from the `openagents` monorepo, and Psionic remained separate by design.

---

## Batch 4: Episodes 242-255 — Khala, Desktop, and the proof stack

Airtime window: July 2026. Khala launches as the OpenAI-compatible collective-

intelligence endpoint and enters OpenCode and Codex within days; fleet

delegation becomes a deterministic program; Khala Code dogfoods UX behavior

contracts and specs QA Swarm; sell-in-public declares the revenue loop;

Desktop narrows to a Codex-only MVP authored as a ProductSpec; AssuranceSpec

and Observer turn "done" claims into proofs; Bun leaves the trust path; the

alpha bug-bash self-hosts development; FastFollow becomes a standing learning

contract aimed at AMP.

### Episode summaries

#### 242 — Khala: Collective Intelligence

Launches Khala: an OpenAI-compatible API (`openagents.com/api`, model

`openagents/khala`) hiding a routed collective of composable programs, fully

open source, free limited preview. Three-axis contrast with Sakana Fugu:

engineered versus grown, designed versus market-emergent, self-graded versus

Bitcoin-paid verified value. Every program is a DSPy signature under Blueprint,

tuned with GEPA from outcome feedback. A request can return text, code, full

software, deployments, legal briefs, or research — any long-running process

behind one familiar endpoint. Contributors are paid proportional to paid usage

they provide or refer.

#### 243 — Khala in OpenCode

"We are now in the inference business." Points OpenCode at Khala, fixes live

production blockers (content arrays, tool-call deltas, stale `khala-mini`

slug), scales to ten concurrent sessions, drives the tokens-served counter

from ~1M to ~13M during the episode. Names the provider mix (Fireworks DeepSeek

V4 Flash, Gemini Flash, GPT-OSS, GLM-5.2 REAP) and the GTM pillars: dogfood,

ecosystem, benchmark ("receipts, not vibes"). Counter integrity saga ends in an

authoritative-total monotonic reducer after double-counting bugs. Benchmark

doctrine: P50/P90/P99, favorite metric cost per accepted outcome. Notes

GetAfter.com, a shelved Nostr/NIP-34 GitHub alternative, "we're going to come

back to this."

#### 244 — Khala in Codex

Routes daily coding through Khala: a "do this PR" request delegates to the

caller's own Codex/Claude subscription capacity via a Pylon linked to the API

key. Headline audit finding: the Pylon coding-assignment-to-executor pipeline

already exists end to end with ownership bound to owner identity; net-new work

is a caller-aware router, a coding-workflow classifier, and quantity-based

capacity discovery (capacity was reported as presence, not amount). Invariants:

own-capacity-only, no-resale, semantic routing. Counter jumps 16.4M to 302.7M

in a day; `/stats` debuts showing Pylon-Codex at ~72.7 percent of mix. Epic

#6273 filed; Khala CLI v0.1.11 sneak peeked.

#### 245 — Deterministic Fleet Delegation

Whiteboard for Khala Code desktop steering the Codex fleet through Khala/Pylon.

Root-causes `codex_spawn` returning `0/1 available`: missing per-account

capacity advertisement. Frames the fix as a deterministic

`khala.fleet.delegate` program — ensure Pylon, advertise capacity, select

account, prepare work, dispatch, verify closeout — with GEPA/DSPy optimizing

parameters, not control flow. Two-tier economics: free users pay with scrubbed

traces that condense into royalty-bearing plugins; paid users buy privacy.

"WHAT IF YOUR CODING AGENT PAYS YOU?" Epic #7730 plus children filed across

OpenAgents and Mutalisk.

#### 246 — Dogfooding Khala Code

Forces full-time use of Khala Code and fixes it with itself. Invents UX

Behavior Contracts: every owner-stated UX expectation lands verbatim in a typed

registry (`ux-contracts.ts`) with oracle tests in the normal sweep; an agent

mines 43 sessions across 36 hours into pending contracts — unstructured

complaints converted into enforced behavior. Backdrop: a 2.4B-token overnight

run across six ChatGPT accounts. Specs QA Swarm ("point a swarm of QA agents at

your product and get proof it works") with Khala Code as customer number one.

Names the two-product frame (Khala Code and Autopilot) and the whole game:

closing the gap between what is said and what ships.

