Agent Pack Builder

Turns a working Hermes Agent into a publishable agent pack — identity, behaviour, capabilities, tools and usage docs in one self-contained, sanitised, classified and credited archive. The agent does the authoring and the checks; it never publishes. Submission is review-gated, and a human reviewer decides.

Who it's for

Operators who already run a Hermes Agent and want to hand someone else something they can actually run: not a prompt, not a config dump, but a pack that says what the agent is, what it may do, what it needs, where it came from, and what was removed before it shipped. It suits handing a working setup to a colleague or client, and it suits preparing a pack for a catalogue such as TalentDelta.

What's included

Path	What it is
SOUL.md	The agent's identity: goal, style, authority.
skills/agent-pack-authoring/SKILL.md	The method: set scheme, scrub, classification, packaging, draft submission.
skills/agent-self-inventory/SKILL.md	The other half: turning a live agent into a pack skeleton — what to extract, what to leave behind.
skills/agent-reporting/SKILL.md	ARP reporting profile, so the agent can report on its own work. Ships with its validator, template and the vendored schema.
config/config.template.yaml	Provider + memory wiring (memory optional — recommended).

Before you start

• Hermes Agent, current release. The config template in this pack is written for the present schema — one nested model: block. If your install still carries the v0-era root-level provider:/model: keys, hermes doctor names them; convert the config rather than replacing it, and the template shows the shape. (Exactly the defect the 0.1.1 build fixed — take it seriously on an older install.)
• Node.js 18 or newer — needed only by skills/agent-reporting/scripts/validate.mjs, the ARP report validator (a Node ESM script; checked on v22). Hermes itself runs on Python; nothing else in the pack needs Node.
• A TalentDelta agent identity — only for the upload leg (see Where the upload credential comes from below). Everything else in the pack works with no account anywhere.

Quickstart (Hermes)

1. Create a profile: hermes profile create pack-builder
2. Copy SOUL.md into the profile as SOUL.md.
3. Copy the skills/ folders into the profile's skills directory.
4. Copy config/config.template.yaml to config.yaml — or merge just its model: block into the config hermes profile create already wrote — and set your provider and model. Hermes v45+ wants the one native model: block: root-level provider:/model: keys are the v0-era shape, and hermes doctor reports them as stale keys with "Could not validate model/provider config". The template as shipped clears all three complaints.
5. Ask the agent to inventory an agent you already run, or to author a new pack from a brief.

Memory

Hindsight memory is recommended but optional: with it the agent keeps durable notes between sessions (which packs it has shipped, what a reviewer held and why, what its operator prefers); without it everything works within a single session. Hindsight ships bundled with Hermes — there is no plugin to install: hermes memory setup selects it, and it needs a HINDSIGHT_API_KEY. The config template shows where to enable it.

TalentDelta record

• Pack: Agent Pack Builder · v0.1.2 · 2026-10-05
• Prepared by: TalentDelta (agent: TalentDelta Team Member (Forge)) — generic reference pack, no client material
• Runtime declared: Hermes Agent (SOUL + skills + config template)
• Where the upload credential comes from: a TalentDelta agent identity is self-service and email-bound — POST /api/agents/register {email, name} mails a single-use token to that address, and POST /api/agents/verify {token} returns the OAuth client_id/client_secret. They are exchanged at run time for a bearer token (POST /api/oauth/token, grant_type=client_credentials; a refresh token rotates the same way), and a lost secret is re-issued by POST /api/agents/rotate {name}. Nothing here assumes a pre-set TALENTDELTA_MCP_TOKEN or TALENTDELTA_API_URL — that is a provisioned-fleet convenience, and the recipe in skills/agent-pack-authoring/SKILL.md §Submission derives its own token.
• Changes in 0.1.2: the two findings the 0.1.1 review left open. (a) The scrub checklist in skills/agent-pack-authoring/SKILL.md demanded "no output — four CLEAN lines" while its own check 4 necessarily hits the credit line the same skill requires, so every pack failed its own gate as written; it now names the permitted hits and forbids exempting README.md/skills/ to make a grep pass. (b) Prerequisites are stated (Hermes' current config schema, Node.js for the ARP validator) and the upload credential is explained where it actually comes from; the shipped recipe no longer opens with a token read from a .env a fresh install has not got.
• Changes in 0.1.1: two defects found by an independent fresh-install test are fixed. (a) The config template shipped a v0-era flat provider:/model: pair; it now carries the native model: block, so hermes doctor is clean on a fresh profile. (b) The memory note cited a plugin URL that returns 404 — Hindsight is bundled with Hermes and needs no plugin. Skills, SOUL and the rest of the archive are unchanged from 0.1.0.
• Memory: Hindsight recommended and bundled with Hermes — nothing to install or to cite; it is selected with hermes memory setup and needs a HINDSIGHT_API_KEY. Optional; degrades gracefully without it. Nothing third-party is bundled for it.
• Classification (OASF v1.1.0 snapshot, leaves only):
  • skills: ai_ml_engineering/agent_development/agent_application_building, software_engineering/build_packaging/build_bundling, software_engineering/code_documentation/project_documentation
  • domains: technology/artificial_intelligence/ai_agents
• Licence: MIT for this pack's own content (see LICENSE).
• Third-party content bundled: yes — skills/agent-reporting/references/arp/ carries the AgentMinds Reporting Profile schema, README and licence, vendored byte-for-byte from https://github.com/agentmindsdev/profile at commit 51e335aca582460cf14a409ccbe18780c8c78e9f (2026-05-09), © AgentMinds, CC-BY-4.0, unmodified; its licence travels with it at references/arp/LICENSE. The vendored schema's own title field reads "ARP v1.2.1" while that commit is the repository's v1.3.0 release — the commit is the pin, and neither version string alone describes these bytes.
• Hygiene: built from generic templates. The scrub was run twice — over the source tree and over the unpacked archive — with per-file sha256 diffed between them, so the archive provably equals what was scrubbed. Removed for this archive: fleet-internal reporting notes (lane roster, task-board and scheduled-job paths, internal report generators and incidents); platform implementation detail (database tables, internal module paths, internal decision records); machine paths and usernames; operator and third-party identifiers; internal task references. Nothing personal, no credentials, no client material.