opencode/opencode_fork_factfinding_report.md

OpenCode Fork Fact-Finding — Master Report ("Better Mouse Trap")

Date: 2026-10-07 · Scope: feasibility, legal perimeter, community demand, competitive landscape, sister-tech synergy, and architectural blueprint for an opinionated OpenCode fork ("OpenCode Prime" / "Oh-My-OpenCode", working titles).

Evidence standard: every claim is cited to a primary source (file path + line, URL, commit/PR/issue number, registry/API observation). Claims that could not be verified are labeled NOT FOUND or UNVERIFIED with the search trail. The mission brief was treated as a set of claims to test, not as evidence. Deep detail lives in the four track files:

Track	Findings file
1 — Legal, licensing, ToS	research/track-1-legal-licensing.md
2 — Community friction	research/track-2-community-friction.md
3 — Harness landscape	research/track-3-harness-landscape.md
4 — Sister projects	research/track-4-sister-projects.md

---

0. Executive verdict

GO — with two non-negotiable constraints and one corrected premise.

1. You must fully rebrand. "OPENCODE" is a registered word mark (USPTO serial 99379413, Anomaly Innovations Inc.). All three working titles — "OpenCode Prime", "Oh-My-OpenCode", "Heretek OpenCode" — embed the registered mark for identical goods (Cl. 9 software, Cl. 42 SaaS) and are HIGH naming risk. MIT licensing of the code is unrestricted; trademark law is the binding constraint (§2.3).
2. You must decouple from opencode.ai defaults for clean-room posture: strip opencode/opencode-go catalog entries and default endpoints, keep Go as a strictly opt-in, etiquette-compliant plugin (§2.5).
3. Corrected premise: the brief's "naive while-loops" characterization of the core engine is wrong. V2 is already a durable event-sourced engine: inbox admission → delivery → Step with Physical Attempts → write-ahead execution claims for recovery (specs/v2/session.md:5-117). The gap is not the engine; it is phase-gated orchestration above the session layer, subagent coordination, and core reliability (§6.3).

The market case is strong and evidence-backed: upstream's highest-demand, longest-ignored work — multi-provider failover (#7602, PR #26292 abandoned), agent teams (#12711/#12730, 8 months unmaintained), deep configurability (founder declined PR #5422), reliability fixes whose PRs were auto-closed (#37852 class) — is exactly this fork's product thesis (§3).

0.1 Five brief premises vs. verified reality

Brief premise	Verified reality	Impact on plan
"Naive while-loops" need replacing with a DAG engine	V2 already has durable Step/Attempt semantics, retry policy, compaction budgeting, crash recovery (specs/v2/session.md)	Layer gating above sessions (plugin-level factory); do not rewrite the runner
awesome-harness = curated ecosystem; classics like Cline, AutoGen, E2B are in it	Autogenerated GitHub-topic index (scripts/generate_list.py); only 6 of 26 brief-named projects listed; flagships (OpenHands 90.2k★, aider 49.4k★) land in the fallback bucket	Steal mechanisms from the 33 deep-dived projects, not taxonomy positions; counts (46/74/39/18/2/49/95 = 323) are all correct
Naming "OpenCode Prime" / "Oh-My-OpenCode" is a branding question	Registered word mark "OPENCODE" exists (Reg. #8267298 per Trademarkia; serial 99379413 verified via Justia)	Full rebrand is a legal deliverable, Phase 0, before any public release
scraper-swarm does "distributed scraping, live intelligence, doc caching"	Multi-engine scraping cluster with scoped-key MCP gateway exists; "distributed scraping"/"live intelligence" NOT FOUND as terms; doc caching is IUMBTEMS's (research/cache.ts), not scraper-swarm's	Market the substance (scoped keys, SSRF fail-closed, rate limits, audit chain); drop the buzzwords
OpenCode Go terms create a legal minefield for swarms	Verified ToS (eff. 2026-08-15) + Go docs: single-member workspaces, "own internal use… not on behalf of any third party", monitored traffic, client etiquette that permits third-party clients	Go is usable by the fork as opt-in BYOK with intact etiquette; pooling/reselling is the bright line (§2.4)

