agent/models.yml

@@ -16,7 +16,7 @@ providers:
     auth: none
     models:
       - id: frank-archon
-        # Model id renamed from frank-archon on 2026-10-08 (official
+        # Model id renamed from frank-medium on 2026-10-08 (official
         # unit-name scheme; the systemd service, scripts and runbooks keep
         # the frank-archon name — service names are unchanged).
         # Server-side compaction routes to compaction.remoteEndpoint in
@@ -82,9 +82,11 @@ providers:
     auth: none
     models:
       - id: frank-dragoon
-        # Model id renamed from frank-dragoon on 2026-10-08 (official
+        # Model id renamed from frank-large on 2026-10-08 (official
         # unit-name scheme; the service/unit/scripts keep the frank-dragoon
-        # name — service names are unchanged).
+        # name — service names are unchanged). The pre-rename id is
+        # `frank-large`, recorded here because the callsign rename overwrote
+        # its own explanation and left "renamed from frank-dragoon".
         name: "frank-dragoon (dankclaw-gpu)"
         # Server-side compaction runs on the compaction.remoteEndpoint in
         # config.yml (192.168.0.143:7771 — frank-zealot's 3090); this steers the
@@ -111,31 +113,80 @@ providers:
         maxTokens: 45056
         reasoning: true
         # VISION ON as of 2026-10-06: the v2 vision tower loads beside the
-        # checkpoint (HALOGEN_VISION_TOWER; red-PNG probe verified on the
-        # card). image_url content parts only — the legacy top-level
-        # images field is the frank-zealot trap, do not repeat it here.
+        # checkpoint (HALOGEN_VISION_TOWER; priced at 0.8 GiB in the boot
+        # ledger, and re-probed 2026-10-21 — a 512 px solid-red PNG comes back
+        # "Red" at 282 prompt tokens). image_url content parts only — the
+        # legacy top-level images field is the frank-zealot trap, do not
+        # repeat it here.
         input: [text, image]
         compat:
           # flash-server emits thinking as `reasoning_content` — same
           # field as every engine this endpoint has run, no harness
           # change needed.
           reasoningContentField: "reasoning_content"
-          # Same Qwen3.8 template family (it ships inside the tokenizer
-          # bundle): any reasoning_effort outside {xhigh, medium, low} is
-          # rejected, not clamped — the fold stays mandatory.
+          # flash-server validates reasoning_effort against its own list and
+          # 400s anything else, naming it: "reasoning_effort 'maximum'
+          # unsupported; use minimal|low|medium|high|xhigh|max|none" (measured
+          # 2026-10-21). Every omp level is already spelled by the engine, so
+          # the map is identity — it exists to record that, not to fold. An
+          # earlier version rewrote high->xhigh and minimal->low to stop
+          # rejections that do not exist, and the rewrite asked the engine for
+          # a different thinking budget than the caller named. Levels really do
+          # differ here: thinking length on one prompt at max_tokens 2048
+          # measured low 766 tokens, xhigh 614, max 1049 with the block closed
+          # at 1024 (reasoning_closed_at). The harness passes an unmapped level
+          # through verbatim (`compat.reasoningEffortMap?.[effort] ?? effort`),
+          # so removing an entry is a pass-through, never a 400.
+          #
+          # `max` still does not reach the engine, and the map is not why: the
+          # harness clamps the requested effort to the model's own effort ladder
+          # before the map applies, and its ladder for this model tops out at
+          # xhigh — captured on the wire 2026-10-21, `:max` left as
+          # `reasoning_effort: "xhigh"` with this identity map in place. The
+          # `max` row below is the engine-side half of the contract; it takes
+          # effect the moment the harness ladder gains `max`, and nothing here
+          # can force it sooner.
+          #
+          # Thinking OFF needs an explicit mode. The default disable mode sends
+          # the LOWEST effort rather than a switch — `:off` left the wire as
+          # `reasoning_effort: "minimal"` (captured), and minimal still thinks:
+          # a canned 26-token block on every "off" request. `none-effort` makes
+          # the harness send `reasoning_effort: "none"`, which this engine
+          # honours as a real switch: `reasoning_tokens: 0`, answer only, no
+          # thinking block in the rendered output (both verified 2026-10-21).
+          #
+          # Re-measure before changing any of it. The allow-list, in the
+          # engine's own words:
+          #   curl -s :7770/v1/chat/completions -H 'content-type: application/json' \
+          #     -d '{"model":"frank-dragoon","messages":[{"role":"user","content":"x"}],
+          #          "reasoning_effort":"maximum"}'
+          #   -> 400 "reasoning_effort 'maximum' unsupported; use
+          #      minimal|low|medium|high|xhigh|max|none"
+          # and read usage.completion_tokens_details.reasoning_tokens at
+          # max_tokens 2048 to see the levels separate. The wire itself is
+          # settled by pointing a scratch HOME at a logging proxy on :7799 that
+          # forwards to :7770 and running `omp --model …:off` — config claims
+          # are worth exactly what their last capture says.
+          reasoningDisableMode: none-effort
           reasoningEffortMap:
-            minimal: low
+            minimal: minimal
             low: low
             medium: medium
-            high: xhigh
-            max: xhigh
-          # The card's thinking recipe, on the wire. The flash-server
-          # engine HAS a server-side sampler (HALOGEN_TEMPERATURE/TOP_P/
-          # TOP_K = 1.0 / 0.95 / 20 in frank-dragoon.env) — these match it,
-          # so per-request values are consistent, not load-bearing.
+            high: high
+            xhigh: xhigh
+            max: max
+          # The card's thinking recipe, on the wire. The flash-server engine
+          # HAS a server-side sampler (HALOGEN_TEMPERATURE/TOP_P/TOP_K =
+          # 1.0 / 0.95 / 20 in frank-dragoon.env), but an omitted parameter is
+          # the engine's choice, not a pinned one: top_k is sent explicitly so
+          # the recipe survives someone editing HALOGEN_TOP_K. The engine
+          # accepts top_k, min_p, presence_penalty and repetition_penalty
+          # (measured 2026-10-21, all 200) — only the values this endpoint is
+          # meant to run are pinned here.
           extraBody:
             temperature: 1.0
             top_p: 0.95
+            top_k: 20
 
