spec+plan: mesh-brain end-state — one heavy lobe per box, cheap gears co-reside (#112) (0.41.2)

[OriNachum]

Spec + plan for the mesh-brain end-state (#112), worked through devague /think → /spec-to-plan on the merged #113 substrate.

What this PR carries

• Spec — docs/specs/2026-07-14-lobes-serves-the-mesh-brain-end-state-one-lobe-per.md (converged frame committed alongside). Four user decisions recorded first-class:
  1. The shape axis is backward compatible and mixable — many specialized machines (some cloud if you want), some multi-role machines, or any mix; machine-as-brain stays the default; one-lobe shapes are the far end of the axis, not a mandate.
  2. Cross-box = direct + honest referral — no data-plane proxying; consumers dial each box directly; with opt-in peer config, capabilities and role_infeasible 404s name the peer that hosts an absent role (Advertised per-role endpoint (:8000) 404s while ready=true — clients dialing it degrade instantly (lobes-cli#87 shape regressed) #92 invariant holds).
  3. Cheap gears co-reside everywhere — "one lobe per box" specializes the heavy generate lobes only; embedder/reranker/stt/tts stay on every box that wants them; consumers keep localhost embed/rerank.
  4. The Qwen-only Spark stays a fleet shape — gateway + one heavy vLLM; "return to 256K-native solo" means 262144 as a shape override, not the legacy scaffold.
• Plan — docs/plans/2026-07-14-lobes-serves-the-mesh-brain-end-state-one-lobe-per.md: 6 tasks in 4 waves — [t1 full-native budgets · t2 orin-small shape · t3 honest referral] → t4 contract-test matrix → t5 physical GB10+Thor acceptance → t6 docs.
• Follow-up filed — proxy-lobes: let a box serve a sleeping lobe by proxying to the peer that hosts it (deferred from #112) #115 proxy-lobes: a box may later serve a sleeping lobe by following its own referral; parked in the frame and cited in the spec.
• Housekeeping — agex PR bookkeeping (Gateway relays SSE frame-by-frame — read1, not read (closes #103) (0.40.1) #104/Deployment shapes: machine-as-brain vs mesh-brain lobe profiles — spec + plan (0.41.1) #113 events) and eidetic memory records (in-repo public store).

Key consequence made explicit

Under the co-residence decision, the end-state shapes have the same role sets as the near-term spark-lobe/thor-lobe — #112 concretely reduces to full-native budgets (cortex 262144 = 2×, senses 131072 = 4×), the declared-but-unvalidated orin-small shape (#108 rule), and the referral layer.

Sequencing

This is a spec+plan-only PR (no runtime source changes). Implementation sequences after the #113 plan lands — recorded as a stated assumption and a plan risk.

• Version: 0.41.1 → 0.41.2 (patch; docs/spec/plan only)

• Tests: 1419 passed, 14 skipped (live-gateway smokes)

• lobes (Claude)

[OriNachum]

/agentic_review

[qodo-code-review]

Code Review by Qodo

New Review Started

This review has been superseded by a new analysis

[qodo-code-review]

PR Summary by Qodo

Spec+plan: mesh-brain end-state shapes (one heavy lobe per box)

📝 Documentation ⚙️ Configuration changes 🕐 20-40 Minutes

AI Description

• Add spec+plan for one-heavy-lobe-per-box mesh-brain end-state and key decisions.
• Export devague frame/plan artifacts and update devague pointers to the new slug.
• Bump version to 0.41.2 and document the release in the changelog.

Diagram

graph TD
  DOCS[("Spec/Plan docs")] --> OP["Operator"]
  OP --> PC["Peer config"] --> GW["Gateway"] --> AG["Consumer"]
  OP --> SP["Spark: cortex"] --> GW
  OP --> TH["Thor: senses"] --> GW
  OP --> OR["Orin: small-model"] --> GW

Loading

High-Level Assessment

The following are alternative approaches to this PR:

1. Split housekeeping vs spec/plan into separate PRs

• ➕ Keeps review focused on the spec/plan content and decisions.
• ➕ Reduces noise from agex/eidetic/devague bookkeeping changes.
• ➖ More PR overhead and coordination.
• ➖ Housekeeping artifacts are often generated alongside spec work anyway.

2. Publish only rendered Markdown (omit devague exported JSON)

• ➕ Less repository churn and smaller diffs.
• ➕ Avoids coupling reviewers to devague schema/version details.
• ➖ Loses machine-readable source-of-truth needed for later tooling or regeneration.
• ➖ Harder to audit exact decision/task structure over time.

3. Keep version unchanged for docs-only changes

• ➕ Avoids churn in release/version tracking for non-runtime changes.
• ➖ If your release process publishes docs/specs as artifacts, versions still matter.
• ➖ May make it harder to reference “which spec set shipped” in downstream ops.

