Changelog¶

All notable changes to the papio daemon and CLI are documented here, keyed to v* release tags. The browser extension is versioned and released independently (ext-v* tags): from extension 0.3.1 onward its changes live in extension/CHANGELOG.md. Through [0.3.0] the two shared a version stream, so older sections below include extension entries. The initial release entry is synthesized from the complete papio and zotio Git histories and the execution records in notes/acquisition-stack-plan.md.

[0.10.1] - 2026-07-23¶

Fixed¶

• Direct-distribution release manifests now record the bundled zotio version, rather than incorrectly repeating the papio release version.

[0.10.0] - 2026-07-23¶

Added¶

• Zotero exception-tag ledger (zotio.exception_tags, off by default): the daemon now answers "is this item coming, or is it mine now?" inside Zotero itself by maintaining two reconciled automatic tags on linked items — papio:needs-action (acquisition parked on a human action) and papio:unavailable (OA and institutional routes exhausted, as of the last attempt). Desired state combines each item's newest job with a live exact-key attachment check, so a PDF attached manually clears the marker. Schema v14 rolls the v13 applied-state ledger forward with personal-library provenance plus pending/owned/foreign/missing-target state; per-item mutations survive partial failures, serialize concurrent passes, and never claim or remove a same-name manual tag. papio zotio tags reconcile runs one pass on demand. Requires zotio ≥ 0.13.0 (items tags add --automatic, items tags remove --automatic-only); personal library only. Turning the feature off cleans up papio-owned tags. Design record: zotio dev/adr/0004.
• Backfill re-checks unavailable items after a cool-down (zotio.unavailable_recheck_days, default 14) instead of retrying them on every watch cadence — availability drifts upward (green-OA deposits, holdings changes, new adapters), and a papio:unavailable saved search doubles as an ILL/manual worklist only if it neither rots nor flaps.

Changed¶

• papio init autodiscovers more of its own answers: the zotio executable is resolved from PATH and labeled with its source, prompt defaults name where they came from (keep current vs from Zotero), the Browser step names the supported browsers and lists those detected on the machine, the Chrome extension ID prompt and --extension-id accept an unpacked extension folder path and compute the ID from it (macOS/Linux; Windows requires the literal ID), and the browser setup instructions are suppressed when doctor reports a healthy, current extension connection.

Fixed¶

• Upgraded golang.org/x/text to 0.39.0, fixing GO-2026-5970 (an infinite loop on invalid normalization input reachable through PDF inspection).

[0.9.2] - 2026-07-22¶

Fixed¶

• papio doctor no longer fails against a healthy daemon whose daily update check has results: the strict ping decode now accepts the optional update_available/latest_version fields (and their zotio counterparts).

[0.9.1] - 2026-07-22¶

Changed¶

• Doctor's extension check is floor-aware: a connected extension below the daemon's minimum version now WARNs with the exact skew and remediation instead of reporting a bare connected PASS while every handoff refuses.
• papio init's guided flow was redesigned: numbered step sections with one-line explanations, one question per prompt, a single [Y/n] grammar, middle-ellipsized long defaults, per-browser setup instructions, a doctor summary that shows only checks needing attention, and a closing Next: action. Behavior and flags are unchanged.

[0.9.0] - 2026-07-22¶

Added¶

• Institution discovery in papio init: paste a library discovery URL or use the resolver configured in Zotero to derive the OpenURL base; scripted setup supports --institution-url.
• papio adapter diagnose <job-id> — sanitized, paste-safe support report for provider/adapter failures (job policy, actions, redacted event timeline, extension connectivity).
• Library-completeness line in papio status: when zotio is configured, status leads with Library: N item(s) missing PDFs — papio acquire --from-zotio fills them (or Library: complete), backed by a new zotio.missing_count RPC over zotio's local mirror.
• papio doctor --start: opts doctor into the same daemon autostart every ordinary command uses, so a first run can bring the stack up instead of reporting it down. Default behavior still diagnoses a stopped daemon.
• papio init defaults the Chrome extension ID to the Chrome Web Store package (npccengdhjmpojpjmjoeeclpdhcjelhf) — a store install needs no ID copying; only unpacked development builds still paste theirs.
• Triage snapshot schema 2 exposes institutional sign-in/open-access guidance only for actions classified by the acquisition waterfall; schema 1 remains available unchanged for older browser extensions.
• Standard bibliographic ingestion for acquire --batch: RIS, BibTeX, CSL-JSON, and MEDLINE/NBIB use one robust interchange pipeline for reference managers, database exports, and systematic-review tools.
• Provider compatibility matrix: a hand-maintained page records the individually live-verified provider routes, adapter status, and limitations without inventing aggregate success rates.

