omp-llm-role 1.0.0 → 1.2.0

package/docs/dev/current-state.md

@@ -1,11 +1,348 @@
-# Current state (2026-10-02)
+# Current state (2026-10-11)
 
-Dated snapshot of the ranking as of the date in the heading — matched/priced/
-eligible counts, per-role leaders, and the open defects. It is refreshed by
+Dated snapshot of the ranking and the landed work as of the date in the heading —
+matched/priced/eligible counts, per-role leaders, the open defects, and the
+per-issue work log below. It is refreshed by
 regenerating the report (`node src/cli/llm-role-rank.ts --out docs/llm-role-rankings.md`) and
 copying the new numbers here; release prose belongs in
-[`../../CHANGELOG.md`](../../CHANGELOG.md), not in this file. The numbers below
-are the 2026-10-02 measurement and are not re-derived on read.
+[`../../CHANGELOG.md`](../../CHANGELOG.md), not in this file. The model-level
+ranking numbers below are the 2026-10-09 measurement (the input-price blend basis
+#21, the #17/#18/#19/#20 wave and the #34–#40 focus-axis/capability-flag work);
+the provider-pin numbers are the 2026-10-11 measurement (spec #71; see the top
+bullet). Neither is re-derived on read.
+
+- Provider ordering: value-for-money route score (2026-10-11, spec #71): the
+  auto-pin route basis changed from the route's own value (`q_route −
+  λ·priceEff_route`) to a min/max-normalized **value-for-money score** `vfm =
+  wT·tNorm − wP·pNorm` (`wT = wTput/(wPrice+wTput)`, `wP =
+  wPrice/(wPrice+wTput)`; `tNorm`/`pNorm` min/max over the candidate set), with a
+  new **soft** `filters.preferredMinThroughput` floor (routes at or above sort
+  first; below-floor routes are kept and sorted last) that the `providerPinning`
+  preset sets to 50. `routeValue` is deleted. The model-level ranking is
+  unchanged — the report is not regenerated (the ranking CLI does not apply
+  capability features and no shipped role sets `preferOwnProvider`). Measured
+  2026-10-11 on the current endpoints cache with the live lock's expanded defs
+  (all four capabilities on, `cacheHitRate 0.9`; `deepseek/deepseek-v4.1-flash`,
+  11 allowed providers, 12 allowed routes):
+  `morph/fp8` ($0.0840,
+  103 tok/s) wins `default`, `smol`, `commit`, `task`, `advisor`; `coreweave/fp8`
+  ($0.1978, 185.5 tok/s) wins `slow`, `tiny`, `designer`; `decart/fp4` ($0.0639,
+  72 tok/s) wins `vision`, `plan`.
+
+- Auto-pin chain expansion (2026-10-10, issue #69): with `providerPinning` on,
+  the fallback chain no longer stops at three providers per model — the primary
+  model contributes every gate-passing route except the pinned best, then each
+  `fallbackChainDepth` fallback model contributes every gate-passing route, in
+  score order (the `vfm` route score — spec #71 replaced the original
+  route-value order) and deduped by provider slug, so a down provider fails over
+  to another provider of the same model before the chain switches models. Chain
+  length is data-driven (`(R_primary − 1) + Σ R_fallback`) instead of the fixed
+  `2 + 3·fallbackChainDepth`; the manual-`providerPin` and no-pin paths are
+  unchanged. The ranking numbers are unchanged — the chain is a write-side
+  shape, not a ranking input.
