lobes-cli 0.75.2__tar.gz → 0.75.3__tar.gz

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/CHANGELOG.md

@@ -4,6 +4,18 @@ All notable changes to this project are documented here. The format is based on
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project
 adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.75.3] - 2026-09-11
+
+### Added
+
+- Evidence `docs/evidence/2026-09-11-accept-thor-reranker-template-senses-unproxy.txt`: the #227 reranker judge-prompt template rolled out to the Jetson AGX Thor (distractors 0.21–0.87 -> 0.000 at 3 dp (raw ~1e-5, strictly positive), relevant 0.995–1.000, `instruction` now honoured, identical to the Spark's #227 scores at the probe's 3-dp precision; 43.2 -> 47.3 ms median on sm_110's eager/TRITON lane), including the first boot's unified-memory race and its clean restart; and the Thor's `senses` -> Orin proxy withdrawn (the Orin does not host it), so `model=senses` 404s `role_infeasible` with no `hosted_by`.
+
+### Changed
+
+- `deployments/jetson-agx-thor__thor-worker/` re-captured from the live Thor at 0.75.1: `docker-compose.yml` gains the two #227 reranker lines, `qwen3_reranker.jinja` is now committed and locked, and the lock's evidence points at the new transcript. `[env]` is unchanged. A restore from the old capture would have booted the uncalibrated reranker.
+- `VARIATION.md` for that entry: senses is infeasible with no peer, the cross-box worker path is now covered (#247), and the `.env`-digest claim is corrected (a catalog entry cannot carry `.env` in `[files]`, because every entry must be committed beside the lock).
+- `docs/qwen3-reranker-0.6b.md` rollout section now shows per-box template status (Spark and Thor templated; the Orin is still untemplated), and CLAUDE.md records that no box in the mesh serves `senses` today.
+
 ## [0.75.2] - 2026-09-11
 
 ### Added

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/CLAUDE.md

@@ -745,10 +745,12 @@ stays byte-identical to the pre-proxy contract.
 Orin. The Thor in turn proxies `cortex` to the Spark. A caller addresses any
 of these on its own box's gateway and never dials the peer. Proxied answers
 carry `X-Lobes-Proxied-By`, image input included
-(`docs/evidence/2026-09-11-accept-worker-proxy-spark-thor.txt`). The Spark's
-former `senses` → Orin proxy was withdrawn because the Orin does not host
-`senses` either, so `model=senses` on the Spark now 404s `role_infeasible`
-with no `hosted_by`. The 2026-07-31 layout proxied `senses` → Orin and
+(`docs/evidence/2026-09-11-accept-worker-proxy-spark-thor.txt`). Both the
+Spark and the Thor withdrew their `senses` → Orin proxies, because the Orin
+does not host `senses` either. `model=senses` on either box now 404s
+`role_infeasible` with no `hosted_by`
+(`docs/evidence/2026-09-11-accept-thor-reranker-template-senses-unproxy.txt`),
+so no box in the mesh serves `senses` today. The 2026-07-31 layout proxied `senses` → Orin and
 `worker` → Thor. **A proxied
 role's `ready` and `context` are the PEER's own advert (issue #220):** the
 background probe reads the peer's `GET /capabilities` and relays that role

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: lobes-cli
-Version: 0.75.2
+Version: 0.75.3
 Summary: lobes — run, assess, and switch the local vLLM model.
 Project-URL: Homepage, https://github.com/agentculture/lobes-cli
 Project-URL: Issues, https://github.com/agentculture/lobes-cli/issues

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/deployments/jetson-agx-thor__thor-worker/VARIATION.md

@@ -4,24 +4,27 @@
 
 A **Jetson AGX Thor** (sm_110, 122.8 GiB unified, L4T R38.2.2, MAXN) running
 the **`thor-worker`** shape over the **`thor`** card profile, captured from the
-live box on 2026-09-11 at lobes 0.74.6.
+live box on 2026-09-11 at lobes 0.74.6. It was **re-captured the same day at
+0.75.1**, after the reranker judge-prompt template (#227) was rolled out to it.
 
