lobes-cli 0.78.4__tar.gz → 0.80.0__tar.gz

{lobes_cli-0.78.4 → lobes_cli-0.80.0}/.agex/data/pr/events.jsonl

@@ -202,3 +202,14 @@
 {"ts":"2026-09-13T08:37:41.066839+00:00","type":"pr_batch_replied","pr":261,"count":3,"resolved":3}
 {"ts":"2026-09-13T08:37:41.819765+00:00","type":"pr_webhook_posted","pr":261,"event":"pr_replied"}
 {"ts":"2026-09-13T08:52:12.707485+00:00","type":"pr_batch_replied","pr":261,"count":0,"resolved":0}
+{"ts":"2026-09-16T11:23:31.383904+00:00","type":"pr_opened","pr":270,"title":"innereye: host ComfyUI as the eleventh Colleague role (#268)"}
+{"ts":"2026-09-16T11:23:32.802979+00:00","type":"pr_review_triggered","pr":270,"command":"/agentic_review"}
+{"ts":"2026-09-16T11:23:33.578369+00:00","type":"pr_webhook_posted","pr":270,"event":"pr_opened"}
+{"ts":"2026-09-16T11:24:36.967038+00:00","type":"readiness_arrived","pr":270,"waited_secs":60}
+{"ts":"2026-09-16T11:24:41.403788+00:00","type":"pr_read","pr":270,"comment_count":2,"threads_unresolved":0,"ci_state":"ok"}
+{"ts":"2026-09-16T12:44:50.051042+00:00","type":"pr_reply","pr":270,"thread_id":null,"in_reply_to":null}
+{"ts":"2026-09-16T12:44:51.569047+00:00","type":"pr_reply","pr":270,"thread_id":null,"in_reply_to":null}
+{"ts":"2026-09-16T12:44:53.179673+00:00","type":"pr_reply","pr":270,"thread_id":null,"in_reply_to":null}
+{"ts":"2026-09-16T12:44:54.522120+00:00","type":"pr_reply","pr":270,"thread_id":null,"in_reply_to":null}
+{"ts":"2026-09-16T12:44:54.534943+00:00","type":"pr_batch_replied","pr":270,"count":4,"resolved":0}
+{"ts":"2026-09-16T12:44:55.310544+00:00","type":"pr_webhook_posted","pr":270,"event":"pr_replied"}

lobes_cli-0.80.0/.devague/current

@@ -0,0 +1 @@
+innereye-lobes-hosts-comfyui

lobes_cli-0.80.0/.devague/current_plan

@@ -0,0 +1 @@
+innereye-lobes-hosts-comfyui

lobes_cli-0.80.0/.devague/deliveries/innereye-lobes-hosts-comfyui.json

@@ -0,0 +1,386 @@
+{
+  "plan_slug": "innereye-lobes-hosts-comfyui",
+  "schema_version": 2,
+  "created": "2026-09-16T09:35:22Z",
+  "updated": "2026-09-16T11:53:36Z",
+  "deviations": [
+    {
+      "id": "d1",
+      "what": "The plan assumed ONE role registry; there are TWO. lobes.roles.ROLES (the Colleague-facing registry, t5's file) and lobes.profiles.schema.ROLES (the profile-machinery registry, 8 names) are separate constants, and schema.py:317 validates [[exclusive_roles]] against the SECOND one. t5 landing therefore does NOT unblock t1, contrary to the sequencing this run briefed. A second blocker sits behind it: shape_render.ROLE_SERVICE, which compose_profile iterates unconditionally whether or not a role is hosted. Both were verified empirically (a temporary local patch to schema.ROLES, reverted) and independently re-checked. Resolution: carve t6a (role vocabulary for the profile machinery: schema.ROLES + ROLE_SERVICE + an audit of KNOB_LANE_ROLES/SPECULATIVE_CONFIG_ROLES) and reorder t1 behind it. t1's committed work is blocked, not wrong",
+      "task_ref": "t1",
+      "reason": "The brief I gave the t1 agent named the wrong registry. The task agent caught it, tested the correction rather than asserting it, and reported it as a reasoning-degradation finding \u2014 which is the mechanism working",
+      "affects": [
+        "t5",
+        "t6",
+        "t7"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "classification": "needs-follow-up",
+      "seq": 1
+    },
+    {
+      "id": "d2",
+      "what": "Registering a role is ATOMICALLY CROSS-CUTTING and cannot be decomposed into file-disjoint tasks. t5 landed correctly within its file boundary and left 9 failures \u2014 all of them the repo's OWN invariant tests asserting a role is registered everywhere or nowhere (test_role_service_map_covers_exactly_the_ten_roles, test_every_proxyable_role_resolves_a_served_name, test_every_config_env_key_reaches_the_gateway_container, test_every_up_target_is_isolated_with_no_deps[innereye]). t14 has the identical problem: adding innereye to schema.ROLES churns the ~41 shape goldens and cannot merge green alone. Resolution: collapse t5 + t14 + t7 and the stub entries currently scoped to t3/t10 into ONE atomic registration merge, declaration-only (stub values, feasible:false, no behaviour), completed on the existing agent/innereye-t5 branch. The behavioural tasks are unchanged: t6 budget/veto, t9 facade, t10 real lifecycle, t2/t3 container, t13 acceptance. Cost accepted: one wide task instead of four narrow ones, less parallelism \u2014 the alternative is a knowingly-red main branch, which the TDD gate forbids",
+      "task_ref": "t5",
+      "reason": "The decomposition was mine and it fought an invariant the codebase already enforces. Third surfacing of the issue #269 root cause: adding a role should fail loudly and completely, or not at all",
+      "affects": [
+        "t7",
+        "t14",
+        "t3",
+        "t10",
+        "t1"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "classification": "needs-follow-up",
+      "seq": 2
+    },
+    {
+      "id": "d3",
+      "what": "The gateway's readiness probe has no owner for the render lane. gateway/_readiness.py hardcodes _HEALTH_PATH='/health' (line 96) and ComfyUI 0.33.2 does NOT serve it \u2014 MEASURED live on the Spark 2026-09-16: /health -> 404, while / -> 200, /system_stats -> 200, /object_info -> 200. So a wired innereye would advertise feasible:true, ready:false forever, directly contradicting success signal c25 ('GET /capabilities reports innereye feasible and ready'). Frame claim c8 anticipated a bespoke probe but its instruction scoped the answer to a COMPOSE healthcheck and explicitly told the implementer NOT to parameterize the shared /health literal \u2014 that covers docker-level health, not the gateway-level ready bit that reaches /capabilities. Resolution: t11 (capabilities and mesh advert) gains the gateway-side readiness probe path for the render lane",
+      "task_ref": "t11",
+      "reason": "Found by the t9 agent, which flagged it as 'likely' rather than asserting it, and deliberately wrote no test encoding the guess. I measured it live against the running ComfyUI and it is true",
+      "affects": [
+        "t9",
+        "t13",
+        "t10"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "classification": "needs-follow-up",
+      "seq": 3
+    }
+  ],
+  "evidence": [
+    {
+      "id": "e1",
+      "obligation_ref": "o1",
+      "test_ref": "tests/test_init_coresidency.py (67 tests) + docs/evidence/2026-09-16-accept-innereye-spark.txt \u00a71",
+      "behavior_text": "lobes init refuses a resolved default shape hosting both cortex and innereye, naming the reason and spark-innereye as the way out",
+      "contract_text": "v1 declares innereye and the heavy generate lanes mutually exclusive via [[exclusive_roles]]",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "suite green at the named commit; the CLI-refusal test monkeypatches resolve_shape at init.py's own call site rather than editing the shared machine-as-brain.toml",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 4
+    },
+    {
+      "id": "e2",
+      "obligation_ref": "o2",
+      "test_ref": "docs/evidence/2026-09-16-t2-comfyui-container-spark.txt + tests/test_comfyui_dockerfile.py",
+      "behavior_text": "the image pins ComfyUI v0.33.2 and stock cu130 torch, renders a FLUX job, and records the aimdo figures beside the bare-venv baseline",
+      "contract_text": "the built image renders at least one FLUX job, proving the containerized stack matches the bare venv",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "LIVE: built, ran with the GPU attached, staged 9318/22700/159 MB and executed in 41.84s against the baseline's 42.05s",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 5