+- Opt-in error reporting (2026-10-08, issue #48): a new `errorReporting` setting
+  (`off` — the default — `ask`, `auto`) turns a genuinely unexpected failure — a
+  command handler throwing, or a `runUpdater` abort whose `RunResult.defect` is
+  set (a `ConfigEditError`/self-check mismatch or the catch-all throw; the
+  enumerated environment aborts stay notify-only) — into one structured report
+  (plugin version, runtime, platform, the failing command, error
+  name/message/stack, timestamp) filed on `Gott50/omp-llm-role`. `ask` surfaces
+  the exact payload and files on confirmation, degrading to a notified prefilled
+  `…/issues/new` URL in a session with no UI; `auto` files without asking; `off`
+  is byte-identical to the pre-#48 behaviour. A token (`GITHUB_TOKEN` → `GH_TOKEN`
+  → `gh auth token`, never stored) uses the API; no token, or a 401/403/404 (a
+  fine-grained PAT without Issues:write answers 404), falls through to the
+  prefilled URL. The payload is redacted (home dir → `~`, token-shaped strings →
+  `<redacted>`) and carries no config contents, model selectors, prompts or
+  history rows. A fingerprint marker dedupes a repeat onto the existing open
+  issue, and the path is capped (one open issue per fingerprint, one creation per
+  fingerprint per UTC day, 3 per UTC day). The ledger of what was filed, offered
+  or suppressed is `llm-role-error-reports.json` under the agent dir. The ranking
+  numbers are unchanged — reporting is a side channel, not a ranking input.
+
+- Empty-collection spellings accepted (2026-10-08, issues #54 + #55): a config
+  whose `task.disabledAgents` is spelled `null` — or whose `retry.fallbackChains`
+  / `modelRoles` body is the serializer's own indented `{}` / `[]` line — used to
+  throw inside `parseConfig`, so **every session start aborted with `config edit
+  refused: disabledAgents: inline value is not supported (block sequence
+  expected)` and no role was ever refreshed**. omp re-serializes `config.yml`
+  through Bun's `YAML.stringify` (`stringifyYamlConfig`), which spells an empty
+  collection exactly that way, so those spellings now read as the empty
+  collection and the patch path rewrites the placeholder line into the block
+  form (`disabledAgents:` + `- "designer"`); any other inline value on a managed
+  key line still throws. Verified against a copy of the real
+  `~/.omp/agent/config.yml` (which carries `disabledAgents: null`): the run wrote
+  and reported 8 decisions instead of aborting — and the real file carries its own
+  production confirmation. Its 2026-10-08T16:34Z `session-start` run wrote the
+  config while `disabledAgents: null` was present (config, `llm-role-state.json`
+  and `llm-role-history.jsonl` are all stamped 18:34:17 local), which pre-fix was
+  impossible: the read threw before any write. The write is recognizably the
+  plugin's surgical one, not omp's serializer: a *changed* role line is re-quoted
+  (`tiny: "…deepseek-v4.1-flash:off"`) while an unchanged one stays byte-identical
+  (`smol: …deepseek-v4.1-flash:off`), the previously plugin-written
+  `openrouter/inclusionai/ling-3.0-flash` chain key is pruned (a dumper never
+  prunes), and the chain keys the plugin does not own (`web:`,
+  `openrouter/~deepseek/…`) keep their unquoted items. The placeholder is left in
+  place because the agent sync had nothing to add or remove — `patchDisabledAgents`
+  early-returns then, by design. The same probe turned up two more defects, one of
+  them tracked as its own issue (#55 — a different error-report fingerprint): the
+  parent-block variant of this class — an empty `task:` + `  {}` / `retry:` + `  {}`
+  used to take the new child key *after* the `{}` line, YAML no parser accepts,
+  with the read-back self-check still passing, i.e. **silent corruption** — now
+  has the child key replace the `{}` line, and a prune-only patch creates nothing;
+  and #55's block-bound bug (`chainEntries` scanned to the end of `retry:`
+  instead of to the end of `fallbackChains:`, so any sibling `retry.*` key
+  aborted every run with `unexpected line in fallbackChains block`) is fixed by
+  bounding the scan at the key's own indent. Ranking numbers unchanged.