 It hosts **`worker`** — `nvidia/Qwen3.6-35B-A3B-NVFP4` at the full native
 262144 window, `gpu_mem_util=0.45`, fp8 KV cache, `max_num_seqs=1`, and DFlash
 k=12 speculation against the `z-lab/Qwen3.6-35B-A3B-DFlash` drafter — plus the
-`embedder` and `reranker` pooling gears.
+`embedder` and `reranker` pooling gears. The reranker serves with
+`--chat-template` pointing at the vendored `qwen3_reranker.jinja`, which is
+committed here.
 
 It does **not** host `cortex`: `PRIMARY_FEASIBLE=false`, with
 `PRIMARY_PEER_ORIGIN` + `PRIMARY_PEER_PROXY` forwarding `model=cortex` to a
 DGX Spark peer. `hand` is declared infeasible on this card (a measured sm_110
-inference defect, not a budget choice) and `senses` is proxied to a Jetson AGX
-Orin peer.
+inference defect, not a budget choice). `senses` is also infeasible, with **no**
+peer: it used to be proxied to a Jetson AGX Orin, but the Orin does not host
+it, so `model=senses` now 404s `role_infeasible`.
 
 An adopter needs, before running it:
 
 * the same card — the lock refuses a variation mismatch by design;
-* a reachable Spark-shaped peer for `cortex` and an Orin-shaped peer for
-  `senses`, or edits to the peer origins;
+* a reachable Spark-shaped peer for `cortex`, or an edit to the peer origin;
 * `VLLM_GDN_DECODE_KERNEL=triton` in the environment — without it, speculative
   decoding does not run on sm_110 at all;
 * the pinned nightly `vllm/vllm-openai@sha256:8bd082c2…`, which the compose
@@ -35,14 +38,23 @@ are absent by construction. Declare them yourself after restoring.
 ## Measured result
 
 > Measured live on 2026-09-11:
-> `docs/evidence/2026-09-10-accept-thor-worker-flip.txt`. Covers the flip
-> itself (worker serving locally, `cortex` answering by proxy with
-> `X-Lobes-Proxied-By`), single-stream decode through the gateway, the
+> `docs/evidence/2026-09-11-accept-thor-reranker-template-senses-unproxy.txt`
+> covers the state captured here. The templated reranker went from 0.21–0.87
+> distractor scores to 0.000 at three decimals (raw values are about 1e-5 and
+> strictly positive), and the `instruction` field is now honoured
+> (43.2 → 47.3 ms median). `senses` now 404s honestly, and the
+> cortex/worker/embedder proxy paths show no regression.
+> `docs/evidence/2026-09-10-accept-thor-worker-flip.txt` covers the flip
+> itself: worker serving locally, `cortex` answering by proxy with
+> `X-Lobes-Proxied-By`, single-stream decode through the gateway, the
 > structured tool-call probe, image intake with negative controls, and an
-> end-to-end agentic run driven by Qwen Code. Not covered: any peer reaching
-> this box's `worker` cross-box (no peer declares `WORKER_PEER_ORIGIN` for it),
-> long-context retrieval at 262144, video intake, and serve-after-restore —
-> this capture has never been restored onto a fresh box and then served.
+> end-to-end agentic run driven by Qwen Code.
+> `docs/evidence/2026-09-11-accept-worker-proxy-spark-thor.txt` covers a DGX
+> Spark reaching this box's `worker` cross-box.
+>
+> Not covered: long-context retrieval at 262144 (115K is the deepest
+> measured), video intake, and serve-after-restore. This capture has never
+> been restored onto a fresh box and then served.
 
 Companion transcripts, same box and dates:
 `2026-09-10-accept-nvidia-35b-a3b-thor.txt` (load, MoE/attention backend
@@ -63,13 +75,21 @@ reproducible are **not** obtainable from a plain render of an older tree:
 add their slots before the shape could express them.
 
 **The compose file here is hand-edited and that is deliberate.** It carries the
-two knob slots described above. It is captured verbatim rather than
-regenerated, which is the entire point of a lock: the bytes that produced the
-measurements, not the bytes a current render would produce.
+two knob slots described above, plus the two #227 reranker lines (the jinja
+bind mount and `--chat-template`), added by hand rather than by re-rendering.
+It is captured verbatim rather than regenerated, which is the entire point of
+a lock: the bytes that produced the measurements, not the bytes a current
+render would produce.
 
