DottaandPaperclip
9e09d8f51f
## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - First-task onboarding requires a saved human proposal decision before child work can start. > - Native proposal tools reject interaction metadata placed outside their advertised schema. > - The existing rejection text focuses on question forms and does not explain the misplaced approval metadata. > - This pull request explains the field placement and keeps approval recovery in its original confirmation or checkbox kind. > - It also retains safe schema guidance in the run log so operators can identify the rejected field. ## Linked Issues or Issue Description Refs: #18025 for native first-task proposal recipes and #16889 for safe native schema guidance. Related open PR #18105 changes separate server API-client guidance. **What happened?** A native first-task confirmation copied `version` and `supersedeOnUserComment` to the tool root. Both belong inside `payload`. The runner correctly rejected the call before execution, but its feedback described question forms. Later calls kept the misplaced field and changed the approval into a question. No proposal card was created. The run log retained only a generic schema error. **Expected behavior** Keep strict validation. Explain the nested field locations. Keep a confirmation or checkbox approval in that kind while correcting its arguments. Save the normal pending card and wait for an actual human decision. Retain bounded schema guidance without storing rejected input. **Steps to reproduce** 1. Send an otherwise valid native confirmation with root-level `version: 1` and `supersedeOnUserComment: false`. 2. Observe the pre-dispatch `invalid_tool_call` rejection. 3. Move both fields into `payload` and keep `interactionKind: "confirmation"`. 4. Verify that the corrected call creates one pending confirmation, exact replay creates no duplicate, and child creation still needs saved human acceptance. **Paperclip version or commit** The observed argument shape and initial regressions came from public source base `c32e1ba6`. The repair is based on maintained master `4a6f6a65`. Public tests use synthetic prompts and identifiers. ## What Changed - Give rejected confirmations and checkboxes approval-specific guidance. It names `payload.version` and `payload.supersedeOnUserComment` without suggesting a question-form recipe. Other human-input errors retain complete question-form guidance and stable question/option ID instructions. - State the allowed native root fields in the first-task skill and put the metadata inside both executable proposal recipes. - Retain the existing safe guidance in `semantic_tool_denied` diagnostics. Keep the original rejection identity and `paperclipExecuted: false`. - Preserve only the exact bounded scalar diagnostic payload after verified event admission. The rest of the envelope and all API request/result bodies keep their existing redaction. - Add a real native Codex transport regression and strengthen advertised-schema, card replay, human-admission, persistence, and privacy controls. Update the execution contract and onboarding maps. ## Verification - Red: the native transport reproduced the malformed confirmation and showed that the old hint omitted the metadata correction. A second red confirmed that approval guidance still included the question-form recipe. The corrected transport case now rejects before execution, retains a safe hint, and dispatches one corrected confirmation that returns a pending card. - Red: both shipped recipe checks lacked explicit nested metadata. All 49 first-task admission tests now pass, including advertised-schema validation, idempotent card replay, actual human acceptance, agent-only rejection, and question-answer controls. - Updated native checks pass on the current repair: 95 native Codex transport tests with 2 existing ignored cases, all 32 provider-bridge tests, and the schema-feedback unit control. - API/case persistence regressions were red before the event-boundary correction. All 71 affected port/redaction tests pass in a clean run at frozen head `3be3a70f`, including the corrected envelope privacy regression. The focused 16-test persistence/privacy group and server typecheck also passed earlier. - Full local `pnpm build` and `pnpm -r typecheck` pass at exact head `3be3a70f` in a clean detached checkout with Node 24.22.0 and pnpm 9.15.4. Exact-head CI passes: [full PR CI](https://github.com/paperclipai/paperclip/actions/runs/40950360142) and [isolated native Runner readiness](https://github.com/paperclipai/paperclip/actions/runs/40950359835). - The full local unit attempt recorded a 15-second timeout in the unrelated Microsoft Teams 100,000-character existing-comment publication case. That failure remains recorded. The exact isolated rerun passed (1 test; 1,814 unselected). The aggregate timeout cause is unproven. That long local run straddled subsequent source edits and also observed the two port regressions during their red phase; it was stopped and is not final-head evidence. The clean affected-suite run above and current-head CI provide the final verification. No timeout or assertion was weakened. - Release-registry CI initially rejected the native test's compile-time schema include outside the release crate's hashed input closure. The guard was reproduced