# Changelog
## [v0.10.0] - 2026-03-12
### Added
- **`:settings` option for `URP.convert/2`:** pass `{path, property, value}`
triplets to configure soffice via `ConfigurationUpdateAccess` before each
conversion. Settings sharing the same nodepath are batched into a single
update call. Useful for tuning cache limits, graphic memory, etc.
- **`:io` option for `URP.convert/2`:** choose between file-based (`:file`,
default) and streaming (`:stream`) I/O independently for input and output.
File I/O uses temp files on soffice's filesystem (~6 round-trips). Streaming
uses XInputStream/XOutputStream over URP (~40-50% slower but no temp disk).
Mixed modes supported: `io: {:file, :stream}` or `io: {:stream, :file}`.
- **`:recv_timeout` and `:max_frame_size` options for `URP.convert/2`:**
per-conversion control over the TCP recv timeout (default 120 s) and maximum
accepted frame size (default 512 MiB).
- **Infinite reconnection with backoff:** the pool retries soffice connections
forever with exponential backoff (`backoff_initial` / `backoff_max`) instead
of crashing. Emits `[:urp, :connection, :retry]` telemetry on each attempt.
- **Frame size guard:** `recv_frame` rejects frames larger than 512 MiB
(configurable) before allocating, preventing OOM from corrupt wire data.
### Fixed
- **Pool query checkin leak:** diagnostic queries (`version`, `services`, etc.)
did not reset `conn.private`, `conn.reply`, or `conn.doc_oid` on checkin,
leaking state across pool reuses. All pool operations now go through
`reset_conversion_state/1`.
- **Default pool ignored backoff config:** `backoff_initial` and `backoff_max`
from `config :urp, :default` were silently dropped. Application now forwards
the full config map to the pool.
- **Large file conversion (>64 MB output):** `gen_tcp.recv` returns `:enomem`
for single reads above ~64 MB. `recv_frame` now reads in 4 MB chunks and
reassembles. Also eliminates redundant binary copies in `send_frame` (iodata
passthrough) and `write_bytes` (`enc_str_iodata/1`).
- **Stale `conn.reply` leaking as conversion result:** `conn.reply` retained
values from bootstrap or diagnostic queries. Early conversion failures (e.g.
file not found) could return stale data as a successful result. Convert now
clears `conn.reply` on checkout with strict type-checked result matching.
- **TID cache loss across conversions:** the URP type cache accumulated during
each conversion is now persisted back to the connection struct, preventing
type desync on subsequent operations.
- **O(n²) `fill_buffer` in enum streams:** chunk accumulation now uses iodata
instead of repeated binary concatenation, flattening once when consumed.
### Changed
- **Iodata frame builders:** `Call` and `Protocol` return iodata instead of
flattened binaries. The single `IO.iodata_to_binary` call happens in
Protocol `send_frame/2`, eliminating redundant intermediate allocations.
- Bridge error paths use `reply: nil` instead of `reply: ""` for consistency.
## [v0.9.1] - 2026-03-06
- **Fix 2 GiB memory spike during conversion:** `read_file/2` now calls
`available()` to get the exact output size before `readBytes()`. Previously
it passed `0x7FFFFFFF` (2 GiB), causing soffice to pre-allocate a 2 GiB
buffer on every conversion regardless of actual PDF size. Peak soffice RSS
drops from ~2.3 GB to ~300 MB.
## [v0.9.0] - 2026-03-04
- Add `:telemetry` events — every pool operation emits `[:urp, :call, :stop]`
with `queue_time`, `service_time`, and `total_time` measurements.
See `URP.Telemetry` for details.
## [v0.8.1] - 2026-03-03
Internal refactor of Bridge and Pool internals. No public API changes.
- **Plug-like error handling:** Bridge functions no longer raise — errors
accumulate on `conn.error` and short-circuit subsequent calls via guards.
This replaces 13+ rescue blocks with a single, predictable data-flow pattern.
Callers piping through Bridge get clean error propagation without try/rescue.
- **Uniform return type:** All Bridge functions return `t()`. Functions that
previously returned tuples (`store_document_write`, `store_to_stream`,
`read_file`) now stash results in `conn.reply`, enabling consistent piping.
- **Pool memory fix:** `reset_conversion_state` now clears `conn.reply`,
preventing large binaries from being held between conversions.
- **Accurate type specs:** Bridge struct fields (`sock`, `desktop_oid`,
`ctx_oid`, `smgr_oid`) now correctly typed as `| nil`; `reply` is `term()`.
- **CI: Debian soffice 26.2** with Docker layer caching — replaces Alpine
image, enables LO 26.2+ tests (markdown export).
- Expanded test coverage: bang variants, Bridge-level `store_to_stream`,
temp file cleanup.