-**`.env` is captured by DIGEST only**, not by content — its digest is in
-`[files]` so `lobes doctor` reports `lock_drift` when the deployed file moves.
-The restorable settings live in `[env]`.
+**`.env` is NOT captured**, not even by digest. `lobes init --from-lock`
+refuses any `.env`-family name in `[files]` outright (`_check_restorable_name`
+in `lobes/cli/_commands/init.py`: "the .env family is never committed"). The
+catalog validator also requires every `[files]` entry to exist in this
+directory, and a deployed `.env` can never be committed. Listing its digest
+would therefore make the entry unrestorable. The restorable settings live in
+`[env]`. The
+`.env`-digest drift check (deviation d4) therefore applies only to a lock kept
+beside a live deployment, never to a catalog entry.
 
 **Re-capture after any change.** There is no capture verb; the writer is a
 library (`lobes/runtime/_lock.py`'s `capture_lock`). Re-run it after a

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/deployments/jetson-agx-thor__thor-worker/deployment.lock.toml

@@ -9,8 +9,8 @@ schema_version = 1
 id = "jetson-agx-thor"
 profile = "thor"
 shape = "thor-worker"
-lobes_version = "0.74.6"
-evidence = "docs/evidence/2026-09-10-accept-thor-worker-flip.txt"
+lobes_version = "0.75.1"
+evidence = "docs/evidence/2026-09-11-accept-thor-reranker-template-senses-unproxy.txt"
 
 [env]
 EMBED_ATTENTION_BACKEND = "TRITON_ATTN"
@@ -64,6 +64,7 @@ WORKER_SPECULATIVE_CONFIG = "\"'--speculative-config={\\\"method\\\": \\\"dflash
 "Dockerfile.vllm-gemma4" = "sha256:a12f2cfdb57c4e1c6495dadf1344d00eeca25eb64546a7a9432760a5311271b1"
 "docker-compose.override.yml" = "sha256:4493642380185ef1f675fbbcfdb0e4b26e1319a17e47cff60781cd32520c5f91"
 "docker-compose.shape.yml" = "sha256:c490e64f3dfe649151769bf0056c7c805447a54c215189543ef001b70ea22fea"
-"docker-compose.yml" = "sha256:2fba2005571fad3ae5348af1886a61a537cd90f0b22d1f776119c5af10a1ea4e"
+"docker-compose.yml" = "sha256:68476d8edd4887d0b58b03ab70d90659351a5a2363c6bd627254996b5cc2f821"
 "mg-logwrap.sh" = "sha256:fd9dfb7d30cee5d8e3045201ecb9c99ee3b1e82c5d397a1b993c99f0006e0ee0"
+"qwen3_reranker.jinja" = "sha256:e1ee98e69aab7b2da366edf1c50efcef37e34b4a0c50fb816336213e68d9047a"
 "qwen3_thinking_tool_parser.py" = "sha256:22a030552d033b897316ca2d0aa75ccac53dc955fe9ddd6c614b75c684338130"

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/deployments/jetson-agx-thor__thor-worker/docker-compose.yml

@@ -594,6 +594,7 @@ services:
       # Durable logs that survive restart/recreate (issue #50) — see mg-logwrap.sh.
       - ${MODEL_GEAR_LOG_DIR:-./logs}:/logs/model-gear
       - ./mg-logwrap.sh:/usr/local/bin/mg-logwrap:ro
+      - ./qwen3_reranker.jinja:/usr/local/share/lobes/qwen3_reranker.jinja:ro
     expose:
       - "8000"
     entrypoint: ["bash", "/usr/local/bin/mg-logwrap"]
@@ -627,6 +628,7 @@ services:
       # mid-request). See #109/#106.
       - ${RERANK_ENFORCE_EAGER:---no-enforce-eager}
       - '--hf-overrides={"architectures": ["Qwen3ForSequenceClassification"], "classifier_from_token": ["no", "yes"], "is_original_qwen3_reranker": true}'
+      - --chat-template=/usr/local/share/lobes/qwen3_reranker.jinja
       - --max-model-len=${RERANK_MAX_MODEL_LEN:-8192}
       - --gpu-memory-utilization=${RERANK_GPU_MEM_UTIL:-0.06}
     healthcheck:

lobes_cli-0.75.3/docs/evidence/2026-09-11-accept-thor-reranker-template-senses-unproxy.txt