+
+- Capability presets (2026-10-07, issue #40): one opt-in flag per capability
+  (`features.<id>` global, `roles.<role>.features.<id>` per role) applies the
+  plugin's recommended settings to a role; every flag ships **off** and no
+  shipped role sets one, so the resolved roles are byte-identical to
+  `DEFAULT_ROLES` until a user opts in. The registry (`FEATURES` in
+  `src/features.ts`) is the single source of truth and `expandFeatures` the
+  single precedence implementation (`effective def = shipped default ⊕
+  preset(enabled flags) ⊕ explicit user keys`; a per-role `false` beats a global
+  `true`; an explicit knob always wins). Grounding on the 2026-10-09 cache
+  (`default` role, 146 eligible; the `providerPinning` row is re-measured on the
+  2026-10-11 cache):
+
+  | Flag | Recommended settings | Measured effect |
+  |---|---|---|
+  | `endpointCeilings` | `filters.tools: true`, `filters.minOutputTokens: 16384` | 146 → 135 eligible, leader unchanged (`deepseek-v4.1-flash`, 0.37 $/M) |
+  | `cachePricing` | `cacheHitRate: 0.5` | 146 eligible, leader unchanged, its `priceEff` 0.3732 → 0.3522 (−5.6 %) |
+  | `providerPinning` | `preferOwnProvider: true`, `filters.preferredMinThroughput: 50` | per-model policy, never changes eligibility; prices each model on its **best route** (max `vfm` among the filter-passing, non-degraded, priced routes, cap-respecting; the soft 50 tok/s floor sorts below-floor routes last; ties by slug then id) — 144 of 144 routed models have a candidate route (2026-10-11), so the soft basis covers the whole routed field |
+  | `costCap` | `filters.maxPriceUsdPerM: 10` | 146 → 137 eligible (the top ≈6 %), leader unchanged; pool p90 7.8, p95 11.1 $/M, every leader 0.20–3.71 $/M |
+
+  `providerPinning`'s recommended setting is a **soft per-model policy**
+  (`preferOwnProvider`), not a hard `providerPin`: a hard pin would make a model
+  with no matching route ineligible and empty most of a role's pool, so the
+  policy prices a model on its **best route** — the max-`vfm` route (the
+  value-for-money score, with the soft 50 tok/s floor sorting below-floor routes
+  last) among the filter-passing, non-degraded, priced routes (thinking-adjusted
+  price ≤ `maxPriceUsdPerM`; ties by slug then id) — and falls back to the
+  `1/price²` blend only when the model has no candidate route. The basis is soft:
+  it never changes eligibility, and on the 2026-10-11 cache every one of the 144
+  routed models has a candidate route, so the capability covers the whole routed
+  field (the old own-lab slug-prefix rule matched only 51 of 149). `minContextTokens`
+  is deliberately not in the `endpointCeilings` bundle (an absolute floor cannot
+  express the video's fraction-of-the-model's-own-window complaint). The flags
+  are settable from `create-role.ts`, `/create-agent`, `/project-roles`
+  (`--feature`/`--list-features`) and the explorer's Features panel; the ranking
+  numbers are unchanged until a flag is turned on.
+
+- Cross-source axis (2026-10-07, issue #38): the focus-metric gate gained the
+  cross-source axis — the zeroeval per-benchmark payload's **second, independent**
+  set of the price/throughput/context quantities (`input_cost_per_million`/
+  `output_cost_per_million`/`speed_rps`/`context_window`, read by
+  `parseBenchmarkPayloadMeta` onto `BenchmarkPayloadMeta.crossSource`, keyed by
+  the entry's `model_id`) compared against the plugin's own `price`/`context`.
+  The join is **direct on the bare id** (the generic llm-stats benchmark's
+  `entries[].model_id` is the llm-stats id, the same space as `Model.id`); the
+  axis reports the median relative divergence per axis against
+  `PRICE_AGREEMENT_TOLERANCE` (0.5, a **chosen, tunable heuristic**).
+  `speedAgreement` is **informational only** and excluded from `status` (the
+  units differ — rps vs tok/s — and re-deriving throughput from `speed_rps` is
+  out of scope). It is an **annotation, never a gate** (`belowBarReason` has no
+  cross-source branch) and it never feeds `price`/`weightPrice` or any ranking
+  input — the `1/price²` blend is unchanged (a test pins `rankRole`
+  byte-identical with the annotation on). It is `unknown` when no model joined or
+  the payload carried none of the four fields. The ranking numbers are unchanged
+  — the axis is a create-agent/explorer annotation, not a weight.