Recommendation: The PR’s approach (shipping both Markdown and the exported devague frame/plan) is reasonable if the repo treats specs/plans as first-class, machine-readable artifacts. If review bandwidth is tight, consider splitting housekeeping/memory/bookkeeping into a follow-up PR next time; for this change, the combined commit is still low-risk since there are no runtime code paths affected.

Files changed (11) +880 / -11

Documentation (8) +877 / -8

events.jsonlRecord PR lifecycle events for #104 and #113 +12/-0

Record PR lifecycle events for #104 and #113

• Adds new agex event records covering PR opened/webhook/readiness/read states for #104 and #113. No runtime behavior changes; this is review/ops bookkeeping.

.agex/data/pr/events.jsonl

lobes-serves-the-mesh-brain-end-state-one-lobe-per.jsonAdd exported devague frame for #112 one-lobe-per-box end-state +389/-0

Add exported devague frame for #112 one-lobe-per-box end-state

• Introduces the full devague frame JSON capturing the end-state claims, decisions, honesty conditions, success signals, and open vagueness items. Encodes the four key user decisions (mixable axis, direct+referral, cheap-gear co-residence, Spark stays fleet shape) and references follow-up #115.

.devague/frames/lobes-serves-the-mesh-brain-end-state-one-lobe-per.json

lobes-serves-the-mesh-brain-end-state-one-lobe-per.jsonAdd exported devague plan JSON with 6 tasks in 4 waves +310/-0

Add exported devague plan JSON with 6 tasks in 4 waves

• Adds the exported plan JSON enumerating tasks t1–t6, acceptance criteria, dependencies, and risks. Establishes sequencing over the merged #113 substrate (assumption that near-term shape schema lands first).

.devague/plans/lobes-serves-the-mesh-brain-end-state-one-lobe-per.json

colleague__public.jsonlAdd public colleague work-lesson memory entries +4/-0

Add public colleague work-lesson memory entries

• Adds several public memory records summarizing prior work lessons (test-first emphasis and scoped diffs). This is archival/supporting material rather than product behavior.

.eidetic/memory/colleague__public.jsonl

lobes__public.jsonlRefresh public lobes memory records and add #112 spec entry +10/-8

Refresh public lobes memory records and add #112 spec entry

• Updates last_recall/recall_count metadata for multiple existing entries and appends new records, including the explicit memory entry for issue #112 being specced. No functional code changes; this is knowledge-base maintenance.

.eidetic/memory/lobes__public.jsonl

CHANGELOG.mdDocument 0.41.2 as spec+plan-only release for #112 +6/-0

Document 0.41.2 as spec+plan-only release for #112

• Adds a 0.41.2 changelog entry describing the new mesh-brain end-state spec+plan, including full-native budgets, orin-small declared shape, and direct+honest referral decision. Clarifies that cheap gears remain co-resident.

CHANGELOG.md

2026-07-14-lobes-serves-the-mesh-brain-end-state-one-lobe-per.mdAdd build plan Markdown for #112 (6 tasks, risks, acceptance criteria) +60/-0

Add build plan Markdown for #112 (6 tasks, risks, acceptance criteria)

• Adds the rendered plan describing six implementation tasks across waves: full-native budgets as shape data, orin-small declared shape, honest referral (no proxy), contract test matrix, physical acceptance on GB10/Thor, and final docs updates. Includes explicit risks about sequencing after #113 implementation and peer-config surface constraints.

docs/plans/2026-07-14-lobes-serves-the-mesh-brain-end-state-one-lobe-per.md

2026-07-14-lobes-serves-the-mesh-brain-end-state-one-lobe-per.mdAdd end-state spec Markdown for one heavy lobe per box +86/-0

Add end-state spec Markdown for one heavy lobe per box

• Adds the rendered spec capturing audience, before/after, requirements, decisions, honesty conditions, and success signals for the mesh-brain end-state. Records the four first-class decisions and explicitly constrains cross-box behavior to direct addressing with honest referral metadata (no data-plane proxying).

docs/specs/2026-07-14-lobes-serves-the-mesh-brain-end-state-one-lobe-per.md

Other (3) +3 / -3

currentPoint devague current frame slug to #112 spec +1/-1

Point devague current frame slug to #112 spec

• Updates the devague pointer to the new frame slug for the mesh-brain end-state spec. This steers devague CLI operations to the new spec context.

.devague/current

current_planPoint devague current plan slug to #112 plan +1/-1

Point devague current plan slug to #112 plan

• Updates the devague pointer to the new plan slug matching the #112 end-state work. Keeps spec-to-plan workflow aligned with the exported plan.

.devague/current_plan

pyproject.tomlBump project version to 0.41.2 +1/-1

Bump project version to 0.41.2

• Updates the package version from 0.41.1 to 0.41.2 to reflect the docs/spec/plan release. No dependency or runtime code modifications are included.

pyproject.toml

[qodo-code-review]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0)

Great, no issues found!

Qodo reviewed your code and found no material issues that require review

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud