lobes-cli 0.77.2__tar.gz → 0.78.0__tar.gz

lobes_cli-0.78.0/.devague/current

@@ -0,0 +1 @@
+orin-associate-at-1m

lobes_cli-0.78.0/.devague/current_plan

@@ -0,0 +1 @@
+orin-associate-at-1m

lobes_cli-0.78.0/.devague/deliveries/orin-associate-at-1m.json

@@ -0,0 +1,557 @@
+{
+  "plan_slug": "orin-associate-at-1m",
+  "schema_version": 2,
+  "created": "2026-09-13T05:41:01Z",
+  "updated": "2026-09-13T08:52:22Z",
+  "deviations": [
+    {
+      "id": "d1",
+      "what": "widen t1 by one test file: recompute ONLY the vllm-associate entry in tests/test_tool_parser_plugin.py _EXPECTED_NON_PRIMARY_HASHES with a dated comment (plan orin-associate-at-1m, issue #260); every other service hash must stay byte-identical",
+      "task_ref": "t1",
+      "reason": "t1's required vllm-associate change (ASSOCIATE_MAX_NUM_SEQS argv + VLLM_ALLOW_LONG_MAX_MODEL_LEN env) moves that service's hash, failing the tool-parser plan's non-primary service hash-lock (1 failed / 4952 passed); the file documents recompute-on-deliberate-change and has precedent (#120, #222, #227, #217). Operator approved 2026-09-13.",
+      "affects": [
+        "c2",
+        "c3"
+      ],
+      "origin": "user",
+      "status": "approved",
+      "classification": "acceptable",
+      "seq": 1
+    },
+    {
+      "id": "d2",
+      "what": "t7 docs quote 1,952 MiB as the operating minimum available host memory of the shipped orin-associate shape (accept transcript) AND keep 2,588 MiB as the A/B NVFP4-arm figure (measure transcript), instead of 2,588 MiB alone",
+      "task_ref": "t7",
+      "reason": "the live rollout (t5 step 9, cold 1,040,073-token request through the Orin gateway) measured minimum available memory 1,952 MiB at 06:11:19Z on the rendered shape with embed, rerank and gateway resident, lower than the A/B's 2,588 MiB; quoting only 2,588 would overstate headroom on a zero-swap board. Operator approved 2026-09-13.",
+      "affects": [
+        "c29",
+        "h22"
+      ],
+      "origin": "user",
+      "status": "approved",
+      "classification": "acceptable",
+      "seq": 2
+    },
+    {
+      "id": "d3",
+      "what": "widen t7: update lobes/profiles/builtin_shapes/orin-associate.toml header and summary from DECLARED/UNVALIDATED to VALIDATED live 2026-09-13 for exactly what docs/evidence/2026-09-13-accept-orin-associate-1m.txt shows, and rewrite tests/test_orin_associate_shape.py::test_claims_no_shape_validation_though_the_lane_and_budget_are_measured to assert the accept citation",
+      "task_ref": "t7",
+      "reason": "the accept transcript booted the rendered shape on the physical Orin, meeting the shape's own 'until one lands' condition; t7's docs correctly call it accepted/validated while the merged shape file and its pinning test still claim no box booted it. Neither t6 nor t7 covered those files. Operator approved 2026-09-13.",
+      "affects": [
+        "t6",
+        "c9",
+        "h1"
+      ],
+      "origin": "user",
+      "status": "approved",
+      "classification": "acceptable",
+      "seq": 3
+    },
+    {
+      "id": "d4",
+      "what": "add an orin-associate-only boot-order override: the generated docker-compose.shape.yml resets vllm-associate's depends_on and makes vllm-embed and vllm-rerank depend on vllm-associate being healthy, so a plain lobes fleet up boots associate first; tests + goldens; short live check on the Orin that lobes fleet up --apply boots associate first and all lanes go healthy",
+      "task_ref": "t6",
+      "reason": "Qodo review on PR #261 (High): the fleet compose gives vllm-associate depends_on vllm-embed/vllm-rerank service_healthy, so the ordinary fleet up path boots gears first \u2014 the order where util 0.70 is unmeasured at 1M and was refused at 128K; the accept run used targeted lobes up starts and never exercised it. Operator approved 2026-09-13.",
+      "affects": [
+        "c17",
+        "c18"
+      ],
+      "origin": "user",
+      "status": "approved",