locally. The test now reads the current canonical schema at execution. All 10 crate-cache controls and the real native approval regression pass afterward; the cache guard and workflow keys remain unchanged. - Current-head Greptile review is 5/5 with no unresolved review threads. All required checks pass (56 success, 4 skipped). The advisory Paperclip Review remains neutral/incomplete after one normal re-review request; no completed assessment or score is claimed. - `cargo fmt --check`, map structural checks, and `git diff --check` pass. - Source and scripted-provider checks do not prove that a model will follow the guidance. The original live onboarding attempt remains failed; acceptance at a deployed revision is separate work. ## Risks - This changes repair guidance and shipped instructions. It does not make malformed arguments valid or supply a human decision. - Diagnostic retention adds no rejected arguments, values, or unknown input keys. The API receipt allowlist is unchanged. The validated-event exception restores only the safe scalar payload; optional debug and other envelope metadata still pass through redaction. - The existing 512-byte guidance bound, strict tool schema, company/run binding, idempotency checks, and proposal admission remain effective. ## Model Used OpenAI GPT-6.1 Sol (`gpt-6.1-sol`) through Codex, with tool use and code execution. The context-window size is not exposed to this session. ## Checklist - [x] I have included a thinking path that traces from project context to this change - [x] I have specified the model used (with version and capability details) - [x] I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work - [x] I have searched GitHub for duplicate or related PRs and linked them above - [x] I have either (a) linked existing issues with `Fixes: #` / `Closes #` / `Refs #` OR (b) described the issue in-PR following the relevant issue template - [x] I have not referenced internal/instance-local Paperclip issues or links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip` URLs) - [x] My branch name describes the change (e.g. `docs/...`, `fix/...`) and contains no internal Paperclip ticket id or instance-derived details - [x] I have run tests locally and they pass - [x] I have added or updated tests where applicable - [x] I have updated relevant documentation to reflect my changes - [x] I have considered and documented any risks above - [x] All Paperclip CI gates are green - [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge --------- Co-authored-by: Paperclip <noreply@paperclip.ing>
15 KiB
navigation
| navigation | ||||||
|---|---|---|---|---|---|---|
|
Onboarding and first work
An operator can start an instance, create a company, configure its first agent, and reach a real first task. Instance setup and company onboarding are separate steps.
Feature ID: onboarding.
Implementation: web wizard, CLI setup.
Agent access: companion.
Sub-features
-
instance:onboarding/instance— choose local instance configuration and start the server. -
company:onboarding/company— create a company with its identity and initial agent. -
adapter:onboarding/adapter— choose an available harness, model, account, and execution environment. -
first-task:onboarding/first-task— leave setup with a task whose execution or missing prerequisite is visible. -
resume:onboarding/resume— return to an incomplete wizard without losing saved state.
The Cloud Product onboarding journey defines portal workspace creation, API-key connection, native Codex creation, and a first downloadable artifact with an independent oracle. Its dependent task-properties journey reuses onboarding state. This is a scripted coverage recipe, not a claim of a completed Cloud run or strict browser-only discovery.
How to get to it (user POV)
web-wizard
Open /onboarding or the onboarding launcher in an empty company context.
Use /<company-prefix>/onboarding to resume an existing company's agent setup;
the unprefixed route starts at organization naming on self-hosted instances.
Cloud-managed instances reuse their existing company for agent setup.
cli-setup
Use paperclipai onboard --help or paperclipai run --help to select an isolated data directory before setup.
Driving it with Playwright and the Paperclip CLI
Preconditions: follow the baseline. Use a new isolated data directory and record deployment mode. A configured adapter may still need account authentication.
web-wizard
Automated: wizard tests cover mocked wizard behavior; onboarding browser suite supplies its own test environment. Live model execution is a separate check.
The published canary smoke checks the exact installed version through Connect a model. Submitting the agent name must save one paused setup record with disabled heartbeats and no task or run. The authenticated release smoke checks the same held state, then completes that record through Connect without creating a second agent. Its Docker harness mocks credential verification; live provider authentication remains separate evidence.
Name-only setup defers the runner field while the provisional adapter remains held in setup. Choosing a provider updates that same setup record with its selected runner. The wizard regression covers a saved native choice followed by OpenAI Subscription selection; it does not prove provider authentication.
