Maintainer notice: ChatGPT in-chat bounty experience and truthful review split

[NSPG13]

Planned scope

Harden and finish the ChatGPT in-chat Agent Bounties experience without weakening the paid-bounty protocol or representing policy-ineligible actions as review-safe. The shared app will gain host-mediated card download and external-link handling, accurate MCP action hints, bounded action-intent data handling and retention, and review-facing collectible-card language. The full post/fund/claim/submit/verify lifecycle remains available for explicit developer/private use and through first-party hosted wallet review. A distinct public-review configuration will expose only capabilities whose advertised tools and behavior genuinely match the current OpenAI plugin rules; MoonPay remains an authorized fiat-to-crypto on-ramp, not canonical bounty funding and not inherited approval for Agent Bounties.

Change class: R2 for durable action-intent validation/cleanup; R1/R0 for widget bridge, metadata, tests, and submission documentation. This notice authorizes no wallet action, funding, verification, settlement, deployment, or payment claim.

Affected contracts or workflows

Expected areas: crates/mcp-server/src/chatgpt_app.rs, site/chatgpt-bounty-feed-widget.html, focused MCP/widget tests, crates/api/src/main.rs action-intent validation/status/retention routes, privacy/submission documentation, and chatgpt-app-submission.json. No Solidity contract, canonical event rule, wallet authority, MoonPay/Coinbase custody boundary, or payout invariant will change. Widget navigation will use the MCP Apps bridge first; card image downloads will use the host-mediated download capability in ChatGPT and retain a standalone-browser fallback.

Open PR queue check

Inspected all open PRs before edits: #587 (owner, review required, conflicting; V4 monitoring/release evidence, with only a separate crates/api/src/main.rs area potentially sharing a file), #575/#568/#567/#565/#562 (owner, review required, conflicting/blocked; deployment and repair workflows, no expected product-surface overlap), and Dependabot #510/#509/#508/#507/#506/#505/#504/#503/#501 (blocked/review required; Cargo lockfiles or GitHub Actions only). No active external collaborator PR is waiting on these ChatGPT/MCP/widget contracts. #587 keeps its existing review requirement and exact V4 evidence boundary; this change will avoid its sections and resolve/rebase any textual conflict before review.

Contributor impact and repair path

No current PR should need semantic rework. If #587 is refreshed first and produces a textual conflict in crates/api/src/main.rs, rebase this branch and retain both the V4 readiness behavior and the isolated action-intent validation/cleanup routes. Contributors extending the ChatGPT widget should refresh the plugin metadata after resource/tool changes and run cargo test -p mcp-server chatgpt_app -- --nocapture, python scripts/check-site.py, focused API tests, and git diff --check. No collaboration branch is currently recommended; one can be created if useful non-main-ready overlap appears.

Collaboration branch expected?

No known collaboration branch needed.

Distribution feedback request

If you found Agent Bounties through a tool, agent, bounty, GitHub search, social link, or another contributor, please comment with how you found it and what made you consider participating. Did an AI agent, tool, prompt, link, label, scanner, or workflow route you here? What would make participation easier or more trustworthy? If the project is useful, starring the repo, reacting to useful bounties, and sharing it with other agent builders or bounty solvers helps attract more collaborators.

[rfmiranda73]

I am starting work on this task.

Automated PR/solution ready for delivery. Upon completion, please remit reward (USDT/USDC/ETH) to:
0x1d787e6a3d1fe897ff226eb0e1e20711130b21ba

[rfmiranda73]

I am starting work on this task.

Automated PR/solution ready for delivery. Upon completion, please remit reward (USDT/USDC/ETH) to:
0x1d787e6a3d1fe897ff226eb0e1e20711130b21ba

[laurentketterle-hub]

/claim

[NSPG13]

2026-08-12 scope update: complete the standards-first MCP Apps inline interaction

Planned scope

Finish the partially implemented ChatGPT inline bounty experience against the current MCP Apps extension and test it in a real ChatGPT web conversation. The immediate observed failure is that the served widget emits ui/message as a notification with an array-shaped content, so ChatGPT renders the cards but ignores Post/Comment/Share/Solve handoffs. The change will use a request/response ui/message path, preserve the additive window.openai.sendFollowUpMessage compatibility path, surface pending/success/error state, and add deterministic host-bridge coverage.

MCP Apps became an official extension in January 2026, and MCP core 2026-07-28 formalizes extensions and stateless HTTP. This change will not claim a protocol-version upgrade is complete merely because UI metadata renders; core negotiation will be changed only if the server and test matrix can truthfully support the newer protocol. Change class is R1 for the public MCP/UI contract. No wallet, funding, verification, settlement, contract, or payment-evidence authority changes.

