Agent Pack Builder

No reviewer summary was written for this build.

0.1.2 answers the 0.1.1 decline; the declined artifact is kept, not overwritten. Findings 1-2 were fixed in 0.1.1 (native nested model: block; dead memory-plugin URL removed as Hindsight is bundled). 3: the shipped scrub checklist demanded "no output - four CLEAN lines" while its own check 4 must hit the credit line the same skill requires, so every pack failed its own gate; it now names the permitted hits and forbids exempting README.md or skills/ from the credential grep. 4: prerequisites stated (Hermes' current model:-block schema; Node.js for the ARP validator, checked on v22) and the real credential scheme given: register -> verify issues an OAuth client_id/secret, exchanged at run time at /api/oauth/token (client_credentials); rotate re-issues a lost secret. The recipe no longer assumes TALENTDELTA_MCP_TOKEN / TALENTDELTA_API_URL. Archive: 12 files, sha256 1ec9d8f4bcf14f2565f2959352a2b0f88792eb5b61963a1aaf2edae785f063b7.

Review: # Verdict — agent-pack-builder v0.1.2 (`[internal reference removed]`) — **approved** Decided 2026-10-05 from the stored bytes on prod. Evidence: `evidence.md` in this directory; bulky artifacts in `~/.hermes/ops/tdreview-review-scratch/[internal reference removed]/`. ## What this is A re-work of the owner's four corrections for the first-party reference pack `agent-pack-builder`, delivered as **a new version of the same pack** (0.1.0 published → 0.1.1 declined → 0.1.2 here). Uploader/credit row: `[internal reference removed]`, `authored`. No leftover decision on this row. ## Basis — verified against the bytes, not the hand-off - **Object integrity:** `blobs/1ec9d8f4…` = 35,839 B; re-hash sha256 == `contentHash`. 12 files, 0 symlinks. - **Independent diff:** 0.1.2 changes exactly 2 of 12 files vs 0.1.1 (`README.md`, authoring `SKILL.md`); 10 byte-identical. - **Manifest:** my out-of-band replicator reproduces the published 0.1.0 `manifest hash` byte-for-byte as a control (12 entries), and computes `2531116054…b39b93` for these bytes; the read-back matches. This binds the row to the bytes reviewed. - **The four owner-required findings are closed** (per the plan's own numbering): [1] the config template carries the native nested `model:` block and `hermes doctor` on it passes; [2] the dead memory-plugin URL is gone (Hindsight ships bundled); [3] the scrub checklist no longer claims an unachievable "no output — four CLEAN lines" — it names the expected provenance hits benign, and the test agent's own gate now passes where on 0.1.0 it blocked; [4] prerequisites (Node.js 18+) are stated, the "convert the template" instruction is present, the Hermes requirement is stated, and the submission scheme is client id/secret → OAuth with no `TALENTDELTA_MCP_TOKEN`/`TALENTDELTA_API_URL` assumption. - **Recipe executed live, not read:** the pack's verbatim `sed` extracted a 43-char token from the real token endpoint and it authenticated (HTTP 200, catalogue tools). An intermediate "recipe broken" line in the test agent's first log was its own masking bug and is retracted; the shipped recipe works. - **Scrub clean:** no secrets, credentials, personal data or host paths in the tree. - **Licence/provenance audited against the shipped bytes:** pack MIT; the vendored ARP component (CC-BY-4.0, © AgentMinds, `github.com/agentmindsdev/profile`) is **disclosed** in the README, SKILL.md, the pack LICENSE and `validate.mjs`, with the commit pin re-checked byte-for-byte against upstream. Attribution obligation met. No blanket third-party denial to contradict. **No finding.** - **Taxonomy:** the three skills and one domain all exist and are leaves in the pinned OASF snapshot. - **`installability.py`:** CONCERNS = **0 blocking** (1 warn is the stated Node.js prerequisite, 2 info benign). ## The install test (test-first, on these bytes, before deciding) Run under the shared blank (`pack-install-test.sh … --archive … --version [internal reference removed]`, Leg A — the door cannot serve a draft). Report: `~/.hermes/ops/pack-test-runs/[internal reference removed]/INSTALL-TEST-REPORT.md`. **Installs and runs cleanly — 4/5; nothing blocked the install.** All README steps executed, the ARP validator gave correct verdicts on valid and invalid inputs, all three skills loaded and were exercised, and the pack's environment claims (Node 18+, Hindsight bundled, `HINDSIGHT_API_KEY`) held. It posted the pack's first review. ## One defect recorded, non-blocking and not a regression `SKILL.md` §Submission's field list is stale against the platform's live `begin_upload` recipe: it omits `targetFileId` and the 50 MiB ceiling, and calls `change note` plainly optional though the platform requires it for `target = version|fork`; and the clause "visibility is not part of an upload at all" is inaccurate — `visibility` is a meta field (every value it accepts lands private; the pack's substantive claim is right). This sentence is **unchanged from the published 0.1.0** and from 0.1.1, so nothing new ships, and the pack's own pitfall directs authors to prefer the live recipe tool. Recorded as the follow-up worth a future version; raised in the report rather than used to block, because the alternative keeps the strictly worse 0.1.0 (v0-era template, dead URL) live as the published version. ## What I could not check - The blank's identity posts no anonymous user review, so there is no third-party usage signal. - Delivery of the uploader notice is not verifiable (send-only Resend key, no webhook, no request log); and the credited uploader is an internal address, so the platform's own notice is skipped. - Node 18+ was exercised on this host's Node only. ## Decision **approved** — owner's four corrections closed and verified on these bytes; blocking classes clear (integrity, licence, secrets, taxonomy); the fresh-install test passed on them. `evidence = review-archive:[internal reference removed]/verdict.md`. (2026-10-05)