Changed¶

• Doctor's daemon-unreachable report collapses the four dependent SKIP lines into one integrations line and names the actual problem ("not running or unreachable") with the --start remediation, instead of a six-line cascade for a stopped daemon.
• Opener failures in papio actions open now name the failing command even when it produced no output (no more bare exit status 1).
• Getting-started, README, and troubleshooting now document the Chrome Web Store install as the default path; unpacked loading moved under a development-install tab. Firefox remains a temporary add-on until the AMO listing clears review.
• BREAKING: Access mode maximal is now delegated in the config value, --access-mode, and wire enum. Existing configs must be updated by hand; stored job policies migrate automatically.

[0.8.0] - 2026-07-21¶

Added¶

• [hooks] on_ready — a generic post-acquisition hand-off: when a job reaches ready, the daemon runs a user-configured shell command once with the job's metadata in PAPIO_* environment variables (PAPIO_DOI, PAPIO_PDF, PAPIO_TITLE, …). Fire-and-forget with a durable hook.on_ready job event as the audit trail; file acquisitions into papis, Calibre, a plain folder, or any script without a bespoke integration (ADR-0004). zotio remains the deep Zotero path.
• zotio is now optional: an empty zotio.executable disables the deep Zotero integration instead of rejecting the config. Ownership lookup degrades to not-owned with a staleness warning so batch submit and watches keep working, doctor reports zotio as "not configured (optional)", and zotio.auto_import = true still requires the executable.
• Browser session arbitration (papio browser sessions / papio browser use <id>|--latest, +browser.sessions/browser.claim RPCs): with papio installed in more than one browser, exactly one session now holds the offer/handoff flow. Later hellos wait as pending (session_busy) instead of silently stealing the session; a holder silent for 10 s yields to a live pending session; a cleanly closing browser releases immediately. Session identity rides the daemon↔native-host envelope, so extensions of every version participate. papio status and doctor report waiting browsers and denied hellos, and the extension-version flap between competing browsers is gone (ADR-0003).
• Handoff actions say whether a login is needed: human actions carry requires_auth and blocked_by (anti_bot, paywall, landing_page), papio actions list prints "open access — no login needed" vs "sign in to your institution first", and batch reports classify handoffs from the structured fields (schema v11 migrates and backfills existing parked actions). The job_offer protocol message carries requires_auth for the extension.
• Open-access sibling fallback: when a DOI's own record yields no legal candidate — or every candidate it did yield fails to fetch — the OpenAlex resolver searches for open-access sibling versions (preprints or repository copies under a different DOI) with strict title/year/author matching and tries those before parking the job — a paywalled paper with a free SSRN/arXiv copy no longer requires a manual re-search.
• OpenAlex works without an API key: the resolver now runs in OpenAlex's free polite pool with just the configured contact email — matching the search/discovery client — and treats sources.openalex.api_key as optional premium capacity. doctor reports the polite-pool state instead of failing.
• Stale-SSO handoff recovery: the daemon records browser.handoff_offered and browser.handoff_failed job events (new handoff_outcome protocol message), and institutional handoff guidance now says to sign in first and simply re-open on a stale-session error — every open mints a fresh link.
• imported terminal job state: a successful zotio apply (manual or auto-import) advances the job ready → imported with the Zotero item keys on the transition, so jobs list and status stop presenting filed work as actionable; batch reports read the keys from the job itself.
• --label now works on single papio acquire (not just --batch), seeding the target collection when --collection is unset.

Fixed¶

• papio actions open failures now name the URL and underlying error (with a papio doctor hint) instead of a bare exit status 1, and explain when open actions exist but none are openable from the CLI.

[0.7.2] - 2026-07-20¶

Fixed¶

• Scheduled backfills now scan the complete zotio missing-PDF queue and skip already-live request IDs without consuming the per-run cap, preventing one persistent entry from starving later work.
• Alert-watch digests now retain consumed identities as a durable seen set, merge title, arXiv, OpenAlex, and DOI aliases, preserve author names losslessly, recheck zotio ownership before acquisition, and serialize clear and acquire operations.
• Long failure reasons remain separate groups until after display truncation, so unrelated failures with a common prefix no longer collapse together.
• Semantic Scholar citation and reference snowballs now enforce year and open-access filters and reject ambiguous query-plus-snowball requests.

