plan(roadmap): roadmap-expansion — Fable 5 planning run (5 backlog features → Road-to-0.0.1-stable)

[rickylabs]

Roadmap Expansion — Fable 5 planning run

Steering input + harness run for the Fable 5 roadmap supervisor to fold five backlog features
into Road to 0.0.1-stable (#301). This PR is the planning charter only — it opens no
epics/issues and touches no framework code until the owner ratifies.

The five topics

• A — Dev Dashboard (killer feature; ships as a plugin at beta.6) — specs/topic-A-dashboard.md
• B — Telemetry production-grade revamp (feeds A) — specs/topic-B-telemetry.md
• C — Tutorial complete rewrites + minimal eis-chat tutorial (beta.7 docs cut) — specs/topic-C-tutorials.md
• D — Per-feature storytelling / positioning docs (beta.7 docs cut) — specs/topic-D-positioning-docs.md
• E — Deno-desktop + unified single-process deployment (beta.8/stable; extends #327) — specs/topic-E-desktop-deploy.md

Charter

• specs/00-mission-and-flow.md — authoritative: mission, the A→G delegation flow (Fable →
  Sonnet-5 deep-search workflow → Opus-4.8 deep-dive agents → WSL Codex adversarial → OpenHands
  PLAN-EVAL), the B1–B4 output contract, the overnight-system + memory directive, hard boundaries.
• specs/01-ratified-decisions.md — milestone train, owner decisions D1–D4, R1/R4, positioning
  lock, delegated calls (D-NSONE, telemetry flow).
• specs/02-eis-chat-reference.md — eis-chat as the working reference; per-topic reading map.
• FABLE-PROMPT.md — the thin launch prompt + PR-discipline mandate.

Each topic-* spec opens with the owner's original bullets preserved verbatim, then the agreed
decisions, eis-chat pointers, delegated calls, dependencies, "what B researches", "what Fable
produces".

Status — ✅ Plan-Gate PASSED (OpenHands run-28716441078-1, minimax M3, 2026-07-04) · owner ratified · filing deferred to Phase 2

• Harness run scaffolded (.llm/runs/plan-roadmap-expansion--seed/) + specs seeded
• A — Supervisor online: charter + specs read; worklog + phase registry committed
• B — Sonnet-5 deep-search corpus (5 concurrent agents; 75 files across 20 topic/folder cells; eis-chat reference staged)
• C — Fable synthesis + both delegated decisions resolved with byte-evidence (analysis/FABLE-STAGE-C-SYNTHESIS.md) — D-NSONE = promote the missing fresh-ui L3 blocks/ layer; grouped-trace flow = two-tier (beta.6 Flow-B framework-native / stable Flow-A duckdb)
• D — 4 Opus-4.8 deep-dive design proposals (design/{A-dashboard,B-telemetry,E-desktop,CD-docs}/): telemetry T1–T9, dashboard DDX-0…19, docs S0+C1–6+D1–9+V, desktop #E1–E8
• D+ — Owner-expanded Topic-A source set folded (see note below)
• E (locked design) — research.md (Plan-Gate checklist + 14-row Findings + jsr-audit surface scan), plan.md (12 locked decisions, 13-fork open-decision sweep, cross-epic DAG, milestone train, risk register, per-epic gate matrix), worklog ## Design
• F1/F2 — WSL Codex adversarial review (verdict FAIL_PLAN, 10 findings) + F2 fixes all 10 (6b12225d)
• G — OpenHands PLAN-EVAL (minimax M3, separate session) — ✅ PASS; verdict-of-record plan-eval.md landed at d22df217 (comment)
• Owner ratification — plan ratified 2026-07-04; forks locked: OF-5 opt-in OTel-SDK on fan-in · OF-10 per-capability sections · OF-11 contract in plugin-dashboard-core (not core axis)
• Filing (Phase 2, owner-authorized) — milestones 0.0.1-beta.5/6/7/8 + epics/sub-issues filed with netscript-pr taxonomy; supersession map owner-approved before any close

Owner expanded the Topic-A source set (mid-run, 2026-07-04). At the owner's direction the
dashboard competitor corpus was extended beyond dev-consoles to the "manage framework features
through the UI" category: Appwrite Console (north-star), Directus (extensibility /
plugin-panel model), Strapi (codegen-from-UI mirroring the CLI + in-dashboard AI). Teardown at
research/A-dashboard/04-baas-admin-console-teardown.md + 17 new matrix rows. This sharpened the
Topic-A IA (flat "Plugin Control list" → cross-cutting panels + per-capability
create→configure→monitor sections), added DDX-17 (DashboardPanelContribution seam /
.withDashboardPanel), DDX-18a-d (per-capability sections), DDX-19 (codegen-from-UI), and
reframed D-NSONE via the Directus panel-contribution precedent. New owner forks OF-10…OF-13.

Draft until the roadmap is complete and PLAN-EVAL passes; the owner ratifies and cuts. Fable
pushes + updates this PR after every stage.

Refs #301

🤖 Generated with Claude Code

https://claude.ai/code/session_012wKHquACkXnWPDgJYhhFjN

[rickylabs]

Supervisor online — run taken (stage A → B)

Fable 5 roadmap supervisor has taken this run (plan-roadmap-expansion--seed).

Done so far

• Charter (FABLE-PROMPT.md) + all specs read in full: 00-mission-and-flow (authoritative), 01-ratified-decisions, 02-eis-chat-reference, and all five topic specs (A dashboard, B telemetry, C tutorials, D positioning, E desktop/single-process). Owner original bullets preserved verbatim in the specs — they remain the source of truth.
• Harness skill activated (netscript-harness); run dir + B1–B4 folder scaffolding verified; worktree on plan/roadmap-expansion confirmed; WSL gh auth verified.

Entering stage B: launching the Sonnet 5 (high) deep-search workflow to fill matrix/ analysis/ research/ context/ with one sub-folder per topic (B1 resource matrix + index, B2 exhaustive codebase analysis, B3 external-source research distilled to .md, B4 other useful context), per the B output contract in specs/00.