+      "classification": "needs-follow-up",
+      "seq": 25
+    }
+  ],
+  "evidence": [
+    {
+      "id": "e1",
+      "obligation_ref": "o1",
+      "test_ref": "tests/test_associate_compose.py::TestDeclaredKnobsRenderNoDeadDeclarations::test_max_num_seqs_set_renders_flag + ::test_max_num_seqs_unset_renders_no_token",
+      "behavior_text": "docker compose config renders --max-num-seqs=2 when set and no token when unset",
+      "contract_text": "c2/h2: conditional --max-num-seqs knob",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "passed in the 187-test obligation run on the merged branch",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 4
+    },
+    {
+      "id": "e2",
+      "obligation_ref": "o1",
+      "test_ref": "tests/test_tool_parser_plugin.py::TestOtherServicesUntouched::test_every_non_primary_service_hash_unchanged",
+      "behavior_text": "only vllm-associate's service hash moved (recomputed with a dated comment); every other non-primary service hash byte-identical",
+      "contract_text": "d1: the associate lane change is scoped to vllm-associate",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "passed in the obligation run after deviation d1",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 5
+    },
+    {
+      "id": "e3",
+      "obligation_ref": "o2",
+      "test_ref": "tests/test_associate_compose.py::TestAssociateAllowLongMaxModelLen",
+      "behavior_text": "raw environment declares the passthrough; default 0; override reaches 1",
+      "contract_text": "c3/h3: allow-long passthrough mirroring vllm-primary",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "3 tests passed in the obligation run",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 6
+    },
+    {
+      "id": "e4",
+      "obligation_ref": "o3",
+      "test_ref": "tests/test_associate_exposure.py (whole file)",
+      "behavior_text": "exposure assertions pass; file unchanged versus main (git diff --stat: empty)",
+      "contract_text": "c4/h11: exposure contract unchanged",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "passed in the obligation run; git diff against main is empty",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 7
+    },
+    {
+      "id": "e5",
+      "obligation_ref": "o4",
+      "test_ref": "tests/test_catalog.py::test_exactly_one_associate_gear + ::test_exactly_one_associate_gear_detects_an_injected_duplicate",
+      "behavior_text": "count of role_hint=='associate' is 1 and names nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4; the helper reports 2 on a copy with a duplicate",
+      "contract_text": "c10/h6/h23: single associate gear, no W4A16 entry",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "2 tests passed in the obligation run",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 8
+    },
+    {
+      "id": "e6",
+      "obligation_ref": "o5",
+      "test_ref": "tests/test_orin_gateway_read_timeout.py (3 tests)",
+      "behavior_text": "orin render emits GATEWAY_READ_TIMEOUT=7200; spark/thor/base do not; the lock allowlist includes the key",
+      "contract_text": "c26: shipped timeout knob",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "3 tests passed in the obligation run",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 9
+    },
+    {
+      "id": "e7",
+      "obligation_ref": "o5",
+      "test_ref": "docs/evidence/2026-09-13-accept-orin-associate-1m.txt steps 9-10",
+      "behavior_text": "non-streamed 1,040,073-token needle 200 PASS wall 2495.1 s; streamed 1,030,073-token needle 200 PASS first byte 1971.42 s; both through the Orin gateway as model=associate with GATEWAY_READ_TIMEOUT=7200",
+      "contract_text": "c26/h19: cold >= 1M completes through the gateway",
+      "evidence_type": "observation",
+      "strength": "execution",
+      "strength_basis": "measured live on the physical Orin during the t5 rollout",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:26:07Z",
