rickylabs · rickylabs · Jul 17, 2026
@@ -0,0 +1,35 @@
+# Session record — RFC single deployment (issue #820) orchestrator run
+
+Date 2026-07-17 · run `.llm/runs/rfc-single-deployment--orchestrator/` · generator Claude Fable 5
+· high (session `7f1fada7-805f-46cb-8ac4-5eb201bdc105`) · evaluator Codex GPT-5.6 Sol · max
+(9 separate daemon threads).
+
+## What happened
+
+Seed/RFC harness run on the #820 charter: eis-chat POC forensics (@ `aeaf2df`, via authenticated
+GitHub API after the local clone was sandbox-blocked), gap analysis G1–G8 against the PM/deploy
+board, and nine revisions of the RFC design under nine adversarial Sol·max PLAN-EVAL cycles. The
+owner authorized loop continuation after the standard two-failure escalation, then bounded the
+loop at cycle 9. Final: **FAIL_PLAN with 6/8 plan-gate boxes PASS** (incl. Decisions-locked);
+run closed as a design record (`closure.md`). No #820 comment (PASS-gated), zero board mutations
+(evaluator-audited each cycle).
+
+## Lessons worth keeping
+
+- **The adversarial loop converges but is expensive**: each Sol·max cycle re-audited prior fixes
+  for *real closure* and expanded into genuinely deeper systems territory (junction atomicity →
+  journal durability → privilege separation → boot-order composability). Design-heavy RFCs
+  should budget 3+ cycles or pre-decompose into evaluable chunks.
+- **Artifact currentness is a first-class gate**: three cycles flagged stale
+  worklog/context-pack/supervisor metadata as blocking. Reconcile resumable artifacts in the same
+  edit batch as the plan rev, every time.
+- **Self-containment rule**: never write "unchanged from rev N" after overwriting rev N — the
+  candidate must stand alone (cycle-3 regression).
+- **Launcher fallback recipe**: `agentic:launch-codex-slice` is Windows-host/implementation-slice
+  shaped; from WSL-native sessions, the wrapper's inner client
+  (`.llm/tools/agentic/codex/app-server-message-cli.ts --model gpt-5.6-sol --effort max --cwd …
+  --message "$(brief)"`) launches the same daemon-attached thread (drift entry 4 pattern;
+  `.llm/tmp/rfc820/launch-eval.ts`).
+- **Evidence-honesty pays**: the evaluator repeatedly accepted designs only when grounded in
+  tree/corpus citations (e.g. rejecting Servy tree-kill and the runtime `AspireResource` as
+  unsupported) — write plans from the code, not from plausibility.
@@ -0,0 +1,68 @@
+# Closure — RFC single deployment (issue #820) — design-record close per owner wrap-up directive
+
+> **Post-closure amendment (2026-07-17, same day):** the owner re-opened the run for exactly the
+> residual revision ("proceed with the revision I'll take care of the final eval myself").
+> **plan.md rev 10** folds the surface classifications (residual 1 below) into §I.2/§E.2/§H; the
+> cycle-10 eval is **owner-launched** (recipe in `context-pack.md` § "How the owner runs cycle
+> 10"). The generator launches no further cycles; the #820 post + `drafts/` remain PASS-gated.
+> This document otherwise stands as the cycle-1..9 record.
+
+Run `rfc-single-deployment--orchestrator` · closed 2026-07-17 · generator Fable 5 · high, session
+`7f1fada7-805f-46cb-8ac4-5eb201bdc105`. Owner directives: loop continuation authorized after the
+cycle-2 escalation; loop bounded at cycle 9 ("one more pass then done").
+
+## Outcome
+
+**9 adversarial PLAN-EVAL cycles (Sol·max, all separate sessions), monotonically converging; the
+final verdict is FAIL_PLAN with 6 of 8 plan-gate boxes PASS** — including, for the first time,
+**"Decisions locked"**, plus board adjustments, risk register, gate set, deferred scope, and the
+research content itself. Per the kickoff's hard gate, the RFC was **NOT posted to #820** and the
+`drafts/` were **not written** (both were PASS-gated). **Zero board mutations occurred at any
+point** — evaluator-audited in every cycle.