Flow ahead: B corpus → Fable analysis → Opus 4.8 per-topic deep-dive design proposals → locked design + harness run docs → WSL Codex adversarial review → fixes → OpenHands PLAN-EVAL (separate session, hard stop).

Boundaries honored: planning only — no issue/label/milestone mutations, no framework code; PR stays draft; owner ratifies and cuts.

Delegated decisions I will resolve with rationale: D-NSONE (NS One promotion into @netscript/fresh-ui vs borrowed reference) and the telemetry grouped-trace showcase flow (derived from eis-chat's real pipeline, incl. cross-language-hop milestone choice).

[rickylabs]

Stage B complete — deep-search corpus landed (entering C → D)

The Sonnet-5 deep-search stage is done: 5 concurrent agents (one per topic), 75 files across the B1–B4 contract (matrix/ analysis/ research/ context/, one sub-folder per topic, every cell + INDEX.md). Committed 3d70ff5a, pushed.

What the corpus changed vs the specs (9 drift candidates → drift.md)

Several are decision-shaping and I'll carry them into the locked design; the ones that touch scope or your delegated calls are flagged for you:

• E1 (significant) — the topic-E "172a-2 service-base-seam" prerequisite is a misattribution. PR Re-architect plugin scaffold surface (#157): thin, typesafe, no plugin-source copy #172 is merged; its 172a-2 sub-slices are CLI plugin type-soundness, unrelated to single-process. The real precursor is much narrower: the server-side in-process .fetch() mount seam (ServiceApp in @netscript/service) already ships; only a client-side ClientLinkPort adapter is missing from @netscript/sdk. E is smaller than framed.
• E2 (significant) — epic: NetScript enterprise deployment framework (cloud-agnostic + bare-metal, CLI + Aspire) #327 currently lists deno desktop as WATCH/reference-only, unscheduled, and feat(aspire): first-party deno desktop app support in the generator (native window as a dev-stack resource) #375 is p3/Backlog. The E deliverable (schedule it as the 4th tier + promote feat(aspire): first-party deno desktop app support in the generator (native window as a dev-stack resource) #375) is a real rescope of the issue's current state, not a note.
• A1 (significant) — D-NSONE premise is weaker than stated. fresh-ui and NS One share a byte-identical L0–L2 layer (NS One's primitives are fresh-ui's own copy-source output). The only real gap is fresh-ui has no L3 "blocks" layer (eis-chat has 9). So the decision is really "promote the missing blocks layer into fresh-ui" vs "borrow per-dashboard" — your promotion lean stays valid, re-costed much cheaper.
• A2 — Aspire pin verified 13.4.6 (clears the ≥9.4 WithCommand gate). But IInteractionService is not exposed in the TS AppHost SDK at any version — the dashboard's interactive prompts must route through command arguments, not interaction-service.
• B1 (significant) — telemetry showcase reality: eis-chat's SigNoz "archaeology ⋈ live telemetry" join is aspirational (not provisioned); the only genuine non-Deno boundary today is a telemetry-dark duckdb.exe subprocess. NetScript already injects TRACEPARENT into subprocesses (7-runtime polyglot) but no example stitches a non-Deno child span back into the parent. So the hardest cross-language hop needs a built demonstration, not just instrumentation — this shapes my flow + milestone pick.
• B2 (significant) — telemetry is worse than "level everyone up to workers": streams F (zero wiring), ai F (seam never invoked), triggers has a real correctness bug (inbound W3C context captured but never used to parent spans → severed trace), and the package itself has a tracked "Refactor" arch-debt mandating a ports/adapters restructure + OTEL-adapter split. Real span-links exist in exactly one place (database Prisma bridge). The revamp includes a restructure + a bugfix, not just parity leveling.
• CD1 (significant) — epic: docs — march to 0.0.1-stable (coverage & accuracy) #232 is a 100% accuracy/coverage umbrella with zero overlap with tutorial-rewrite or positioning scope. Landing C+D "under epic: docs — march to 0.0.1-stable (coverage & accuracy) #232" needs an explicit rescope (or a new docs-cut child epic) — I'll draft both options for your call.
• C1 — there are 5 live tutorial tracks, not 4 (chat landed separately and teaches against mid-flight @netscript/ai, publish:false today); tutorial chapter URLs are wired into 8 capability-hub nav sections. No beta.6/beta.7 GitHub milestones exist yet — issue-filing per AGENTS.md will need you to create them (no mutations this run).
• D1 — two unreconciled docs IAs (capabilities/ ~15 pages vs the 9 pillar folders) block per-feature authoring; an IA-reconciliation slice must precede the D fan-out. Only 2 named-competitor mentions exist site-wide today.

Next

Stage C (my synthesis; verifying the decision-critical B2 files) → stage D (Opus 4.8 per-topic deep-dive design proposals). Still planning-only; PR stays draft; no issue/milestone mutations. I'll resolve D-NSONE and the telemetry grouped-trace flow in the locked design with rationale, and bring the scope forks above (CD1 rescope choice, missing milestones) back to you.

[rickylabs]

Stage E locked — integrated roadmap ready for adversarial review (F1)

Design is locked. The five owner topics are folded into Road-to-0.0.1-stable (#301) as 2 NEW epics + 2 rescopes, fully decomposed with acceptance criteria, labels, milestones, a cross-epic DAG, a risk register, and per-slice agent briefs.

Both supervisor-delegated decisions resolved with byte-level evidence:

• D-NSONE → promote the missing @netscript/fresh-ui L3 blocks/ layer (L0–L2 are byte-identical copy-source; the only real gap is fresh-ui has no L3 at all). data-grid NOT promoted (collides with the existing DataGrid<T>); MCP components out for beta.6. Reframed as a panel-contribution registry via the Directus precedent.
• Telemetry grouped-trace flow → two-tier: beta.6 flagship = Flow B (framework-native multi-process: eischat→workers-api→workers→oRPC callback→streams fan-in, real span-links); stable = Flow A cross-language duckdb.exe hop.