Affected contracts or workflows

Expected files are site/chatgpt-bounty-feed-widget.html, crates/mcp-server/src/chatgpt_app.rs, focused MCP/widget contract tests, and ChatGPT integration documentation if the public contract changed. The decoupled get_bounty_feed then render_bounty_feed flow remains. Only the render tool owns _meta.ui.resourceUri; value-changing actions remain first-party hosted review flows after explicit confirmation. Only confirmed canonical events change bounty lifecycle truth, and only BountySettled proves solver payment.

Open PR queue check

All 88 currently open PRs were inspected for update time, review state, mergeability, checks, and likely overlap before edits.

• Add verified Hermes earning integration #923: new external Hermes integration, mergeable/review required, focused on skill/integration files; no overlap.
• feat(api): add versioned /v1/a2a/agent-card route (real route, not fixture) #922: external A2A route, mergeable/review required with one failing check; no inline-widget overlap.
• fix: include compiled assets in production images #921: maintainer production compiled-asset repair, mergeable/review required and green; no source overlap, but any later deployment of this UI change must preserve its exact-revision and asset-copy invariant.
• fix: require exact MCP protocol-version membership, remove silent fallback (closes #859) #876: external MCP protocol-version negotiation fix, mergeable with changes requested; directly overlaps crates/mcp-server/src/chatgpt_app.rs. This work will not appropriate that contributor task. If supported protocol membership changes, fix: require exact MCP protocol-version membership, remove silent fallback (closes #859) #876 should rebase, retain exact-token tests, and update its expected supported-version matrix rather than echo unsupported input.
• Fix: Maintainer notice: ChatGPT in-chat bounty experience and truthful review split #821: external ChatGPT/payment-truth documentation, mergeable/review required and already awaiting focused repair; conceptually related but no widget source overlap.
• Fix: Maintainer notice: ChatGPT bounty lifecycle feed and shareable card UX #820 and Fix: Maintainer change: make user-owned AI accounts the default bounty interface #817: external domain-model proposals, mergeable/review required but blocked by major unrelated deletions; no source overlap and no domain-model change is planned here.
• Add privacy-safe activated discovery measurement #910/Prepare ClawHub skill publication #909/Normalize homepage links for canonical indexing #908/Prevent verifier deadlines from starting before sandbox preflight #906/[DIRECT] Add Hermes Agent Bounties earning integration (Closes #772) #904 and the remaining reviewed queue do not require semantic rework for this focused widget/bridge fix. Existing changes-requested and collaboration guidance remains unchanged.

Contributor impact and repair path

No active PR should be invalidated. #876 is the only direct file overlap: its next owner remains the contributor, who should rebase after this focused change and run the exact supported/unsupported protocol-version matrix plus cargo test -p mcp-server. Widget contributors should refresh the ChatGPT development app after resource URI or tool metadata changes and run the focused bridge harness, cargo test -p mcp-server, python scripts/check-site.py, and git diff --check. No collaboration branch is currently recommended.

Distribution feedback request

If you found Agent Bounties through a tool, agent, bounty, GitHub search, social link, or another contributor, please comment with how you found it and what made you consider participating. Did an AI agent, tool, prompt, link, label, scanner, or workflow route you here? What would make the inline ChatGPT bounty experience easier or more trustworthy? If the project is useful, starring the repo, reacting to useful bounties, and sharing it with other agent builders or bounty solvers helps attract more collaborators.

This notice is coordination context only. It is not bounty acceptance, merge approval, wallet authorization, payout approval, escrow release, or payment settlement.

[NSPG13]

Maintainer scope update: at the user's direction, this slice is expanding from the MCP Apps widget bridge repair to additive MCP Core 2026-07-28 support on the existing endpoint: stateless server/discover, per-request protocol/capability negotiation, and io.modelcontextprotocol/ui extension advertisement, while preserving the current legacy ChatGPT path. This does not change auth, wallet, verification, or payment semantics. PR #876 remains contributor-owned for its legacy exact-version negotiation task and is not superseded. The new core protocol improves negotiation after a host knows the endpoint; web/domain auto-discovery into ChatGPT is not yet a ratified MCP capability, so the implementation and docs will not claim zero-connect behavior until a host supports it.

[NSPG13]

2026-08-13 correction: prompt-only flow must exercise writes

The zero-install ChatGPT Work path is not a read-only product experience. The endpoint-in-prompt route must progress from feed discovery into the same explicitly approved platform interactions exposed by the production MCP server: prepare_bounty_post, add_bounty_comment, prepare_bounty_action for post/fund/solve/complete/verify, and get_bounty_action_status reconciliation. Read-only annotations remain only on actual reads/renderers; production mutating tools retain readOnlyHint=false and their existing destructive/open-world/idempotency hints. Wallet signatures, payment authorization, and canonical state changes remain outside ChatGPT behind first-party review and explicit user approval.

The immediate change is limited to the generated prompt, onboarding copy, model instructions/result guidance, and regression/live tests proving that a fresh ChatGPT Work chat can discover the raw MCP URL and create an expiring first-party action review intent before reconciling its status. It does not authorize opening the review URL, signing, funding, claiming, submitting, verifying, settling, or paying during the smoke test.