@@ -0,0 +1,258 @@
+ACCEPTANCE — Jetson AGX Thor: reranker judge-prompt template (#227) rolled out, senses proxy withdrawn
+Date: 2026-09-11
+Box:  Jetson AGX Thor (thor, gateway :8000, lobes-cli 0.75.1), thor-worker shape
+Catalog: deployments/jetson-agx-thor__thor-worker/ re-captured from the live box afterwards
+
+Two operator changes, each touching ONE service:
+  A. vllm-rerank: serve Qwen3-Reranker-0.6B with the vendored judge-prompt
+     chat template (#227 had rolled to the Spark only; `lobes doctor` on the
+     Thor reported `scaffold_files: missing qwen3_reranker.jinja`).
+  B. gateway: stop proxying `senses` to the AGX Orin, which does not host it.
+
+================================================================================
+A0. BASELINE — untemplated reranker (scripts/probe_reranker_calibration.py,
+    run ON the Thor with `env -u GATEWAY_API_KEY … --url http://localhost:8000`)
+================================================================================
+# reranker calibration probe (#227)
+date:       2026-09-11T10:51:51+0300
+host:       thor
+url:        http://localhost:8000
+model:      Qwen/Qwen3-Reranker-0.6B
+git sha:    
+container:  vllm/vllm-openai@sha256:8bd082c274fae025b7079498fe1da65182ba1d4c2188c0f5a68c1042c38c3695
+args:       ["/usr/local/bin/mg-logwrap","vllm","serve","Qwen/Qwen3-Reranker-0.6B","--served-model-name=Qwen/Qwen3-Reranker-0.6B","--host=0.0.0.0","--port=8000","--runner=pooling","--convert=classify","--attention-config={\"backend\": \"TRITON_ATTN\"}","--enforce-eager","--hf-overrides={\"architectures\": [\"Qwen3ForSequenceClassification\"], \"classifier_from_token\": [\"no\", \"yes\"], \"is_original_qwen3_reranker\": true}","--max-model-len=8192","--gpu-memory-utilization=0.06"]
+templated:  False
+vllm:       0.26.1rc1.dev942+g5a4c8d992
+
+## probe set — relevance_score per document, usage.prompt_tokens
+
+[sky]  Q: What colour is the sky on a clear day?
+  prompt_tokens total=67 per_pair=22.3  latency=164ms
+  [0] 0.971  On a clear day the sky is blue because air scatters short wavelengths.
+  [1] 0.783  Cats purr when they are content.
+  [2] 0.267  The invoice is due on the last business day of the month.
+  ranking: [0, 1, 2]
+
+[ports-ledger]  Q: Which file lists the ports the gateway exposes?
+  prompt_tokens total=71 per_pair=23.7  latency=65ms
+  [0] 0.937  The gateway port ledger is docs/gateway-fleet.md, which lists 8000 and 8001.
+  [1] 0.540  Cats purr when they are content.
+  [2] 0.521  Bananas are a good source of potassium.
+  ranking: [0, 1, 2]
+
+[france (assess probe)]  Q: What is the capital of France?
+  prompt_tokens total=51 per_pair=17.0  latency=60ms
+  [0] 0.970  Paris is the capital and most populous city of France.
+  [1] 0.209  The Amazon rainforest spans several South American countries.
+  [2] 0.872  Bananas are a good source of potassium.
+  ranking: [0, 2, 1]
+
+[toolbatch inversion]  Q: How does the tool batcher group tool calls into one request?
+  prompt_tokens total=81 per_pair=27.0  latency=58ms
+  [0] 0.795  NOTICE: this file is generated; do not edit by hand.
+  [1] 0.988  toolbatch collects consecutive tool calls and issues them as a single batched 
+  [2] 0.733  Cats purr when they are content.
+  ranking: [1, 0, 2]
+
+[graded relevance]  Q: How do I stop the lobes fleet without deleting the deployment directory?
+  prompt_tokens total=89 per_pair=29.7  latency=52ms
+  [0] 0.968  Run `lobes stop --apply`; it runs `docker compose down`, removing the containe
+  [1] 0.713  `lobes stop` stops the fleet.
+  [2] 0.298  Cats purr when they are content.
+  ranking: [0, 1, 2]
+
+## instruction probe — same pair with and without a top-level `instruction`
+  /v1/rerank: without=[(0, 0.9358), (1, 0.5226)] tokens=53
+  /v1/rerank: with   =[(0, 0.9358), (1, 0.5226)] tokens=53
+  /v1/rerank: instruction changes scores: False