The four epics

Epic	Milestones	Slices	Merge-gate
telemetry-revamp (NEW, Spine-1 enabler)	beta.5 → stable	T1–T9	T8 real Flow-B e2e
dev-dashboard (NEW, Spine-1 headline, ships as a plugin)	beta.6 core → stable depth	DDX-0…19	DDX-16 scaffold.runtime join
docs-cut (C+D)	beta.7	S0 + C1–6 + D1–9 + V	OpenHands per-domain
#327 rescope (E, desktop/single-process)	beta.8 → stable	#E1–#E8	#E7 deploy-e2e

Owner expanded the Topic-A source set mid-run (Appwrite/Directus/Strapi "manage-through-UI"): sharpened the dashboard IA (flat list → cross-cutting panels + per-capability create→configure→monitor sections) and added DDX-17 (DashboardPanelContribution / .withDashboardPanel), DDX-18a-d, DDX-19 (codegen-from-UI). Teardown + 17 matrix rows committed.

Consolidated decision register: 12 locked decisions (LD-1…12) + 13 owner forks (OF-1…OF-13). Only OF-5 (telemetry SDK-dep posture) is "must-resolve-now" in the rework sense, and it is technically resolved (LD-4) — it needs only the owner's dependency-posture ratification. No FAIL_PLAN open decision.

Next: F1 WSL Codex adversarial review of the locked design → F2 fixes → G OpenHands PLAN-EVAL (minimax M3, separate session). PR stays draft; the owner ratifies (creates milestones, picks forks) and cuts.

Artifacts: plan.md, research.md, design/{A-dashboard,B-telemetry,E-desktop,CD-docs}/, worklog.md § Design.

[rickylabs]

Stage F2 — adversarial-review fixes applied (commit 6b12225d)

F1 verdict was FAIL_PLAN (WSL-Codex, gpt-5.5, thread 019f2e6c…): 3 BLOCKER, 5 MAJOR, 1 MINOR, 1 NIT. All 10 findings are now resolved in the planning artifacts — planning-only: no framework code, no epic/issue/label mutation, PR stays draft.

#	Sev	Finding	F2 fix
F1-01	BLOCKER	LD-2 claimed fresh-ui/eis-chat L0–L2 universally byte-identical, but only 5/37 pairs were sampled	LD-2 + research #1 softened to "5/37 sampled-identical; remaining 32 = DDX-0 scripted full-tree-diff gate"; DDX-0 acceptance is the proving gate
F1-02	BLOCKER	OF-10 marked safe-to-defer, but the drafted beta.6 issue graph already bakes in per-capability → deferring forces rework	OF-10 reclassified must-resolve-now (with a documented flat-list fallback re-draft path); closing note corrected to "OF-5 and OF-10 must be ratified"
F1-03	BLOCKER	DDX-8 child issue named only fan-in + triggers bugfix; omitted T6 (oRPC span-creation) + T7 (query surface)	DDX-8 and DDX-16 now hard-depend telemetry T4/T5/T6/T7 by ID; acceptance fails if the oRPC callback span is mock-only / not queryable
F1-04	MAJOR	DDX-16 dependency line used a range that omitted DDX-0/1/2/4/14/15/17/18a-d	DDX-16 given an explicit full beta.6 dependency list + cross-epic T4–T7
F1-05	MAJOR	Drafts use wave:* labels but .github/labels.yml has no wave: block; wave:v2 non-canonical	Owner action to sync wave:v1/v1-min/defer folded into OF-1; DDX-19 wave:v2 → wave:defer
F1-06	MAJOR	gate:jsr missing on public-surface slices	Added gate:jsr + doc:lint/publish --dry-run acceptance to DDX-0, DDX-2, DDX-4, DDX-17, #E1
F1-07	MAJOR	"20 slices" undercount (DDX-18a-d counted as 1)	Corrected to 23 (17 core + DDX-17 + 4×DDX-18 + DDX-19); still < 30
F1-08	MAJOR	DDX-2 called public/ + telemetry/ "doctrine layering" — not doctrine-05 roles	telemetry/ → middleware/ (drift corrected); public/ reframed as harness-observed CLI-scaffolder divergence, tracked #305/#306 — dropped the "doctrine-clean" claim
F1-09	MINOR	#E2 put Closes #375 in the issue body	Moved to the resolving-PR body; issue draft = "Part of #327; folds #375"
F1-10	NIT	research #14 covered only beta.6/beta.7	Extended to beta.5/beta.8; owner creates any missing milestones (OF-1)

Consolidated decision register after F2: LD-1…12 (unchanged) + OF-1…13. Two forks are ratify-now (not silently deferrable): OF-5 (telemetry SDK-adapter dependency posture) and OF-10 (per-capability IA shape) — both resolved technically with a locked recommendation and a documented override/fallback, so no FAIL_PLAN open decision remains.

Next: dispatching Stage G — OpenHands PLAN-EVAL (minimax M3, separate session, hard stop before implementation).

[rickylabs]

@openhands-agent model=openrouter/minimax/minimax-m3 provider=openrouter output=pr-comment iterations=800

use harness — run PLAN-EVAL (planning gate only) for the roadmap-expansion program plan. Hard stop before implementation: read and judge the planning artifacts only. Do NOT write framework code, do NOT scaffold, do NOT mutate any GitHub issue/label/milestone, do NOT run deno task/deno check/deno publish and do NOT touch deno.lock. You are the independent evaluator in a separate session; the plan was authored by a different (Fable) session — evaluate it adversarially, do not defend it.