+
+## What stands (evaluator-confirmed sound across the final cycles)
+
+- **PM-first sequencing** (§A): the Tier-4 split keeping the PM-independent single-runtime lane
+  (#451/#453/#454/#455 + #452/#456a/#457a) complete in beta.11 while singleton-graph work
+  consumes the beta.12 PM engine in beta.13 — repeatedly endorsed as "sound, avoids
+  over-blocking".
+- **Aspire-stack citizenship** (§B.4): generator-emitted `PackagingModel` → pure manifest
+  compiler → the canonical `deploy build` verb + a named TS-AppHost publish pipeline step, all in
+  SD-2; PM-20 stays pure extraction.
+- **The installer** (§B): `DeployTargetPort` adapters + `MaintenancePort` (no rival axis), the
+  §B.1a operation state machines (incl. deterministic four-state purge + §B.1b canonical
+  machine-state root), the §B.3 identity/privilege matrix with updater least-privilege grants,
+  and the §B.2a install-graph digest rule.
+- **The update lifecycle** (§C): one journaled release-snapshot mechanism from beta.11 (real
+  Windows apply; `Deno.autoUpdate` never the authority), crash-recoverable journal with migration
+  barriers, `maintenance(rollback-failed)` terminal, replay high-water, pinned trust, and the
+  §C.3b three-phase ownership that resolves boot-recovery composition.
+- **Composition** (§D): both modes kept; typed manifests + the SD-7 cross-mode conformance suite
+  as the anti-fork enforcement; transport seam scoped to transport only.
+- **The draft board** (§E): 22 bounded, dependency-ordered drafts; G1–G8 exactly-one ownership;
+  owner forks OF-A..OF-K.
+
+## Residuals (the cycle-9 items — small)
+
+1. **Public-vs-internal classification for three surfaces** (the only real design residual):
+   PM-5's `RuntimeCommandSpec` additions (`clearEnv` + strip list), PM-15's renderer knobs, and
+   SD-1's host-side supervisor/broker/proxy/Job-Object types. *Generator recommendation (non-
+   normative, unevaluated):* PM-5 additions = public contract extension (rubric at PM-5); PM-15
+   knobs = internal render config, classification re-decided at the PM-20 move; SD-1 host types =
+   internal (only #451/SD-6 client surfaces are public).
+2. **Resume-artifact metadata lag** — reconciled as part of this closure (context-pack,
+   supervisor, worklog, drift now agree on the final state).
+
+## Path to done (for a future session, if the owner wants the RFC posted)
+
+One small rev folding residual 1 into §I.2/§E.2 + one PLAN-EVAL cycle would plausibly PASS; then
+execute kickoff deliverable 5 (the #820 comment + `drafts/` files) and stage-H ratification of
+OF-A..OF-K. Resume via `context-pack.md`; the eval launch recipe is `.llm/tmp/rfc820/
+launch-eval.ts` + `plan-eval-brief.md` (route: Sol·max through the agentic app-server client).
+
+## Record map
+
+`plan.md` rev 9 (the RFC) · `research.md` (forensics + gaps + re-baseline + debt) ·
+`plan-eval-cycle{1..9}.md` (verdict trail) · `escalation.md` (cycle-2 loop-limit stop) ·
+`supervisor.md` / `worklog.md` / `context-pack.md` / `drift.md` (6 entries) · `corpus/`
+(evidence) · session record `.llm/2026-07-17-rfc-single-deployment-run.md`.