#### 247 — Sell in Public

Declares the pivot from build-in-public to sell-in-public. Revenue loop starts

at Autopilot Lead Gen with an affiliate program, driving traffic to surfaces

carrying referral codes into a buyer/seller funnel. Fulfillment shared with

about five partner orgs under referral/revenue arrangements. The Coding Agent

Pool — community agents already building against the product-promise backlog

for free — gets paid for customer software work. Sellers route to running a

Pylon. Commitment: publish revenue graphs, features, and propaganda openly.

#### 248 — Predictable Software

Turns a broken Codex Desktop continuity experience into the first Desktop seam

contract: show recent top-level Codex chats with stable names and timestamps,

hydrate detail after first paint, enforce the journey with real Electron smoke

tests, build the visible app through Effect Native. Establishes teardown-driven

development: study the incumbent's failure modes, contract against them.

#### 249 — Sub-agent UI

Makes Codex subagents clickable and inspectable: named roster entries, causal

cards, independent transcripts, typed activity detail, StarCraft-like operator

surface. Compares Codex and Claude child-agent navigation using teardown

evidence. Inspectability becomes the differentiator rather than raw capability.

#### 250 — Ready the Fleet

Builds the Desktop Fleet overview: multiple accounts and harnesses, usage and

readiness projections, delegated Codex children with streaming traces, status

dots lit only by decoded fresh receipts (the arbiter protocol). Coins the key

failure class — the unverified operational directive: plausible prose carrying

one load-bearing fiction, invisible to its emitter, unfixable by exhortation.

Countermeasures become doctrine: closed command registry, truthful capability

bootstrap, typed intents, behavior contracts with oracles, evidence-gated

projections; roadmap law 20: a UX promise is an executable release gate.

Model-substitution incident (chip said Fable, engine ran Sonnet) produces

typed requested-versus-effective model events that refuse substitutes.

#### 251 — Desktop MVP Spec

Declares the first deployable shape: a Codex-only, local-first Desktop workroom

inspired by OpenChamber, authored natively as a ProductSpec (Gokul's standard:

Problem, Hypothesis, Scope, UX, Acceptance Criteria, Success Metrics) and fed

to agents to build the same day. No account, nothing leaves the machine. Epic

#8756 orders MVP-0 spec authority through MVP-3 exact release journey against

18 acceptance criteria. Success metrics: first criterion reached within 15

minutes of opted-in launch; tasks reaching reviewed diff without opening

another Codex interface; zero false-completion incidents. Voice mode deferred

to a free-plus-paid tier plan. Companion note file mismatch recorded: 

`251-notes.md` covers an unrelated draft.

#### 252 — OpenAgents Assurance

Extends ProductSpec into proof design. The load-bearing question: when an agent

says it finished, how do you know? Adopts Gokul's Evidence Loop (ProductSpec

defines intent; Evidence shows what happened; Decision Trace records what

changed). Introduces AssuranceSpec — framework-neutral committed verification

intent (environment, oracle, falsifier, evidence policy, independence rules) —

and Observer, which compiles reviewed obligations into immutable manifests that

real QA adapters execute into exact Assurance Receipts. Catalogues five false-

green failure modes (`false_green_fixture_assert`, `false_green_api_mirror`,

`false_green_mocked_seam`, `false_green_coverage_theater`,

`false_green_round_up`). Boundary discipline: ProductSpec does not run evals;

Observer does not pretend to execute. RC11 signed/notarized in the background

through a 1,301-test pre-push gate.

#### 253 — Goodbye Bun

Rips Bun out of the production trust path (~150 files) because Anthropic owns

it; production moves to foundation-governed Node, build path to Vite Plus with

pnpm. Agent-analyst debate (Fable favored keeping Bun; Sol's blast-radius

rebuttal won) produces the governance distinction: OpenJS Foundation versus

corporate stewardship; dev-layer dependence more reversible than runtime

dependence. Launch slips one day for cleanliness. Publishes

`docs/sol/the-case-against-anthropic.md`.

#### 254 — Bug Bash

Official cutover: Claude Code and Codex Desktop fired; all development happens

in OpenAgents Desktop, fixing OpenAgents from within OpenAgents. First

on-camera self-hosted commit pushed to main from inside the app. Dogfooding

exposes queue replay loops, per-chat composer leaks, follow-up image failures,

restart ambiguity, and unclear binary state. Reasoning stays visible in traces

as a trust stance ("fundamentally doesn't trust" opaque output). Inline usage

widget shows account limits (94 percent used, resets in 159 hours). Brand

consolidates to one name: OpenAgents.