SKILL

Activate these repo skills before judging (skill-first is mandatory; skipping is the #1 eval-fail):

• netscript-harness — primary. The Plan-Gate checklist, PLAN-EVAL protocol, verdict format (.llm/runs/<run-id>/plan-eval.md), and evaluator separation live here. Follow it exactly.
• openhands-handoff — your own run/output contract: write the verdict to OPENHANDS_RUN_DIR and post one pr-comment; do not reuse legacy scratch; never mutate the lock.
• netscript-doctrine — to check the DDX-2 package folder vocabulary against doctrine-05 role folders and the ARCHETYPE-2/5 archetype claims, plus the thin-plugin/fat-core direction.
• netscript-pr — to verify the label taxonomy (namespaced colon labels, exactly one status:, wave: mapping), milestone assignment, and the closing-keyword-in-PR-body obligation used throughout the drafts.
• jsr-audit — to sanity-check the gate:jsr surface-scan claims (public surface has doc:lint + publish --dry-run acceptance) on DDX-0/2/4/17 and #E1.
• netscript-tools — for gate-evidence and lock-hygiene expectations you assert against (you must NOT run the gates yourself; judge that the plan names the right ones).

What to evaluate

Target: branch plan/roadmap-expansion (this PR, #397). Run artifacts under .llm/runs/plan-roadmap-expansion--seed/:

• plan.md — the integrated program plan: locked decisions LD-1…12, open-decision sweep OF-1…13, cross-epic DAG, milestone train (beta.5→beta.6→beta.7→beta.8→stable), risk register, per-epic gate matrix, jsr-audit surface scan, commit-slice framing.
• research.md — the re-baselined findings the plan depends on (esp. S0: initial public repo genesis #1 fresh-ui evidence, Wave 3 — @netscript/plugin (A4 plugin host) [umbrella · integration branch] #14 milestones).
• design/A-dashboard/{proposal.md,epic-and-issues.md,open-questions.md} — Topic A (dev-dashboard, 23 slices DDX-0…19).
• design/B-telemetry/, design/CD-docs/, design/E-desktop/epic-and-issues.md — the other epics.
• worklog.md — the design section + progress log (includes the F1→F2 history).

This plan already went through an adversarial review (F1, WSL-Codex) that returned FAIL_PLAN with 3 BLOCKER + 5 MAJOR + 1 MINOR + 1 NIT; commit 6b12225d (F2) claims to fix all ten. Verify the F2 fixes actually hold — in particular:

• (F1-01) LD-2 / research S0: initial public repo genesis #1 no longer overstate the fresh-ui byte-identical evidence and DDX-0 carries the full-tree-diff proving gate;
• (F1-02) OF-10 is correctly classified so no rework-forcing decision is left silently deferred;
• (F1-03) DDX-8 and DDX-16 hard-depend telemetry T4/T5/T6/T7 by ID with acceptance that fails on a mock-only oRPC span;
• (F1-05/06/07/08) wave-label gap, gate:jsr coverage, the 23-slice count (<30), and the DDX-2 doctrine folder vocabulary.

Method

Run the harness Plan-Gate checklist and fill the plan-eval.md template:

| Plan-Gate item | PASS/FAIL | evidence |
| research present and current | | |
| decisions locked | | |
| open-decision sweep (any deferred decision that FORCES REWORK ⇒ FAIL_PLAN) | | |
| commit slices (<30, each names gate + files) | | |
| risk register | | |
| gate set selected | | |
| deferred scope explicit | | |
| jsr-audit surface scan (pkg/plugin) | | |

Then run your own open-decision sweep — independently look for any decision that, if deferred, forces rework (do not just trust the plan's own OF table). Give a final verdict: PASS or FAIL_PLAN. If FAIL_PLAN, enumerate the required fixes precisely (finding → fix), each tied to a file+location.

Output

• Write the completed evaluator artifact to .llm/runs/plan-roadmap-expansion--seed/plan-eval.md (the verdict of record — OPENHANDS_RUN_DIR), committed on this PR branch. Do not commit anything else; do not touch deno.lock (if any tool dirties it, git checkout -- deno.lock before committing).
• Also write OPENHANDS_SUMMARY_PATH and post a single pr-comment summarizing the verdict, the per-item checklist results, and (if FAIL_PLAN) the required fixes.
• Keep PR plan(roadmap): roadmap-expansion — Fable 5 planning run (5 backlog features → Road-to-0.0.1-stable) #397 in draft. This is planning-gate evaluation only.

[rickylabs]

OpenHands Agent — Completed

Model: openrouter/minimax/minimax-m3
Provider: OPENROUTER
Job status: failure
Trace: .llm/tmp/run/openhands/pr-397/run-28716441078-1 (committed to the PR branch)
Artifact: openhands-agent-28716441078-1

PLAN-EVAL — plan-roadmap-expansion--seed

Summary

Independent PLAN-EVAL pass on the plan/roadmap-expansion branch (PR #397, draft). Verified
all 8 Plan-Gate checklist boxes PASS and ran an independent open-decision sweep that found no
additional rework-forcing deferrals beyond what the plan already names. Verdict: PASS.

All ten F1 adversarial findings (3 BLOCKER + 5 MAJOR + 1 MINOR + 1 NIT) from WSL-Codex
F1-adversarial-review.md have been verified fixed in the F2 artifacts (commit 6b12225d),
without introducing new gaps. The two ratify-now forks in the OF table (OF-5 SDK-adapter
dep posture, OF-10 per-capability IA) are both technically pre-resolved in the drafted
artifacts with documented fallback re-draft paths — the plan does not ship an unresolved
decision that would force rework.

Changes

• .llm/runs/plan-roadmap-expansion--seed/plan-eval.md — full PLAN-EVAL verdict
  artifact (overwrote the harness stub template with the completed 8-row checklist +
  independent evaluator-run open-decision sweep + PASS verdict + 4 non-blocking
  observations).

No other files modified. No framework code, no plugin/package surface, no deno task /
deno check / deno publish runs. deno.lock untouched. No GitHub issue / label /
milestone / PR-comment mutations (the PR-comment will be posted via the workflow owner per
the trigger contract; the runner did not post one itself).