---

1. Evidence ledger (mission-brief claims)

1.1 Legal & product claims (Track 1)

Claim	Verdict	Source
Core is MIT, "Copyright (c) 2025 opencode"	VERIFIED	LICENSE:1-22
MIT permits fork/modify/rebrand/sublicense/sell/close-source	VERIFIED	LICENSE:5-8 (grant clause)
Only conditions: notice retention + AS-IS disclaimer	VERIFIED	LICENSE:9-11, LICENSE:13-22
Explicit patent grant	NOT FOUND — MIT has zero patent language; no CLA/DCO/patent policy in repo or upstream	grep patent → 0; CONTRIBUTING.md; issue search "CLA" → 0 results
"OpenCode Go" $10/mo, "Go Plus" $40/mo	VERIFIED	services/www/src/docs/content/console/go.mdx:10-11
Endpoint https://opencode.ai/zen/go/v1	VERIFIED	console/go.mdx:391-420; packages/console/function/src/log-processor.ts:17-20; live 200 (2026-10-07)
Models Grok 4.7, Kimi K3, GLM-5.3, Qwen 3.8, DeepSeek V4	VERIFIED (Qwen3.8 Max/Flash; DeepSeek V4.1 Flash/V4 Pro/V4 Flash)	console/go.mdx:36-65
Go limited to single-member workspaces	VERIFIED	console/go.mdx:33-34; console/members.mdx:20
OpenCode ToS exists	VERIFIED — https://opencode.ai/legal/terms-of-service, eff. Aug 15 2026, ANOMALY INNOVATIONS, INC.	live read 2026-10-07; repo source packages/console/app/src/routes/legal/terms-of-service/index.tsx
Separate AUP / Cloud API agreement	NOT FOUND (grep + web search trail in Track 1 §3.2)	—
"OPENCODE" trademark	VERIFIED application (serial 99379413, Anomaly Innovations Inc., filed 2025-09-08, Cl. 9+42); registration #8267298 UNVERIFIED (Trademarkia aggregator only; USPTO TSDR needs API key)	justia.com/993/79/opencode-99379413.html; trademarkia.com/opencode-99379413
Upstream is anomalyco/opencode (ex-sst/opencode)	VERIFIED	GitHub API 301 sst→anomalyco (repo id 975734319); README badge README.md:14; CONTRIBUTING.md:12
Fork Heretek-AI/opencode status	VERIFIED — GitHub fork created 2026-10-07, parent anomalyco/opencode; v2 head 119cb15d1d = upstream v2 head = local HEAD → zero divergence; GitHub default branch on both repos is dev (v2 is this repo's working branch per AGENTS.md)	GitHub API; git ls-remote; git remote -v

1.2 Community claims (Track 2)

Claim	Verdict	Source
Upstream is a very-high-volume project	VERIFIED — 212,153★ / 28,254 forks; issue/PR numbers >53k	repo_view 2026-10-07
Users demand multi-agent/swarm	VERIFIED — #12711 design + 3-PR working stack #12730-12732 (~11.7k lines) untouched by maintainers 8 months; #11012 subagent-enclosure closed NOT_PLANNED with 14 sibling issues	full reads of #12711, #12730, #11012
Users demand multi-provider failover	VERIFIED — #7602 (top-reaction, open since 2026-01), #42513, #42502; PR #26292 abandoned after rebase churn	#26292 thread quotes
Container-sandbox demand (Docker/E2B/Firecracker) as a named feature cluster	NOT FOUND — issue search sandbox container e2b firecracker isolation → 0 results; demand exists as generic sandboxing (#2242, open 13+ months) and permission-escape bugs (#39931, #42979, #43336)	Track 2 §2(e)
Critical reliability bugs persist upstream	VERIFIED — SQLite corruption class (#14970, open since 2026-02), silent stream-drop recorded as success (#37852, 19 comments), compaction infinite loops (#48827), runaway token burn (#45442/#47253)	full reads
Fixes get auto-closed	VERIFIED — 5 tested fix PRs for #37852 class closed under automated-pr-cleanup (#39473, #43881, #42566, #44611, #40032)	PR search states 2026-10-07

1.3 Landscape claims (Track 3)

Claim	Verdict	Source
awesome-harness: 323+ repos, 7 categories, counts 46/74/39/18/2/49/95	VERIFIED exactly — 323 entries, all counts match, 323/323 URLs HTTP 200	README rows + data/harnesses.json + badge
List is curated	REFUTED — autogenerated topic-intersection index with fallback bucket	scripts/generate_list.py categorize_repositories(); config/topics.yaml
Brief-named classics (Cline, Roo-Code, Mentat, GAIA, Terminal-Bench, HumanEval, Cybench, OSWorld, BrowserGym, AutoGen/AG2, CrewAI, ChatDev, Swarms, MetaGPT, LangGraph, E2B, Modal, Firecracker, Daytona, Devcontainers) are listed	NOT LISTED (20 of 26); only Aider, SWE-agent, OpenHands, SWE-bench, AgentBench, WebArena are entries. Mentat (AbanteAI/mentat) and GAIA's GitHub repo 404	Track 3 §3 set comparison + URL checks

1.4 Sister-project claims (Track 4)

Claim	Verdict	Source
IUMBTEMS = "AI Build Factory for OpenCode v2"	VERIFIED	IUMBTEMS/README.md:3-5; packages/core/src/factory/machine.ts
State machine Grill → Research → Spec → Build ⇄ QA → Release	VERIFIED — 8 states incl. DONE/HALTED, every transition stage+seat gated	factory/state.ts:3; machine.ts:245-248 and transitions
Mechanical gating refuses out-of-order calls	VERIFIED	machine.ts:245-248 → model-facing ToolError via packages/opencode/src/tools.ts:66-70
Human-only signed approvals agents cannot fake	VERIFIED	approval/keystore.ts (HMAC-SHA256, 0600), approval/record.ts:1-6, CLI human.ts:33-66, TUI /es-approve preview→redeem (rpc.ts:183-227)
Evidence verification before phase advancement	VERIFIED mechanically for red-first TDD, acceptance criteria, research quote/location witnesses; ADVISORY (prompt-level) for QA note citation format	machine.ts:538-652; claims/witness.ts:1-51; vs machine.ts:663-677
Dual-QA + independent auditors	VERIFIED — dual QA seats both required (machine.ts:700-704), thesis/antithesis auditors block advancement (machine.ts:995-1003), human-only dismissal	agents/registry.ts:267-366
scraper-swarm: swarmd, panel API, @scraper-swarm/opencode-plugin	VERIFIED	apps/swarmd/, apps/panel-api/, packages/opencode-plugin/package.json
scraper-swarm: "distributed scraping", "live intelligence", doc caching	NOT FOUND (zero grep hits); doc cache is IUMBTEMS's research/cache.ts	Track 4 §0
scraper-swarm scoped keys, rate limits, fail-closed SSRF, hash-chained audit	VERIFIED	apps/gateway/src/gateway/server.py:51,226-412,473,515-522