+  /v1/score: without=[(0, 0.9358), (1, 0.5226)] tokens=53
+  /v1/score: with   =[(0, 0.9358), (1, 0.5226)] tokens=53
+  /v1/score: instruction changes scores: False
+
+## latency — 1 query x 5 docs, median of 5 (1 warm-up discarded)
+  samples_ms=[43.2, 42.6, 43.2, 46.9, 43.4]
+  median_ms=43.2
+
+================================================================================
+A1. DEPLOY RECORD
+================================================================================
+The Thor's docker-compose.yml is the hand-edited #244 capture, so it was NOT
+re-rendered (same decision as the Spark's #227 deploy, deviation d2 there).
+  cp docker-compose.yml docker-compose.yml.bak-20260911-pre-227
+  lobes doctor --fix            -> "would write qwen3_reranker.jinja" (nothing else)
+  lobes doctor --fix --apply    -> wrote qwen3_reranker.jinja
+    sha256 e1ee98e69aab7b2da366edf1c50efcef37e34b4a0c50fb816336213e68d9047a
+    == lobes/templates/fleet/qwen3_reranker.jinja == the Spark's deployed copy
+  diff bak -> live: exactly two added lines under vllm-rerank:
+    596a597  - ./qwen3_reranker.jinja:/usr/local/share/lobes/qwen3_reranker.jinja:ro
+    629a631  - --chat-template=/usr/local/share/lobes/qwen3_reranker.jinja
+  env -u GATEWAY_API_KEY lobes up reranker --apply
+    -> docker compose … up -d --no-deps vllm-rerank   (worker/embed/gateway untouched)
+
+  FIRST BOOT FAILED, second succeeded (RestartCount=1, OOMKilled=false):
+    rerank-20260911T075242Z.log:
+      AssertionError: Error in memory profiling. Initial free memory 46.42 GiB,
+      current free memory 46.43 GiB. This happens when other processes sharing
+      the same container release GPU memory …
+    That is the Thor's unified-memory boot race (CLAUDE.md support table,
+    "concurrent first boot can fail on a memory race"), not the template.
+    rerank-20260911T075345Z.log (the restart):
+      non-default args: … 'chat_template': '/usr/local/share/lobes/qwen3_reranker.jinja' …
+      GPU KV cache size: 64,880 tokens, Maximum concurrency for 8,192 tokens per request: 7.92x
+      Detected the chat template content format to be 'string'.
+      Application startup complete.
+    healthy 115 s after `up` (including the failed attempt).
+
+================================================================================
+A2. AFTER — templated reranker (same script, same box)
+================================================================================
+# reranker calibration probe (#227)
+date:       2026-09-11T10:55:05+0300
+host:       thor
+url:        http://localhost:8000
+model:      Qwen/Qwen3-Reranker-0.6B
+git sha:    
+container:  vllm/vllm-openai@sha256:8bd082c274fae025b7079498fe1da65182ba1d4c2188c0f5a68c1042c38c3695
+args:       ["/usr/local/bin/mg-logwrap","vllm","serve","Qwen/Qwen3-Reranker-0.6B","--served-model-name=Qwen/Qwen3-Reranker-0.6B","--host=0.0.0.0","--port=8000","--runner=pooling","--convert=classify","--attention-config={\"backend\": \"TRITON_ATTN\"}","--enforce-eager","--hf-overrides={\"architectures\": [\"Qwen3ForSequenceClassification\"], \"classifier_from_token\": [\"no\", \"yes\"], \"is_original_qwen3_reranker\": true}","--chat-template=/usr/local/share/lobes/qwen3_reranker.jinja","--max-model-len=8192","--gpu-memory-utilization=0.06"]
+templated:  True
+vllm:       0.26.1rc1.dev942+g5a4c8d992
+
+## probe set — relevance_score per document, usage.prompt_tokens
+
+[sky]  Q: What colour is the sky on a clear day?
+  prompt_tokens total=279 per_pair=93.0  latency=898ms
+  [0] 0.998  On a clear day the sky is blue because air scatters short wavelengths.
+  [1] 0.000  Cats purr when they are content.
+  [2] 0.000  The invoice is due on the last business day of the month.
+  ranking: [0, 1, 2]
+
+[ports-ledger]  Q: Which file lists the ports the gateway exposes?
+  prompt_tokens total=283 per_pair=94.3  latency=65ms