Validation

This was a planning-only evaluator pass; no implementation validation is in scope. The
evaluation was driven by reading:

• .llm/harness/gates/plan-gate.md (the checklist contract)
• .llm/harness/evaluator/plan-protocol.md + verdict-definitions.md (evaluator discipline)
• .llm/harness/gates/archetype-gate-matrix.md (gate family authority)
• .llm/runs/plan-roadmap-expansion--seed/plan.md (LD-1…12 + OF-1…13 + DAG + milestones
  • risk register + per-epic gate matrix + deferred scope)
• .llm/runs/plan-roadmap-expansion--seed/research.md (14-row Findings table, post-F2
  softened Finding S0: initial public repo genesis #1 and post-F2 extended Finding Wave 3 — @netscript/plugin (A4 plugin host) [umbrella · integration branch] #14)
• .llm/runs/plan-roadmap-expansion--seed/worklog.md (Design section + F1→F2 history)
• .llm/runs/plan-roadmap-expansion--seed/F1-adversarial-review.md (10 F1 findings
  enumerated; the basis for the F2 verification)
• .llm/runs/plan-roadmap-expansion--seed/design/A-dashboard/{proposal,epic-and-issues, open-questions}.md (Topic-A — DDX-0…19 = 23 slices; DDX-8/16 hard-dep T4/T5/T6/T7 by
  ID; DDX-16 explicit full beta.6 dep list; gate:jsr on DDX-0/2/4/17; per-capability IA
  thesis; .withDashboardPanel seam; codegen-from-UI DDX-19 stable-tier)
• .llm/runs/plan-roadmap-expansion--seed/design/B-telemetry/epic-and-issues.md (T1…T9;
  beta.6 crit path T1→T2→T3→T5→T8; T6 pulled onto flagship gate)
• .llm/runs/plan-roadmap-expansion--seed/design/CD-docs/epic-and-issues.md (S0 + C1–6
  • D1–9 + V; epic: docs — march to 0.0.1-stable (coverage & accuracy) #232 fork Opt-2 recommended)
• .llm/runs/plan-roadmap-expansion--seed/design/E-desktop/epic-and-issues.md (#E1…E8;
  #E1 gate:jsr + #E2 Closes #375 in PR body not issue body; [Deploy-S13] WATCH: RFC-14 unified-mode + Nitro deno_server (tier-3 serverless) #349 WATCH)
• .github/labels.yml (verified wave:* block is absent — OF-1 owner action; verified
  the four wave:* draft labels reduce to three canonical wave:v1 / wave:v1-min /
  wave:defer per netscript-pr skill)
• .agents/skills/netscript-pr/SKILL.md (label taxonomy + closing-keyword obligation
  • wave→milestone map)

The runner used grep / cat -n / find only — no deno task / deno check /
deno publish / deno doc / deno fmt / git push / gh mutations. deno.lock is
unchanged (mtime + size identical to the rest of the unmodified tree).

Responses to review comments or issue comments when relevant

The trigger comment on PR #397 (@openhands-agent ... use harness — run PLAN-EVAL ...) is
the only request. The single PR-comment summary required by the trigger's
output_mode=pr-comment is being staged at OPENHANDS_COMMENT_PATH (the workflow owner
posts it; the runner did not post one). The verdict of record lives at
.llm/runs/plan-roadmap-expansion--seed/plan-eval.md and was not authored or defended by
this evaluator — it is an independent adversarial judgment.

Remaining risks

Non-blocking observations recorded in plan-eval.md § "Notes (non-blocking)" — none of
these affects the Plan-Gate verdict but a future F-cycle may want to:

1. Drift Watch R2 only names a fallback if the Aspire app kind slips; it does not
   name a fallback if the command kind itself becomes unavailable. Current mitigation =
   Aspire pin 13.4.6 ceiling. Low-likelihood; only act if Aspire's roadmap signals
   command-kind removal.
2. epic: NetScript AI Stack — first-class AI runtime, chat & plugin seams (anchor #219) #238 AI-on-codegen is a cross-epic edge to a flagship AI plugin. The plan correctly
   defers DDX-19 to stable; IMPL-EVAL should verify the epic: NetScript AI Stack — first-class AI runtime, chat & plugin seams (anchor #219) #238 owner is looped in before
   any beta.6 stretch attempt.
3. .github/labels.yml wave:* block absence is correctly tracked as OF-1 owner
   action (post-F2 fix to F1-05). The plan is not in a position to mutate labels.yml
   from this evaluator session; the obligation is on the owner at ratification time.
4. Per-slice agent-briefs.md are referenced in the per-epic drafts but live one
   level deeper (design/*/agent-briefs.md) than this evaluator's read. Lane discipline
   (WSL Codex / Opus workflows / OpenHands validation) is consistent across LD-11,
   worklog.md, and the DDX Lane columns; the per-slice handoffs are not re-examined
   here.

PR #397 stays draft per the trigger contract.

Run: https://github.com/rickylabs/netscript/actions/runs/28716441078

[rickylabs]

Stage G — PLAN-EVAL verdict: ✅ PASS

Run: OpenHands run-28716441078-1 (minimax M3 via OpenRouter) — a separate session from the Fable author and the WSL-Codex F1 adversarial reviewer, per evaluator-separation discipline.

Verdict of record: .llm/runs/plan-roadmap-expansion--seed/plan-eval.md (landed at d22df217).

Result — all 8 Plan-Gate boxes PASS

Plan-Gate item	Result
Research present and current	✅ PASS
Decisions locked (LD-1…LD-12)	✅ PASS
Open-decision sweep (OF-1…OF-13 + independent evaluator sweep)	✅ PASS
Commit slices < 30 (gate + files each)	✅ PASS
Risk register	✅ PASS
Gate set selected	✅ PASS
Deferred scope explicit	✅ PASS
jsr-audit surface scan (pkg/plugin)	✅ PASS

• All 10 F1 findings (3 BLOCKER + 5 MAJOR + 1 MINOR + 1 NIT) verified fixed at F2 (6b12225d) with no new gaps.
• The evaluator ran an independent open-decision sweep (8 candidate deferrals beyond the OF table) and found no additional rework-forcing decisions. No FAIL_PLAN open decision.
• 4 non-blocking observations recorded (Aspire command-kind fallback, #238 AI-on-codegen cross-edge, .github/labels.yml missing wave:* block = OF-1, per-slice agent-briefs out of eval scope).

