); backup & snapshot retention windows; append-only soft-caps whose threshold is set to the harness's actual load mechanism, never a hardcoded line count (some harnesses silently truncate over-budget memory files); and an external-clock clause — periodic upkeep must hang off a scheduler because an AI has no habit loop. - Metadata & relations layer (
team-skeleton§11). Optional YAML frontmatter: a small three-drawer field set (requiredtype/owner/status, auto-derivedcreated/updated, optionaltags/links), a four-relation vocabulary (references/supersedes/part_of/derived_from), a deliberately-pinned three-kind light ontology (no RDF/OWL), the rule that frontmatter is enforced only in tandem with a broken-link validator, and a field-keep test ("does a consumer actually read it?"). - Design lesson 5 (
team-skeleton§7) — human associative recall vs. AI index recall. A human patches missing structure by association; a cold-started agent navigates purely by explicit indexes/links. This asymmetry is the root reason the §9–§11 mechanisms earn their keep on an AI team specifically.
Changed
- Section renumbering in
team-skeleton: placeholder glossary §9 → §12, bilingual terminology §10 → §13 (the three new sections take §9–§11). Themounting-guidecross-reference to the placeholder glossary was updated to §12 (both language versions). - Terminology cross-reference table extended with the new concepts.
[0.3] — 2026-07-14 · Node-layer scripts included
- Node-layer scripts (
scripts/) shipped alongside the docs: session-start / stop / precompact hook wiring plus a concurrency-safe append tool; selftest 101/101, verified on a real team folder (both start and close-out hooks fired). - The md layer remains the required foundation; the node layer is optional acceleration that degrades gracefully when Node.js isn't present.
[0.1–0.2] — Docs-first foundation
- Initial reverse-engineered specification: team-skeleton (structure +
team-config.jsonparameterization + md/node split + collaboration protocol + fourteen hardcoding points + design lessons), migration-protocol, and mounting-guide — each in English and Chinese, fully desensitized.