[0.7.1] - 2026-07-20¶

Fixed¶

• Browser-download adoption now waits for Firefox .part writes and their zero-byte target placeholders to settle before validating the final file, preventing valid PDFs from being quarantined mid-download.
• Accepting an identity review for a browser-adopted PDF now binds the override to that file's SHA-256 candidate key, so the scheduler's safe re-resolution path reaches ready for the same bytes instead of parking the same PDF for review again.

[0.7.0] - 2026-07-20¶

Added¶

• papio jobs failures (+jobs.failures RPC): groups jobs that did not complete on their own (failed, unavailable, needs-review, awaiting-human) by state, provider host, and terminal reason, with --since (Go durations, Nd days, or RFC3339) — so adapter and resolver work can be prioritized from real failure data.
• papio acquire --from-digest <watch> [--keys ...] (+watch.digest_acquire RPC): queues pending alert-watch digest entries as acquisition jobs through the same submission path watches use, preserving the recorded identifiers so deduplication holds; entries are removed only after their manifest is persisted and submission succeeded. papio watch digest clear <id> (+watch.digest_clear) empties a digest.
• Structured webhook notifications: webhook payloads now carry event, watch_id, watch_label, and count alongside message for routable receivers; plain notices keep the exact previous {source, message, sent_at} shape.
• Click-to-acquire support: the daemon accepts the new page_acquire browser-protocol message (advertised via the page_acquire hello_ack feature) and queues the page's DOI through the normal acquisition pipeline, answering with page_acquire_ack (job id or duplicate marker).

Fixed¶

• papio acquire --from-digest --keys now treats every --keys value as one opaque work key: keys containing commas (title-derived digest keys) are no longer split into fragments that miss valid entries. Repeat the flag for multiple keys.
• papio jobs failures --since rejects negative durations (-1h, -1d) instead of silently computing a future cutoff that hid every failure.

[0.6.0] - 2026-07-20¶

Added¶

• Backfill watches: papio watch add --kind backfill schedules the existing acquire --from-zotio queue on a cadence, so a growing Zotero library steadily self-completes its missing PDFs. Runs are bounded by --limit-per-run and idempotent (deterministic per-item request IDs), and the watch is force-runnable with papio watch run like any other.
• Alert-only watches: papio watch add --mode alert runs the scheduled discovery search and library-ownership filter but reports new works instead of acquiring them. New finds are recorded once per watch (re-runs never re-report) and browsable with the new papio watch digest <id> command; notifications point at the digest.
• Webhook notifications: a new notify.webhook_url config field delivers every daemon notification (watch results, human-action handoffs, imports) as a JSON POST — Slack/Discord/ntfy-style receivers work out of the box — in addition to the local desktop channel. Optional notify.webhook_secret is sent as a bearer token. Delivery is best-effort and never fails the work that triggered it.
• Semantic Scholar discovery backend: discovery is now pluggable behind a source seam. discovery.sources = ["openalex", "semanticscholar"] in config fans searches (and watches) across both backends with DOI/title deduplication in preference order; papio search --source selects one explicitly. Citation snowball (--cites, --cited-by) is supported on both; arXiv-only Semantic Scholar results now carry their identifier through to acquisition. API key (optional) lives at sources.semanticscholar.api_key.

Changed¶

• The MCP read resources (papio://jobs, papio://artifacts, papio://exports, …) now return {"<name>": [...], "truncated": bool} envelopes instead of bare arrays, making the 100-row cap honest. Filtered and paginated access remains available through the command facade (jobs list --state --limit).

[0.5.0] - 2026-07-19¶

Added¶

• Broader browser reach for the native-messaging connector. papio native-host install now registers the host with every installed Chromium browser it detects — Chrome, Edge, Vivaldi, Brave, Opera (Chromium too) — plus Firefox, each at its own per-user location (directory on macOS/Linux, registry key on Windows), so the same extension works across them. A new browser.extension_ids config field lists additional Chrome-family extension IDs (e.g. an Edge Add-ons build) alongside extension_id; the daemon accepts any of them and the manifest's allowed_origins lists them all.