Job-status note (corrected)

The workflow reports "Job status: failure", but this is not an evaluation failure — the evaluator agent exited 0 and produced the PASS verdict above. The failing steps are "Commit changes back to PR branch" and "Commit run trace to PR branch" (a CI commit-back failure), which is why the branch previously carried the empty plan-eval.md harness template. I recovered the evaluator's filled artifact verbatim from the run log (provenance noted in-file) and landed it at d22df217, restoring the verdict of record the failed CI step should have pushed.

Owner ratification (2026-07-04)

Owner has ratified the plan ("absolute bar", authorized). The three locked forks are recorded in the register:

• OF-5 → allow opt-in @opentelemetry/sdk-* on fan-in (recommendation adopted).
• OF-10 → per-capability sections at beta.6 (recommendation adopted).
• OF-11 → DashboardPanelContribution contract owned by plugin-dashboard-core, not a core definePlugin axis (overrides the earlier core-axis lean; now locked).

PR #397 stays draft. Issue-filing (milestones + taxonomy) is deferred to the owner-authorized filing phase.

[rickylabs]

@openhands-agent model=openrouter/minimax/minimax-m3 output=pr-comment

use harness

Run a PLAN-EVAL (the harness first evaluator pass) on the Topic F-ai (AI-suite) leg of this planning run, on this PR branch plan/roadmap-expansion. You are a separate evaluator session from the Fable author and the WSL-Codex F1 reviewer — do NOT author, defend, or re-edit the plan. Judge the plan, not code. Do not run any implementation gate or comment on slices that do not yet exist.

SKILL

Activate before any work:

• .agents/skills/netscript-harness — harness workflow + evaluator protocol (MANDATORY; read evaluator/plan-protocol.md).
• .agents/skills/netscript-doctrine — archetype / gate / debt vocabulary for the planned surfaces.
• .agents/skills/netscript-deno-toolchain — deno doc / jsr-audit surface inspection.
• .agents/skills/netscript-pr — label / milestone taxonomy (for the label-sync check only; do NOT mutate GitHub).

Scope (F-ai leg only)

This is a follow-on leg. The A–E plan already PASSED PLAN-EVAL (verdict of record plan-eval.md, OpenHands run-28716441078-1, minimax M3). Evaluate ONLY the Topic F-ai addition:

• plan.md → the "Topic F-ai — post-ratification integration" section (LD-F1…F6, OF-F1…F8, the F-ai DAG lane, supersession map, F-ai risk rows).
• design/F-ai/{proposal,epic-and-issues,open-questions,agent-briefs}.md.
• Confirm the F1-ai adversarial fixes landed at F2 (commit 1d3ca080). F1-ai-adversarial-review.md lists 8 findings (F1AI-01…08, verdict FAIL_PLAN); confirm each is fixed and no new gap was introduced.

Inputs (read, in this order)

1. .llm/harness/gates/plan-gate.md — the checklist you enforce.
2. .llm/harness/evaluator/verdict-definitions.md — verdict meanings incl. FAIL_PLAN.
3. .llm/harness/evaluator/plan-protocol.md — your procedure.
4. .llm/harness/gates/archetype-gate-matrix.md — the gate matrix.
5. .llm/runs/plan-roadmap-expansion--seed/research.md, the F-ai ## Design worklog rows (n4/n5/n6) in worklog.md, and drift.md (entries FAI-1, FAI-2).
6. .llm/runs/plan-roadmap-expansion--seed/plan.md (Topic F-ai section) + design/F-ai/*.
7. .llm/runs/plan-roadmap-expansion--seed/F1-ai-adversarial-review.md.
8. .github/labels.yml — for the F1AI-04 label-sync check (READ-ONLY).

Procedure

Walk the Plan-Gate checklist box by box; cite the exact plan location that satisfies each box or mark it unchecked. Run your OWN independent open-decision sweep for any F-ai decision that would force rework if deferred — if you find one the plan did not flag, that is an automatic unchecked box. Confirm the F-ai commit slices (FAI-0…17, 18 slices) are ordered, sized < 30, and each names its proving gate + files. Confirm the jsr-audit surface scan covers the F-ai public-surface deltas (plugins/ai publish:false→publishable, the @netscript/ai/otel GenAI adapter == Topic-B T9, @netscript/fresh/ai).

Specifically verify the eight F1-ai fixes: F1AI-01 (no invented beta.6 Topic-A "dashboard AI panel"; FAI-0…3 reframed as a parity floor for the OF-6 telemetry seam + stable DDX-19 handshake, not a Topic-A hard-dep), F1AI-02 (FAI-17 deps = Topic-B T3 + T6, T1 transitive — matches design/B-telemetry/epic-and-issues.md:156), F1AI-03 (single authoritative supersession headline: 15 KEEP · 1 FOLD (#257→#379) · 0 close · 3 NEW (FAI-4/10/11)), F1AI-04 (label-sync owner action for epic:/wave: into .github/labels.yml; beta slices wave:v1 not wave:defer), F1AI-05 (archetype gate matrix present), F1AI-06 (OF-F3 safe-to-defer), F1AI-07 (FAI-9 = beta.6 capability merge gate + per-milestone gates), F1AI-08 (line-ref precision).

Output (this is the verdict of record)

Fill the template .llm/harness/templates/plan-eval.md and write it to a NEW file:
.llm/runs/plan-roadmap-expansion--seed/plan-eval-F-ai.md
Do NOT overwrite the existing plan-eval.md — that is the A–E verdict of record. Emit exactly one verdict: PASS or FAIL_PLAN (if FAIL_PLAN, list each unchecked box + the specific fix required). Commit it to this branch and post your summary as the PR comment.

