model-registry-refresh

Name: model-registry-refresh
Author: Raishin

17stars3forks

How to install model-registry-refresh?

raishin/vanguard-frontier-agentic/model-registry-refresh

$npx -y skills add Raishin/vanguard-frontier-agentic --skill model-registry-refresh

Installs into the current project.

Summary

Re-verify and extend catalog/model-registry.json — the fail-closed model-name and reasoning-effort matrix scripts/model-policy.mjs validates against — via delegated Context7-backed research, orchestrator-owned registry edits, and the full validation chain; use when a policy check

Files · 1

View on GitHub

SKILL.md

1	# Model Registry Refresh
2
3	## Doctrine
4
5	`catalog/model-registry.json` is the single source of truth `scripts/model-policy.mjs` fails closed
6	against. Every model name and reasoning-effort value it accepts must trace to official documentation
7	with a citation — never to memory, never to a plausible-sounding guess at a slug. This skill is the
8	repeatable workflow for keeping that registry accurate without letting research cost dominate the
9	orchestrator's context.
10
11	## When to run
12
13	- `npm run model-policy:check` fails with an error naming a model "not in the verified model
14	registry" — the registry is missing a model the policy (or an operator) wants to use.
15	- A provider (OpenAI, Anthropic, Cursor) ships new models or retires old ones and the catalog
16	needs to reflect current reality.
17	- Quarterly staleness check — `last_refreshed` in `catalog/model-registry.json` is more than
18	~3 months old.
19
20	## Step 1 — delegate research to Haiku Explore agents
21
22	Fan out one Haiku `Explore` agent per harness (or per namespace, for codex) using the Context7 MCP
23	tools (`mcp__Context7__resolve-library-id` then `mcp__Context7__query-docs`) plus official docs
24	URLs already cited in the registry. Each research task must:
25
26	- Ask for exact slugs/IDs, not families — `gpt-5.5` not "the gpt-5 line".
27	- Ask for reasoning-effort support per model, not per harness — some models in a family
28	predate newer effort levels (see `o1`/`o3`/`o4-mini` lacking `none`/`minimal`/`xhigh` in the
29	current registry).
30	- Ask for failure-mode evidence — what error shape a bad model name or unsupported effort
31	actually produces (HTTP status, error code/type), so `docs/model-policy-matrix.md`'s failure
32	table stays accurate.
33	- Require a source citation per claim — a Context7 library ID + section, or an official docs
34	URL. A finding without one is not actionable.
35	- Require an explicit `UNVERIFIED` flag on anything the agent could not confirm from a primary
36	source (e.g. inferred from a changelog mention, or contradicted between two docs). Do not let
37	an agent silently round an uncertain claim into a confident one.
38
39	Example prompt template (adapt per harness/namespace):
40
41	```
42	Research current [codex OpenAI models \| codex Ollama routing \| codex OpenRouter routing \|
43	claude-code subagent model/effort fields \| cursor subagent model field] using Context7
44	(resolve-library-id then query-docs) and official docs. Report, for each model/field:
45	exact slug or ID, supported reasoning-effort values (if any), and the error shape observed
46	or documented for an invalid value (HTTP status + error code/type). Cite the Context7 library
47	ID + section or the exact docs URL for every claim. If you cannot confirm a claim from a
48	primary source, prefix it UNVERIFIED and say why. Do not guess slugs from training data.
49	```
50
51	Run these Explore agents in parallel; each is scoped to one harness or namespace so the
52	citations stay traceable to a narrow question.
53
54	## Step 2 — orchestrator updates the registry
55
56	The orchestrator, not a delegate, edits `catalog/model-registry.json`:
57
58	- Add new models with `last_verified` (today's date) and a `source` where the schema allows it;
59	update the relevant namespace's `sources` array if a new canonical URL was used.
60	- Bump the registry-level `last_refreshed` date.
61	- Never remove a model still referenced by `catalog/model-policy.json` without first
62	migrating the policy rule(s) that reference it to a replacement model — check with
63	`npm run model-policy:report` before deleting anything.
64	- Treat every `UNVERIFIED`-flagged finding from Step 1 as a blocker, not a data point to
65	merge as-is — either verify it directly or leave the registry unchanged for that item.
66	- Validate the edit against `schemas/model-registry.schema.json` structurally (required fields,
67	anchored `match` patterns, `last_verified` date format) before moving on.
68
69	## Step 3 — sync the human-readable matrix
70
71	Delegate to a Sonnet writer subagent to update `docs/model-policy-matrix.md` so its tables match
72	the registry exactly (namespace tables, verified-model tables, failure modes, enforcement
73	boundaries). Give the delegate the exact diff you made to `catalog/model-registry.json` in Step 2
74	and instruct it to touch only `docs/model-policy-matrix.md` — no other file, no commits.
75
76	## Step 4 — verify
77
78	Run in order, orchestrator-owned:
79
80	```bash
81	npm run model-policy:check # registry schema + policy resolves against it
82	npm run validate # full gate suite
83	npm run asset-integrity:writ