- Create company and resume. Start a fresh empty instance and run doctor using
the verification skill. Run
node .agents/skills/verify-paperclip/scripts/onboarding-smoke.mjs --evidence-dir "$VERIFY_EVIDENCE_DIR". The helper opens/onboardingin a fresh self-hosted instance, enters a unique name ine.g. Northwind Labs, and choosesContinue. The agent-name field must appear and the saved company must have no invented goal. Reopen the company-prefixed onboarding route and reload; the agent step must appear without creating another company. Screenshots and a trace capture the real UI action and result. This proves company creation/resumption, not hiring or a first model run.
Manual: Create a disposable company and first agent. Leave midway, reopen the company's onboarding route, and finish; drafts save automatically. Follow the created first task, verify the selected agent and company, and inspect its output or explicit authentication gate. Reload both company and task.
For Codex or Claude, choose the harness and leave runner selection automatic to prove the first-agent default. On a qualified target it saves Paperclip Runner; Advanced also permits Legacy. Do not switch runners after onboarding when claiming proof of that default. Test the selected runtime and account, then complete a bounded useful task and follow-up with an independently checked file artifact. Saved execution configuration and recorded run identity must match after reload. The Codex first-task eval applies operator QA bounds before Get started through the existing agent configuration API. It verifies that only timeout and concurrency change; automatic runner, account and environment choices remain. These campaign bounds are separate from production creation defaults. Use runner validation and environment qualification for the prerequisites; a successful login alone is insufficient.
For managed Cloud setup, an outdated default Claude CLI is prepared automatically in a private versioned runtime directory before its hello test. Verify that the selected model and account stay unchanged and that the compatible executable is actually used. A failed installation still blocks hiring. Cloud failures must not display Manual debug shell commands or legacy operator hints; self-hosted failures retain that help. Wizard controls and runtime controls cover these boundaries. They do not establish live Cloud onboarding success.
For a new Claude agent on an explicit immutable sandbox image or a Daytona named snapshot, use Connect to qualify the selected environment through the existing permission-gated probe. Do not preprobe outside onboarding when claiming this first-use path. Fresh scoped discovery determines the runner for both the account test and hire; the measured Legacy result remains valid. An instance-admin denial, incomplete cleanup, missing lease or changed configuration stops setup before the adapter test or hire. Daytona resolves the selected snapshot to its immutable provider ID before creating the measured lease; native execution pins that same ID. Mutable image tags and other providers’ named templates remain unqualified. Helper controls and mounted wizard/new-agent tests use mocks; live managed acceptance remains a separate gate. For a new Codex agent with Auto on a saved immutable OCI sandbox, Connect first uses the existing saved environment probe when authoritative qualification is absent. The board actor needs the existing instance-environment administration permission. Refresh company/environment-scoped discovery before testing the selected runtime. A denied probe, unconfirmed cleanup or setup failure must be visible; selecting Legacy remains an explicit choice. Creation and discovery themselves allocate no target. A valid measured unsupported target follows the server's Legacy default. Named Cloud snapshots remain a separate unqualified boot-identity and fresh-signup gap.
Selected native setup derives its target isolation as task execution does. The hello probe retains its read-only permission profile, approval-never policy and bounded provider turn. It skips Codex's project-instruction discovery in its new, empty private workspace; ordinary tasks still load their project instructions. Local and SSH transports retain Codex isolation. An injected environment flag cannot establish an outer sandbox identity. Verify these boundaries with the existing native setup controls and a real managed task/follow-up; source checks alone are insufficient.
The explicit manual Product E2E cell
confirmation-replies.runner-codex.daytona.task-card-accept covers this fresh
managed wizard path without fixture agent creation or a fixture preprobe. Its
existing attachment oracle checks a downloaded physical file and follow-up
continuity. Setup evidence must correlate both lease IDs with the same company,
environment and provider; missing responses or unmatched rows keep cost coverage
unknown. Run this cell only on an eligible target with separately authorized
provider, runtime and cleanup reservations. Its source and mock controls are
prepared. The literal ea6c8eb5e16d61111968f633a66748b397b64a3e live attempt
failed: the child document existed, but the physical file was attached to the
parent, and recovery stopped the completion callback before a provider turn.
The failed report remains evidence; the revised source still needs live proof.
Local cases and automatic selection remain unchanged.