---

2. Track 1 — Legal, licensing & ToS boundaries

2.1 MIT permissions (what you may actually do)

Permission	Clause	Notes
Fork, modify, rebrand code	LICENSE:5-7 "use, copy, modify, merge"	No field-of-use limits
Publish, distribute, sublicense	LICENSE:7-8	Derivatives may carry different terms
Sell / commercialize / close-source	LICENSE:8 "and/or sell copies"	Allowed; only duty is notice retention
Obligation 1	LICENSE:9-11	Keep "Copyright (c) 2025 opencode" + permission notice in every copy — a LICENSE/NOTICE file suffices; no trademark license implied
Obligation 2	LICENSE:13-22	Keep the AS-IS warranty/liability disclaimer

Patent posture: MIT contains no patent grant and the project has no CLA/DCO/patent policy (verified: greps 0 hits, upstream issue search 0 results). Practically low risk for a dev tool; structurally unmitigated. Option: put new Heretek-authored code under Apache-2.0 (grants stack additively; upstream code stays MIT).

2.2 The ToS boundary (hosted services only)

The ToS explicitly carves out repo code: "our open source software that is not provided to you on a hosted basis is subject to the open source license… such license and terms will exclusively govern" (terms-of-service source index.tsx:48-51). The MIT fork is therefore outside the ToS; users who connect Go/Zen accept Anomaly's ToS directly.

Boundary question	Governing clause	Assessment
(a) Multi-user via one key	"own internal use, and not on behalf of or for the benefit of any third party" (index.tsx:131-132); Go single-member workspaces (go.mdx:33-34)	One human running many agents through their own key is not barred by the seat rule; pooling one key across humans/operators = breach
(b) Client alteration / proxying	No clause names reverse proxies or headers (NOT FOUND); etiquette doc invites third-party clients: own User-Agent + stable x-opencode-session (go.mdx:74-84; validated-clients table lists Claude Code, Codex, …)	Modified clients permitted if etiquette intact; spoofed identity or stripped session IDs engage restriction (b) "not reasonably intended" + abuse monitoring
(c) Headless / high-frequency swarms	Restriction (j) unreasonable load + off-session processes; (e) programmatic extraction; (k) crawling/scraping (index.tsx:158-184)	Unattended agent traffic is the product's design center, but a 50-agent continuous loop plausibly trips (j); hard caps exist anyway (5h/weekly/monthly windows, go.mdx:113-119)
(d) Reselling / pooling	Same "third party" clause; plus (d) bulk accounts to circumvent limits; (g) using Output to train competing models (index.tsx:158-167)	Reselling Go endpoints = breach. Training/scraping swarms must run on BYOK/self-hosted backends, never Go/Zen Output

2.3 Trademark & naming risk

Name	Risk	Rationale
"OpenCode Prime"	HIGH	Embeds registered word mark in full for identical goods (Cl. 9/42)
"Oh-My-OpenCode"	HIGH	Prefix convention doesn't avoid source-identifier use of OPENCODE
"Heretek OpenCode"	HIGH	House-mark + third-party-mark composite is the classic confusion pattern
Full rebrand + "a fork of OpenCode" in docs + unaffiliation disclaimer	LOW	Nominative text use; MIT notice retention is textual and does not require trademark use

Required regardless: retain MIT LICENSE verbatim; add "not affiliated with, sponsored by, or endorsed by Anomaly Innovations Inc."; never reuse upstream logos or opencode.ai URLs in branding.

2.4 Clean-room decoupling checklist (file-grounded)

Action	Location
Strip/replace catalog entries opencode, opencode-go; repoint catalog source	packages/core/src/models-dev/snapshot.txt; packages/core/src/models-dev.ts:300-301 (source already parameterized)
Remove default console server + free "public" key path	packages/core/src/plugin/provider/opencode.ts:21-22, 319-341
Make Zen provider opt-in only	packages/ai/src/providers/opencode-zen.ts:6,21 (baseURL, OPENCODE_API_KEY)
Own User-Agent identity (do not spoof upstream's)	packages/core/src/app.ts:24-26 (opencode/<channel>/<version>/<name> — set app.name)
Keep etiquette headers on any Go traffic	packages/core/src/session/model-request.ts:264-275 (x-opencode-session et al.)
De-feature opencode-go from UI ordering	packages/app/src/providers/connect/dialog.tsx:232; packages/cli/src/commands/handlers/auth/login.ts:21
Delete/quarantine bundled SaaS (allowed by MIT)	packages/console/**, services/www/**, infra/, sst.config.ts
BYOK surfaces to keep (already exist)	packages/schema/src/config/provider.ts:33-95; custom OpenAI-compatible providers via npm: "@ai-sdk/openai-compatible" + baseURL (docs providers.mdx:375-405) — covers OpenRouter, Ollama, vLLM, LiteLLM, LM Studio, custom Heretek gateways