Note on job status: if the run's agent exits 0 and the committed/summary artifact says PASS, that IS the verdict of record even if a later CI "Commit changes back to PR branch" / "Commit run trace" step reports the job as failed — that is a commit-back step failing, not a self-post decline. Do not re-run on that basis.

Do NOT create, close, or label any GitHub issue, epic, or milestone. This run is Phase-2-filing HARD-STOPPED and PR #397 stays DRAFT.

[rickylabs]

OpenHands Agent — Completed

Model: openrouter/minimax/minimax-m3
Provider: OPENROUTER
Job status: failure
Trace: .llm/tmp/run/openhands/pr-397/run-28722662042-1 (committed to the PR branch)
Artifact: openhands-agent-28722662042-1

OpenHands run summary — run-28722662042-1

Run: plan(roadmap): roadmap-expansion — Fable 5 planning run
Trigger: @openhands-agent model=openrouter/minimax/minimax-m3 output=pr-comment use harness
Triggered at: PR #397 (plan/roadmap-expansion, draft) issue comment
Session kind: PLAN-EVAL (evaluator only, separate from Fable 5 author and WSL-Codex F1-ai reviewer)
Scope: Topic F-ai (AI suite) leg only — the A–E plan already PASSED PLAN-EVAL (verdict of
record plan-eval.md, OpenHands run-28716441078-1, minimax M3). This is the F-ai follow-on
evaluator pass; this session did NOT author, defend, or re-edit the plan.

Summary

Verdict: PASS for the F-ai leg of plan-roadmap-expansion--seed.

Walked the Plan-Gate checklist (.llm/harness/gates/plan-gate.md) box by box. Ran an
independent open-decision sweep for any F-ai decision the plan left out that would force rework
if deferred — found none. Verified the F2 commit 1d3ca080 (Fable 5, 2026-07-05) closed all
eight F1-ai findings (3 BLOCKER + 3 MAJOR + 1 MINOR + 1 NIT) and did not introduce a new gap.
Confirmed the 18-slice F-ai DAG (FAI-0…17, well under 30) is ordered, sized, and each slice
names its proving gate + files + dep edges. Confirmed the F-ai jsr-audit surface scan covers
all public-surface deltas (plugins/ai publish:false→publishable, the @netscript/ai/otel
GenAI-span adapter == Topic-B T9, the @netscript/fresh/ai FA3/FA4 runtime surfaces,
@netscript/fresh-ui ai registry, etc.) with the F-ai archetype gate matrix (§8.1) selected
per surface. The V3 commit-trail cleanup in agent-briefs is correct (F-ai agent-briefs has
zero commits.md references; A/B/E agent-briefs all use V3-correct "no commits.md" framing).

Open decisions (OF-F1…OF-F8): the only "must resolve now" fork is OF-F4 (milestones +
.github/labels.yml epic:/wave: sync — a pure owner action that blocks issue-filing, not
design). OF-F1 (E9/#248 co-ownership) is pre-resolved by LD-F5 with a documented single-issue
path. OF-F3 (Topic-A AI surface at beta.6) is correctly reclassified as safe-to-defer — the
F1-ai review removed the invented beta.6 Topic-A AI panel, so deferring OF-F3 changes no F-ai
slice or milestone. No FAIL_PLAN open decision in the F-ai set.

The A–E verdict of record (plan-eval.md) is unchanged. This is the F-ai leg verdict of
record (plan-eval-F-ai.md). Plan is ready for owner ratification; PR #397 stays draft. No
GitHub issues/epics/milestones/labels created, closed, or mutated (Phase-2 filing HARD-STOPPED).

Changes

• Created: .llm/runs/plan-roadmap-expansion--seed/plan-eval-F-ai.md (138 lines, the F-ai
  leg verdict of record, filled from the .llm/harness/templates/plan-eval.md template).
  • Plan-Gate checklist table: 8/8 PASS (Research, Decisions locked, Open-decision sweep,
    Commit slices < 30, Risk register, Gate set selected, Deferred scope explicit,
    jsr-audit surface scan).
  • F1-ai fix verification table: 8/8 FIXED (F1AI-01…F1AI-08), with file:line evidence per fix.
  • Evaluator-run open-decision sweep: 11 candidate deferrals considered, 0 rework-forcing.
  • Verdict: PASS, with 6 non-blocking notes.
• Committed: a1e66534 on plan/roadmap-expansion (1 file changed, 138 insertions; local
  branch ahead of origin/plan/roadmap-expansion by 1 commit; NOT pushed — the workflow's
  "Commit changes back to PR branch" CI step will pick it up per the trigger protocol).
• Git config set locally (per AGENTS.md "openhands" identity default, no global config):
  user.name openhands, user.email openhands@all-hands.dev.