First-task task creation enforces proposal acceptance at the shared server
boundary. Verify that native create_task and both REST creation endpoints reject
an authenticated onboarding run before its latest proposal has a saved human
acceptance, including an explicit standalone parent. Then accept a confirmation
or the current plan revision and verify creation succeeds. Opening answers,
agent-only acceptance, stale plans and newer pending/rejected proposals must not
admit new work. Admission controls
exercise native and REST calls against disposable PostgreSQL; live model and
Cloud onboarding remain separate evidence.
For a native proposal, reject root-level version or supersedeOnUserComment
without creating a card. Correct them inside payload, preserve the confirmation
or checkbox approval kind, and verify one pending card after idempotent replay.
An answered question form still cannot authorize the proposed child task.
After acceptance, reject a new task whose blockers include its onboarding source. The rejected command must leave no task, dependency, wake or creation receipt; a corrected retry must still create the ready child. Preserve valid blockers for other work. Check equivalent UUID spellings and a later child PATCH as well as native dependency edits, including the child's own run with a waiting parent and no first-class dependency yet. Preserve normal API attempt auditing and allow blocker corrections without a new proposal acceptance. The shipped handoff recipe copies the actual approved inputs and constraints into the child, then makes the parent depend on the child. Its schema and persistence controls do not prove that a live model follows it.
Native completion also requires a saved human proposal decision. An opening answer alone must leave the first task in review. Verify that premature native completion is refused, including a result already saved before finalization. Accepted proposals and acknowledged human rejections can close the first task; rejection still does not authorize new tasks. The native completion controls exercise these boundaries against disposable PostgreSQL. These controls do not identify a historical Cloud failure or establish live onboarding success.
For accepted file-producing work, verify exactly one self-assigned child with a self-contained description. The onboarding run must yield through the existing dependency/block disposition so the child's own execution can publish its file. Native file registration remains bound to the executing task. On completion, the original conversation must receive the actual child and downloadable file links. Document-only child work may finish inline; a plan-only request must not start implementation. Existing first-task asset controls check the instructions, not model compliance. Inspect the child's execution, document, downloaded file and source-correlated completion reply in the real journey.
Leave and resume during harness/runner setup. Confirm the draft retains the explicit choice and compatible account bindings, old native Codex setup links recover the harness and runner, and a changed harness or runner invalidates stale test results. Existing Legacy agents remain Legacy when their onboarding/setup configuration is reopened; this journey does not migrate them.
Reload both /onboarding and the company-prefixed route after the first agent is
saved but before Get started. The authorized draft must retain its current step
and that agent's execution choice. An explicit new-company opening still starts
fresh. Opening another company must clear the previous company's resource IDs;
an unauthorized account must not restore its draft. Existing component controls
cover these boundaries; they do not substitute for the managed first-task eval.
The existing mocked-login browser case reopens and reloads Connect directly. It
checks the saved company/name/harness/runner, exact active authorization session,
one login start and no hire while login is pending. This is login UI evidence,
not a provider or managed-environment qualification.
cli-setup
Automated: CLI onboarding and isolated test drive cover their setup contracts; they do not establish that every model account is usable.
Manual: Run setup in the disposable directory, open the URL it prints, and confirm the instance identity and company. Restart using that same configuration and verify saved state. Exercise an unavailable adapter and confirm an actionable error.
Gotchas
- An empty company, an empty instance, and an unauthenticated browser are different states.
- Creating an agent does not prove its provider login or first execution succeeded.
- See access for authenticated deployments and agents for later hiring.
Storybook verification
resume
Capability ID: onboarding/resume.
Scope: ui Storybook: Restored incomplete wizard; invalid or unauthorized saved company; temporarily withheld draft; account switch; storage failure. Production: Reload company-prefixed onboarding without duplicating company; restore only authorized account/company draft and retain input after recoverable loading failure. Sources: Source: OnboardingWizard.tsx, onboarding-state.ts. Tests: OnboardingWizard.test.tsx, onboarding-state.test.ts, onboarding.spec.ts
instance
Capability ID: onboarding/instance.
Scope: cli Storybook: No authored UI story required for host installation and startup. Production: Configure isolated data directory, start instance, confirm printed URL/identity/health, restart with saved configuration. Sources: Source: onboard.ts, onboard-service.ts. Tests: onboard.test.ts, onboard-service.test.ts, test-drive.test.ts
Documentation scope: Login, invitations, and access
Deployment modes and core login/invite/member/token paths are mapped; exhaustive identity-provider, expiry, role, and resource-policy combinations remain.