+        "commit": "6eeacca"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 10
+    },
+    {
+      "id": "e8",
+      "obligation_ref": "o6",
+      "test_ref": "tests/test_orin_associate_shape.py::test_declares_the_full_associate_override_matching_the_card_documentation + ::test_associate_renders_feasible_with_its_full_declaration_despite_the_card_veto",
+      "behavior_text": "shape override values and rendered ASSOCIATE_* env match the 1M budget",
+      "contract_text": "c7/h4/h16: measured 1M budget in the shape",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "passed in the obligation run",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 11
+    },
+    {
+      "id": "e9",
+      "obligation_ref": "o6",
+      "test_ref": "docs/evidence/2026-09-13-accept-orin-associate-1m.txt steps 3-5",
+      "behavior_text": "live render showed ASSOCIATE_MAX_MODEL_LEN=1048576 / util 0.7 / 8192 / max_num_seqs 2 and the engine booted with that argv, KV 2,899,067 tokens",
+      "contract_text": "c7/c17: the rendered budget boots",
+      "evidence_type": "observation",
+      "strength": "execution",
+      "strength_basis": "measured live on the physical Orin",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T06:06:36Z",
+        "commit": "6eeacca"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 12
+    },
+    {
+      "id": "e10",
+      "obligation_ref": "o7",
+      "test_ref": "tests/test_orin_associate_shape.py::test_hosts_associate_plus_pooling_gears_and_drops_hand_and_every_other_role",
+      "behavior_text": "hosts set and dropped roles asserted; goldens regenerated",
+      "contract_text": "c25/h10: three hosted roles",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "passed in the obligation run",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 13
+    },
+    {
+      "id": "e11",
+      "obligation_ref": "o8",
+      "test_ref": "docs/evidence/2026-09-13-accept-orin-associate-1m.txt steps 0 and 7",
+      "behavior_text": "capabilities associate context 128000 before, 1048576 after; lane /v1/models max_model_len 1048576",
+      "contract_text": "c11/h7: advertised and live window agree at 1M",
+      "evidence_type": "observation",
+      "strength": "execution",
+      "strength_basis": "read live from the Orin gateway and the lane",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T06:10:02Z",
+        "commit": "6eeacca"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 14
+    },
+    {
+      "id": "e12",
+      "obligation_ref": "o9",
+      "test_ref": "docs/evidence/2026-09-13-accept-orin-associate-1m.txt steps 5-12",
+      "behavior_text": "healthy after 360 s; probes 3/3 via gateway + direct 42; needles PASS; agentic 24/24 tools 6/6 recall; associate restarts 0; no new kernel OOM",
+      "contract_text": "c18/h1/h13/h17: acceptance success signal",
+      "evidence_type": "observation",
+      "strength": "execution",
+      "strength_basis": "full acceptance run on the physical Orin, 07:29:26Z ALL DONE",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:29:26Z",
+        "commit": "6eeacca"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 15
+    },
+    {
+      "id": "e13",
+      "obligation_ref": "o9",
+      "test_ref": "docs/evidence/2026-09-13-accept-orin-associate-1m.txt step 4 (compose diff)",
+      "behavior_text": "lobes init --force brought the stale Orin compose current, adding the #227 judge-prompt --chat-template and its bind mount to vllm-rerank",
+      "contract_text": "t5 re-scaffold side effect recorded in the accept transcript",
+      "evidence_type": "observation",
+      "strength": "execution",
+      "strength_basis": "pre/post compose diff captured by the rollout driver",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T06:00:31Z",
+        "commit": "6eeacca"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 16
+    },
+    {
+      "id": "e14",
+      "obligation_ref": "o10",
+      "test_ref": "docs/evidence/2026-09-13-measure-associate-budget-orin-1m.txt",