Validation

• F2 commit verified: 1d3ca080 at origin/plan/roadmap-expansion HEAD (synced);
  subject "plan(F-ai): resolve F1-ai adversarial findings + V3 commits.md cleanup"; by Fable 5,
  2026-07-05. Working tree clean after this run's commit.
• Read-only label-file check (F1AI-04): .github/labels.yml (182 lines) contains
  0 epic: label definitions and 0 wave: label definitions (only 1 comment-header
  mention). The F-ai epic-and-issues header (lines 20-30) names the owner action to add
  epic:ai-stack, epic:telemetry-revamp, and the wave:v1/wave:v1-min/wave:defer
  block. The F-ai slice labels themselves are correct (FAI-10/11/14 = wave:v1, FAI-17 =
  wave:defer for stable; no other F-ai slice carries wave:defer on a beta milestone).
• F1AI-02 fix cross-verification: Topic-B T9's own declared deps at
  design/B-telemetry/epic-and-issues.md:156 and DAG :168 read Deps: T3, T6 /
  T3, T6 → T9. F-ai's FAI-17 dep statement now matches (T3 + T6, T1 transitive) across
  proposal, epic-and-issues, agent-briefs, OQ-1, and plan.md LD-F5.
• F1AI-03 fix cross-verification: supersession headline 15 KEEP · 1 FOLD ([AI-stack FB4] fresh-ui: mcp-ui-widget (themed sandboxed ui:// iframe) #257→[AI-stack FA4] @netscript/fresh/ai: createMcpAppCallHandler route (widget action -> tools/call, allowlist + stdio fallback + OTel) #379)
  · 0 close/supersede · 3 NEW (FAI-4/10/11) is single-source and mirrored verbatim across
  plan.md:177-181 and design/F-ai/epic-and-issues.md:331-334 + the table :323-333.
  [AI-stack FB6] @netscript/fresh-ui: interactive MCP-App bridge (widget-action → tools/call → re-render, deferred) #272 is KEEP stable (not a fold).
• F1AI-05 fix cross-verification: archetype gate matrix §8.1 is present in
  design/F-ai/proposal.md:306-321 (surface → archetype → required gates); cross-referenced
  from plan.md:133-141.
• F1AI-06 fix cross-verification: OF-F3 reclassified safe-to-defer in plan.md:161 and
  plan.md:171-173; OQ-3 rework classification in design/F-ai/open-questions.md:46-54.
• F1AI-07 fix cross-verification: FAI-9 renamed "beta.6 capability merge gate";
  per-milestone gate table at design/F-ai/epic-and-issues.md:370-376; cross-referenced
  from plan.md:139-141.
• F1AI-08 fix cross-verification: exact file:line citations throughout
  (stream-proxy.stub.ts:16-64, sandbox.ts:71-77, memory.ts:70, aiContractV1:377-379,
  deno.json:25, packages/fresh/deno.json:15-16).
• V3 bookkeeping verified: design/F-ai/agent-briefs.md has zero commits.md refs;
  remaining commits.md mentions in design/{A-dashboard,B-telemetry,E-desktop}/agent-briefs.md
  are V3-correct ("no commits.md" framing). Drift V3 entry recorded at drift.md:197-204.
• No code validation run — this is a planning-only evaluator session; per the trigger
  protocol, no implementation gate was run. Runtime, lint, e2e, and deno doc --lint are
  deferred to per-slice IMPL-EVAL.
• No framework code changes — .llm/runs/plan-roadmap-expansion--seed/plan-eval-F-ai.md
  is the only file in this run's commit.

Responses to review comments or issue comments

• Trigger comment (@openhands-agent … use harness): this session. The PR comment
  summary is in the "Summary for PR comment" block at the bottom of
  .llm/runs/plan-roadmap-expansion--seed/plan-eval-F-ai.md. The workflow owns the actual
  PR-comment posting (the OpenHands API tool is not invoked directly per the operational
  contract).
• No other review comments are present on PR plan(roadmap): roadmap-expansion — Fable 5 planning run (5 backlog features → Road-to-0.0.1-stable) #397 (the PR is draft and the prior
  PLAN-EVAL session's "Commit changes back to PR branch" step was pending — this run's
  verdict is the F-ai follow-on).

Remaining risks

These are non-blocking observations captured in the verdict's "Notes" section, surfaced here
for downstream IMPL-EVAL sessions and the next F-cycle (if any):

1. Generative-UI sandbox primitive is implicitly locked to eis-chat's
   iframe sandbox="allow-scripts" + srcdoc + zero-src pattern (web-platform,
   wrap-don't-reinvent). FAI-6 IMPL-EVAL should record the exact sandbox= attribute set
   in the slice's acceptance evidence so FAI-8 (which shares the iframe primitive) inherits
   a documented posture, not an implicit one.
2. TanStack AI exact-pin lint rule (OF-F7) should be added to the
   .llm/tools/fitness/-equivalent TanStack-touching slice's deno task deps:why evidence
   before FAI-0 lands, so the pin discipline is part of the FAI-2 e2e acceptance (drift at
   the FAI-0/FAI-1 boundary is otherwise easy to miss).
3. Supersession map's NEW issue list (FAI-4 / FAI-10 / FAI-11) requires explicit
   owner approval before any Phase-2 filing. The map says "DRAFT — owner approves before
   any Phase-2 filing; NO closes until approved; NEVER a closing keyword on epic: NetScript AI Stack — first-class AI runtime, chat & plugin seams (anchor #219) #238." The
   next run that actually files must surface the approval back to the owner explicitly,
   not as a single "drafts are ready" line.
4. Per-milestone gate table (design/F-ai/epic-and-issues.md:370-376) is the
   authoritative gate-set source for the F-ai epic; IMPL-EVAL must use this table, not
   the F-ai epic entry's single "merge-gate" column, to avoid the F1AI-07 conflation
   (FAI-9 gates beta.6 only, not the whole epic).
5. No wave:defer on F-ai beta slices — verified. Only FAI-17 (stable) carries
   wave:defer; this is the canonical label per .agents/skills/netscript-pr/SKILL.md:190-210
   for stable-tier slices.
6. The "Commit changes back to PR branch" CI step for this run may report failure if it
   has a race with the prior run-28716441078-1's commit-back step (a known CI fragility, not
   a verdict issue). Per the trigger's "Note on job status" instruction, if the agent exits
   0 and the committed verdict artifact says PASS, that IS the verdict of record — the
   later CI failure is a commit-back step failing, not a self-post decline. Do not re-run
   on that basis.
7. The A–E plan's two "must resolve now" forks (OF-5 and OF-10) are already owner-ratified
   per the plan.md header. If the owner reverts either ratification, the F-ai plan is not
   affected directly, but the cross-topic edges (OF-F1 to Topic-B T9, OF-F2 to Topic-A DDX-19
   stable handshake) need to be re-verified for the reratified A–E plan. Not a current
   concern — the A–E ratifications stand as of 2026-07-04.

Run: https://github.com/rickylabs/netscript/actions/runs/28722662042