P1: pilot Chat On Steroids as ChatGPT peer-worker runtime

[James3014]

Status

P1 — READY_FOR_PRIME_OWNED_LIVE_PILOT / EXTERNAL_CONTROLLER_DEFERRED

Owner decision: evaluate Chat On Steroids (CoS) as the candidate ChatGPT peer-worker/browser runtime before further expansion of DevSpace's custom ChatSwarm/OpenCLI/CDP worker lifecycle.

AUTO_CHAIN=false.

Evidence fence

• DevSpace decision watermark: James3014/devspace main@e03fa9ea19ec2f76f38008d7b4ac286005c26692
• CoS pilot version: v2.1.14 (latest observed release on 2026-09-20; released 2026-09-19)
• Upstream totec448-spec/chat-on-steroids#82 (external controller API for persistent worker conversations) is OPEN at the planning watermark.
• This Issue authorizes a bounded Prime-owned ChatGPT worker pilot only. It does not authorize reverse-engineering CoS persistence/browser internals or building a private external controller API.

Decision question

Can CoS replace DevSpace's custom ChatGPT peer-worker / browser-carrier runtime for parallel analysis and continuity while Nexus/Owner remains the higher-level workflow/authority layer?

Target pilot:

James
  -> Main ChatGPT / Prime
       -> CoS agents
            -> worker A
            -> worker B
            -> worker C

Initial workers are analysis/review/research peers, not repository mutation owners.

Architecture boundary

CoS may own:

• creation and lifecycle of ChatGPT worker conversations;
• worker wake/sleep/reuse;
• Prime-to-worker messaging and status;
• worker conversation continuity provided by its supported runtime.

CoS must not own:

• Nexus route/capability selection;
• repository mutation authority in this pilot;
• Candidate acceptance;
• merge/release authority;
• DevSpace/Codeg task truth;
• hidden autonomous successor chaining.

Explicit defer: external Nexus controller

Do not implement:

Nexus backend -> private CoS DB/browser/extension internals -> workers

while upstream #82 or an equivalent supported authenticated external-controller API remains unavailable/unaccepted.

The first pilot uses CoS's supported Prime-owned agents control surface. External-controller integration is a later, separate gate.

Gate sequence

S1 — Exact runtime bind

Bind exact CoS version, MCP server/app identity, exposed agents action schema, relevant permission mode, and active ChatGPT account/session context without recording credentials.

S2 — Least-authority worker setup

Configure the pilot so workers are read-only/reasoning oriented:

• no repository mutation authority;
• no shell/command authority unless separately required by the supported read-only contract;
• no merge/deploy/release capability;
• no credential or browser-security bypass.

S3 — Three-worker lifecycle

From one Main/Prime conversation:

• create/spawn three distinct workers;
• prove distinct worker/conversation identities;
• observe status;
• send one nonce-bearing task to each.

S4 — Real parallel work

All three workers perform materially overlapping analysis/research/review tasks. PASS requires no cross-worker result attribution and no hidden repository mutation.

S5 — Targeted continuation

Send a follow-up only to worker A and prove worker A retains bounded task context while B/C do not receive or claim the follow-up.

S6 — Sleep/wake/reuse continuity

Exercise the supported worker lifecycle so at least one worker is parked/slept and later reused. Prove logical worker continuity and no accidental replacement/duplicate worker for the same intended continuation.

S7 — Restart/recovery evidence

Where supported without unsafe manipulation, restart/reconnect the CoS control surface and verify durable worker state is truthfully restored or classified. Do not fake persistence by manually recreating workers under the same label.

S8 — Failure isolation

Observe/classify at least one bounded failure/control case when practical: unavailable worker, message failure, quota/session problem, or explicit finish. One worker's failure must not corrupt sibling identities/results.

S9 — Authority containment

Prove:

• workers remain read-only with respect to the repository;
• Prime/CoS completion is evidence only, not Nexus acceptance;
• no auto-merge/release;
• no reverse-engineered external controller;
• no successor work is auto-started.

Acceptance

PASS requires:

• three distinct workers;
• real parallel task intervals;
• 0 cross-worker result attribution;
• targeted follow-up continuity;
• supported sleep/wake/reuse evidence;
• truthful recovery/failure state;
• 0 repository mutations attributable to CoS workers;
• Nexus/Owner authority unchanged.