+
+- Provenance axis (2026-10-07, issue #37): the focus-metric gate gained the
+  provenance axis — the benchmark's owner (the catalog row's `dataset_org_id`,
+  falling back to `dataset_slug`) and the source's own per-entry org mix (the
+  dominant lab and its share, read from each row's `organization_id`, falling
+  back to `provider_id`), below-bar above `FOCUS_PROVENANCE_DOMINANT_SHARE` (0.5,
+  a **chosen, tunable heuristic** — the per-entry org mix is not measurable from
+  the repo's caches at implementation time). It is the **source's own claim**
+  about who publishes the benchmark and whose scores it carries: an independent
+  cross-check on the pool-derived `composition` axis (whose rule is unchanged),
+  with `compositionAgreement` recording whether the two concur (`disagree` when
+  exactly one flags `below-bar`, `agree` when both or neither, `unknown` when the
+  payload carries no org mix). It is an **annotation, never a gate**
+  (`belowBarReason` has no provenance branch), so a vendor-populated source is
+  warned but never drops a discovered candidate; it is `unknown` when there is
+  neither an owner nor a payload org mix. The three catalog fields (`datasetId`/
+  `datasetOrgId`/`datasetSlug`) ride `BenchmarkCatalogEntry` behind the cache
+  shape guard (a pre-#37 cache is refetched); the per-entry org mix rides
+  `BenchmarkPayloadMeta.provenance`. The ranking numbers are unchanged — the axis
+  is a create-agent/explorer annotation, not a weight.
+
+- Maintenance axis (2026-10-07, issue #36): the focus-metric gate gained the
+  maintenance axis — the catalog row's `updated_at` age in months against
+  `FOCUS_DATASET_STALENESS_MONTHS` (12), with `version_count`/`star_count`
+  carried as supporting detail. This is the **dataset-date** axis, distinct from
+  the model-date `freshness` axis: a benchmark whose dataset was last touched
+  years ago but whose rows include one recent model passes freshness and is
+  caught here. It is an **annotation, never a gate** (`belowBarReason` has no
+  maintenance branch), so a stale dataset is warned but never drops a discovered
+  candidate; it is `unknown` when the catalog row is absent or its `updated_at`
+  is missing/unparseable. The four fields (`updatedAt`/`versionCount`/
+  `latestVersionRowCount`/`starCount`) ride `BenchmarkCatalogEntry` behind the
+  cache shape guard (a pre-#36 cache is refetched). The ranking numbers are
+  unchanged — the axis is a create-agent/explorer annotation, not a weight.
+
+- Modality axis (2026-10-07, issue #35): the focus-metric gate gained the
+  modality axis — the catalog row's `modality` (text 514, multimodal 194, image 19,
+  audio 9, video 2, null 7 of 745) and the payload's per-entry `multimodal`
+  share. It is an **annotation, never a gate**: there is no role modality field,
+  so the axis is never `below-bar` and never a drop reason; it is `unknown` only
+  when neither the catalog row nor the payload carries a modality signal. The
+  catalog row's `modality` rides `BenchmarkCatalogEntry` behind the cache shape
+  guard (a pre-#35 cache is refetched); the per-entry `multimodal` rides
+  `BenchmarkPayloadMeta.modality`. The ranking numbers are unchanged — the axis
+  is a create-agent/explorer annotation, not a weight.
+
+- Trust axis (2026-10-07, issue #34): the focus-metric gate gained a fifth axis —
+  the per-entry `self_reported` share of a benchmark payload, below-bar above
+  `FOCUS_SELF_REPORTED_MAX_SHARE` (0.5). Measured live: `verified` is uniformly
+  `false`; `self_reported` varies (gpqa 19/20, deepswe-1.1 14/20,
+  terminal-bench-4.0 10/20, aa-omniscience-index 2/3, automationbench-aa 0/1, the
+  general benchmarks ~20/20), so a strict-majority bar flags the mostly
+  vendor-submitted sources while passing an independently measured one. The
+  catalog row's `is_community` (20/745 true) rides `BenchmarkCatalogEntry` behind
+  a cache shape guard and is surfaced on the trust axis as `community` (issue
+  #43), so a community-submitted benchmark is visible even before its scores are
+  loaded. The ranking numbers are unchanged — the axis is a
+  create-agent discovery gate, not a weight.
