# pi-ext 1.0 readiness ## Status and evidence boundary Non-npm source, documentation, and useful-demo cleanup is implemented. Npm artifact work, registry/account configuration, release automation changes, version bump and publication are deferred by the human. Physical audio, native Windows/macOS/Android acceptance, and actual IDE/desktop clipboard use are not certified by Linux automated checks. This is a static readiness audit and proposed implementation order, not a release certification or binding specification. All 45 source extension manifests and READMEs were inspected through two independent read-only audits, with targeted runtime, contract-test, demo-scene, and packaging inspection. No complete runtime, platform, browser, native-audio, or release-artifact matrix has been executed for this audit. Root `//:check` and `//:lint` pass after the current safety/accounting batch; focused verification is recorded below. Concurrent `extensions/memory/` changes and unrelated root changelog entries were preserved. The previously approved Mermaid authoring guidance remains uncommitted in `.pi/skills/pi-ext-readme/SKILL.md`. ## Product direction Independent generated extension packages share the repository version and release cadence. Keep source workspaces private; consumer distributions, not development directories, become publishable. Luci remains the sole CI and publication system. No npm account changes, staging, approval, publication, Git push, or infrastructure changes are authorized by this report. Version remains `0.1.10`; bump to `1.0.0` after the release contract and verification are settled. Consumer release notes remain separate from internal root `CHANGELOG.md`. The injected root-only installation law and existing Git-only publishing instructions conflict with the agreed independent-package/npm direction; propose exact maintained-instruction changes for approval before changing them. ## Publication inventory Public registry search returned 25 packages maintained by `bugabinga` on 2026-10-26. Each package's public registry metadata was fetched separately: each has only `0.1.0`, `latest` points there, and no deprecation message was present. All 25 have a registry signature; no `dist.attestations` field was observed. These observations do not establish account permissions, package security settings, or build provenance. The durable metadata inventory is `1.0-npm-inventory.json` beside this report. Nine names match current extension slugs: `context-size`, `cost`, `footer`, `git-safe`, `nu-bang`, `prompt-autopsy`, `quota`, `runtime`, `session-rename`. Sixteen historical names are absent from current source: `angel`, `builtin-tools`, `ctx`, `cwd`, `devil`, `diff-review`, `ghost`, `git-status`, `inline`, `llmiterate`, `lucifiber`, `neko`, `prune`, `pulse`, `q`, `ssh-tools`. All names above use the prefix `@bugabinga/pi-ext-`. The other 36 current package names were absent from this maintainer search; that is not proof of name availability. Do not automatically resurrect, rename, deprecate, or unpublish historical packages. ## Shared release blockers 1. `scripts/build/manifest.ts` hardcodes private generated packages and does not emit public repository, discovery, or publication metadata. 2. `scripts/build/build.ts` copies declared runtime resources, not human READMEs and their linked assets. 3. Source manifests declare MIT, but no repository or extension license text was found; bundled dependency notices need a deliberate delivery policy. 4. `scripts/build/release.ts` creates ordinary tar archives of private directories plus a root collection, not consumer-tested npm tarballs. 5. `scripts/release/index.ts` requires Git destinations and per-extension eligibility/scenario declarations; no current extension opts into that path. 6. Snapshot receipts currently bind bytes, not successful consumer execution; including verifier files is not equivalent to running verification. 7. `ephemlets` bakes an external `jiti` tree while `activity` resolves a native PTY dependency at consume time; npm packing/installation must prove both behaviors rather than relying on directory-copy tests. 8. `.ci/ci.kdl` distributes archives inside its successful child before the later tag-only compatibility job completes; these are staging artifacts, not evidence that every release gate has passed. 