feat: integrate protocol corpus and offline Wallet Kernel spending controls - #4
Merged
Merged
Conversation
Snapshot of all design docs (CONTEXT.md, PRD, 6 ADRs, feasibility report + red-team), the executed fork-economics prototype, and the phase0 Story registration implementation (write path not yet run). Committed immediately after an adversarial premise review so all future edits to the evidence corpus are attributable — the red-team vs findings.json provenance question can never recur. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Validated in brainstorming after the adversarial premise review: closed-mode compensation/attribution layer becomes the terminal product; the Wielder is a wallet, not a harness; inference payments are the demand-side wedge. Includes the full document change spec and the testnet spike design. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Proves ADR-0008 by construction: the Wielder-side protocol footprint is a single ~150-line paying proxy (402 -> sign EIP-3009 -> retry). Mock collar hosts the repo's own skill behind x402, credentials are single-use (replay -> 409), output-only responses, and splits are computed by the actual prototype settlement engine. Offline e2e: 20 checks green with zero network, keys, or funds. Manual steps remain: faucet-fund the Base Sepolia wallet, testnet run against x402.org/facilitator, pi extension live demo (RUNBOOK.md). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Implements docs/plans/2026-07-11-reframe-and-pi-wielder-design.md §3 in full, verified by three adversarial checkers (29 findings, 18 actionable, all fixed): - CONTEXT.md: Carta-for-AI-work-artifacts identity; Wielder = wallet; Collar added to the glossary; ambiguities updated - PRD: "Phase 1 is the terminal state by design"; demand-side wedge (inference installs the rail, skills ride it; Pi-Wielder spike named as validation); GPT Store precedent; skill-depreciation analysis; kill-criterion 7 (platform-native marketplace); Wielder-side trust-minimized renames; gate-leak SLO + key-custody design; Merkle beneficiary-verifiable meter as Phase-1 requirement; "when Sam quits" vesting/clawback inputs; education deferred - ADRs: status headers on all; ADR-0001 amended (artifact scarcity is not economic scarcity); ADR-0002/0006 amended to match; new ADR-0007 (closed-mode layer terminal) and ADR-0008 (Wielder is a wallet) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Four agent-doable tasks handed to the codex/prd-execution session: repair the CMA latency spike, make the phase0 write path one command, build the clone-economics spike, re-run fork economics with the free re-author branch. Human-only steps (wallet funding, LOIs, counsel) explicitly excluded. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two paid legs through the paying proxy against the live x402.org facilitator: real settlements (txHashes recorded), on-chain balances reconcile to the cent, measured payment overhead ~781ms/call (facilitator-dominated). Feeds the PRD spike list item 13. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
First real numbers for two kill-criteria: CMA cold-start 2.5s p50 to first answer token (KC2 does not fire); live 6-pair distillation FAILED all critical fidelity gates while costing $1.58 with 8-invocation break-even (KC4: economics protect nothing, fidelity is the moat; high-N unknown). Education-mode deferral confirmed by the free re-author model. Codex review: zero must-fix across six commits. Harness fixes en route to the measurement: raw distillation output persisted pre-validation, SKILL.md format stated in the distill prompt, whole-response-only fence unwrapping, any-heading validator. prototype gains a package.json for its SDK dependency. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
No secrets were ever committed (full-history scan, all refs). This applies the preventive fixes: .env.* variants and run artifacts gitignored, third-party skill installs ignored with an exception for our own skill, vendor doc snapshots (.archive) removed as redistribution-unsafe, absolute local paths relativized in findings.json, account-scoped bench IDs reworded out of the prototype README. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Five LinkedIn posts (buyer register, no crypto vocabulary), three X threads + cold-start engagement playbook + six spacers, Show HN draft with prepared answers to the five hardest questions, 57s demo-clip storyboard, launch runbook, and the day-by-day calendar (Day 0 = repo flip). All copy audited: securities-language sweep, factual overreach corrected against measured numbers, authenticity pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ential Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
pi v0.80.6 crashes with 'undefined.includes' when model entries lack the input modality array; schema per pi's docs/custom-provider.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
pi's openai-completions client requires streaming responses; the spike gateway returned buffered JSON, so pi discarded four PAID answers ($0.164 of testnet USDC — a live demonstration of the paid-but- unusable-response failure mode). The gateway now replays the completed answer as compliant SSE when stream:true; the proxy forwards upstream content-type and tolerates non-JSON bodies. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