## [v0.8.0] - 2026-03-02
- Add `URP.services/1` — list all registered UNO service names
- Add `URP.filters/1` — list all export filter names (discover available filters at runtime)
- Add `URP.types/1` — list all document type names
- Add `URP.locale/1` — query soffice locale setting
- Add Protocol `parse_string_sequence_reply/1` for `sequence<string>` replies
- Add Diagnostics section to module docs with examples
## [v0.7.0] - 2026-03-02
- **File-based I/O:** load and store documents via XSimpleFileAccess instead of
XInputStream/XOutputStream streaming — eliminates thousands of TCP round-trips
per conversion, ~8x faster for typical documents
- **Breaking:** `close_document!/2` now returns `t()` instead of `{binary(), t()}`
- **Pipeline-friendly Bridge helpers** — all conn-threading helpers (`call`, `sfa_call`, `qi`)
return just `conn`; enables `conn |> qi(...) |> call(...)` style
- Add `last_reply` and `last_error` fields to Bridge conn for introspection/debugging
- Add `:filter_data` option for export-specific settings
(e.g. `[UseLosslessCompression: true, ExportFormFields: false]` for PDF)
- Add XSeekable support — ZIP-based formats (docx, xlsx, pptx, odt) stream
without buffering the entire file first
- Reuse connections across conversions (no reconnect per document)
- Pre-compute static URP frame bodies at compile time
- Fix TID cache for cross-thread XSeekable replies
- Replace inline magic numbers with named constants throughout
- Add PERFORMANCE.md with benchmarks and container recommendations
- Add Benchee benchmark suite
- Switch Docker images to libreoffice-*-nogui packages
## [v0.6.1] - 2026-03-01
- Add `URP.version/1` — query soffice version string over URP (no CLI access needed)
- Hide URP.Pool from hex docs (internal module)
- Simplify README
## [v0.6.0] - 2026-03-01
- **Breaking:** `:filter` option is now required — no default export filter
- Remove PDF-centric language from docs — URP is a generic document conversion tool
## [v0.5.0] - 2026-03-01
- **Breaking:** unified API — single `URP.convert/2` replaces convert_stream/2, convert_file_stream/2, and convert/3
- **Breaking:** remove URP.convert_stream/2, URP.convert_file_stream/2, URP.convert/3, URP.Pool.convert_stream/3, URP.Pool.convert_file_stream/3, URP.Pool.convert_url/4
- **Breaking:** output destination is now `:output` option (path, `:binary`, or `fun/1`) instead of `:sink`
- **Breaking:** default output writes to temp file (returns `{:ok, path}`) instead of accumulating bytes in memory
- Add enumerable input support — pass any `Enumerable` (e.g. `File.stream!/2`, S3 download streams) to `URP.convert/2`
- Add lazy enumerable streaming (enum reader + Bridge stream loader)
## [v0.4.0] - 2026-02-28
- **Breaking:** remove `use URP, otp_app: :my_app` macro — call URP.convert_stream/2, URP.convert_file_stream/2, URP.convert/3 directly
- **Breaking:** `URP.Test.stub/1` replaces URP.Test.stub/2 — stubs are global, no module name needed
- **Breaking:** remove URP.Test.start/0 — ownership server starts automatically
- Add URP.Application — default pool, DynamicSupervisor, and ownership server start automatically
- Add named pools via `config :urp, :pools` — started on first use via DynamicSupervisor
- Handle soffice DisposedException with reactive retry on reconnect (matches C++ callers' approach)
- Make nimble_ownership a required dependency (no longer optional/test-only)
## [v0.3.1] - 2026-02-26
- Exclude Mix.Tasks.Bump from hex package to avoid module conflicts
## [v0.3.0] - 2026-02-26
- Fix protocol correctness: validate block count, support FUNCTIONID16/14, skip MOREFLAGS byte
- Fix one-way detection: use func_id (only `release` is one-way), not unreliable header flags
- Add `parse_exception/1` — extract human-readable messages from UNO exception replies
- Include exception details in error messages from `load_document!` and `store_to_url!`
- Add protocol unit tests (28 tests covering header parsing, encoding, reply classification)
- Add error handling integration tests
- Simplify mix bump task (remove network dependencies, fix editor hang with gpg signing)
## [v0.2.0] - 2026-02-26
- **Breaking:** remove URP.Connection — all calls go through URP.Pool via wrapper modules
- Make `use URP` route all calls through a supervised Pool
- Change default pool_size from 4 to 1 (matches single soffice instance)
- Rewrite README: clarify setup steps, explain function differences, add design tradeoffs
- Add Kubernetes scaling note
- Fail hard when soffice is not reachable in tests (instead of silently skipping)
- Hide mix bump from hexdocs
## [v0.1.2] - 2026-02-26
- Add otp_app config support for URP.Pool
- Add soffice service to CI for integration tests
- Add `:mix` to dialyzer PLT apps
- Add release workflow and mix bump task
- Remove stale urp_convert.exs script
- Fix license year
## [v0.1.1] - 2026-02-25
- Add streaming conversion via XInputStream/XOutputStream (no shared filesystem needed)
- Add file-backed streaming (convert_file_stream/2)
- Add sink option for streaming output to file or callback
- Add URP.Pool (NimblePool) for connection pooling
- Add `use URP` macro and `URP.Test` for stubbable wrapper modules
- Add typespecs to all public functions
- Add Nix flake dev shell
- Add dialyzer
- Make README the main page on hexdocs
## [v0.1.0] - 2026-02-25
- Initial release
