The Proposed roles page (#/agent-catalog) reads 18 source-controlled role
definitions from shared/agent-catalog.mjs. It performs no model call, account
write, provider probe, or tool execution. Each role has its own stable ID, purpose,
instructions, expected output, model family, future tool grants, limits, fixture,
and activation requirement. A display name is never treated as an API model ID.
Search by name, purpose or provider, and filter by documented or unresolved model
ID; an empty result is explicit. Filtering never activates a role or changes
saved agent configurations. On narrow screens, details appear above the list;
selecting a role moves focus and the viewport back to those details. Desktop
selection keeps its side-by-side layout.
Each selected role also shows concrete synthetic fixture inputs and unscored
future evaluation criteria. Download fixture JSON exports that packet locally;
it starts no model, tool, task or account write.
- Documented: an exact API ID appears in the linked provider's own catalog, or a local candidate has a pinned repository revision, file and digest. This does not show account access, adapter behavior, or suitability for the role.
- Unresolved model ID: no exact API ID or chosen local artifact has been
verified.
cz-local-coderon Qwen roles is the existing application's local alias, separate from their upstream repository/file identity. - Contract tested:
node --test tests/agent-catalog.test.mjschecks unique, disabled definitions and rejects mismatched, stale, oversized, or fabricated output envelopes. This is deterministic host validation, not model evaluation. - Configured / runtime verified: neither applies to the new catalog. The existing saved local agent path is separate and retains its earlier evidence.
As of 2026-10-19, 12 entries reference documented API IDs, two Qwen roles share one pinned local artifact, and four need exact model selection. All 18 remain disabled. No access to Anthropic, OpenAI, Google, Mistral, Moonshot, or open-model inference was tested in this mission.
| Source (accessed 2026-10-19) | Relevance and adopted decision |
|---|---|
| Anthropic models | Haiku 5.5, Sonnet 5.5, Opus 5.5, and Fable 5.1 IDs. Keep role names distinct from the model IDs and require a later account and adapter check. |
| OpenAI API models | Luna, Sol, and Astra IDs. Record them as documented only; no API credential or availability is inferred. |
| Gemini API models | Gemini 3.8 Flash endpoint for Explorer. The role remains disabled until adapter and evaluation work. |
| Mistral Small 4 | Current mistral-small-2603 ID for Editor. Earlier mistral-small-2503 was rejected because it appears under deprecated models. |
| Pinned Qwen GGUF and model card/license | Explicitly choose the application's existing Qwen2.5-Coder 7B Q4_K_M candidate for both local roles. Record the official revision, file SHA-256 and Apache-2.0 license. Reuse the established pin; do not download or substitute new weights. File identity is independent of role or runtime qualification. |
Local Llama, Gemma, and DeepSeek roles still need an explicitly selected model artifact, license, hardware check, and qualified adapter. Kimi needs a verified Moonshot API ID and account access. No silent substitute is allowed.
The two Qwen roles declare repository Qwen/Qwen2.5-Coder-7B-Instruct-GGUF,
revision 13fb94bfda8c8cf22497dc57b78f391a9acb426a, file
qwen2.5-coder-7b-instruct-q4_k_m.gguf and digest
509287f78cb4d4cf6b3843734733b914b2c158e43e22a7f4bf5e963800894d3c.
The file is 4,683,073,536 bytes; runtime assets and KV-cache/compute overhead are
additional. Actual RAM/VRAM fit remains unqualified. The starter's 4,096-token
window motivates a planned 2,496-token complete-input / 1,600-token output budget
and a 6,000-character outer input ceiling. Character count does not prove token
fit: a later host must tokenize the entire formatted prompt before activation.
The catalog shows declared identity only and never inspects the user's files.
npm run model:inspect -- --hash performs the separate read-only byte check;
it does not read credentials, launch the provider, or qualify either role.
The selected candidate is versioned; changing an installer pin requires an
explicit role/fixture review and does not silently replace this identity.
Every candidate receives a bounded task, project/run ID, source revision and
untrusted references. Host validation compares the output's run ID, project ID
and source revision with the exact expected context, rejects unknown fields,
and requires matching profile and artifact types, a bounded summary, evidence
references and uncertainties. It must state runtimeVerified: false and
merged: false.
The expected host context includes a bounded allowedEvidence list. Outputs may
cite only those exact IDs; the output cannot grant itself new evidence. This
checks reference provenance, not whether the summary correctly understands a
source. A future web tool must supply observed source IDs through the host.
shared/agent-evaluation.mjs contains task fixtures for conflict, stale base,
disagreement, invalid output, missing documents, and large repository context.
They define validation inputs, not scored model results.
npm run agents:fixtures
npm run agents:fixtures -- --profile qwen-local-reviewershared/agent-fixtures.mjs supplies 12 concrete input packs shared by the 18
roles: source disagreements, missing reproduction steps, stale code, test cases,
narrow focus, review disagreement, migration decisions, missing documents,
malformed content, outdated startup, complexity and bounded repository mapping.
Each has supplied text/reference IDs, future acceptance criteria and an authored
example envelope. The runner checks that envelope and seven invalid variants
per role: stale revision, wrong run/project, invented reference, claimed runtime
verification, claimed merge and empty output. It writes a host-only report to
.cache/verification/latest-agent-fixtures.json. These are synthetic fixtures,
not conversations, model contributions, benchmark scores or runtime receipts.
The browser download contains the same declared scope and zero model calls.
The independent future evaluation set contains 12 different cases and no example outputs. It is not imported by the demo exporter or contract runner. A later evaluation host must keep criteria and demonstration answers out of provider input, supply only the selected case's input references, and restrict repository tools accordingly. This versioned public set is separate from the examples; it does not prove a model has never seen similar material. No role-quality evaluation ran during this mission.
Tool grants in the catalog are proposed permissions. repository.read,
proposal.read, proposal.write, and web.read have no catalog executor yet.
There is no shell, secret, network write, merge, or sandbox grant. Before
activating any profile, implement a provider adapter with explicit access checks,
host-enforced permission and cost ceilings, timeout and cancellation receipts,
revision checks, test fixtures, and a real runtime qualification. Record the
provider response and failures without upgrading another profile by association.
The disconnected hosted-provider drafts prepare text-only
request shapes and inspect offline response fixtures for four documented
families. They supply no credential or transport and do not change the disabled
or runtime-unverified status of any role.
All nine distinct documented API IDs were rechecked in their primary catalogs during this mission; twelve role entries refer to them. Mistral now links directly to its model page rather than the overview. A documented API identifier is not an immutable weight digest, account entitlement or successful response. Recheck model lifecycle, exact response identity and access before a later activation.
The existing Commons Builder/Reviewer workflow is a separate, locally configured path and still does not execute generated JavaScript. Its parser and review do not establish runtime correctness.
An actual browser capture of the proposed-role view shows the source-controlled catalog in a disposable scenario fixture. The scenario badge is part of the deterministic example; none of the proposed roles ran to produce this image.
The narrow Qwen details capture is an actual Chrome element capture of declared artifact, alias, license and planned limits. It reports no local file availability or runtime qualification.
The fixture-section capture is a real Chrome element capture from Fable Architect's read-only details. It shows authored input, unscored future criteria and the local JSON download; it is not model output.