---

3. Track 2 — Community friction: the fork's market

3.1 Failure clusters (representative real issues)

Bucket	Headline evidence	State
State & persistence	#14970 SQLite corruption over NFS/containers → total history loss (open since 2026-02, recurring to 2026-09); #53146 cross-process seq collisions; #49221 startup brick on bad schema	open, unfixed
Compaction & context	#48827 compaction infinite loop when irreducible context > usable window (unattended tool calls until rate-limited); #2945/#6535 instruction loss; #36682 compaction as prompt-injection vector	mixed
Multi-agent	#11012 subagent enclosure closed NOT_PLANNED; #12711/#12730-12732 full agent-teams stack ignored 8 months; #37852 subagents silently return empty on stream drop; #43996 depth-2 permission hangs	open
TUI & DX	#42264 TreeSitter crash-loop freezing streamed text (unfixed v1.18.15→v1.18.31, 3 OSes); #41477 30s+ startups; #37794 53s service block; #4714/#4997 long-standing UX asks	open
Tooling & sandboxing	#39931 bash permission escape ("the bash allowlist is advisory only"); #42979 redirection bypass; #43336 symlink bypass; #2242 standing sandbox ask (13+ months). Named container-sandbox demand (E2B/Firecracker): NOT FOUND	open
Provider & streaming	#37852 silent stream-drop as clean completion (19 comments, maintainer acked 2026-08-06, unfixed); #38458/#42950 SSE drops; #49039 429 gaps; #7602 failover demand	open

3.2 Stalled high-value PRs (fork capital)

PR	Adds	Why stalled (evidence)	Age
#26292	Provider fallback chain w/ cooldowns (closes #7602)	Author quit after a month of rebase churn ("probably lost"); scope contested in-thread, no maintainer ruling	~5 mo
#5422	Provider cache config; measured 44–73% token-cost reduction	Founder @thdxr: "i don't know if we want to go as far as making all this deeply configurable yet"; promised follow-up never happened	~10 mo
#12730 (+#12731/#12732)	Agent teams: messaging, recovery, events (~11.7k lines w/ tests)	Zero maintainer engagement; community begging through 2026-06; merge state DIRTY	~8 mo
#4865	Subagent sidebar navigation	"our designer had some opinions on this… once we have UX thought out" (2025-12-02); nothing since; a fork rebased it as #46109	~10.5 mo
#5245	OpenTelemetry integration	Maintainer claims existing experimental_opentelemetry suffices; two users contradict; PR left in limbo	~10 mo
#39473/#43881/#42566/#44611/#40032	Tested fixes for the #37852 silent-drop class	Closed unmerged by automated-pr-cleanup sweeps	—

3.3 Day-1 cherry-pick list

Priority	Items	Rationale
P0	#26292 (failover chain)	Top-demand feature; author-tested; dead upstream
P0	#39473 + #43881 (stream-finalization fixes)	Fixes silent-drop class that breaks subagent orchestration — prerequisite for swarm reliability
P0	#5422 or community fork johnnymo87/opencode-cached	44–73% token-cost reduction, measurable
P1	#12730/#12731/#12732 (agent teams substrate)	Messaging/recovery substrate for Epistemic Swarm (expect rework against current dev)
P1	#4865 / #46109 (subagent sidebar)	Highest-demand orchestration view; two implementations to choose from
P1	#8963 (data-dir isolation) + #41711 (DB maintenance)	Mitigates the #14970 storage-corruption class
P2	#5245 (OTel); #31638 (skip full-history hydration); #14743 (cache-hit rate); #39015 (model-gated auto-approve)	Observability + perf + permissions

3.4 Why upstream's loss is the fork's gain

Upstream triage is automated (60-day issue auto-close; automated-pr-cleanup sweeps that killed tested fixes), contribution gates are enforced (issue-first bot) while reviewed PRs bit-rot 5–10 months, and founder-stated preferences — less configurability, subagent enclosure, "use a dedicated proxy" for failover — are the exact opposite of this fork's thesis. The demand exists, the code exists, and much of it is orphaned.

---

4. Track 3 — Harness landscape (awesome-harness, 323 repos)

Count audit: the brief's numbers are exactly right — 323 entries, split 46/74/39/18/2/49/95 across the seven named categories; all 323 URLs live (2026-10-07). The nature of the list is wrong in the brief: it is a generated GitHub-topic index (scripts/generate_list.py), so only 6 of 26 brief-named projects are entries and the fallback "Emerging & General" bucket (95) contains the flagships (OpenHands 90.2k★, aider 49.4k★, SWE-agent 20.5k★, AgentBench, WebArena). A stale seed (thu-coai/Safety-Bench, 404) partly explains the 2-entry Safety category.