+      "behavior_text": "rows for NVFP4 and W4A16: boot, KV pool, probes, needles cold/repeat, concurrency, agentic 1/2 sessions, tj, min memory, OOM/restarts; verdict quotes the >= 10% rule (1M decode -7%, agentic -1%/-0.7%, keep NVFP4)",
+      "contract_text": "c19/h18: A/B recorded and decided by the stated rule",
+      "evidence_type": "manual",
+      "strength": "fidelity",
+      "strength_basis": "transcript assembled from the raw driver results and reviewed against compare.md",
+      "outcome": "pass",
+      "run": null,
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 17
+    },
+    {
+      "id": "e15",
+      "obligation_ref": "o11",
+      "test_ref": "docs/evidence/2026-09-13-accept-orin-associate-1m.txt step 12 + docs/orin-associate-deployment.md",
+      "behavior_text": "RestartCount embed 1, rerank 0 quoted; race-doc grep hits: 1",
+      "contract_text": "c27/h20",
+      "evidence_type": "observation",
+      "strength": "execution",
+      "strength_basis": "docker inspect read live after the agentic run; doc grep at the merged commit",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:29:26Z",
+        "commit": "6eeacca"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 18
+    },
+    {
+      "id": "e16",
+      "obligation_ref": "o12",
+      "test_ref": "lobes/profiles/builtin_shapes/orin-associate.toml ROLLBACK note + docs/evidence/2026-09-13-accept-orin-associate-1m.txt step 1",
+      "behavior_text": "rollback note present (grep hits 1); dated .env backup path quoted (grep hits 2)",
+      "contract_text": "c28/h21",
+      "evidence_type": "manual",
+      "strength": "coverage",
+      "strength_basis": "grep at the merged commit",
+      "outcome": "pass",
+      "run": null,
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 19
+    },
+    {
+      "id": "e17",
+      "obligation_ref": "o13",
+      "test_ref": "docs/orin-associate-deployment.md",
+      "behavior_text": "grep hits 1,925 MiB: 2, 2,588 MiB: 2, capped side-process rule: 2",
+      "contract_text": "c29/h22 as amended by d2",
+      "evidence_type": "manual",
+      "strength": "coverage",
+      "strength_basis": "grep at the merged commit",
+      "outcome": "pass",
+      "run": null,
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 20
+    },
+    {
+      "id": "e18",
+      "obligation_ref": "o14",
+      "test_ref": "tests/test_orin_associate_shape.py::test_claims_validation_only_by_citing_the_accept_transcript + tests/test_gateway_fleet_doc.py",
+      "behavior_text": "shape claims VALIDATED only with the accept citation; stale 'no box has booted this SHAPE' gone; doc-alignment tests pass",
+      "contract_text": "c9/h5 and deviation d3",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "passed in the obligation run",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T07:52:44Z",
+        "commit": "b6a287b"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 21
+    },
+    {
+      "id": "e19",
+      "obligation_ref": "o14",
+      "test_ref": "grep 0.56|1,524,000|11.91|16384 over the five docs excluding lines marked historical/superseded and the Spark 0.30+0.14+0.06+0.06 budget sum",
+      "behavior_text": "unmarked retired literals: CLAUDE.md:222:0.56` on the 128 GB GB10. These are the **machine-as-brain** (default) | docs/machine-profiles.md:246:  refused at boot and `0.56` accepted at `max_model_len=128000` | docs/orin-associate-deployment.md:66:`--max-num-seqs`, and `--max-num-batched-tokens 16384` \u2014 kept here for the | docs",
+      "contract_text": "h5: retired literals only historical",
+      "evidence_type": "manual",
+      "strength": "coverage",
+      "strength_basis": "grep at the merged commit",
+      "outcome": "fail",
+      "run": null,
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 22
+    },
+    {
+      "id": "e20",
+      "obligation_ref": "o15",
+      "test_ref": "grep for both transcript paths in the five docs",