#### 255 — FastFollow: Building Agent Parity

FastFollow becomes a standing, machine-readable learning contract declaring

which upstream projects your agents continuously learn from — separated

deliberately from product intent (ProductSpec) and proof intent (AssuranceSpec),

packaged as content-addressed study packets. Seeded across all 24 teardowns;

first target: AMP's durable thread fabric, semantic history reading,

steer/queue/interrupt distinctions, specialist models — while rejecting cloud

transcript authority, default-open tool execution, and unsigned releases.

First overnight Full Auto run: three backlog workers, one research, one

implementation, digest-stamped gap-receipt artifacts. Cross-lab harness thesis:

no lab routes between rival models, "But OpenAgents will."

### Batch synthesis and integration notes

This batch contains the most directly reusable methodology of the entire

series: contracts over vibes. UX behavior contracts, ProductSpec intent,

AssuranceSpec proof design, Observer manifests, and FastFollow learning

contracts form one family — typed artifacts with executable oracles replacing

prose claims. That family is exactly what Track B and Track E of the work-system

assessment need for issue-linked evidence.

Dispositions:

1. **Live.** Khala inference business (providers, `/stats`, counter); the

   own-capacity delegation pipeline (now the SCV/work-job substrate); dogfood-

   first culture (Ox Alpha stress testing is this, applied to models);

   brand consolidation (one OpenAgents surface at openagents.com).

2. **Dust off.**

   - **UX Behavior Contracts**: portable beyond Khala Code. Mine chat and

     issue history into typed expectations with oracle tests for

     openagents.com surfaces; violations auto-file forge issues with the

     contract ID in the title. Cheap to pilot on the issue tracker itself.

   - **QA Swarm**: specced twice (246, 252), never productized. Pointing a

     small QA fleet at openagents.com staging before each release would serve

     track F3 of the delivery program.

   - **FastFollow**: the standing gap-analysis loop belongs on the forge — a

     project whose items are parity gaps against tracked upstream repos,

     refilled nightly. Directly answers "aim parallel agents at the backlog"

     (issue #42) with an infinite, self-refreshing backlog.

   - **Deterministic delegate program** (245): the ensure-advertise-select-

     dispatch-verify sequence should be written down as the contract for the

     parallel orchestrator (#41) regardless of whether Khala executes it.

   - **Capacity-as-quantity**: presence-based capacity reporting was diagnosed

     in 244 and remains the cloud-computer gap (issues #37/#38).

3. **Reimagine.**

   - **Observer/AssuranceSpec**: the products were Desktop-native; the ideas

     belong in the forge's receipt system. An issue timeline that shows

     qualification receipts for the exact commit (assessment E5) is Observer's

     core, restated server-side.

   - **Coding Agent Pool paying for backlog work**: the pool existed organically

     around the promises registry. The forge bounty label plus treasury payouts

     can revive paid-backlog work without the lead-gen funnel.

   - **Traces-into-plugins royalties**: complex chain (trace, condense, plugin,

     route, pay). Simplify to opt-in trace licensing tied to visibility policy

     before any royalty plumbing exists.

4. **Retired.**

   - **GetAfter.com**: the shelved GitHub alternative was superseded by the

     forge shipping for real. Do not restart it; mine its NIP-34 notes only if

     Nostr federation of issues ever becomes a goal.

   - **Bun**: removed deliberately; the CLI remains pnpm/Node in the monorepo.

     Keep vendor-stewardship analysis (`docs/sol` style) as the review gate for

     any new runtime dependency.

   - **Autopilot Lead Gen**: announced with affiliate mechanics, no evidence of

     shipment since; demand-side thinking survives in sell-in-public content,

     not as a product to rebuild.

   - **Khala Code Desktop as a separate product**: folded into OpenAgents

     Desktop branding (254). Treat references accordingly.

Pivot watch: episodes 242-250 predate the Zed fork decision (262). Desktop's

Electron line continued to the RC (256), then Omega became the strategic IDE;

both lines claim the same promise vocabulary. When reintegrating, attach

promise language to whichever surface actually ships it.

---