4.1 Category summary & steal-list

Category (count)	Standouts verified	What to steal for the fork
💻 SWE & Coding (46)	openinterpreter 68.5k★ (runtime-switchable /harness profiles), jcode 20.3k★, SWE-bench 6.0k★, harnessrouter 2.9k★, Tianshu-harness 1.1k★	Aider's tree-sitter+PageRank repo map (repomap.py); SWE-agent's lint-gated windowed ACI; Tianshu's evidence-gated delivery ("model-says-done ≠ runtime-confirms-done") — same thesis as IUMBTEMS; jcode's server-mediated swarm file-conflict mediation; harnessrouter's secret-redaction-by-reference
📊 Evaluation (74)	SWE-bench Verified, AgentBench, lm-evaluation-harness 14.1k★	--gold harness self-test (validate pipeline against reference patches); grading separated from execution (swebench report re-scores saved logs) → fork self-benchmark mode
🌐 Web/Browser/OS (39)	Mind2Web 1.0k★, WebArena 1.6k★ (fell to fallback)	Screenshot-driven correction + session isolation patterns; scraper-swarm already covers grounding transport
🐝 Multi-Agent/Swarm (18)	jcode swarm mediation, anything-llm (misfiled), agentops 5.9k★ (integrates AutoGen/AG2/CrewAI — frameworks themselves not listed)	Role topologies + token-conservation patterns from the deep dives; conflict mediation (read/edit tracking + stale-reader notify) is the key coherence primitive
🛡️ Safety (2)	OpenART 231★ (default target is OpenCode already, AGPL-3.0 caveat), proofagent-harness 30★	OpenART as ready-made safety-CI adversary (images/Dockerfile.opencode); proofagent's live proof watch session screening + control crosswalk
📦 Sandboxes/Runtimes (49)	smolvm 6.6k★ (branchable CoW checkpoints, <200ms boot), Containarium 296★ (byte-exact resumable run logs), OverlayBD 472★ (lazy-pull images), sprites-js (detach/reattach sessions), ECC 274.6k★ (multi-harness lifecycle install)	smolvm-style checkpoint/branch = substrate for speculative multi-path swarm execution; Containarium resumable logs for long-horizon runs; OverlayBD record-trace prefetch for instant workspace restore; Smolfile declarative sandbox manifests
🚀 Emerging/General (95)	OpenHands 90.2k★ (per-conversation Docker isolation, ACP agent registry), deepagents 30.0k★ (context offload to disk, sub-agents with isolated windows), QwenPaw 35.5k★ (three-layer memory, Scroll Context: persist+index evicted turns)	Context offload to filesystem instead of lossy summarization; per-agent allow/deny/ask/sandbox governance; "interface must never assert state the runtime cannot prove" (provable-state UI law)

4.2 Top-5 adoptable mechanisms → fork seams

#	Mechanism	Fork seam	Invasiveness
1	Evidence-gated delivery (Tianshu) / factory gates (IUMBTEMS)	Plugin tool transforms (packages/core/src/tool.ts:26-32)	Plugin-only
2	Repo map for context economy (Aider)	Tool package or skill; competes with compaction budgeting (specs/v2/session.md compaction section)	Plugin-first
3	Context offload to disk (deepagents) / Scroll Context (QwenPaw)	Tool output spooling seam already exists (specs/v2/tools.md "Producers And The Registry Own Different Limits")	Plugin-first, some core
4	Checkpoint/branchable sandboxes (smolvm)	New sandbox backend trait (§6.9)	Core seam (new)
5	Resumable run logs + detach/reattach (Containarium/sprites)	Persistent-pty/session surfaces (packages/core/src/shell/)	Core patch

---

5. Track 4 — Sister projects: verified state

5.1 IUMBTEMS ("Epistemic Swarm") — production-usable today

~32.7k LOC TypeScript, 40 test files, SHA-pinned CI, published npm @heretek-ai/[email protected]. The claimed factory is real code:

stateDiagram-v2
    [*] --> GRILL
    GRILL --> RESEARCH: frontier approval (HMAC-signed, human-only)
    RESEARCH --> SPEC: REPORT.md + all deferred facts answered + dossier scored
    SPEC --> BUILD: spec approval + roadmap within maxPhases
    BUILD --> QA: red-first TDD + acceptance criteria + full gates pass
    QA --> BUILD: failPhase (3rd failure → manager replan)
    QA --> RELEASE: dual QA seats agree (manager tiebreak) + audits clear
    RELEASE --> DONE: release gates + draft PR (human merges)
    GRILL --> HALTED: spend/runtime ceiling
    BUILD --> HALTED: spend/runtime ceiling

Every transition is stage+seat gated (machine.ts:245-248); out-of-order tool calls hard-fail as model-facing ToolErrors. Approvals are HMAC-signed records with typed-code CLI dialogs and TUI /es-approve preview→redeem tokens — no agent or MCP path can approve (approval/record.ts:1-6, packages/cli/src/mcp.ts:1-6). Evidence verification is mechanical (red-first tests machine.ts:555-572, criteria machine.ts:622-652, quote/location witnesses claims/witness.ts:1-51); only QA note citation format is prompt-level. Dual QA (functional + adversarial, both required) and thesis/antithesis auditors that block advancement are shipped code. Integration with OpenCode is one additive plugin against @opencode/plugin 2.0.24 (agent/skill/tool/websearch transforms, hook bridge, 4 TUI panels) + es CLI/stdio-MCP fallback.

5.2 scraper-swarm — deployable, with honest gaps

~11.4k LOC Python/TS: gateway (Streamable-HTTP /mcp exposing web_search, fetch_page, stealth_scrape, deep_research; scoped swarm_sec_ keys with per-tool scopes, sliding-window rate limits, fail-closed SSRF pre-denial, hash-chained audit with verify endpoint), panel-api + React console (AES-256-GCM vault; key issuance returns ready-to-paste OpenCode config), swarmd (Docker reconcile sidecar, desired-state wanted.json). Gap list (verified): "distributed scraping"/"live intelligence" don't exist as terms or components; doc caching belongs to IUMBTEMS (research/cache.ts); the opencode plugin hand-rolls its hook API (no @opencode/plugin dependency) — runtime compatibility with the fork's plugin host is UNVERIFIED, one smoke test required; docs/plugin name swarm_fetch/swarm_search tools the gateway doesn't expose (web_search/fetch_page are real) — naming must be reconciled.