9. The declared compatibility matrix is Linux with Pi 1.4.0; Windows/macOS and native-device verification remain explicit gaps. 10. Ultra fanout failed before auditing because a configured child extension referenced absent `dist/klaus`; independent advisors completed the static audit instead, without changing user configuration. ## Extension coverage All rows include README and manifest inspection plus sampled public contracts and relevant implementation paths. "No additional blocker" means none demonstrated by this bounded static inspection; it does not mean runtime certification. "Existing demo" means scene/cast/GIF exist, not that the recording was visually verified or re-executed. Shared release blockers apply to every row and are not repeated. | Extension | Concrete follow-up or evidence gap | Demo treatment | | --- | --- | --- | | activity | Relocated package load and real-Pi process/PTY lifecycle passed; external npm-install proof deferred | Recorded lifecycle with observed completion/stop states | | afk | Relocated load passed; real-Pi bounded checklist/reviewer/return flow passed | New recording; meaningful napping/settled frame inspected | | ask | No additional blocker demonstrated | Existing chooser demo | | bak | Fixed invalid-setting fallback for storage and automatic uploads; documented complete-session backup content | Diagram useful; no forced terminal recording | | canvas | Relocated assets/authenticated state passed; 27 Chromium browser tests passed; first-use guide improved | Genuine Firefox drawing screenshot, saved native scene | | chrome-cdp | Unknown arguments rejected; close completion protocol verified | Existing browser guidance retained | | clipboard | Real desktop integration remains unrun | Existing synthetic-clipboard demo | | context-size | Corrected headless-status claim | Reuse footer demo | | continuity | Attribution corrected; selected real-Pi scene passed | Existing useful recording retained | | cost | Fixed full-session native usage and timestamp-based monthly accounting; documented bounded response-only monthly scope | Reuse footer demo | | dilemma | Caption reflects agent-requested investigation; scene passed | Existing recording retained | | ephemlets | Relocated guidance/native compilation/starters/delegated builder passed | New real-Pi choice/Submit/follow-up recording | | fast | Enabled/default request bodies verified at registered transport with synthetic account token and zstd decoding | No forced demo | | firefox-bidi | Dead ledger removed; raw BiDi behavior retained and focused tests passed | Browser protocol remains exercised by browser verification | | footer | Scene distinguishes cumulative cost from context occupancy | Updated recording; meaningful frame inspected | | git-safe | Safety fixes retained; negative-control check rejects false execution condition and verifies independent checkout side effects | Real enabled-hook control versus safe clone; refreshed recording | | good-job | Acquisition/read barriers and shared module-reload ownership protected; reclaim ENOENT handled | No extra recording needed | | gratis | Provider availability/billing remain external conditions, already disclosed | No forced demo | | handoff | Failures/cancellation protected; real TUI changed session IDs and retained reviewed text without automatic submission | Refreshed recording and workflow diagram; replacement-editor frame inspected | | horst | Added model-provider disclosure for injected host facts | No forced demo | | intellij | Launch completion verified through registered command; actual IDE launch remains external acceptance | Setup guide retained | | masks | No additional blocker demonstrated | Existing picker demo | | memory | Unrelated implementation/demo work preserved; relocated package load passed; source SQLite/FTS5 tests remain distinct from npm-install proof | Existing separately authored recording preserved | | mockup | Relocated load plus Chromium/Firefox select/pin/iterate/accept passed | Genuine annotated Firefox screenshot embedded | | model-info | No additional blocker demonstrated | Reuse footer demo | | notes | Caption now accurately describes idle parking/review/send; scene passed | No fabricated busy-flow claim | | nu-bang | Corrected `!!` claim: context exclusion, not command repetition | Existing recording remains useful | | nushell | Scene and recording synchronized; actual off/on notifications and fresh Nu pipeline output checked | Re-recorded with readable holds | | open | Corrected editor and terminal prerequisites | Pane screenshot optional | | os-notifier | Relocated module URL/resource test fixed and passed on Linux; no Android-device execution claimed | No forced demo | | pfui | Actual bundled fallback policy linked and described | No forced demo | | prompt-autopsy | Real Neovim displays a synthetic prompt; post-exit unique assistant reply proves resumed Pi input; editor-failure warning rejected | Repeatable focused editor-check task | | quota | Dismissal, A→B→A ordering, cache fallback, and headers-versus-poll regressions passed | New clearly synthetic allowance recording; meaningful frame inspected | | rtk | README describes supported registered shell tools instead of claiming Bash-only interception | Existing recording retained | | runtime | Native home fallback/Windows profile boundary covered with platform-simulated tests, not native Windows | Reuse footer demo | | session-rename | Added background model-call privacy/usage disclosure | No forced demo | | skillz | No additional blocker demonstrated | Inclusion/exclusion example optional | | the-system | No additional blocker demonstrated by sampled contracts | Existing lifecycle demo; board example and lifecycle diagram useful | | tps | Registered-runtime throughput/reset/empty/error