+
+- Explorer scope switcher (2026-10-06, issue #28): the explorer is no longer
+  user-level only. A **scope** is a role-config source — the user-level lock
+  file, or one project where `/project-roles` was used. `startExplorer` resolves
+  the scope list (the user-level scope, the projects in the new global registry
+  `llm-role-projects.json`, and the session's project when its cwd has a project
+  role config) and opens on the session's project when present, else user-level.
+  A project scope reads the project lock merged over the user-level lock — the
+  updater's own read — and Export writes that scope's lock file through the same
+  validate → backup → atomic path. The ranking dataset is loaded once as the
+  union of every known scope's resolved roles, so a metric only one scope
+  weights (a project-only declared `bench:<id>`) is fetched and switching is
+  free. The registry is written by the project-aware `runUpdater` immediately
+  after `resolveProjectDir()` and before any early return (guarded only by
+  `dryRun`), so a day-stamped no-op session start still registers; entries whose
+  project lock file is gone are pruned on the next write. New `POST /api/scope`;
+  `GET /api/bootstrap` gains `scopes`/`activeScope`. Verified live (headless
+  Chromium): the header lists both scopes, the project is selected by default,
+  switching to user drops the project role and repoints the lock path, and a
+  registry entry with no lock file renders disabled with a reason tooltip.
+
+- Measurement wave (2026-10-06, issues #31–#33): three candidate axes were
+  measured and **not** landed — guardrail/alignment (#31: the semantically
+  relevant catalog entries sit at one model, the best-covered candidate covers
+  5.0% of the field vs the 35% floor, and the safety/security entries that carry
+  data are collinear with `general`, r 0.89–1.00), per-task cost (#32: no
+  per-task or per-role token profile exists — the only per-model profile the
+  plugin already has is OpenRouter's platform-wide traffic, ~10–14× the plugin's
+  own input/req, so the $/M price stays the price axis), and agentic multi-agent
+  axes (#33: no catalog benchmark measures delegation/teams/handoffs/swarms/
+  recovery/a2a, the one multi-agent eval is below the 3-model floor, and the
+  20-entry endpoint cap makes the coverage bar unreachable). Numbers and verdicts
+  in `scoring.md`.
+
+- Endpoints cache shape guard (2026-10-06, issue #30): `readEndpointsCache` now
+  also rejects a cache whose records lack `cacheReadPrice` (alongside
+  `weightPrice`/`contextLength`), so a cache written before the cache-read
+  pricing landed is refetched instead of leaving every route's cache-read price
+  `undefined` and a role's `cacheHitRate` silently inert. The fix forces a
+  refetch on the first run after the upgrade; the refreshed measurement is
+  400 models, throughput-matched 152/400, priced 151.
+
+- Wave (2026-10-06): the `1/price²` blend now weights routes by the **input
+  (prompt) price** — the router's sort key — while reporting the billed 3:1 blend
+  under that distribution (issue #21), so prices moved; endpoint capability
+  filters (#18), the four-axis benchmark-quality gate (#17), provider pinning
+  (#19) and route-aware pricing (#20) landed. The domain-knowledge (#23) and
+  truthfulness (#24) axes were measured and **not** landed (`scoring.md`). Fresh
+  measurement: 400 models, throughput-matched 152/400, priced 151. Eligible:
+  `default`/`smol`/`slow`/`plan`/`commit`/`tiny`/`task`/`advisor` 143, `vision`
+  75, `designer` 88. Leaders: `default`/`task`/`designer` DeepSeek-V4.1-Flash,
+  `smol`/`commit` Muse Spark 1.1 (was DeepSeek-V4.1-Flash), `slow`/`plan`/
+  `advisor` Muse Spark 1.3, `vision` Qwen3.8 Flash, `tiny` Muse Spark 1.1 (was
+  Gemini 3.1 Flash-Lite).