People who liked this also viewed

6 picks

Type

model-registry-refresh

17stars3forks

View on GitHub

How to install model-registry-refresh?

raishin/vanguard-frontier-agentic/model-registry-refresh

$npx -y skills add Raishin/vanguard-frontier-agentic --skill model-registry-refresh

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Model Registry Refresh
2
3	## Doctrine
4
5	`catalog/model-registry.json` is the single source of truth `scripts/model-policy.mjs` fails closed
6	against. Every model name and reasoning-effort value it accepts must trace to official documentation
7	with a citation — never to memory, never to a plausible-sounding guess at a slug. This skill is the
8	repeatable workflow for keeping that registry accurate without letting research cost dominate the
9	orchestrator's context.
10
11	## When to run
12
13	- `npm run model-policy:check` fails with an error naming a model "not in the verified model
14	registry" — the registry is missing a model the policy (or an operator) wants to use.
15	- A provider (OpenAI, Anthropic, Cursor) ships new models or retires old ones and the catalog
16	needs to reflect current reality.
17	- Quarterly staleness check — `last_refreshed` in `catalog/model-registry.json` is more than
18	~3 months old.
19
20	## Step 1 — delegate research to Haiku Explore agents
21
22	Fan out one Haiku `Explore` agent per harness (or per namespace, for codex) using the Context7 MCP
23	tools (`mcp__Context7__resolve-library-id` then `mcp__Context7__query-docs`) plus official docs
24	URLs already cited in the registry. Each research task must:
25
26	- Ask for exact slugs/IDs, not families — `gpt-5.5` not "the gpt-5 line".
27	- Ask for reasoning-effort support per model, not per harness — some models in a family
28	predate newer effort levels (see `o1`/`o3`/`o4-mini` lacking `none`/`minimal`/`xhigh` in the
29	current registry).
30	- Ask for failure-mode evidence — what error shape a bad model name or unsupported effort
31	actually produces (HTTP status, error code/type), so `docs/model-policy-matrix.md`'s failure
32	table stays accurate.
33	- Require a source citation per claim — a Context7 library ID + section, or an official docs
34	URL. A finding without one is not actionable.
35	- Require an explicit `UNVERIFIED` flag on anything the agent could not confirm from a primary
36	source (e.g. inferred from a changelog mention, or contradicted between two docs). Do not let
37	an agent silently round an uncertain claim into a confident one.
38
39	Example prompt template (adapt per harness/namespace):
40
41	```
42	Research current [codex OpenAI models \| codex Ollama routing \| codex OpenRouter routing \|
43	claude-code subagent model/effort fields \| cursor subagent model field] using Context7
44	(resolve-library-id then query-docs) and official docs. Report, for each model/field:
45	exact slug or ID, supported reasoning-effort values (if any), and the error shape observed
46	or documented for an invalid value (HTTP status + error code/type). Cite the Context7 library
47	ID + section or the exact docs URL for every claim. If you cannot confirm a claim from a
48	primary source, prefix it UNVERIFIED and say why. Do not guess slugs from training data.
49	```
50
51	Run these Explore agents in parallel; each is scoped to one harness or namespace so the
52	citations stay traceable to a narrow question.
53
54	## Step 2 — orchestrator updates the registry
55
56	The orchestrator, not a delegate, edits `catalog/model-registry.json`:
57
58	- Add new models with `last_verified` (today's date) and a `source` where the schema allows it;
59	update the relevant namespace's `sources` array if a new canonical URL was used.
60	- Bump the registry-level `last_refreshed` date.
61	- Never remove a model still referenced by `catalog/model-policy.json` without first
62	migrating the policy rule(s) that reference it to a replacement model — check with
63	`npm run model-policy:report` before deleting anything.
64	- Treat every `UNVERIFIED`-flagged finding from Step 1 as a blocker, not a data point to
65	merge as-is — either verify it directly or leave the registry unchanged for that item.
66	- Validate the edit against `schemas/model-registry.schema.json` structurally (required fields,
67	anchored `match` patterns, `last_verified` date format) before moving on.
68
69	## Step 3 — sync the human-readable matrix
70
71	Delegate to a Sonnet writer subagent to update `docs/model-policy-matrix.md` so its tables match
72	the registry exactly (namespace tables, verified-model tables, failure modes, enforcement
73	boundaries). Give the delegate the exact diff you made to `catalog/model-registry.json` in Step 2
74	and instruct it to touch only `docs/model-policy-matrix.md` — no other file, no commits.
75
76	## Step 4 — verify
77
78	Run in order, orchestrator-owned:
79
80	```bash
81	npm run model-policy:check # registry schema + policy resolves against it
82	npm run validate # full gate suite
83	npm run asset-integrity:writ

People who liked this also viewed

6 picks

Type