status coverage added and passed | Reuse footer demo | | ultra | First workflow foregrounded; trust/cost/persistence/cooperative-cancel limits preserved; checkpoint/fresh diagram validated | Existing fanout recording retained | | usage | Accounting fixed; in-memory sessions now read configured store instead of cwd; token formula corrected | Real synthetic-provider/model dashboard recording, meaningful table frame inspected | | uv | Quoted Python argument literals and Windows executable spellings covered by interception regressions | No extra recording needed | | vcs-status | Per-refresh generation prevents stale publication; controlled overlap regression passed | Reuse footer demo | | voice | Portable URL conversion fixed; native speech/service/hardware acceptance still requires approved setup, not simulated proof | No fake acoustic recording | | web | No additional blocker demonstrated in sampled fetch/search/map paths | Search-to-source example optional | ## Prioritized implementation order ### 1. Safe preparation, started - Correct source-disproved README claims and add omitted root README entries without cosmetic rewrites. - Make release planning genuinely read-only and validate an entire snapshot selection before building. - Reject manifest-listed symlinks before reading their targets and remove temporary stages on failure. - Preserve existing user work and record remaining scope honestly. ### 2. Safety and truthful observable behavior - `bak`: replace invalid-value defaulting with explicit invalid state; invalid opt-out must not upload, invalid destination must not select another bucket. - `git-safe`: preserve intended replacement of marked extension-owned disposable clones; fail closed on inspection errors, bound output, retain queued-target overlap protection, and distinguish checkout limits from download-size guarantees. Safety fixes, negative-control demonstration, and independent filesystem postconditions are implemented and passed. - `handoff`: implemented rejection of failed/empty generation before review or session replacement; actual loader abort now ignores all late provider settlement. Registered-command tests verify failures/cancellation; a separate real TUI scene verifies distinct native session IDs, reviewed text in the replacement editor, and no extra model submission. The README recording remains an unsaved-session example. A separate saved-session check verifies an unsent reviewed draft at two model calls, then explicitly submits it and inspects the persisted child's reference to its original parent file. - `cost` and `usage`: implemented agreed scope: full-file native session usage in the cost footer, persisted assistant responses for monthly/dashboard statistics, identity-based fork deduplication, no repeated parent tool aggregates. - Implemented controlled good-job acquisition/module-ownership, quota provider/cache/header freshness, and VCS overlapping-result regressions. - Add `fast` dispatch-boundary coverage, then address bounded Windows/Android/editor and command-classification findings. Concrete source evidence includes `extensions/bak/index.ts` settings resolution, `extensions/git-safe/marker.ts` and `clone.ts`, `extensions/handoff/index.ts` response handling, `extensions/cost/index.ts` cost/month traversal, and `extensions/usage/implementation.ts` response deduplication. The queued-target `git-safe` race was reproduced with controlled subprocesses before fixing it. Good-job acquisition barriers, quota A→B→A/cache/header ordering, and VCS overlapping-refresh behavior now have controlled regression coverage. This remains test evidence, not a claim of observed production incidents. Do not convert this list into speculative rewrites or new features. ### 3. One npm artifact canary, then generalize - Settle release selection, package names, license text, and the conflicting maintained architecture instructions. - Use a fresh release staging directory, one root version, and an explicit public distribution manifest while keeping source workspaces private. - Include consumer README, working image/reference links, runtime resources, license and required notices, and `pi-package` discovery metadata. - Produce a real npm tarball and install/discover/use it outside the checkout with the supported Pi host. - Exercise one simple package first, then external/native dependencies, browser assets, skills/guidance, and worker entrypoints. - Bind successful verification receipts to source revision, package name/version, exact tarball identity, and check outcomes. - Publish the already verified bytes later; do not rebuild/repack during eventual registry delivery. ### 4. Human docs and demonstrations - Preserve existing accurate first-use guides; simplify genuinely dense guides such as canvas and ultra. - Correct existing scene claims and recipes before generating replacement recordings. - Prioritize useful missing examples for activity, AFK, canvas, ephemlets, mockup, usage, quota, and memory when its active work settles. - Use browser examples for browser behavior and