   "llamacpp (dankclaw-3090)":
     baseUrl: http://192.168.0.143:7771/v1
@@ -143,7 +194,7 @@ providers:
     auth: none
     models:
       - id: frank-zealot
-        # Model id renamed from frank-zealot on 2026-10-08 (official
+        # Model id renamed from frank-small on 2026-10-08 (official
         # unit-name scheme; the service/unit/scripts keep the
         # frank-zealot name — service names are unchanged).
         name: "frank-zealot (dankclaw-3090)"

omp-docs/log.md

@@ -20,6 +20,21 @@ This page is the bundle's **only** work log: one file, at the root. Never start
 second one in a subdirectory or under another name — `scripts/kb-lint.sh` reports any
 `log.md` that is not this one.
 
+## 2026-10-21 — an effort map folding against rejections that never existed
+
+`agent/models.yml` rewrote `minimal→low`, `high→xhigh` and `max→xhigh` for
+frank-dragoon to dodge 400s the engine does not send: flash-server's own error names
+its allow-list (`minimal|low|medium|high|xhigh|max|none`) and the harness passes an
+unmapped level through verbatim, so the fold only asked for a different thinking
+budget than the caller named. The map is identity now, and
+`reasoningDisableMode: none-effort` was added because the default disable mode does
+not disable — `:off` left as `reasoning_effort: "minimal"` and paid for a canned
+thinking block, where `none` returns `reasoning_tokens: 0`. Both settled by capturing
+the wire through a logging proxy rather than reading the config: a claim about what
+the harness sends is worth exactly its last capture. The same pass restored the three
+pre-callsign model ids (`frank-medium` / `frank-large` / `frank-small`) that the
+2026-10-08 rename had overwritten inside its own explanatory comments. Refs #62
+
 ## 2026-10-21 — a documented mode the panel never had
 
 The home ultrawide had been running at 60 Hz for weeks while `desk-1-home.kdl`