@@ -0,0 +1,76 @@
+# Context Pack: RFC single deployment (issue #820) — rev 10 ready; final eval OWNER-RUN
+
+## Run Metadata
+
+| Field          | Value                                        |
+| -------------- | -------------------------------------------- |
+| Run ID         | `rfc-single-deployment--orchestrator`        |
+| Branch         | `feat/beta10-cli-integration` @ `f391190f` (re-baselined vs `origin/main` @ `47cc2fa9`) |
+| Current phase  | **rev 10 staged; cycle-10 PLAN-EVAL pending, launched BY THE OWNER** (generator launches no further cycles per the 2026-07-17 owner instruction) |
+| Archetype      | N/A — seed/RFC run (downstream drafts span Archetypes 2/3/4/5/6/7 + all three overlays) |
+| Session        | generator `7f1fada7-805f-46cb-8ac4-5eb201bdc105` (Fable 5 · high) |
+| Evaluator route | Codex · GPT-5.6 Sol · max via the agentic app-server client (drift entry 4); threads c1 `019f6fa1…` c2 `019f6fb5…` c3 `019f6fd5…` c4 `019f6feb…` c5 `019f7006…` c6 `019f701c…` c7 `019f7034…` c8 `019f7052…` c9 `019f7078…` |
+
+## Current State (history in `closure.md`; this file is post-re-open)
+
+plan.md **rev 10** is the staged candidate: after the cycle-9 verdict (FAIL_PLAN with **6/8
+boxes PASS**, incl. Decisions-locked) and the design-record closure, the owner re-opened with
+"proceed with the revision, I'll take care of the final eval myself". Rev 10 folds cycle 9's
+only design residual — the surface classifications (PM-5 additions PUBLIC; PM-15 knobs INTERNAL
+@ beta.12, re-decided at PM-20; SD-1 host surface INTERNAL, public = #451/SD-6) — into §I.2/§E.2
+with §H dispositions. The RFC remains **unposted** (PASS-gated) and **zero board mutations have
+ever occurred** (evaluator-audited cycles 1–9). Owner authorizations on record: loop
+continuation (drift 5), cycle-9 bound (drift 6), owner-run final eval (drift 7).
+
+## How the owner runs cycle 10
+
+From the repo root:
+
+```
+deno run --no-lock --allow-read --allow-run --allow-env .llm/tmp/rfc820/launch-eval.ts
+```
+
+It sends `plan-eval-brief.md` (already carrying the cycle-10 note) to a fresh Codex Sol·max
+thread via the agentic app-server client; the verdict overwrites `plan-eval.md` (cycles 1–9 are
+archived as `plan-eval-cycle{1..9}.md`). On PASS: execute kickoff deliverable 5 — the #820 RFC
+comment (public-repo hygiene: eis-chat PR/issue numbers OK, no internal process details) +
+`drafts/` files (NOT filed) — then stage-H ratification of OF-A..OF-K.
+
+## Key Decisions (rev 9 — evaluator-PASSed as locked)
+
+| Decision | Source |
+| --- | --- |
+| PM-first Tier-4 split; single-runtime lane complete in beta.11; graph mode behind PM in beta.13 | plan §A.1 |
+| PM bar PM-1..14 (+15/16/18 per-machine); PM-A/PM-B separate drafts; #543 stays beta.12 | plan §A.2/A.3 |
+| Universal containment: harness+guardian pipe-EOF (PM-B core) + Job Objects; per-machine = PM-15 `KillMode` + Windows Job-Object wrapper (OF-K); SD-H Linux backstop @ stable (OF-H) | plan §A.3 |
+| Installer = `DeployTargetPort` adapters + `MaintenancePort`; §B.1a op state machines; four-state deterministic purge; §B.1b canonical machine-state root | plan §B.1 |
+| Schemas CLI-internal through beta.12 → SD-2 moves+publishes; pinned trust key; sequence high-water + explicit epoch reset (OF-J); `minBootstrapVersion` | plan §B.2 |
+| Install-graph digest match-or-refuse (OF-I) | plan §B.2a |
+| Identity/privilege matrix; updater sole writer; per-user read tokens via OS-auth broker | plan §B.3 |
+| Per-user dynamic ports + `/_svc` proxy w/ per-launch token; per-machine port-reservation registry | plan §B.3a |
+| `PackagingModel` (CLI-internal) + build verb + named Aspire publish step, all SD-2 | plan §B.4 |
+| One journaled update mechanism incl. Windows apply (L0.7); stable bootstraps + journal-first resolution; §C.3b three-phase ownership + single confirmer | plan §C |
+| Both composition modes kept; manifests + SD-7 conformance suite enforce | plan §D |
+| Owner forks **OF-A..OF-K** resolve at stage-H filing | plan §F |
+
+## Gates
+
+| Gate | Status | Evidence |
+| --- | --- | --- |
+| PLAN-EVAL c1–c9 | FAIL_PLAN ×9; c9 = 6/8 boxes PASS | `plan-eval-cycle{1..9}.md` |
+| PLAN-EVAL c10 (gate of record) | **PENDING — owner-launched** on rev 10 | recipe above; verdict → `plan-eval.md` |
+
+## Open Questions
+
+- OF-A..OF-K (plan §F, 11 forks) — stage-H owner forks.