real service evidence for speech; no decorative terminal demos. - Reuse footer demonstration for telemetry rather than manufacturing near-identical per-extension recordings. - Use Mermaid only for user workflows, decisions, or observable behavior. ### 5. Luci staging and delivery design - Keep build/test execution separate from narrowly scoped publishing credentials. - Configure no secrets or account settings until exact changes are approved and npm 2FA enrollment is available. - Resolve stage approval interaction and recovery behavior without promising bulk/atomic publication. - Track per-package results, registry integrity, retries/backoff, and resumable partial success. - Stop on an existing name/version with different bytes rather than skipping blindly. - Treat signed release manifests, registry signatures, and native npm provenance as separate claims. - No GitHub-hosted publishing migration. ### 6. Final release boundary - Finish intended work, commit logical changes, and verify clean index/worktree without absorbing concurrent unrelated work. - Bump the root source version once to `1.0.0`; verify every selected generated distribution carries it. - Write separate user-facing release notes covering features, config, bugs, and required actions. - Run only the verification gates explicitly requested at integration; full CI/E2E requires its own authorization. - Require successful Luci evidence for the exact selected revision before final registry promotion. - Announce only packages whose publication and registry identity were confirmed. ## Proposed 1.0 acceptance criteria - [ ] Every selected extension is independently installable and discoverable through its delivered Pi package, without source-checkout imports or undeclared sibling dependencies. - [ ] Every selected distribution has the same root-owned `1.0.0` version; source workspaces stay private. - [ ] Registry names and existing-package ownership/security requirements are explicitly verified before any publication. - [ ] Actual npm tarballs contain only intended runtime/resources/docs/licenses/notices; consumer links and gallery metadata are usable. - [ ] Exact packed artifacts pass meaningful relocated consumer checks, including the external/native/browser/worker/skill cases relevant to their declarations. - [ ] Confirmed safety, accounting, generation-error, and supported-platform defects are fixed with decisive regressions; unconfirmed races are reproduced or explicitly resolved. - [ ] READMEs explain purpose, first use, necessary setup/config/privacy/limits and reflect source behavior; generated config remains authoritative. - [ ] Useful demonstrations are real, reproducible, and accurately described; omitted examples have a user-value reason rather than a blanket quota. - [ ] Voice is either backed by an obtainable compatible service and truthful acceptance evidence or explicitly held from the release selection by the human. - [ ] Luci delivery is credential-separated, integrity-checked, resumable, and tested for partial success without live registry mutation during development. - [ ] Consumer release notes are agent-authored, human-readable, and written at release time rather than copied from the internal engineering changelog. - [ ] Final commits, exact-revision CI evidence, clean worktree/index, and individual publication receipts are confirmed before declaring the release complete. ## Executed verification - Latest complete root `//:test`, after all implementation workers settled: 3,429 passed, 73 skipped; 270 test files passed, 10 skipped; 18 modernization tests passed separately. - Latest root `//:check` and `//:lint` passed, including all 45 README/config references and current TypeScript/static checks. - `//extensions/git-safe:demo-check` rejected the deliberately false execution condition, then verified the enabled-hook control marker, safe-clone files/ownership marker, and absence of safe-clone hook side effects. - `//extensions/prompt-autopsy:editor-check` displayed a unique synthetic prompt in real Neovim, exited it, then received a unique scripted assistant reply from fresh Pi input; `editor failed` is rejected. - Real handoff TUI evidence recorded two distinct native session IDs, the edited constraint in the replacement editor, and exactly two model calls with no automatic submission. - `//extensions/handoff:demo-check` additionally verifies saved parent linkage after an explicit test submission; the pre-submission checkpoint remains at two calls, and the persisted child contains the reviewed constraint. - `//extensions/clipboard:demo-check` rejects a missing fixture before model/tool input under a desktop-disabled negative control, then passes the isolated copy/read round-trip. - All 20 configured README demo scenes passed the complete assert-only scene run after fixture-source binding was restored. - Selected real-Pi scenes passed for activity, AFK, ephemlets, usage, quota, git-safe, handoff, continuity, dilemma, footer, notes and nushell; changed/new useful recordings were regenerated. - Meaningful GIF