+
+- `plan` cost posture (2026-10-03, issue #16): `plan`'s `price` raised 0.12 → 0.25
+  (λ 0.00682 → 0.01667) so the posture binds on the **reachable** pool — the pool
+  the selector is chosen from. Measured reachable leader-flip w_price 0.2175 (Hy4
+  preview → DeepSeek-V4.1-Flash) vs full-pool 0.311 (Muse Spark 1.3 → Qwen3.8
+  Flash); 0.25 flips the actionable pick while the full-pool #1 is unchanged.
+  Non-price weights rescaled by 0.852273: `general` 0.6, `long_context` 0.116676,
+  `throughput` 0.033324 (Σ(non-price) = 0.75 = 1 − price; `q` unchanged). The
+  reachable flip rides on `long_context` imputation (39% coverage; DeepSeek's
+  index is null → capability fill 0.195, not a measured weakness). Full-pool
+  `plan` #1 stays Muse Spark 1.3 (value 0.837 → 0.801); reachable #1 Hy4 preview →
+  DeepSeek-V4.1-Flash. The other nine roles are untouched.
+
+- Keyed-catalog availability fast path (2026-10-03, issue #15): the updater now consumes
+  the shared `KeyAvailability` primitive — when the account's OpenRouter "Filter the
+  model catalog for API keys" setting is on, allowed candidates skip the probe and
+  key-blocked candidates are pruned from the walk order (the budget counts candidates
+  examined, so pre-seeding a blocked verdict would not save it); candidates in neither
+  catalog are still probed, and `no-filter`/`unavailable` degrade to the probe walk with
+  no abort path. Measured on this account (setting on): keyed `GET /api/v1/models`
+  returned 108 rows vs 466 public; after `rankingIdOf` normalization 107 keyed ranking
+  ids ⊆ 454 public, 347 blocked. A 6-blocked/6-allowed probe cross-check agreed 6/6
+  blocked and 5/6 allowed (the sixth probed `unknown`, which counts as usable).
+
+- Role-weight rebalance (2026-10-03, issue #13): `math` dropped from `plan` and
+  `slow` (collinear with `reasoning`, r 0.919, and missing for 18% of the pool —
+  a coverage penalty, not a quality signal) and its share moved to the
+  consolidated `general` axis (capability-preserving, 100% coverage), not
+  `throughput` (which would have tripled `slow`'s speed weight and flipped both
+  leaders to a cheap/fast model). `general`+`reasoning` merged into `general`
+  in `default`/`slow`/`plan`/`advisor` (a pure relabel: `general`~`reasoning`
+  r 0.984, so the combined capability share is unchanged). `smol`/`commit`
+  `price` raised 0.30/0.35 → 0.45 (the measured binding threshold; non-price
+  weights rescaled proportionally, so it is a posture change, not a capability
+  change). `plan`'s description drops the now-unweighted `math`. New weights:
+  `default` general 0.4926/code 0.1705/agents 0.1137/tool_calling 0.0947/
+  throughput 0.0285/price 0.10; `slow` general 0.60/code 0.18/agents 0.13/
+  throughput 0.04/price 0.05; `plan` general 0.704/long_context 0.1369/
+  throughput 0.0391/price 0.12 (superseded by issue #16, above); `advisor`
+  general 0.5947/long_context 0.1004/
+  price 0.20/throughput 0.1049; `smol` price 0.45/throughput 0.22/general
+  0.1571/code 0.0943/tool_calling 0.0786; `commit` price 0.45/throughput
+  0.2285/general 0.2369/code 0.0846. Leaders: `default` DeepSeek-V4.1-Flash
+  (unchanged), `slow`/`plan` → Muse Spark 1.3, `advisor` Muse Spark 1.3
+  (unchanged), `smol`/`commit` → DeepSeek-V4.1-Flash. `required`, `thinking`,
+  `filters`, `enabled` and the `vision`/`tiny`/`task`/`designer` weights are
+  unchanged.