Fixed¶

A triaged audit pass (33 confirmed findings fixed, each with a regression test where behavior changed):

• OA resolver identity verification: CORE, Europe PMC, and OpenAlex title searches now verify the normalized title — plus publication year and the full author list whenever the request supplies them — before trusting a result; Unpaywall requires the returned DOI to match the requested one; arXiv compares exact version-stripped IDs instead of substring matching. Cuts wrong-paper acquisition risk across the discovery plane.
• PDF identity matching scopes DOI and supplementary-material signals to the document's front matter, so a bibliography citing other DOIs — or a body mention of "supplementary material" — no longer rejects a correct article.
• Download safety: caller-supplied headers (Authorization, API keys) are stripped on cross-origin and HTTPS→HTTP redirects, and the body-reader goroutine no longer leaks when a response ignores cancellation.
• Storage integrity: bundle export and Zotero plan staging copy artifacts instead of hard-linking the immutable store (a consumer mutating the copy could corrupt it); concurrent same-hash promotions converge atomically; failed exports roll back the files they created; failed SQLite backups no longer strand a partial destination file; promotion and backup fall back gracefully on filesystems without hard-link support.
• Concurrency: RPC calls on separate IPC connections no longer serialize daemon-wide behind one slow call; the browser bridge releases its session lock during PDF validation on download adoption; the serial auto-importer releases its lock during retry backoff; concurrent zotio plan applies of the same plan are now mutually exclusive, and a claim abandoned by a crash or cancellation heals after a 15-minute lease instead of wedging the plan.
• Job lifecycle: context cancellation during auto-import stays retryable instead of recording a permanent failure; crash recovery clears abandoned quarantine files and the quarantine sweep continues past individual cleanup failures; a validation-persistence failure can no longer orphan a just-promoted artifact; watch-discovered works keep their OpenAlex identifier; pending notifications flush on a timer instead of waiting for the next event; ping answers from cache instead of blocking on the daily update check.
• Protocol strictness: strict JSON decoding rejects trailing documents; batch submissions reject unknown fields; Zotero item keys are validated fail-closed at the zotio integration boundary (the published v1 protocol contract is unchanged); batch identity hashes widened from 32 to 128 bits, with manifests from earlier releases still readable.

[0.4.0] - 2026-07-19¶

Added¶

• First-class Windows support. The daemon's local RPC runs over a named pipe on Windows — restricted to the current user via an explicit SDDL, the analog of the Unix socket's 0600 — while macOS and Linux keep their Unix-domain socket; the transport is chosen at build time. papio init, papio native-host install/uninstall/status, and papio doctor register the browser connector through the per-user registry (HKCU\Software\...\NativeMessagingHosts) instead of a manifest directory, and — because Windows has no unprivileged symlinks — install a copy of the papio binary as the native host (rerun papio init after upgrading to refresh it). Configuration lives at %APPDATA%\papio and data at %LOCALAPPDATA%\papio, and the update hint recognizes Scoop (scoop update papio). macOS and Linux behavior is unchanged.

[0.3.0] - 2026-07-18¶

Added¶

• Automatic, institution-agnostic library-resolver access. When a library's OpenURL resolver shows a "full text options" menu instead of direct-linking to the provider, papio follows the institution's top-ranked electronic service link itself — gated on a host permission for that resolver origin. The daemon advertises its configured resolver origins in the hello_ack handshake (new optional resolver_origins, backward compatible within papio-browser/1); the extension requests exactly those origins, so the popup surfaces a one-click "Allow library access" prompt (and the toolbar badge counts them) whenever a configured resolver isn't granted yet, and the options page lists the user's own resolvers under "Your library". Custom resolver domains outside the built-in Ex Libris hosts are reached through an optional https://*/* pattern that is never granted in bulk — only the exact configured origin is ever requested. Institution identity lives only in config.toml, never in extension code.

• Update discovery, without auto-install and without silent network calls. Store-delivered extension builds are stamped with the daemon version they shipped with, so the popup can show a calm "papio X.Y is available" line when the connected daemon is older — papio itself performs no network activity for this. Separately, an opt-in [updates] check = true setting (offered by the papio init prompt, default yes) has the daemon consult the papio and zotio GitHub releases APIs independently at most once a day. papio status appears in daemon status; both targets surface in papio doctor and a once-daily stderr hint. Configurations without the setting never check.