+    },
+    {
+      "id": "e3",
+      "obligation_ref": "o3",
+      "test_ref": "docs/evidence/2026-09-16-accept-innereye-spark.txt \u00a72 + tests/test_comfyui_compose.py",
+      "behavior_text": "the service publishes no host port; 8188 is refused from the host AND from the Thor over the tailnet while the gateway reaches it on the compose net",
+      "contract_text": "ComfyUI keeps binding loopback and lobes terminates auth in front of it \u2014 restated as: publishes no host port",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "LIVE negative controls, both re-run after a CONFOUND (the bare venv was still listening on 8188) was found via ss and removed; positive control in-container /object_info 200",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 6
+    },
+    {
+      "id": "e4",
+      "obligation_ref": "o4",
+      "test_ref": "docs/evidence/2026-09-16-accept-innereye-spark.txt \u00a71,\u00a75 + tests/test_comfyui_volumes.py",
+      "behavior_text": "models mounted read-only from an env knob, output writable, user 1000:1000; artifacts host-owned and host-deletable, models mount rejects an in-container write",
+      "contract_text": "the model directory is a HOST BIND MOUNT and the host keeps read/write access",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "LIVE through the compose service: Config.User 1000:1000, in-container write to models refused 'Read-only file system', rendered artifact on disk owned spark:spark",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 7
+    },
+    {
+      "id": "e5",
+      "obligation_ref": "o5",
+      "test_ref": "the full suite (5151 passed) incl. tests/test_roles.py + docs/evidence/2026-09-16-accept-innereye-spark.txt \u00a74",
+      "behavior_text": "innereye is registered in every role-keyed table; the repo's own all-or-nothing invariants pass, and capabilities reports the role on both surfaces",
+      "contract_text": "innereye is registered everywhere or nowhere; an unhosted innereye is honestly refused, never half-served",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "the invariant tests named in the task (test_role_service_map, test_every_proxyable_role_resolves_a_served_name, test_every_config_env_key_reaches_the_gateway_container, test_up_no_deps_isolation[innereye]) all pass; lobes capabilities --json agrees with the wire field-for-field",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 8
+    },
+    {
+      "id": "e6",
+      "obligation_ref": "o6",
+      "test_ref": "tests/test_init_coresidency.py::test_declared_peak_gib_is_never_summed_or_compared_in_shape_render",
+      "behavior_text": "the declared peak's only dereference outside schema.py is a presence check in overcommitted_groups; it is never summed nor compared against a card total",
+      "contract_text": "the peak's ONLY consumer is the exclusive_roles co-residency veto; it meters nothing",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "a source-scanning test rejects '+', 'sum(', 'total' and 'gpu_mem_util' on any line mentioning the knob, and asserts the two specific statements expected",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 9
+    },
+    {
+      "id": "e7",
+      "obligation_ref": "o7",
+      "test_ref": "tests/test_gateway_get_relay.py (6 tests)",
+      "behavior_text": "open_upstream carries a method without changing POST callers; _relay_streaming is reachable from do_GET and relays binary byte-for-byte",
+      "contract_text": "a method-general upstream opener and a GET-side streaming relay",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "byte-for-byte proven against a stdlib stub upstream (256 KiB non-UTF-8 blob, sha256 equality); the real-ComfyUI half was proven separately live in the t13 transcript \u00a75",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 10
+    },
+    {
+      "id": "e8",
+      "obligation_ref": "o8",
+      "test_ref": "tests/test_gateway_render_facade.py (41 tests) + docs/evidence/2026-09-16-accept-innereye-spark.txt \u00a73,\u00a75,\u00a76,\u00a77",
+      "behavior_text": "the facade is job-scoped: a foreign job id and a counter-named artifact of another job are both refused without dialing ComfyUI; no outward /history, /queue or parameterised /view exists; unauthenticated is 401",
+      "contract_text": "the gateway issues the job id and serves only ids it issued; enumeration is refused",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "LIVE: 401 on POST and GET with no bearer (and from the Thor); with bearer an unissued id 404s render_job_not_found; job A asking for job B's real artifact 404s; artifact bytes byte-identical to disk",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 11
+    },
+    {
+      "id": "e9",
+      "obligation_ref": "o9",
+      "test_ref": "tests/test_gateway_no_lifecycle_actuator.py + tests/test_cli_up.py + docs/evidence/2026-09-16-accept-innereye-spark.txt \u00a71",
+      "behavior_text": "lobes up innereye starts exactly comfyui with --no-deps and --down stops only it; a cold backend yields 503 + Retry-After naming the manual remedy; no _compose import exists under lobes/gateway/",
+      "contract_text": "manual lifecycle only; the data plane never starts or stops anything",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "LIVE dry-run then apply; the AST-based import guard has a control test proving the scanner still finds real hits elsewhere, and its blind spots (a string-built importlib call) are documented not implied",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 12
+    },
+    {
+      "id": "e10",
+      "obligation_ref": "o10",
+      "test_ref": "tests/test_gateway_readiness.py + tests/test_mesh_routes.py + docs/evidence/2026-09-16-accept-innereye-spark.txt \u00a74,\u00a77",
+      "behavior_text": "a peer discovers the lane from GET /capabilities alone; the lane never enters a mesh announcement; the render lane's readiness probes /object_info so a wired-and-up innereye reports ready:true",