+
+- Designer weight reconciliation (2026-10-03, issue #12): `designer` `code`
+  0.10 → 0.06 (collinear with `website`, r 0.890, so it double-counted one
+  capability axis) and the freed 0.04 moved to the independent axes (`price`
+  +0.02, `throughput` +0.02), per Scoring rule 5. New weights: general 0.26/
+  code 0.06/vision 0.18/throughput 0.15/price 0.17/website 0.18 (Σ = 1.00,
+  Σ(non-price) = 0.83 = 1 − 0.17); λ 0.00882 → 0.010241. The `design` archetype
+  now derives its `weights`/`required` from `DEFAULT_ROLES.designer` by
+  reference (the `general` archetype's pattern), so the two design weightings
+  cannot diverge. `required`, `thinking`, `filters.image`, `enabled` and the
+  description are unchanged; the leader is unchanged (a refinement, not a
+  re-ranking).
+
+- Coverage-safe benchmark discovery + agent-write hardening in `/create-agent`
+  (2026-10-03, issue #11): discovery now refuses a benchmark that would turn `q`
+  into a coverage score — a fill-0 metric whose catalog coverage is below 35% of
+  the ranking field is dropped (non-fatal, with a reason), while a capability-
+  filled metric (`writing`/`website`/`bench:<id>`, fill 0.195) stays weightable at
+  any coverage. The focus set is capped at 3 (priority order preserved) so one
+  benchmark keeps a decisive share. The handler loads the ranking universe once
+  (the same daily-cached chain the updater/explorer use) for the field size,
+  per-metric coverage, and the differentiation comparison; a failed load skips
+  the share rule (fill rule still applies, coverage annotated `unknown`). The
+  report prints each focus metric's coverage and warns below the bar; `--dry-run`
+  and `--json` carry the same signal. Free-text extraction no longer folds
+  `general`/`price`/`throughput` (every archetype weights them) but keeps
+  `reasoning`. The agent file is written **before** the role, and a failed role
+  write rolls it back (restoring the prior file on a `--force` re-create, else
+  removing the new one), so a partial failure never leaves a dangling role, a
+  retry needs no `--force`, and the user's prior agent is never destroyed. The
+  step-3 benchmark prompt previews the request's resolved weights (`resolveRole`),
+  so the "already in this role's weights" marker appears live. The vendored
+  architect prompt carries a source-version marker (`ARCHITECT_PROMPT_VERSION`)
+  asserted by a test. Unit-tested in `tests/create-agent.test.ts` and
+  `tests/architect-provenance.test.ts`.
 
 - Benchmark discovery in `/create-agent` (2026-10-02, issue #10): when a purpose
   is given, the command fetches the llm-stats benchmark catalog (745 rows,
@@ -44,13 +381,15 @@ are the 2026-10-02 measurement and are not re-derived on read.
   explorer's `/api/bootstrap` listed the external metric with its derived
   metadata. Unit-tested in `tests/benchmark-sources.test.ts`.
 
-- Release (2026-10-02): `Gott50/omp-llm-role` is public and `main` is pushed
-  (`cc120bb`). The git and marketplace install routes were re-verified against
-  the real URLs in throwaway HOMEs (omp 18.4.10): both installed 1.0.0, and each
-  post-install session run wrote `modelRoles` + `retry.fallbackChains` into that
-  HOME's `config.yml` and landed the three daily caches in the installed copy.
-  npm is not published yet (name unclaimed, no credential on the workstation) —
-  see Releasing.
+- Release (2026-10-03): `Gott50/omp-llm-role` is public, `main` is pushed, and
+  `omp-llm-role@1.0.0` is published to npm (`npm view omp-llm-role version` →
+  `1.0.0`). All three install routes were verified against the real URLs in
+  throwaway HOMEs (omp 18.4.12): npm (`omp plugin install omp-llm-role`), git
+  (`github:Gott50/omp-llm-role`) and marketplace
+  (`omp-llm-role@gott50-plugins`) each installed 1.0.0, and each post-install
+  session run wrote `modelRoles` + `retry.fallbackChains` into that HOME's
+  `config.yml` and landed the three daily caches in the installed copy. See
+  Releasing.