+- (cycle-9 residual 1 — surface classification — resolved in rev 10 §I.2/§E.2/§H.)
+
+## Drift and Debt
+
+- Drift: **7 entries** (effort raises; no dedicated PR; sandbox-blocked clone → API corpus;
+  evaluator via wrapper's inner client; owner-authorized loop continuation; owner wrap-up bound;
+  owner-run final eval). Debt: none created.
+
+## Commits
+
+- None — kickoff scoped this run to run-dir artifacts + a post-PASS #820 comment (never earned).
@@ -0,0 +1 @@
+[]
@@ -0,0 +1,90 @@
+{
+ "url": "https://api.github.com/repos/rickylabs/eis-chat/issues/135",
+ "repository_url": "https://api.github.com/repos/rickylabs/eis-chat",
+ "labels_url": "https://api.github.com/repos/rickylabs/eis-chat/issues/135/labels{/name}",
+ "comments_url": "https://api.github.com/repos/rickylabs/eis-chat/issues/135/comments",
+ "events_url": "https://api.github.com/repos/rickylabs/eis-chat/issues/135/events",
+ "html_url": "https://github.com/rickylabs/eis-chat/pull/135",
+ "id": 4802560256,
+ "node_id": "PR_kwDOTIdMCM7tiekE",
+ "number": 135,
+ "title": "fix(kb): shared Garnet queue fixes multi-process split-brain Deno KV (#133) + desktop perms (#118)",
+ "user": {
+  "login": "rickylabs",
+  "id": 129366361,
+  "node_id": "U_kgDOB7X5WQ",
+  "avatar_url": "https://avatars.githubusercontent.com/u/129366361?v=4",
+  "gravatar_id": "",
+  "url": "https://api.github.com/users/rickylabs",
+  "html_url": "https://github.com/rickylabs",
+  "followers_url": "https://api.github.com/users/rickylabs/followers",
+  "following_url": "https://api.github.com/users/rickylabs/following{/other_user}",
+  "gists_url": "https://api.github.com/users/rickylabs/gists{/gist_id}",
+  "starred_url": "https://api.github.com/users/rickylabs/starred{/owner}{/repo}",
+  "subscriptions_url": "https://api.github.com/users/rickylabs/subscriptions",
+  "organizations_url": "https://api.github.com/users/rickylabs/orgs",
+  "repos_url": "https://api.github.com/users/rickylabs/repos",
+  "events_url": "https://api.github.com/users/rickylabs/events{/privacy}",
+  "received_events_url": "https://api.github.com/users/rickylabs/received_events",
+  "type": "User",
+  "user_view_type": "public",
+  "site_admin": false
+ },
+ "labels": [],
+ "state": "closed",
+ "locked": false,
+ "assignees": [],
+ "milestone": null,
+ "comments": 0,
+ "created_at": "2026-07-03T10:08:50Z",
+ "updated_at": "2026-07-03T10:48:14Z",
+ "closed_at": "2026-07-03T10:43:13Z",
+ "assignee": null,
+ "author_association": "OWNER",
+ "active_lock_reason": null,
+ "draft": false,
+ "pull_request": {
+  "url": "https://api.github.com/repos/rickylabs/eis-chat/pulls/135",
+  "html_url": "https://github.com/rickylabs/eis-chat/pull/135",
+  "diff_url": "https://github.com/rickylabs/eis-chat/pull/135.diff",
+  "patch_url": "https://github.com/rickylabs/eis-chat/pull/135.patch",
+  "merged_at": "2026-07-03T10:43:13Z"
+ },