5.3 Maturity

	IUMBTEMS	scraper-swarm
Tests	40 files + 6 recorded evals + nightly real-host suite vs OpenCode v2 HEAD	11 files (auth/scope/rate/SSRF/reconcile)
CI	SHA-pinned actions, lint+typecheck+test, pack smoke, docs drift	Present, unpinned actions
Production-ready	OpenCode path: yes (npm-published, dogfooded with QA tiebreak/escalation artifacts)	Gateway/panel/swarmd: yes (installer, compose, tunnels, MFA); single-process rate limiting is the known scale seam (Valkey)
Aspirational	Claude Code/Pi/Antigravity adapters (roadmap 1.1–1.3); some integrity limits (bwrap fallback, advisory MCP identity)	Multi-replica buckets; plugin binding fix; buzzword features don't exist

5.4 Integration seam map (already validated in Track 4)

Capability	Seam	Invasiveness
Factory agents/tools (es_*)	plugin transforms: packages/core/src/plugin/module.ts, tool Editor packages/core/src/tool.ts:26-32	Plugin-only
Per-agent permissions	Permission.Service packages/core/src/permission.ts:1-38 (default-deny wildcard)	Plugin/config
scraper-swarm gateway as web tools	MCP remote + headers: packages/core/src/mcp/client.ts:272, tool/mcp.ts:30-45	Config-only
Fetch guard (block raw fetch/exfil)	tool.execute.before is the only hook that may fail: packages/core/src/plugin/hooks.ts:17-25	Plugin-only
Evidence cache / citable research	websearch transform + plugin storage; IUMBTEMS already wraps it (research.ts:1-45)	Plugin-only
Human approval dialogs	TUI plugin surface packages/plugin/src/tui/ + storybook (packages/tui/src/feature-plugins/system/storybook)	Plugin-only
7 of 9 capabilities overall	—	Plugin-only; core patches limited to tool re-routing, sandbox backends, OS-isolated seats

---

6. Track 5 — The "Better Mouse Trap" blueprint

6.1 Value proposition & brand strategy

Product thesis: the opinionated reliability-and-autonomy distribution of OpenCode — the power features upstream declined, wrapped in a gated autonomous build factory no competitor in the 323-repo landscape combines (Tianshu has evidence gates but no factory; ECC has lifecycle but no cryptographic human gates; nothing has dual-QA + signed approvals + scoped-key grounding).

Concrete switch reasons, each mapped to verified demand:

Promise	Evidence anchor
Multi-provider resilience out of the box (failover chains, cache tuning)	#7602/#42513/#42502 demand; #5422's measured 44–73% savings
It doesn't silently lose your work (storage isolation, stream-finalization, loop guards)	#14970, #37852, #48827, #45442 classes
Real multi-agent (teams, messaging, subagent visibility)	#12711/#12730 stack; #4865 sidebar; #11012 grief
Gated autonomous builds: phases refuse to advance without evidence and human signatures	IUMBTEMS machine.ts (shipped)
Grounded web with scoped keys and audited egress	scraper-swarm gateway (shipped)

Name strategy (decision needed before Phase 0 ends): full rebrand is the only low-risk option (§2.3). Recommended pattern: standalone product mark + permanent nominative line "a fork of OpenCode" + unaffiliation disclaimer + verbatim MIT notice. Keep "Prime" as an edition suffix if desired (e.g. <Mark> Prime); do not attach it to "OpenCode". Candidate directions are a branding call for Heretek; the legal constraint is fixed: no product name may contain "OpenCode" as an identifier.

6.2 Target architecture

flowchart TB
    subgraph L1["Distribution layer (rebranded)"]
        TUI["TUI (split views, subagent graph, storybook fixtures)"]
        CLI["CLI / service / desktop / web"]
    end
    subgraph L2["Opinionated layer (the fork's core patches — keep this thin)"]
        REL["Reliability: storage isolation, stream-finalization, loop guards"]
        SBOX["Sandbox backends: local / docker / microVM (pluggable)"]
        ROUT["Tool re-routing & fetch guard core seam"]
    end
    subgraph L3["Plugin layer (7/9 integrations live here today)"]
        FACT["IUMBTEMS factory plugin: es_* tools, gates, approvals, dual-QA"]
        SWARM["Swarm coordination plugin: teams, budgets, messaging"]
        GROUND["Grounding plugin: scraper-swarm MCP, evidence cache"]
    end
    subgraph L4["Upstream core (track anomalyco/opencode, rebase weekly)"]
        SESS["Session engine: inbox/steps/attempts/recovery (specs/v2)"]
        TOOL["Tool registry + permissions + MCP client"]
        PROTO["Protocol / Schema / Server / generated SDK"]
    end
    L1 --> L2 --> L3 --> L4
    BYOK["BYOK providers: OpenRouter / LiteLLM / vLLM / Ollama / Heretek gateways"] -. opt-in .-> L4
    GO["OpenCode Go (opt-in plugin, etiquette-compliant)"] -. opt-in .-> L4

6.3 Engine & execution — what actually changes

The runner stays. Per specs/v2/session.md, each Step already reloads history, resolves agent/model, budgets compaction against the context window, and uses narrow retry with jittered backoff (≤4 retries); recovery is driven by write-ahead execution claims. What the fork adds above it:

1. Phase-gated DAG orchestration (IUMBTEMS's factory machine, promoted from plugin to Prime mode): tool execution refuses out-of-order stage calls; phase advancement requires evidence (red-first tests, acceptance criteria) and HMAC-signed human approvals. This is a supervisor over sessions, not a runner rewrite — preserving upstream compatibility.
2. Progress guards on the known loop classes: no-progress detection on compaction cycles (#48827 math: usable = context - min(limit.output, OUTPUT_TOKEN_MAX)), identical-tool-call loop detection (#45442), runaway-step ceilings (#49042).
3. Stream finalization semantics: a stream ending with finish=unknown, zero usage, and no content must be a retriable failure, never a clean stop (#37852; adopt auto-closed PRs #39473/#43881).

6.4 Subagent & swarm coordination

• Substrate: cherry-pick/rework agent-teams stack #12730-12732 (name-addressed messaging, auto-wake, shared task list) — it is exactly the peer-to-peer layer IUMBTEMS seats need and upstream ignored for 8 months.
• Budget discipline: per-agent token/dollar quotas with hard halts — IUMBTEMS already has spend-ceiling halts (machine.ts:213-241 guard → FactoryHalted); extend to per-seat quotas.
• Coherence: adopt jcode's file read/edit tracking + stale-reader notification so parallel workers can't silently diverge.
• Completion semantics: fix the notification/empty-result classes (#50751, #44931, #37852) before claiming swarm reliability — these are the bugs that make parent/child workflows untrustworthy today.

6.5 Grounding & perception

Bundled default: scraper-swarm remote MCP (config-only integration, §5.4) with scoped swarm_sec_ keys per agent type, gateway rate limits, fail-closed SSRF, hash-chained audit. Evidence layer: IUMBTEMS's content-addressed SourceCache + quote/location witnesses so research claims are mechanically re-verifiable. Resolve the swarm_fetch/swarm_search naming drift (docs say it; gateway exposes web_search/fetch_page/stealth_scrape/deep_research) before publishing docs.

6.6 Isolation & safety

• Pluggable sandbox backend (new core seam, §6.9): local (bubblewrap-style), container (Docker), microVM (smolvm/E2B-class). Motivation is bug-driven (#39931 "bash allowlist is advisory only", #42979, #43336) — the permission model needs an OS-level boundary beneath it, not more string parsing.
• Adopt smolvm's checkpoint/branch primitive for speculative multi-path swarm runs and OverlayBD-style lazy-pull for instant workspace restore (§4.2).
• Human-signed waivers: reuse IUMBTEMS's HMAC approval records (approval/keystore.ts) as the security-waiver channel — agents cannot mint waivers (TUI/CLI human-only paths).
• Safety CI: OpenART already defaults its target to OpenCode (AGPL-3.0 — keep it an external CI harness, don't embed its runtime in the permissively-licensed fork).

6.7 TUI & UX

Four-panel split: Active Agent | Diff Preview | Subagent Graph | Live Logs. Realistic base: the plugin TUI surface + storybook fixture discipline (AGENTS.md V2 TUI Stories rules) already support panels (IUMBTEMS ships 4). Take #4865/#46109 (subagent sidebar) and the provable-state UI law (never assert state the runtime cannot prove — e.g., "stream dropped" must display as failure after §6.3(3)).

6.8 Fork maintenance strategy (anti-fork-death)

Facts: fork created 2026-10-07; v2 head == upstream v2 head 119cb15d1d → divergence starts at zero; upstream ships at high velocity (53k+ PRs) and its default branch is dev while this repo works on v2 (AGENTS.md).

Rule	Mechanism
Core stays upstream-shaped	Only the §6.2 "thin" patch set touches packages/core; everything else is plugin/config layer. Respects repo layering (Schema→Core/Protocol→Server; client never imports Core/Server)
Patch set is enumerated and rebasable	Each core patch = one commit in prime/core-patches/* with a linked upstream issue/PR; weekly automated rebase vs origin/v2 (or dev when cutting over), git rerere for conflict memory, CI green gate before updating the tracking ref
Cherry-pick pipeline	Curated backlog (§3.3) with a weekly triage pass; treat abandoned-but-tested upstream PRs as assets; expect rework (all five headline PRs are DIRTY after months)
Stay API-compatible	Protocol/HttpApi changes still flow through bun run generate from packages/client (AGENTS.md); keep generated SDK surface stable so community plugins keep working
Brand/legal hygiene in-repo	NOTICE file (MIT retention), unaffiliation disclaimer, no upstream logos; telemetry/headers carry the fork's own identity (app.ts:24-26)

6.9 Proposed seam signatures (designed, not yet code)

// Sandbox backend seam (new, packages/core): one trait, three adapters.
export interface SandboxBackend {
  readonly id: "local" | "container" | "microvm"
  create(spec: SandboxSpec): Promise<SandboxHandle>   // workspace mount + egress policy
  checkpoint(handle: SandboxHandle): Promise<CheckpointRef>   // smolvm-style CoW
  restore(ref: CheckpointRef): Promise<SandboxHandle>
  destroy(handle: SandboxHandle): Promise<void>
}

// Provider resilience (shape adopted from upstream PR #26292): fallback chain + cooldowns.
export interface ProviderFallbackConfig {
  fallbacks: ProviderRef[]            // ordered chain per agent
  cooldown_seconds: number            // tripped provider penalty window
  respect_retry_after: boolean        // honor provider-supplied delays
}

Both slot into existing surfaces: tool execution already takes a ruleset snapshot (packages/core/src/tool.ts:39 snapshot(permissions?)); provider config already lives in packages/schema/src/config/provider.ts.