 
 - `/create-agent` free-text form (2026-10-02): the command now accepts a
   plain-language request as well as flags. `parseCreateAgentInput` splits the
@@ -314,18 +653,18 @@ are the 2026-10-02 measurement and are not re-derived on read.
   0.663, margin 0.039 > `switchMargin`), because the incumbent has no Design
   Arena data and takes the neutral fill while MiMo carries a measured 0.97.
   The other nine roles are untouched (only `designer` weights `website`).
-- 400 llm-stats models; OpenRouter matched 152/400 (throughput), 151 priced.
-- Eligible per role: 143 (vision 75, designer 88, image-input filter).
+- 408 llm-stats models; OpenRouter matched 155/408 (throughput), 153 priced (the
+  2026-10-08 snapshot read 155/407 — the daily cache moved, not the code; the
+  #34–#40 axes and flags leave every ranking byte-identical until a flag is
+  turned on).
+- Eligible per role: 146 (vision 75, designer 89, image-input filter).
 - Value-ranking leaders (this report, thinking-adjusted prices; the *ranking*
   leaders, before the account's provider whitelist drops the blocked ones):
-  `default` DeepSeek-V4.1-Flash (0.810), `smol` Muse Spark 1.1 (0.740, blocked
-  → DeepSeek), `slow` GLM-5.3 (0.816), `vision` Qwen3.8 Flash (0.731, blocked
-  → DeepSeek), `plan` Qwen3.8 Flash (0.751), `commit` Muse Spark 1.1 (0.768,
-  blocked → DeepSeek), `tiny` Muse Spark 1.1 (0.779, blocked → Ling Fin),
-  `task` DeepSeek-V4.1-Flash (0.759), `advisor` Muse Spark 1.3 (0.788,
-  blocked → Hy4), `designer` DeepSeek-V4.1-Flash (0.787). The suffix pass moved
-  `default` and `advisor` onto their reachable leaders (Muse Spark 1.3 and
-  GPT-5.6 Sol were blocked).
+  `default` DeepSeek-V4.1-Flash (0.810), `smol` DeepSeek-V4.1-Flash (0.729),
+  `slow` DeepSeek-V4.1-Flash (0.832), `vision` DeepSeek-V4.1-Flash (0.730),
+  `plan` Muse Spark 1.3 (0.793), `commit` DeepSeek-V4.1-Flash (0.762), `tiny`
+  Muse Spark 1.1 (0.791), `task` DeepSeek-V4.1-Flash (0.754), `advisor`
+  Muse Spark 1.3 (0.787), `designer` DeepSeek-V4.1-Flash (0.787).
 - Thinking-adjusted pricing landed (2026-09-27): the suffix table moved into
   `DEFAULT_ROLES` as a per-role `thinking` field, and the price axis scales by
   the level's factor for thinking-capable models — `slow` (`:max`, ×7.86)
@@ -373,8 +712,9 @@ are the 2026-10-02 measurement and are not re-derived on read.
   cannot re-add keys outside `settings.roles` `patchModelRoles` selects. The
   owner-written `retry.fallbackChains` key
   `openrouter/~deepseek/deepseek-v4-flash-latest` is now unreferenced by any
-  role; it is left in place deliberately (SPEC §6.4: unwritten, unreferenced
-  keys are never pruned), so removing it is a manual call.
+  role; it is left in place deliberately (unwritten, unreferenced keys are never
+  pruned — see [`writes-and-state.md`](writes-and-state.md)), so removing it is a
+  manual call.
 - Plugin verified live (2026-09-23) with the provider-allowlist probe: the
   account's allowed-providers whitelist excludes first-party openai/azure/
   anthropic endpoints, so the probe gate rewrote `slow` → GLM-5.3 (`:max`),