+  [0] 1.000  The gateway port ledger is docs/gateway-fleet.md, which lists 8000 and 8001.
+  [1] 0.000  Cats purr when they are content.
+  [2] 0.000  Bananas are a good source of potassium.
+  ranking: [0, 1, 2]
+
+[france (assess probe)]  Q: What is the capital of France?
+  prompt_tokens total=264 per_pair=88.0  latency=64ms
+  [0] 0.995  Paris is the capital and most populous city of France.
+  [1] 0.000  The Amazon rainforest spans several South American countries.
+  [2] 0.000  Bananas are a good source of potassium.
+  ranking: [0, 2, 1]
+
+[toolbatch inversion]  Q: How does the tool batcher group tool calls into one request?
+  prompt_tokens total=292 per_pair=97.3  latency=61ms
+  [0] 0.000  NOTICE: this file is generated; do not edit by hand.
+  [1] 1.000  toolbatch collects consecutive tool calls and issues them as a single batched 
+  [2] 0.000  Cats purr when they are content.
+  ranking: [1, 0, 2]
+
+[graded relevance]  Q: How do I stop the lobes fleet without deleting the deployment directory?
+  prompt_tokens total=301 per_pair=100.3  latency=52ms
+  [0] 1.000  Run `lobes stop --apply`; it runs `docker compose down`, removing the containe
+  [1] 1.000  `lobes stop` stops the fleet.
+  [2] 0.000  Cats purr when they are content.
+  ranking: [1, 0, 2]
+
+## instruction probe — same pair with and without a top-level `instruction`
+  /v1/rerank: without=[(0, 0.9999), (1, 0.0)] tokens=194
+  /v1/rerank: with   =[(0, 0.9996), (1, 0.0)] tokens=200
+  /v1/rerank: instruction changes scores: True
+  /v1/score: without=[(0, 0.9999), (1, 0.0)] tokens=194
+  /v1/score: with   =[(0, 0.9996), (1, 0.0)] tokens=200
+  /v1/score: instruction changes scores: True
+
+## latency — 1 query x 5 docs, median of 5 (1 warm-up discarded)
+  samples_ms=[48.5, 44.5, 47.3, 49.4, 45.3]
+  median_ms=47.3
+
+lobes assess --probes --role reranker --json:
+{"passed": true, "probes": {"reranker": {"role": "reranker", "probe": "rerank_relevance", "ok": true, "evidence": {"top_index": 0, "expected_index": 0, "ranking": [0, 2, 1], "prompt_tokens": {"total": 264, "per_pair": 88.0}}, "latency_ms": 82.2, "error": null}}}
+
+A. SUMMARY (Thor, before -> after)
+  prompt tokens per pair   17–30  ->  88–100
+  distractor scores        0.209–0.872 (bananas 0.872 for "capital of France")  ->  all 0.000 at 3 dp
+                           (raw values are ~1e-5, STRICTLY POSITIVE — see note below)
+  relevant documents       0.937–0.988  ->  0.995–1.000
+  toolbatch inversion      NOTICE 0.795 beside the right answer  ->  0.000
+  `instruction` honoured   False  ->  True (/v1/rerank and /v1/score)
+  median latency (1q x 5d) 43.2 ms  ->  47.3 ms
+Scores are identical to the Spark's #227 acceptance numbers AT THE PROBE'S
+3-DECIMAL PRECISION (docs/evidence/2026-08-30-accept-reranker-template-spark.txt),
+same probe set.
+
+Note — "0.000" is rounding, not an exact zero. Raw /v1/rerank relevance_score,
+"What is the capital of France?", 2026-09-11 (PR #248 review follow-up):
+  thor : Paris 0.9951247572898865 | Bananas 1.9502436771290377e-05 | Amazon 6.019133252266329e-06
+  spark: Paris 0.9953681230545044 | Bananas 1.8491691662347876e-05 | Amazon 5.395844254962867e-06
+A consumer that treats scores as strictly positive (e.g. eidetic-cli's
+`--rerank-threshold`, where a non-positive cutoff means "no cutoff") is
+unaffected; pick a small positive cutoff such as 0.01 to drop distractors.
+Latency diverges from the Spark: the Spark got FASTER (28.0 -> 18.1 ms); the
+Thor got slightly slower. The Thor lane runs `--enforce-eager` + TRITON_ATTN
+(validated sm_110 divergences), so the ~4x longer templated prompt is not
+offset the way it was on the GB10. The same saturation caveat as #227
+applies: two relevant documents both score 1.000 with the weaker one first.
+
+================================================================================