• Version-skew awareness across every surface. The hello_ack handshake now carries the daemon's version and a feature list (optional, backward compatible within papio-browser/1), so the auto-updating extension can degrade gracefully against an older daemon instead of failing opaquely. The popup reports daemon health directly: a quiet version line when healthy, and actionable states for daemon-unreachable, daemon-out-of-date, and extension-out-of-date; the toolbar badge shows ! when attention is needed and stays clear otherwise. The options page footer shows extension and daemon versions at a glance. The daemon records the connected extension's version and rejects extensions below a minimum floor with a clear update-the-extension message.

• papio doctor now walks the whole integration chain in one report: the Phase-1 readiness checks (config paths, database, PDF tooling, credentials) followed by integration checks — daemon reachability and version match, browser-extension connectivity, native-messaging-host manifests for Chrome and Firefox, and the zotio preflight — each failure with a concrete fix: line. The same diagnostics are exposed to agents as a read-only papio_doctor MCP tool.

• Every CLI command now warns on stderr (once per invocation, never on stdout) when the running daemon's version differs from the CLI binary, with the exact recovery command.

• Release engineering: release_metadata.py compat mechanically verifies the cross-artifact compatibility floors (daemon↔extension minimums, zotio minimum version, extension manifest/package version agreement) as a release.sh step and a source-only CI check; release.sh now also packages the Firefox extension archive alongside the Chrome one. A shared release runbook lives at .agents/skills/papio-release/SKILL.md and is cross-referenced from zotio.

• Extension store submission path for Chrome Web Store and Firefox Add-ons (AMO). extension/scripts/submit-firefox.sh signs and submits the built Firefox package via web-ext (AMO API credentials from extension/.env), and extension/scripts/submit-chrome.sh uploads the Chrome package via chrome-webstore-upload-cli; both are exposed as bun run submit:firefox and bun run submit:chrome. Paste-ready store listing kits (name, summary, full description, per-permission rationale, data-collection disclosure, and reviewer build instructions for the bundled source) live at extension/docs/amo-listing.md and extension/docs/chrome-web-store-listing.md.

• The bundled-zotio compatibility floor (internal/zotio/client.go MinimumVersion) now targets a released zotio line (0.10.0) instead of an unreleased 1.0.0; a built zotio 0.10.0 satisfies every capability, operation, and write-target papio's preflight requires. release.sh now stamps the bundled zotio binary with zotio's own version rather than papio's, so the cross-artifact compatibility check reflects the real zotio being shipped.

• Documentation: a Version skew and updates troubleshooting section (update flow, popup states, config-newer-than-binary errors), sister-project cross-references between papio and zotio in both READMEs and docs, and regenerated command reference.

• The MCP papio_status tool now surfaces the same actionable category and guidance as the CLI for parked and no-file jobs (including the config-aware institution_not_configured), so agents driving papio over MCP get the same diagnosis and next step as a human. The category catalog moved to a shared internal/errcat package consumed by both the CLI and the MCP server, so the two surfaces cannot drift.

• papio init now captures the browser extension IDs during first-run setup, so the native messaging host installs on the first run instead of failing with browser.extension_id is not set and forcing a config hand-edit and re-run. The Firefox add-on ID defaults to the built extension's fixed gecko id (papio@orgmentem.com) so Firefox works out of the box; the Chrome ID is prompted (paste the value from chrome://extensions). New --extension-id and --firefox-extension-id flags cover non-interactive setup. Unit-tested (flag and interactive paths, including that the captured Chrome ID reaches the native-host install) and smoke-verified end to end.

• Actionable error categories in papio status. Every parked or settled-without- a-file job now shows a short, stable category and a one-line next step instead of a raw internal reason (or nothing, for failed/unavailable). The catalog is config-aware: a job that found no copy under assisted/maximal mode with no institution configured surfaces as institution_not_configured pointing at papio init, rather than a silent unavailable. Categories/guidance are added to the status JSON (category, guidance) for agents. The same category and next-step now print under papio acquire --wait when a job parks or settles without a file, and the desktop human-action notification tells the user to run papio status to see why instead of a bare count. Unit-tested; the status view and acquire --wait guidance are smoke-verified against the live daemon.

• Per-institution access profiles and guided institution onboarding. Named resolver profiles under [browser.resolvers.<name>] are now full institution tables (openurl_base_url plus optional shibboleth_entity_id and proquest_account_id), so a multi-institution user routes each job's login to the right library. This lifts the earlier "default profile only" limitation on federated login-routing and the ProQuest account-id unlock: the daemon now wires login_entity_id/proquest_account_id per selected profile, and a named institution never inherits the default institution's identity. papio init gained an "Institution" step (and --openurl-base, --shibboleth-entity-id, --proquest-account-id flags); the ProQuest prompt accepts a pasted resolver URL and extracts accountid= for users who don't know their numeric id. Config validation, per-profile offer wiring, and the account-id extractor are unit-tested; the interactive flow is smoke-verified end to end. Older single-base configs keep loading: a resolver profile may still be a bare name = "https://…" string (shorthand for openurl_base_url), so no config migration is required.