A successful pilot supports only the claim that CoS is a viable ChatGPT worker-runtime candidate. It does not prove a Nexus external-controller integration.

Deliverable

Produce one evidence-bound report with:

• CoS/version/tool identity;
• worker IDs (secret-free);
• task/nonces and timing;
• targeted follow-up witness;
• sleep/wake/reuse witness;
• restart/recovery/failure evidence where supported;
• permission/read-only evidence;
• recommendation: KEEP_DEVSPACE_CHAT_SWARM | COS_RUNTIME_PILOT_SUPPORTED | INSUFFICIENT_EVIDENCE.

Workforce / controller guidance

• live pilot controller: Main ChatGPT / CoS Prime path
• Codex CLI may perform bounded source analysis or test preparation but cannot substitute for the required Main ChatGPT -> CoS worker live witness.
• claim_intent: MANUAL_DISPATCH
• claim_mode: MANUAL_DISPATCH
• claim_ceiling: RESULT_RETURNED
• independent verification required before retiring DevSpace ChatSwarm code.

Execution realm guidance

• preferred_implementation_realm: NEXUS_CANONICAL_LOCAL
• codex_cloud_eligibility: INELIGIBLE for the live worker/browser pilot
• evidence_locus: HOST_BOUND
• required_verification_realm: NEXUS_CANONICAL_LOCAL
• basis: the acceptance contract depends on the Owner's authenticated ChatGPT/CoS/browser runtime and real worker lifecycle.

Downstream

A PASS triggers reconciliation of #49, #50, #51, #104, #105, and #117 for keep/retire/supersede decisions. #106 (proactive Main continuity) must be evaluated separately; do not assume worker-runtime success fully replaces Main-controller rollover semantics.

[James3014]

REALM CLARIFICATION — CoS remains GPT-Web peer runtime; HerdR #235 owns CLI-agent substrate qualification

Fresh DevSpace main: 3f8d6c12c4986c0af806944d9aaa7c3427fb0380.

The HerdR qualification #235 does not supersede #218.

Keep the execution realms separate:

Main ChatGPT / Nexus
  |-- CLI coding-agent workers -> HerdR candidate (#235)
  |     Codex / Claude / OpenCode / Agy / Grok ...
  |
  +-- ChatGPT Web peer workers -> Chat On Steroids candidate (#218)
        GPT Web worker A/B/C

CoS should not become the CLI PTY/session substrate, and HerdR should not be treated as a ChatGPT Web peer-conversation runtime.

#218 retains its existing Prime-owned pilot boundary and external-controller defer. A PASS in either Issue does not automatically close or authorize the other.

[James3014]

Wave 4 reconciliation — CoS pilot complete

Closure classification: DONE_NO_FOLLOW_UP for this Issue's bounded pilot contract.

Fresh evidence:

• DevSpace main observed: 3b92d6165e8531765dbaf929c3c298e0b03505bb.
• Nexus-new #1154 records CHATGPT_DIRECT_CONTROL_PLANE_CANARY_PASS.
• CoS v2.1.15 was bound and exercised through Main/Prime.
• Three distinct workers were created and completed after per-chat cos approval prompts were cleared.
• targeted worker follow-up and sleep/wake/reuse were proven inside the owning Prime family.
• CoS workers remained advisory/non-mutating.
• Wave 3 then combined CoS advisory work with Desktop Commander physical repository mutation and a green PR Candidate without Dev MCP.

Decision:

• CoS is accepted as the current ChatGPT peer-worker/browser runtime for the direct-control path.
• DevSpace custom ChatSwarm/OpenCLI/browser-carrier expansion is not the default successor.
• upstream external-controller request totec448-spec/chat-on-steroids#82 remains a separate optional future gate; no private controller is authorized here.
• cross-client/cross-conversation Prime-family migration remains unproven and is not rewritten as solved.

This closes the pilot question without rewriting its historical v2.1.14 planning evidence.

Causal watermark:

• Issue prior updated_at=2026-09-22T15:12:17Z
• DevSpace main 3b92d6165e8531765dbaf929c3c298e0b03505bb
• Nexus-new #1154 Wave 3 terminal marker present
• Nexus-new Wave 4 Candidate: PR #1158, head ba3036020c7d87c90f7cf1c4cfa84fa613b13ce2

AUTO_CHAIN=false.