frames were inspected, including actual AFK settled cat, nonempty usage provider/model table, quota widget, ephemlet selection submission, handoff replacement editor, real hook-comparison output, activity lifecycle, footer and Nu pipeline output; no frame-by-frame or universal visual certification is claimed. - Canvas: 27 Chromium browser tests plus real isolated Firefox drawing/download evidence; Mockup: 2 Chromium/Firefox select/pin/iterate/accept tests plus real annotated Firefox evidence. - Focused relocated checks passed for canvas browser assets/state, ephemlet guidance/compiler/starters/delegated builder, and os-notifier Android resource/module resolution on Linux; activity/AFK/memory/mockup/clipboard package loading passed outside the checkout. - New quota controlled runtime tests protect A→B→A publication/cache fallback and response headers against older polls; event payloads still do not identify a hypothetical late same-provider native response after A→B→A. - Good-job barriers protect acquisition before rename completion, pending reads, failed reads, independently evaluated storage modules, and concurrent completed-claim reclaim; actual two-cwd Pi runtime execution remains distinct from storage-module re-evaluation proof. - Handoff `mise run //extensions/handoff:test`: 16 tests passed, plus 18 modernization tests. - Handoff regressions first reproduced four invalid-draft cases and two late-cancellation error notifications; fixed code passes the unchanged deferred rejection/error/success scenarios and ordinary provider-failure reporting. - Replacement-editor delivery was first verified through a callback double, then through real TUI execution with two distinct native session IDs, the edited constraint, and exactly two scripted model calls. - Safety/accounting focused suite `mise run //:test extensions/cost extensions/usage extensions/bak extensions/git-safe`: 184 tests passed, plus 18 modernization tests. - Final root `//:check` and `//:lint`: passed, including typecheck, static extension checks, 45 README/config checks, Biome, rumdl, Taplo, frontmatter, and skill checks. - Empty-repository behavior is verified against real Git, not only a fake subprocess fixture; subprocess coverage also preserves UTF-8 characters split across output chunks. - New regressions first exposed the queued-target overlap gap, checkout after failed metadata inspection, failed empty-repository checkout, and unbounded subprocess output; all now pass. - Backup settings tests assert invalid configuration causes no storage requests; cost tests protect full-session history after branching and a nonzero month below lifetime session spend. - Baseline and post-cleanup `mise run //:check`: passed, including typecheck and all 45 README/config checks. - Earlier post-cleanup root `//:lint` was blocked by concurrent `extensions/memory/__live__/review.ts:72` (`noConsole`); that file was left untouched. - An initial new `releasePlan()` return-inference type error was fixed with its explicit exported return type; the subsequent root check passed. - Inventory JSON formatting was corrected through `//:format-files`; at that point the remaining root lint failure was solely the unrelated memory file. - `mise tasks validate`: passed after release-plan task change. - `mise run //:format-files` on changed release code/tests: passed. - `mise run //:test scripts/release scripts/readme`: 59 tests passed, plus 18 modernization tests. - Snapshot regressions exercise `snapshot()` with internal, external, and dangling manifest-listed symlinks; assert no target content reads, no retained stage, and no receipt. - CLI regressions run the real command in a temporary repository with a controlled build boundary; cover successful read-only plan and eligible-first/ineligible-second preflight without artifact changes. - `mise run //:check-readmes`: all 45 passed after README corrections. - Handoff Mermaid validated through installed renderer: 54 columns, no warnings, within explicit 80-column validation budget; GitHub rendering not independently inspected. Full CI, full browser matrix, native-audio/device acceptance, actual npm installation, account changes, version bump, and publication were not run. Focused canvas/mockup browser checks and selected real-Pi demo regeneration were run. A locked dependency reinstall was necessary after concurrent task installs left incomplete bindings; no manifest/lockfile/dependency-version change was made. The complete dependency audit retains exit code `1`, stderr and affected node paths in `1.0-dependency-audit.json`: 21 advisories, including one critical. `1.0-dependency-risk.md` records static DOMPurify delivery/mode analysis, including a prebundled copy in The System; the specific attached-shadow-root exploit combination was absent in inspected call paths, not proven generally safe. Security findings remain open; suggested Excalidraw downgrades were not applied and tests are not security clearance. The green focused suite does not close the shared release blockers or certify all extension behavior.