• ProQuest account-id unlock: on ProQuest's "Find your institution" wall, papio appends ?accountid=<id> to the current URL, which unlocks Example University's institutional access with no sign-in at all (verified live 2026-07-18 — resolves the wall cold, "Access provided by EXAMPLE UNIVERSITY"). New per-institution config [browser] proquest_account_id (digits); the daemon passes it as the optional job-offer field proquest_account_id (default profile only); the ProQuest adapter gains accountIdParam: "accountid", and on a login verdict papio appends it (latched, once) — preferred over the federated route since it needs no credentials. This is the fix for the ProQuest openurl-handler blocker (the Shibboleth-DS route authenticated only ProQuest's main context, not the link-resolver handler). Config + protocol + adapter + bridge are unit-tested; the full download still needs a live pass on a ProQuest-held title.

• Institution auto-selection ("login routing"): on a provider login wall, papio navigates the handoff tab straight to the institution's federated login, skipping the provider's institution picker — selection is deterministic config (which institution you're at), not a secret, so only credential entry stays with you. New per-institution config [browser] shibboleth_entity_id (the Shibboleth IdP entityID, e.g. Example University's https://idp.example.edu/entity); the daemon passes it to the extension as the optional job-offer field login_entity_id (default resolver profile only, to avoid mis-routing another institution's job); and an adapter gains an optional federatedLogin template ({entityID} placeholder). On a login verdict papio navigates once (latched) to <federated-login>?entityID=<configured>. ProQuest ships the route (verified live 2026-07-17: Example University's entityID via ProQuest's Shibboleth DS URL routes straight to idp.example.edu login, skipping the WAYF picker). Config + protocol + adapter classify + bridge routing are unit-tested; the full post-sign-in download on a ProQuest-held title still needs a live pass.

• ProQuest institution-wall handling (proquest adapter v0.2.0): a login classify rule (ordered before article) now recognizes ProQuest's "Find your institution" wall (form#institutionForm + input#institutionName, fixture-backed fixtures/proquest/login-return.html captured live via CDP). papio surfaces it as a human sign-in step (login → auth_pending) instead of silently staying assisted/unknown. Matters disproportionately because Example University's OpenURL resolver routes many titles (incl. SAGE/T&F journals) to ProQuest rather than the publisher. Classify verified by fixtures; the full post-sign-in download recovery (authenticate ProQuest → re-drive → entitled docview → download) still needs a live pass.

• SAGE Journals adapter (journals.sagepub.com), fixture-backed (fixtures/sage/success.html, captured live via CDP from a Example University-authenticated article). SAGE emits no Highwire metas; classifies on publication_doi + the downloadPdfUrl anchor (same shape as ACM) and downloads that anchor's /doi/pdf/<doi>?download=true href. Classify is fixture-verified; the end-to-end download is not yet live-exercised because Example University's resolver routed the SAGE test title to ProQuest rather than sagepub (the adapter fires when a title routes to journals.sagepub.com).

• Wiley Online Library adapter (onlinelibrary.wiley.com), fixture-backed (fixtures/wiley/success.html, captured from a Example University-authenticated article). Classifies via the Highwire citation_pdf_url/citation_title metas, then builds and fetches Wiley's direct /doi/pdfdirect/<doi>?download=true file through the privileged downloads API — citation_pdf_url (/doi/pdf/) and the /doi/epdf/ link both return an HTML viewer wrapper, only pdfdirect returns the file (verified live end-to-end: 1.15 MB PDF → ready). Closes the gap where Wiley pages classified unknown and stayed assisted (browser- agnostic; affected Chrome too). tandfonline/psycnet remain unimplemented — permissioned but not yet fixture-backed (both paywalled in the dev session; psycnet also emits no standard metadata).