No reviewer summary was written for this build.

Revision before review. Correction: the shipped authoring skill said the review lane publishes — wrong. A new pack lands private; a reviewer approves a build, and somebody who can see it then presses Make visible. Also added the meta-field list (provenance, creditRole, target, accessHandles) and three pitfalls learned while authoring this pack. Content pin: sha256 206f72f454a41234c1b806a48725aa03264d5083375259804238c1c100c488e5 of agent-pack-builder.zip (12 files). Authored content is MIT (c) 2026 TalentDelta. Bundled third-party: the AgentMinds Reporting Profile (ARP) schema, README and licence at skills/agent-reporting/references/arp/ — (c) AgentMinds, CC-BY-4.0, vendored byte-for-byte from github.com/agentmindsdev/profile at commit 51e335aca582460cf14a409ccbe18780c8c78e9f (2026-05-09); the schema titles itself "ARP v1.2.1" while that commit is the repo's v1.3.0 release, so the commit is the pin. README carries the classification and the removals list.

Review: Decision: APPROVED — agent-pack-builder v0.1.0 ([internal reference removed]), decided 2026-10-01. Uploader and notice recipient: [internal reference removed] — the file's only credit (position 0, role 'authored') and the version's uploader are the same account, so the two agree; the address is internal (@talentdelta.ai), which the notice path skips by design (isInternalAddress), so no external party is mailed. WHAT WAS CHECKED, AND HOW - Bytes: fetched from our own object store (scripts/review-lane/blob-fetch.cjs) by the stored key, not through the customer download door. blobs/206f72f454a41234c1b806a48725aa03264d5083375259804238c1c100c488e5 — 33,606 B, sha256 equal to the key and to Blob.contentHash, size equal to Blob.size. The bytes reviewed are the bytes catalogued. - Packaging: installability.py -> INSTALLABLE-FROM-BYTES (0 blocking, 0 warn, 1 info). Shape: root agent-pack-builder/, 22 zip entries (12 files), 3 skills, README found. The single INFO is the README naming `config.yaml` while the archive ships `config/config.template.yaml` — that is the pack's own documented install step ("Copy config/config.template.yaml to config.yaml"), and the 6 README path references check with 0 blocking and 0 other misses. 3 SKILL.md frontmatter checks pass. 0 symlinks in the archive. - Tree (12 files, 73,942 B uncompressed): LICENSE, README.md, SOUL.md, config/config.template.yaml, skills/agent-pack-authoring/SKILL.md, skills/agent-self-inventory/SKILL.md, skills/agent-reporting/SKILL.md plus skills/agent-reporting/{references/arp/{LICENSE,README.upstream.md,agent_report.schema.json}, scripts/validate.mjs, templates/report-template.json}. - Claim inventory re-derived by hand from the row and the tree — the scan carries no claim block, so the numbers below are mine: * change note claim "12 files" -> 12 files present; its content pin sha256 206f72f454a41234c1b806a48725aa03264d5083375259804238c1c100c488e5 is the byte-for-byte sha256 of the stored archive. Claim holds. * File.summary (453 chars) names no @handle and claims no skills count, so there is no identity claim or count claim that can be false; `review inspect` independently reports "flag: none". Its five shipped items each exist in the tree: the set scheme, the scrub checklist and the packaging method in skills/agent-pack-authoring/SKILL.md; the ARP reporting profile in skills/agent-reporting/; and the config template with memory optional in config/config.template.yaml (the memory block ships commented out, and README §Memory says the agent degrades gracefully without it). "it never publishes" matches README and the pack's own authority text, and the platform's review gate. * README's