+B. SENSES — proxy to the Orin withdrawn
+================================================================================
+The Orin does not host senses — its own GET /capabilities, 2026-09-11:
+  orin senses: {'feasible': False, 'ready': False, 'proxied': None, 'hosted_by': None}
+{"error": {"message": "Invalid API key. Pass this gateway's configured key as 'Authorization: Bearer <key>'.", "type": "invalid_api_key", "code": "invalid_api_key"}}
+HTTP/1.1 401 Unauthorized
+X-Lobes-Proxied-By: http://orin.tail0be7e0.ts.net:8000
+{'feasible': False, 'ready': False, 'proxied': True, 'hosted_by': 'http://orin.tail0be7e0.ts.net:8000', 'context': 262144}
+
+Change (~/.lobes/.env, backup .env.bak-20260911-pre-senses-unproxy):
+  MULTIMODAL_PEER_ORIGIN=http://orin.tail0be7e0.ts.net:8000  -> commented out
+  MULTIMODAL_PEER_PROXY=true                                 -> commented out
+  (MULTIMODAL_FEASIBLE=false unchanged; the stale "deviation d1" topology
+   header comment rewritten to the #244 topology)
+  env -u GATEWAY_API_KEY lobes up gateway --apply     (gateway only)
+  post-recreate: inbound key UNSET (unchanged posture), no MULTIMODAL_PEER_ORIGIN in env
+
+### AFTER — Thor model=senses
+{"error": {"message": "The model `senses` is not feasible on this machine \u2014 its backend (`multimodal`) is declared hardware-infeasible by this deployment's per-machine profile and will never be served here.", "type": "role_infeasible", "code": "role_infeasible"}}
+HTTP/1.1 404 Not Found
+capabilities senses: {'feasible': False, 'ready': False, 'proxied': None, 'hosted_by': None, 'context': 32768}
+thor /v1/models: ['nvidia/Qwen3.6-35B-A3B-NVFP4', 'Qwen/Qwen3-Embedding-0.6B', 'Qwen/Qwen3-Reranker-0.6B', 'unsloth/Qwen3.8-27B-NVFP4']
+### regression — Thor model=cortex (-> Spark)
+  'pong'
+X-Lobes-Proxied-By: http://spark.tail0be7e0.ts.net:8001
+### regression — Spark model=worker (-> Thor)
+  'pong'
+X-Lobes-Proxied-By: http://thor.tail0be7e0.ts.net:8000
+### regression — Spark model=embedder (-> Thor)
+  dims: 1024
+X-Lobes-Proxied-By: http://thor.tail0be7e0.ts.net:8000
+### Thor reranker via Thor gateway /v1/rerank
+  [(1, 0.995), (0, 0.0)]
+
+(senses `context: 32768` is the Thor's own operator-declared
+MULTIMODAL_MAX_MODEL_LEN — #234's rule: a declared window is still reported.
+Before the change it was 262144, relayed from a peer that does not host it.)
+
+================================================================================
+C. HEALTH AFTER BOTH CHANGES
+================================================================================
+lobes doctor (Thor): healthy — every check [ok], including
+  scaffold_files: all 8 expected scaffold files present
+  gateway_version_match: gateway and CLI both report lobes-cli 0.75.1
+LOBES_SMOKE_BASE_URL=http://thor.tail0be7e0.ts.net:8000 pytest tests/test_live_capabilities.py:
+  PASSED test_advertised_ready_roles_are_reachable
+  PASSED test_advertised_models_are_reachable
+  PASSED test_colleague_discovers_and_dials_generate_roles
+  FAILED test_cli_and_gateway_capabilities_agree   — harness artifact: it shells the
+         LOCAL `lobes capabilities --port 8000` on the Spark, which is not this box
+  FAILED test_deployed_gateway_version_matches_cli — main is 0.75.2 (docs/test-only
+         release), the deployed gateway 0.75.1; no runtime change between them
+
+================================================================================
+Honest scope
+================================================================================
+VALIDATED: the templated reranker on sm_110 (calibration, instruction, assess
+probe), senses honestly infeasible with no referral, no regression on the
+cortex/worker/embedder proxy paths. NOT validated: rerank ordering among
+several relevant documents (saturation, as on the Spark); serve-after-restore
+of the re-captured catalog entry.