• Firefox dev loop: bun run dev runs build.ts --watch (rebuilds firefox/ on any src/, icons/, or manifest.json change) alongside web-ext run, which hot-reloads the add-on in a dedicated Firefox Developer Edition instance. web-ext-config.mjs pins an absolute, gitignored dev profile (.ff-dev-profile) so permissions and institutional logins persist across reloads — and, being path-based, boots straight in without Firefox's profile-chooser modal. web-ext installs and hot-reloads over the devtools RDP (not WebDriver/Marionette), so it does not set navigator.webdriver — but that live RDP connection makes Firefox show its remote-control indicator and is itself an automation surface a bot wall could fingerprint. Two modes: bun run dev for fast iteration and fixture testing; for real Cloudflare- walled providers, bun run build then load firefox/ manually via about:debugging (one-shot install, no persistent connection, no indicator, navigator.webdriver false).

• Brand: a papio logo — an oblique lowercase p (coral #E85D4A) inside a broken ink ring (#2B2D42); the p's descender becomes a download arrow that exits through the ring's bottom gap. Structural sibling of the zotio badge with its own palette. Vector sources live in docs/assets/ (logo.svg, logo-dark.svg for dark surfaces, logo-tile.svg for theme-agnostic toolbar icons, logo-wordmark.svg, logo-wordmark-dark.svg) and are used in the README wordmark header, the docs site logo/favicon (mkdocs.yml), the Chrome extension toolbar/action icons (extension/icons/, wired in manifest.json), and the extension popup header.

• Brand: the README header wordmark (logo-wordmark.svg, logo-wordmark-dark.svg) is now an animated SVG. The mark builds in on a calm ~10s loop — the broken ring draws on, the coral p and download arrow drop into place, the wordmark rises in — then a cheeky little papio (baboon) head peeks over the wordmark to blink, tilt, and wave before ducking away, leaving a long clean hold on the finished logo. Pure CSS (no script/SMIL, self-contained for GitHub's <img> rendering); the resting state is byte-for-byte the prior static logo and prefers-reduced-motion: reduce shows it with no animation.
• Background work window: papio now does its browsing in one dedicated minimized, unfocused Chrome window instead of the user's tab strip. Every broker handoff tab (first, queued, and download-fallback) and the keepalive tab route there; provider-spawned viewer tabs inherit it via their opener. A tab surfaces (window restored + focused, tab activated) only when the human is needed: on the IdP transition (auth_pending), on keepalive reauth, and from the popup's Focus button — which now also restores a minimized window. Opt out anytime via the options page ("Keep papio tabs in a background window"); disabling restores the legacy visible-handoff behavior, as does any runtime without chrome.windows.
• Firefox support, day one: bun run build now emits a second complete extension at extension/firefox/ (MV3 event-page background as a classic iife bundle, browser_specific_settings.gecko.id = papio@orgmentem.com, strict_min_version 128) generated from the same manifest.json source of truth. The native-host installer registers a Firefox manifest (allowed_extensions, Mozilla NativeMessagingHosts dir) alongside Chrome's when the new [browser] firefox_extension_id config is set, and the host accepts Firefox's bare-ID invocation with the same exact-match, fail-closed validation as Chrome's origin. The options page gained a "Library resolver access" grant section because Firefox treats MV3 host_permissions as runtime-optional; on Chrome it simply shows the install-time grants. No behavior change for Chrome users. The provider section also gained "Grant all providers" / "Revoke all" — one click issues a single permissions.request for every publisher origin (one Firefox doorhanger) instead of ten separate grants.

Changed¶

• Rewrote the README on the zotio template: centered wordmark + tagline + badges + docs nav, a "Why papio" section with the hard boundaries, a hand-drawn two-row serpentine architecture diagram in the brand palette (docs/assets/architecture.svg + -dark.svg, theme-switched via a <picture> element; replacing the mermaid flowchart, which rendered poorly on GitHub) with the access-mode table, the research loop, validation/provenance and zotio-boundary sections, the MCP tool surface, and install paths (brew, scoop, signed releases, source). Brand style: papio italic in prose, zotio plain.
• Redesigned the wordmark's baboon cameo: the abstract head is now a recognizable hamadryas baboon (cape mantle, long muzzle, heavy brow) that peeks up holding a stack of papers instead of waving. Light and dark wordmark variants stay in sync; in dark mode the paper stack renders navy against the cream mantle for contrast, and the face details (eyes, brows, muzzle) stay navy on the coral face in both modes.
• Config unknown-field errors now explain that the config was likely written for a newer papio and name the offending fields, instead of surfacing a raw TOML parse error. zotio preflight failures name the installed version, the configured executable path, and the action that fixes the mismatch.
• MCP tool surface now derives from the papio CLI command tree instead of a parallel set of hand-maintained typed tools, so the CLI is the single source of truth and the two can no longer drift. The default surface is a command facade — papio_command_search to discover commands and papio_command_run to execute one (JSON output, command-local flags only, inherited globals rejected); PAPIO_MCP_SURFACE=mirror instead exposes one papio_<command> tool apiece. Setup and lifecycle commands (init, config, daemon, native-host, mcp) are hidden via mcp:hidden annotations. Two composite tools with no single-command equivalent stay first-class — papio_acquire_batch (bulk work input) and papio_batch_wait (bounded polling) — alongside the five read resources. Migrated the server library from modelcontextprotocol/go-sdk to mark3labs/mcp-go for parity with zotio.