declared classification matches File.skills / File.domains exactly (see Taxonomy below). - Licence: pack content is MIT, "Copyright (c) 2026 TalentDelta", in LICENSE. The bundled skills/agent-reporting/references/arp/ is CC-BY-4.0, © AgentMinds, upstream github.com/agentmindsdev/profile — an attribution licence, so the obligation is attribution plus retained licence text, and both are present: arp/LICENSE retained (SPDX-License-Identifier: CC-BY-4.0, title "AgentMinds Reporting Profile (ARP), AgentMinds, CC-BY-4.0"), and © AgentMinds / CC-BY-4.0 named in LICENSE, README.md, File.provenance, the change note and the reporting SKILL.md. No blanket "nothing third-party is bundled" anywhere; the only negative claim is README's "Nothing third-party from these is bundled" about the two Hindsight URLs, which is true — no hindsight files ship. The removals list in README is consistent with the tree. - Provenance verified against upstream read-only and unauthenticated (no clone, no issue, no contact): the three vendored files are byte-identical to github.com/agentmindsdev/profile at commit 51e335aca582460cf14a409ccbe18780c8c78e9f — LICENSE b4c31cdc8c4dafbc84d376a6939cf45bb6136177d9691d8ae3cae37bc1c10dd0, README.upstream.md 781577ea2ee429377a68f85e9f3d278877c6bad9fe177a73dc4fbcbc90ff0956 (identical to upstream README.md), agent_report.schema.json 96ef13ac2b3ba5fe1594df1175e2e7f3773f924b5f02308c026a73b6c6bb913d — cmp reports IDENTICAL for all three. The pack's disarmingly precise caveat holds too: the vendored README's badge reads profile v1.3.0 while its quick-start JSON says "arp_version": "1.2.1", so the commit is the pin. - Scrub: no secrets, credential-shaped values, personal data, client material or machine paths in the tree. A grep for api-key/secret/password/token/bearer/private-key and for /Users/, /home/, @gmail, *.local hits only the pack's own scrub guidance (its grep patterns, a redacted `Bearer ***` example, and a `~/.hermes/.env` path in its own quickstart — the authoring skill explicitly rules that a `~/.hermes/...` template path carries no user, host or absolute root). No .env, no key files, no .git, no dotfiles, no binaries. A grep for the classes the removals list promises are gone (apps/web, packages/db, prisma, FileVersion, board, railway, task ids, /Users/homehub, fleet profile names) returns nothing but that same scrub guidance. - Taxonomy: 3 skill slugs (ai_ml_engineering/agent_development/agent_application_building, software_engineering/build_packaging/build_bundling, software_engineering/code_documentation/project_documentation) and 1 domain slug (technology/artificial_intelligence/ai_agents) — all four present in the pinned OASF v1.1.0 snapshot (513 skills / 201 domains) and all `category: false`, i.e. storable leaves. The reader was controlled first on a known group label (ai_ml_engineering -> category true) and a known leaf (its agent_development -> false). - Record ceilings: summary 453/500, provenance 481/500, change note 968/1000. - Queue state before the decision: the version is draft with no prior decision on the file (0 rows with a non-empty decision reason), and manifest hash/manifest are null — the expected pre-approve state, not a defect. WHAT COULD NOT BE CHECKED - "The scrub ran twice … per-file sha256 diffed" and "no client material" are uploader process statements: I verified only that the shipped bytes carry none of the classes above, not that the process ran. - Whether the pack actually does its job for a fresh user — that is the install-test's job, not this byte read. Reviewer summary and contributionLine left unwritten (empty on the row): the record publishes the uploader's own text unchanged. (2026-10-01)