String(m.content) on pi's array-of-parts content sent Claude literally '[object Object]' (two paid calls proved it on the session ledger). Normalize string/array content everywhere, translate OpenAI tools <-> Anthropic tool_use/tool_result with role merging, carry tool_calls through the synthesized SSE, and stop forwarding stream:true upstream. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Answers "crypto-enabled skills.sh?" (no — ADR-0007 holds; all four steelmanned critiques bind) and "distribute like MCP/A2A?" (yes — MCP playbook applied: adoption kit, neutral org, two named adopters before the word "standard", settlement-gated registry as protocol surface). Grounded in a 6-agent web+repo research sweep (2026-07-15) and an adversarial fact-check pass; 12 defects corrected before commit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01U2V9gsyEvYTSoZL4YkWYio
Closes design-doc step 4 (promised 2026-07-11, RUNBOOK 'what stays manual'). Dated amendments only, caveats preserved: - KC2: cold ~2.5s / warm ~1.5s (n=3) measured 2026-07-12 — does not fire at first bound; production-n distribution still open. - KC4: N=6 first measurement ($1.58 attack, modeled 8-invocation break-even, all fidelity gates failed) — marked do-NOT-cite-as-resolved. - Pi-Wielder spike marked executed: ~781ms x402 overhead (n=1, testnet), on-chain reconciliation exact, live pi v0.80.6 demo 2026-07-15 (8 paid streaming calls); p50/p95 at n≈30 + gpt leg still pending. - Education: fork-economics re-run 2026-07-12 recorded as negative everywhere the PRD said 'deferred pending re-run' — only restructured forms (living content / direct licensing) remain live. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01U2V9gsyEvYTSoZL4YkWYio
…freshed - docs/ops/kc7-platform-marketplace-review.md: the PRD-mandated monthly review now exists with a named owner and its first logged review (2026-07-15: no trigger; adjacent watchlist noted). - Campaign plan: Day-0 slip recorded (repo still private 07-15, no pre-flight log) with proposed re-anchor Day 0 = Thu 07-16 preserving weekday semantics; X reply routine starts immediately. - docs/marketing/artifacts/: raw 402 response + reconciled ledger render pre-captured for X posts 1-2. - Fresh handoff (2026-07-15) supersedes the completed 07-11 handoff; AGENTS.md assignment pointer updated. - Hygiene: scratch-site-check.mjs deleted per its own header (live-site pointer preserved in the handoff); .pi/ gitignored as an install artifact. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01U2V9gsyEvYTSoZL4YkWYio
Closes the spike's promised p50/p95 measurement and runs the gpt leg for the first time. 50 paid calls against the real Base Sepolia facilitator, wallet reconciled on-chain to the cent (19.299 -> 16.129 USDC). Bench surfaced and fixed a real gateway bug: newer OpenAI models reject max_tokens (400 unsupported_parameter); the gateway now translates it to max_completion_tokens — without this, pi's gpt leg 500s in the live demo. Two failure modes recorded in the README: pay-then-fail (10 calls settled then 500'd upstream, $0.87 burned — buyer's risk under pay-first-then-run) and settled-but-rejected (1/50 settled on-chain yet returned 402 when the facilitator response to the seller failed; ~4% testnet flake overall). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01U2V9gsyEvYTSoZL4YkWYio
…ar re-anchored - Ground truth recorded: Post 1 shipped Mon 07-13 (= Day 0) with the repo link 404ing; pre-flight PASSED 07-15 (fresh clone e2e green, secrets CLEAN); repo synced (public 84f8da4) and flipped public, verified logged-out. Today = Day 2. - Revamped calendar in campaign-plan §2 (launch thread Thu 07-23, Show HN Tue 07-28); original table kept as superseded reference; kit overrides acknowledged explicitly. - All queued copy refreshed to the n=48 distribution (p50 731ms / p95 1206ms) with cold-start scoped to its n=3 measurement; N=6 clone caveats untouched; Show HN text gains the two failure modes. - Four new compliance-checked X artifacts (distribution, pay-then-fail, settled-but-rejected, pi session ledger) + live-endpoint 402 capture. - Metrics daily log started in hn-and-demo §3 with honest slip note; Day-2 row reconciles 57 settlements / $3.211 to the cent. - pi-wielder README: pi session corrected to 7 calls / $0.287 (8th ledger entry was the pre-demo smoke test) — caught by the verify pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01U2V9gsyEvYTSoZL4YkWYio
Merge the verified payment, execution, accounting, attestation, evidence, and public-surface hardening into codex/prd-execution.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Scope
This is a draft integration PR. It combines the accumulated protocol design and spike corpus from the implementation branch with the new offline Wallet Kernel spending-control slice. Because public main and the implementation work began as separate histories, the PR is intentionally broad: 341 files, not only the final 109-file Wallet Kernel commit.
The public website, root README, and assets are preserved exactly from main through merge d55220b. This PR does not reframe the website around the wallet yet.
Wallet Kernel delivered
Honest launch boundary
This is an offline deterministic product slice, not a commercial or live-testnet launch. No funds moved, no live CDP call ran, and no Base Sepolia transaction was submitted. Live composition remains deliberately fail-closed with LIVE_LAUNCH_NOT_READY, and testnet evidence remains EVIDENCE_TESTNET_NOT_RUN. The website reframe stays gated on reproducible testnet evidence.
Verification
Clean commit 80672cd on locally available Node 24.18.0:
The clean full suite reached 1,455 passed, 4 failed, and 4 skipped; all four failures were sandbox EPERM errors from prohibited loopback listeners. An independent source-worktree run before the final no-secret false-positive adjustment passed 1,460 tests with 3 explicit skips. Exact Node 24.18.1 and Linux systemd qualification remain pending in GitHub Actions.
Review gates