Fixed¶

• Reliability: overlapping extension state writes are now persisted through a serialized save chain, so a reordered chrome.storage write can no longer resurrect a stale snapshot after a service-worker restart.
• Reliability: concurrent queued-handoff fallback timers no longer drop each other's forced releases; a single drain loop consumes every pending release, so queued jobs can no longer be stranded invisibly with tab_id -1.
• Reliability: a failed native-host idle-poll write now tears the bridge down instead of leaving the process alive but no longer polling (which starved the extension of offers and cancels).
• Reliability: fetchCandidates propagates the OpenHumanAction write error before parking a landing-page-only job, matching exhaustedCandidates, so a transient write failure can no longer strand a job with no human-action row.
• Concurrency: removed a redundant drain goroutine in readBodyWithContext that doubled leaked goroutines when a response body read hung.
• MCP acquire.report now classifies failures — missing batch as not_found, malformed batch ID as invalid_argument, and other failures as internal — instead of collapsing every error into not_found.
• Batch settlement is now a single source of truth (batch.Report.Settled), removing a stale duplicate outcome list in papio_batch_wait that carried legacy outcome spellings.
• Docs/schema for papio_batch_wait timeout_seconds now state that 0 or an omitted value defaults to 300, matching the implementation.

[0.2.0] - 2026-07-15¶

Phase 0 — contracts and prerequisite¶

• Established the papio Go/Bun workspace, fail-closed shared protocol fixtures, and draft work-request, acquisition-bundle, and browser contracts.
• Added zotio's stored-attachment upload path with reconciliation and retry-safe Web API registration, which is the import prerequisite for papio exports.

Phase 1 — durable open-access acquisition¶

• Added private configuration, SQLite migrations, daemon IPC, durable job and lease recovery, source budgets, redacted events, quarantine, and content-hash artifact storage.
• Added normalized work identity, deterministic candidate ranking, bounded HTTPS acquisition, PDF validation, OCR fallback, and review/rejection paths.

Phase 2 — institutional browser handoff¶

• Added the native-host bridge, versioned bounded browser protocol, native-host install/status commands, and a least-privilege MV3 extension for one requested institutional download per job.
• Added adoption confinement and validation for browser downloads, with restart-safe daemon and extension lifecycle handling.

Phase 3 — provider adapters and protocol lock¶

• Added declarative, permission-gated adapter execution and sanitized fixture capture for ProQuest, JSTOR, EBSCO, and Springer flows.
• Locked work-request/1, acquisition-bundle/1, and papio-browser/1 with strict cross-runtime fixtures; retained Go as the core after the reversal review.

Phase 4 — zotio, MCP, and human resolution¶

• Added zotio capability/version preflight, preview/apply plans, confirmation hashes, import-ledger idempotency, missing-PDF intake, and stored attachments.
• Added MCP tools and resources over the same application service, plus bounded human identity-review resolution and action lifecycle cleanup.
• Added extension session recovery across daemon restarts and startup wake-up.

Post-Phase 4 — autonomous acquisition¶

• Added OpenAlex discovery, batch acquisition, serialized retry-safe auto-import, session keepalive, observed-provider fixture capture, library-aware batches, OA browser fallback, snowball search, status/reporting, notifications, watchlists, MCP loop closure, and first-run onboarding.
• Updated zotio integration with collection-aware missing-PDF scopes, item-type valid container-title mapping, exact-key enrichment, and transactional workflow execution.

Phase 5 — release preparation¶

• Added local release artifacts for papio and zotio binaries, the extension ZIP, dependency inventories, license reports, hashes, and a machine-readable release manifest.