+      "contract_text": "discovery without advertised-but-unreachable reach",
+      "evidence_type": "automated",
+      "strength": "execution",
+      "strength_basis": "LIVE: capabilities reported feasible/ready/loaded true on the hosting box and was fetched from the Thor over the tailnet; mesh non-advert is asserted at BOTH announcement entry points after the first fix was found to be dead code for the live heartbeat path",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 13
+    },
+    {
+      "id": "e11",
+      "obligation_ref": "o11",
+      "test_ref": "docs/comfyui-innereye.md + docs/colleague-stack.md (markdownlint clean)",
+      "behavior_text": "the doc states the two exposures side by side, names the three non-goals, assigns output growth to the operator, states the rollback, and labels every lobes-side surface DECLARED/UNVALIDATED",
+      "contract_text": "the reference doc and the catalog statement",
+      "evidence_type": "manual",
+      "strength": "coverage",
+      "strength_basis": "READ, not executed: this is prose. A dedicated deployments/ catalog entry was deliberately NOT fabricated since no box's lock has been captured for this shape; the required not-a-renderable-deployment statement lives in the doc instead",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 14
+    },
+    {
+      "id": "e12",
+      "obligation_ref": "o12",
+      "test_ref": "docs/evidence/2026-09-16-accept-innereye-spark.txt (all nine sections)",
+      "behavior_text": "all four success signals pass with negative controls, the overlapping-submission measurement is recorded, and the cross-box fetch from the Thor is byte-identical",
+      "contract_text": "one transcript carries all four success signals with negative controls and an honest statement of the innereye dependency",
+      "evidence_type": "manual",
+      "strength": "execution",
+      "strength_basis": "LIVE on the DGX Spark with the Jetson AGX Thor as a real cross-box peer. TWO honesty caveats are IN the transcript: the gateway under test was merged SOURCE not the packaged image (0.79.0 unpublished), and innereye-the-client never drove the facade \u2014 every request was curl",
+      "outcome": "pass",
+      "run": {
+        "timestamp": "2026-09-16T14:18:01",
+        "commit": "8d04f09"
+      },
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false,
+      "seq": 15
+    }
+  ],
+  "deltas": [
+    {
+      "id": "b1",
+      "kind": "amended",
+      "behavior_text": "The isolation property is 'the service publishes no host port', NOT 'ComfyUI binds 127.0.0.1'. A container bound to loopback is unreachable from the gateway; every fleet lane binds 0.0.0.0 in-container and isolates via expose-only. The announcement, after-state and success-signal claims were all amended to say so.",
+      "caused_by": [
+        "c28"
+      ],
+      "evidence_refs": [
+        "e3"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false
+    },
+    {
+      "id": "b2",
+      "kind": "added",
+      "behavior_text": "A per-backend readiness probe path. gateway/_readiness.py hardcoded /health, which ComfyUI 404s, so a wired innereye would have advertised feasible:true, ready:false forever \u2014 contradicting the plan's own success signal. probe_backend_ready now takes a health_path and the render lane probes /object_info.",
+      "caused_by": [
+        "d3"
+      ],
+      "evidence_refs": [
+        "e10"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false
+    },
+    {
+      "id": "b3",
+      "kind": "added",
+      "behavior_text": "declared_peak_gib \u2014 the first non-fractional budget knob in the codebase, lane-gated to innereye alone. Every prior role budget is a gpu_mem_util fraction; ComfyUI has no equivalent flag, so a fake fraction would have misrepresented a per-graph footprint as a fixed server budget.",
+      "caused_by": [
+        "c17"
+      ],
+      "evidence_refs": [
+        "e6"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false
+    },
+    {
+      "id": "b4",
+      "kind": "added",
+      "behavior_text": "user: on a fleet compose service \u2014 comfyui is the first. Without it every artifact would land root-owned in a bind-mounted directory the host user could not clear, silently breaking the guarantee that the weights and outputs stay reachable outside Docker.",
+      "caused_by": [
+        "c30"
+      ],
+      "evidence_refs": [
+        "e4"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false
+    },
+    {
+      "id": "b5",
+      "kind": "amended",
+      "behavior_text": "Registering a role is ATOMICALLY cross-cutting: t5 was re-scoped to absorb the retired t14 and t7 because the repo's own invariant tests assert a role is registered everywhere or nowhere. The planned file-disjoint decomposition was not achievable for this change class.",
+      "caused_by": [
+        "d2"
+      ],
+      "evidence_refs": [
+        "e5"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false
+    },
+    {
+      "id": "b6",
+      "kind": "removed",
+      "behavior_text": "No websocket proxy was built. The plan originally owed a raw-socket WS tunnel; innereye never opens a websocket (its stdlib-only rule makes /ws impractical), so the most expensive piece of gateway work was dropped and named a non-goal instead.",
+      "caused_by": [
+        "c11"
+      ],
+      "evidence_refs": [
+        "e8"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false
+    },
+    {
+      "id": "b7",
+      "kind": "added",
+      "behavior_text": "The lobes up remediation for an opt-in core role now resolves the real hosting shape from each built-in shape's declared hosts, instead of a hardcoded thor-{role} guess that was ALREADY wrong for associate (orin-associate) before innereye existed.",
+      "caused_by": [
+        "c16"
+      ],
+      "evidence_refs": [
+        "e9"
+      ],
+      "origin": "llm",
+      "status": "approved",
+      "superseded": false
+    }
+  ],
+  "supersessions": []
+}