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/docs/qwen3-reranker-0.6b.md

@@ -213,11 +213,23 @@ given box is templated; the probe's PASS rule is unchanged (ordering only).
 
 ### Rollout — by-box divergence, accepted
 
-Validated on the **DGX Spark only** (decision q3). The Jetson AGX Thor and the
-Jetson AGX Orin also host `vllm-rerank` and **serve untemplated scores until
-their next `lobes init --apply`**. That divergence window is accepted, not
-overlooked: during it, two boxes in the same mesh answer the same rerank request
-on different score scales.
+This was first validated on the **DGX Spark only** (decision q3), with the
+other boxes serving untemplated scores until they were rolled. That divergence
+window is accepted, not overlooked: during it, boxes in the same mesh answer
+the same rerank request on different score scales.
+
+| box | templated | evidence |
+|---|---|---|
+| DGX Spark GB10 | yes, 2026-08-30 | `docs/evidence/2026-08-30-accept-reranker-template-spark.txt` |
+| Jetson AGX Thor (sm_110) | yes, 2026-09-11 | `docs/evidence/2026-09-11-accept-thor-reranker-template-senses-unproxy.txt` |
+| Jetson AGX Orin | **no**, still untemplated as of 2026-09-11 (no `--chat-template` in its args, no jinja in its deploy dir) | — |
+
+On the Thor the scores are identical to the Spark's on the same probe set.
+Latency differs, though: 43.2 → 47.3 ms median, where the Spark went
+28.0 → 18.1 ms. The Thor lane's validated sm_110 divergences (`--enforce-eager`,
+`TRITON_ATTN`) do not offset the ~4× longer templated prompt. Both boxes
+were rolled by adding the two lines by hand, because both deployed compose
+files are hand-edited and were not re-rendered.
 
 ### Gateway note
 

lobes_cli-0.75.3/lobes/templates/fleet/qwen3_reranker.jinja

@@ -0,0 +1,11 @@
+<|im_start|>system
+Judge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be "yes" or "no".<|im_end|>
+<|im_start|>user
+<Instruct>: {{ instruction | default(instruct | default(messages | selectattr("role", "eq", "system") | map(attribute="content") | first | default("Given a web search query, retrieve relevant passages that answer the query", true), true), true) }}
+<Query>: {{ messages | selectattr("role", "eq", "query") | map(attribute="content") | first }}
+<Document>: {{ messages | selectattr("role", "eq", "document") | map(attribute="content") | first }}<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "lobes-cli"
-version = "0.75.2"
+version = "0.75.3"
 description = "lobes — run, assess, and switch the local vLLM model."
 readme = "README.md"
 license = "Apache-2.0"

{lobes_cli-0.75.2 → lobes_cli-0.75.3}/uv.lock

@@ -440,7 +440,7 @@ wheels = [
 
 [[package]]
 name = "lobes-cli"
-version = "0.75.2"
+version = "0.75.3"
 source = { editable = "." }
 
 [package.optional-dependencies]