6.10 Implementation roadmap

Phase	Content	Acceptance criteria
0 — Baseline & brand isolation (legal gate)	Rebrand + NOTICE + disclaimer; strip opencode.ai defaults & catalog entries (§2.4 table); Go becomes opt-in plugin with etiquette intact; upstream tracking CI (weekly rebase pipeline, git rerere store); clean build from v2	No code path dials opencode.ai by default (grep-provable); LICENSE/NOTICE shipped; automated sync job green
1 — Reliability core (the market wedge)	Stream-finalization semantics (#39473/#43881); progress guards (#48827/#45442 classes); storage isolation (#8963/#41711); cherry-pick P0s (#26292 failover, #5422 cache config)	#37852-class silent drop becomes a retriable error (regression test); compaction loop can no longer run unattended; per-process data dirs documented
2 — Factory & memory (IUMBTEMS integration)	Ship IUMBTEMS as bundled Prime plugin (already compatible: @opencode/plugin 2.0.24); /factory command + gated phases; evidence cache; instruction/memory upgrades (Scroll-Context-style indexed eviction)	A full Grill→Release run completes with dual-QA + signed approvals on the fork; out-of-order tool calls hard-fail (existing IUMBTEMS tests run against fork host)
3 — Swarm & grounding	Agent-teams substrate (#12730 rework) + token quotas + jcode-style conflict mediation; scraper-swarm gateway bundled default (fix plugin binding smoke test + tool naming); fetch guard plugin	≥3 parallel workers with scoped keys complete a shared task without divergence; gateway audit chain verifies
4 — Sandbox & enterprise hardening	SandboxBackend trait + local/container adapters (+ microVM where viable); checkpoint/branch for speculative runs; OTel (#5245) for fleet ops; OpenART safety-CI in the pipeline	Permission escapes (#39931/#42979 class) blocked by OS boundary; safety-CI red on known exploit fixtures
5 — Developer preview & community release	Storybook-driven TUI surfaces (split views, subagent graph); benchmark mode (SWE-bench --gold self-test); docs with honest limits appendix (IUMBTEMS "Known enforcement limits" model)	Preview installs cleanly on all three OSes; self-benchmark numbers published with reproducible runs

---

7. Risks & open questions

1. Trademark exposure is the one hard legal risk; registration #8267298 is aggregator-sourced (TSDR confirmation pending). Even on the conservative reading, product names embedding "OpenCode" are avoidable risk — avoid them.
2. Stalled-PR bit-rot: every cherry-pick candidate is months DIRTY against a fast-moving dev; budget real rework for #12730 and #26292, or fork the community-maintained variants (johnnymo87/opencode-cached, #46109).
3. scraper-swarm plugin binding is unverified against the fork's @opencode/plugin host (hand-rolled PluginContext, no dependency); one runtime smoke test is required before Phase 3. Tool naming drift (swarm_fetch/swarm_search) must be reconciled.
4. models.dev dataset licensing is UNVERIFIED — the fork vendors/represents its own catalog snapshot; confirm terms before redistribution.
5. MIT patent gap: no contributor patent grants exist; acceptable risk for now, mitigate with Apache-2.0 on new code if commercialization proceeds.
6. Naming decision (product mark) is a Heretek call; legal constraint is fixed (§6.1).
7. AGENTS.md designates v2 as the working branch while GitHub's default branch on both repos is dev; the fork-tracking pipeline should decide explicitly which upstream line it follows (v2 today; dev when feature work needs it) and document that choice in CONTRIBUTING.

---

8. Sources

Full citation lists per claim: research/track-1-legal-licensing.md (38 sources incl. LICENSE lines, ToS source + live page, USPTO/Justia/Trademarkia records, GitHub API observations), research/track-2-community-friction.md (23 searches, 12 full issue/PR reads, verbatim maintainer quotes with comment URLs), research/track-3-harness-landscape.md (323/323 URL batch check, generator source, 33 per-project deep dives), research/track-4-sister-projects.md (local repo paths with lines, npm registry, GitHub repo_view).

Primary anchors used in this report: LICENSE:3-22; specs/v2/session.md (prompt admission :5-11, execution :53-55, steps/attempts :70-72, retry :82-84, compaction :98-100, recovery :114-116); specs/v2/tools.md (tool contract :5-7, registry limits :148-150, laws :167-170); packages/console/app/src/routes/legal/terms-of-service/index.tsx; services/www/src/docs/content/console/go.mdx:10-11,33-34,74-84; packages/core/src/{app.ts:24-26, models-dev.ts, plugin/provider/opencode.ts, permission.ts, plugin/hooks.ts:17-25, mcp/client.ts:272, tool.ts:26-39}; packages/schema/src/config/provider.ts; IUMBTEMS packages/core/src/factory/{machine,state}.ts, approval/{record,keystore}.ts, claims/witness.ts; scraper-swarm apps/gateway/src/gateway/server.py, docs/opencode-integration.md. Network observations (GitHub API, USPTO aggregators, opencode.ai pages, npm registry) all taken 2026-10-07.