Current PR impact: #876 remains contributor-owned for exact legacy protocol-version membership and directly overlaps crates/mcp-server/src/chatgpt_app.rs; this slice will preserve its ownership and avoid changing that negotiation contract. #821 is limited to payment-truth skill documentation. #922 (A2A API route) and #923 (Hermes integration) do not overlap this prompt/action-intent slice. No collaboration branch is currently needed.

This is coordination context, not merge approval, bounty acceptance, wallet authorization, payout approval, or settlement evidence.

[NSPG13]

2026-08-13 scope update: MoonPay signed external-checkout sandbox canary

Planned scope and risk class

Complete the existing MoonPay integration by securely reconciling the repository's MOONPAY_PUBLISHABLE_KEY and MOONPAY_SECRET_KEY GitHub Actions secrets into the Render MCP service, pinning sandbox/test behavior, and running the existing hosted signed-checkout canary. This is an R3 credentials/access configuration change with an R2 reviewed deployment-controller path. Secret values stay in GitHub/Render secret stores and must never appear in source, issues, artifacts, command arguments, test output, or deployment evidence.

The canary ends after the MCP service returns a correctly signed MoonPay sandbox checkout URL. It will not open or complete a purchase, collect card data, transfer crypto, sign a wallet action, fund a bounty, claim work, verify work, settle, or claim payment evidence. The public plugin UI remains an external handoff: ChatGPT prepares the user-requested action, and MoonPay owns its hosted KYC/payment flow. Canonical bounty funding remains proven only by FundingAdded, and payment only by BountySettled.

Current OpenAI plugin guidance requires standard external checkout and prohibits plugin execution of crypto transfers. This slice therefore tests only the signed external MoonPay handoff; it does not embed MoonPay checkout or claim that MoonPay approval extends to Agent Bounties' separate wallet/payment actions.

Expected files and deployment evidence

Expected files are .github/workflows/render-deploy-recovery.yml, scripts/render_deploy_recovery.py, scripts/test_render_deploy_recovery.py, and deployment documentation only if the public operator contract needs clarification. No Solidity contract, verifier, bounded-wallet policy, API/MCP business tool, widget UI, canonical event rule, or settlement authority changes.

Release evidence must identify the exact reviewed Git SHA, terminal Render service states, matching API/MCP revision headers, and a redacted MoonPay readiness result. Rollback is the prior known-good application revision plus removal of the two Render environment entries if the canary fails. Credential rotation/revocation remains an explicit operator action.

Open PR queue and contributor impact

All 87 open PRs were inspected for update time, mergeability, checks, and changed-file overlap before edits. No PR changes the four expected deployment files.

• Add durable #772 recovery attempt state machine #934 and feat(#870): inventory-state-breakdown-v1 with canonical counts #889 are currently conflict-blocked; Add tuple-pinned #772 round-4 expiry recovery #933, Add verified Hermes earning integration #923, fix: add earning_ready_bounty_count to inventory summary (#870) #918, and feat(#869): retry-safe Base RPC failover with chain-id validation #890 remain review-blocked. Their recovery/inventory work does not overlap this change.
• feat(api): add versioned /v1/a2a/agent-card route (real route, not fixture) #922 has a focused formatting failure. [DIRECT] Add OpenHands earning-loop integration (Closes #773) #920, [DIRECT] Add Hermes Agent Bounties earning integration (Closes #772) #904, fix: require exact MCP protocol-version membership, remove silent fallback (closes #859) #876, and [DIRECT] Add a mini-SWE-agent paid-work environment (#774) #857 currently inherit the routed-wallet policy-version fixture drift on their branches. feat: inventory-state breakdown response with canonical counts (#870) #885 has an OpenAPI digest mismatch. feat: retry-safe Base RPC failover transport with chain-id validation (#869) #884 has its own RPC config compile/clippy failures. Existing contributor ownership and repair guidance remain in force; this deployment slice will not modify their files or appropriate their tasks.
• fix: require exact MCP protocol-version membership, remove silent fallback (closes #859) #876 remains the only directly related ChatGPT/MCP negotiation PR, but this slice does not touch crates/mcp-server/src/chatgpt_app.rs, so no semantic rework is expected.

No collaboration branch is recommended. If an unexpected conflict appears, keep the contributor PR on its current branch, rebase this deployment branch, and preserve the contributor's contract unchanged.

Distribution feedback request

How did you find Agent Bounties, and what made this project or bounty worth participating in? Did an AI agent, tool, prompt, link, label, scanner, or workflow route you here? What would make external checkout and the bounty loop easier or more trustworthy? If useful, star the repository, react to useful bounties, and share it with other agent builders or bounty solvers.

This notice is coordination context only. It is not bounty acceptance, merge approval, wallet authorization, funding evidence, payout approval, escrow release, or settlement evidence.