+      "behavior_text": "per-file (measure, accept) citation counts: {'CLAUDE.md': ('1', '1'), 'docs/deployment-shapes.md': ('2', '2'), 'docs/nemotron-3.5-lightning-30b-a3b-nvfp4.md': ('1', '1'), 'docs/machine-profiles.md': ('1', '1'), 'docs/orin-associate-deployment.md': ('5', '3')}",
+      "contract_text": "c12/h8",
+      "evidence_type": "manual",
+      "strength": "coverage",
+      "strength_basis": "grep at the merged commit",
+      "outcome": "pass",
+      "run": null,
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 23
+    },
+    {
+      "id": "e21",
+      "obligation_ref": "o14",
+      "test_ref": "context review of the six lines e19 flagged: CLAUDE.md:219-223, docs/machine-profiles.md:240-247, docs/orin-associate-deployment.md:62-90 and :183-212",
+      "behavior_text": "all six hits are compliant: CLAUDE.md:222 continues the Spark GB10 budget sum 0.30+0.14+0.06+0.06 = 0.56 (not an Orin associate figure); machine-profiles.md:246 sits in the sentence opened by 'HISTORY (superseded 2026-09-13)'; orin-associate-deployment.md:66 and :90 sit under the heading '### HISTORY (superseded 2026-09-13): the original 128,000-token docker run'; :191 and :209 quote the compose template's generic :-16384 fallback, which the doc states is not the shape's number. e19's fail came from a line-scoped grep filter that could not see an enclosing heading or a wrapped sentence.",
+      "contract_text": "h5: retired associate budget literals appear only where marked historical (criterion read at sentence/section scope)",
+      "evidence_type": "manual",
+      "strength": "coverage",
+      "strength_basis": "each flagged line read in context at the merged commit",
+      "outcome": "pass",
+      "run": null,
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 24
+    },
+    {
+      "id": "e22",
+      "obligation_ref": "o16",
+      "test_ref": "tests/test_orin_associate_boot_order.py (4 tests incl. docker compose config merge) + tests/test_orin_associate_shape.py::test_end_to_end_init_render_activates_associate",
+      "behavior_text": "override resets vllm-associate depends_on and adds service_healthy edges from both gears; no other built-in shape gains an edge; real compose merge resolves the reversed order",
+      "contract_text": "d4: associate-first boot enforced by the shape override",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "4970 passed full suite at the d4 commit",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T08:35:00Z",
+        "commit": "4b9be8c"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 26
+    },
+    {
+      "id": "e23",
+      "obligation_ref": "o16",
+      "test_ref": "Orin ~/lightning-spike/bootorder/{driver.log,restore.log,restore-samples.log}",
+      "behavior_text": "rendered override resolved associate depends_on=None and gears depends_on associate service_healthy; plain docker compose up -d --no-build started associate+gateway 08:42:38Z with embed/rerank held in created, gears started 08:48:34Z after associate healthy, all four healthy 08:51:32Z, associate restarts 0, capabilities context 1048576",
+      "contract_text": "d4 live: associate-first order on a plain compose up",
+      "evidence_type": "observation",
+      "strength": "execution",
+      "strength_basis": "sampled every 5 s on the physical Orin; lobes fleet up --apply itself failed at the gateway --build step (unpublished MODEL_GEAR_VERSION 0.78.0), so the order check used compose up --no-build with the same file set",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-13T08:51:32Z",
+        "commit": "4b9be8c"
+      },
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false,
+      "seq": 27
+    }
+  ],
+  "deltas": [
+    {
+      "id": "b1",
+      "kind": "amended",
+      "behavior_text": "the vllm-associate service hash lock in tests/test_tool_parser_plugin.py was recomputed for the new knobs; all other service hashes unchanged",
+      "caused_by": [
+        "d1"
+      ],
+      "evidence_refs": [