+ "body": "Closes the mechanism half of #133. Live-verified via the Aspire CLI/MCP.\n\n## Problem\nThe Aspire generator no-ops the `DenoKv` cache engine, so in **multi-process** mode workers-api (enqueue) and the workers processor (listen) each open their OWN per-process Deno KV → the `embed-document` job is enqueued in one KV and listened on another, so it never runs (KB doc pins `pending`). This is the real root cause of the stalled KB ingestion (deeper than the previously-suspected release blocker).\n\n## Fix (Docker-less Windows bare metal)\nShared Deno KV Connect (`denokv`) needs Docker, which this host lacks, so run **Garnet** (RESP-compatible, .NET-native) as an Aspire **executable** and point the workers at it. The SDK's `@netscript/kv/redis` + `@netscript/queue` `RedisAdapter` then give a real cross-process queue (`LPUSH`/`BRPOPLPUSH`).\n\n- `aspire/.helpers/register-infrastructure.mts` — `addExecutable('garnet','garnet-server', …)` (hand-edit, keep on regen).\n- `aspire/.helpers/register-plugins.mts` + `register-background.mts` — inject `GARNET_URI`/`REDIS_URI`/`CACHE_PROVIDER=garnet` + `waitFor(garnet)` into workers-api + the workers processor.\n- `workers/runtime.ts` — `import jsr:@netscript/kv/redis` to self-register the KV adapter before `getKv()`.\n- `apps/dashboard/deno.json` — `--allow-all` on `desktop:dev`/`desktop:build` so the deno-desktop runtime can read env (was crashing on `PORT`). (Partial #118; full desktop render is a follow-up PR.)\n\n## Verification (Aspire CLI/MCP only — shared stack never hacked)\ngarnet Running/Healthy; workers-api + workers both `WaitFor(garnet)` + carry `GARNET_URI`; a job triggered on **workers-api** was **executed by the separate workers processor** (cross-process delivery over Garnet). See #133 for logs.\n\n## Requires / upstream\n`dotnet tool install -g garnet-server`. Proper generator fix + self-provisioning: rickylabs/netscript#372 (Garnet executable) + #371 (Deno KV Connect).\n\n## Reconciliation\nOverlaps #134 (NetScript beta.1 bump) on `register-plugins.mts` (different hunks) and `deno.lock`, and this branch's `workers/runtime.ts` redis import is pinned to `alpha.19`. To be rebased onto master after #134 merges, with the redis import bumped to `beta.1`.",
+ "closed_by": {
+  "login": "rickylabs",
+  "id": 129366361,
+  "node_id": "U_kgDOB7X5WQ",
+  "avatar_url": "https://avatars.githubusercontent.com/u/129366361?v=4",
+  "gravatar_id": "",
+  "url": "https://api.github.com/users/rickylabs",
+  "html_url": "https://github.com/rickylabs",
+  "followers_url": "https://api.github.com/users/rickylabs/followers",
+  "following_url": "https://api.github.com/users/rickylabs/following{/other_user}",
+  "gists_url": "https://api.github.com/users/rickylabs/gists{/gist_id}",
+  "starred_url": "https://api.github.com/users/rickylabs/starred{/owner}{/repo}",
+  "subscriptions_url": "https://api.github.com/users/rickylabs/subscriptions",
+  "organizations_url": "https://api.github.com/users/rickylabs/orgs",
+  "repos_url": "https://api.github.com/users/rickylabs/repos",
+  "events_url": "https://api.github.com/users/rickylabs/events{/privacy}",
+  "received_events_url": "https://api.github.com/users/rickylabs/received_events",
+  "type": "User",
+  "user_view_type": "public",
+  "site_admin": false
+ },
+ "reactions": {
+  "url": "https://api.github.com/repos/rickylabs/eis-chat/issues/135/reactions",
+  "total_count": 0,
+  "+1": 0,
+  "-1": 0,
+  "laugh": 0,
+  "hooray": 0,
+  "confused": 0,
+  "heart": 0,
+  "rocket": 0,
+  "eyes": 0
+ },
+ "timeline_url": "https://api.github.com/repos/rickylabs/eis-chat/issues/135/timeline",
+ "performed_via_github_app": null,
+ "state_reason": null
+}
@@ -0,0 +1 @@
+[]