+        "e2"
+      ],
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false
+    },
+    {
+      "id": "b2",
+      "kind": "amended",
+      "behavior_text": "the Orin associate docs quote 1,925 MiB as the shipped shape's minimum available memory at 1M alongside the A/B's 2,588 MiB, instead of 2,588 MiB alone",
+      "caused_by": [
+        "d2"
+      ],
+      "evidence_refs": [
+        "e17"
+      ],
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false
+    },
+    {
+      "id": "b3",
+      "kind": "amended",
+      "behavior_text": "the orin-associate shape header and summary say VALIDATED live 2026-09-13 citing the accept transcript, and the pinning test asserts that citation",
+      "caused_by": [
+        "d3"
+      ],
+      "evidence_refs": [
+        "e18"
+      ],
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false
+    },
+    {
+      "id": "b4",
+      "kind": "added",
+      "behavior_text": "the Orin reranker now serves with the #227 judge-prompt chat template, a side effect of the t5 re-scaffold bringing the stale compose current",
+      "caused_by": [
+        "c18"
+      ],
+      "evidence_refs": [
+        "e13"
+      ],
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false
+    },
+    {
+      "id": "b5",
+      "kind": "added",
+      "behavior_text": "shapes hosting associate boot it before the pooling gears on a plain compose up (shape override reverses the base template's gears-first depends_on)",
+      "caused_by": [
+        "d4"
+      ],
+      "evidence_refs": [],
+      "origin": "llm",
+      "status": "proposed",
+      "superseded": false
+    }
+  ],
+  "supersessions": []
+}

{lobes_cli-0.77.2 → lobes_cli-0.78.0}/.devague/frames/lightning-on-thor.json

@@ -4,7 +4,7 @@
   "schema_version": 6,
   "status": "drafting",
   "created": "2026-09-12T21:27:30Z",
-  "updated": "2026-09-13T01:12:07Z",
+  "updated": "2026-09-13T01:48:40Z",
   "claims": [
     {
       "id": "c1",
@@ -357,6 +357,42 @@
       "links": [],
       "instruction": "",
       "revisions": []
+    },
+    {
+      "id": "c28",
+      "kind": "decision",
+      "text": "operator approved a measure-only 1M-context spike on the Jetson AGX Orin 64GB (2026-09-13): DSpark x5 + bf16 KV at the live util 0.70 on vllm/vllm-openai:v0.27.1, --max-num-seqs 2, standalone container (no compose/.env edits), associate lane offline for the run and restored in boot-safe order afterwards",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [],
+      "hard_questions": [],
+      "links": [],
+      "instruction": "",
+      "revisions": []
+    },
+    {
+      "id": "c29",
+      "kind": "decision",
+      "text": "the Orin 1M test is an A/B: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 (modelopt) vs useful-quants/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-W4A16 (compressed-tensors GPTQ INT4 g64), both at 1,048,576 with DSpark x5, bf16 KV, util 0.70, max-num-seqs 2 on vllm/vllm-openai:v0.27.1; W4A16 pre-downloaded first; associate offline ~3-3.5 h and restored in boot-safe order; worker-on-Orin estimate deferred (operator 2026-09-13)",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [],
+      "hard_questions": [],
+      "links": [],
+      "instruction": "",
+      "revisions": []
+    },
+    {
+      "id": "c30",
+      "kind": "decision",
+      "text": "goal plan scope (operator 2026-09-13): build and implement a plan for the Orin associate lane at 1,048,576 only (adopt the NVFP4-vs-W4A16 A/B winner, add the needed lane knobs such as --max-num-seqs and VLLM_ALLOW_LONG_MAX_MODEL_LEN, measured shape budget, live acceptance); the Thor keeps worker; workflow = waves via assign-to-workforce, /deviate recorded on GitHub issues, /validate-delivery, /summarize-delivery, then PR via /cicd",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [],
+      "hard_questions": [],
+      "links": [],
+      "instruction": "",
+      "revisions": []
     }
   ],
   "open_vagueness": [