Repository navigation
Releases
RSO Context Alpha 0.11.0
Leaner packets, cheaper writes, better recall per byte.
- Defaults: compact byte budget 12 KB -> 4 KB, query limit 8 -> 4 (rso_check keeps 8).
- Graph nodes are off by default in query packets (
include_graph/--graphto opt in). - CLI
queryprints the compact packet by default;--fullreturns rso-context-packet/v2. packet_hashno longer covers budget counters (run), so budgeted and unbudgeted runs hash alike.- MCP server instructions drop the "call rso_use first" rule.
rso_useon a known project is a
read-only freshness check plus a small targeted refresh, never a whole-project ingest. - Per-file dedup (max 2 items per file), adjacent span merging, cross-file identical-chunk dedup.
- Compact evidence items carry hashes and project id only inside
expand. - Claims: capped at 4, short/code-like noise dropped, no unconditional verified/proposed boosts;
coordination-board paths rank after other hits. - Very long spans (one-line or minified files) are windowed around the match in packets.
- Ingest skips vendored and generated content (third_party, worktrees, nested checkouts, lock
files, *.min.js, *.map) and indexes a repeated chunk inside one file once. - Edited files no longer return empty evidence: the query re-ingests just those files and retries;
with refresh off, the indexed copy is returned labelled stale. - Schema revision 8 (compatible with schema_version 7, so running 0.10 servers keep working): FTS5 porter stemming with a prefix index (rebuilt from the existing index on
migration, orphan rows dropped), larger stop-word list, bm25 column weights, AND-first query
planning with OR fill. - Ingest reads and hashes files outside the write lock, writes only changed files, and commits in
batches of 200 files. - New
prune-orphanscommand removes chunks of superseded file versions (protected claim evidence
kept); dry run unless--apply --yes. - A query performs at most one write transaction (budget, cache and run record together).
RSO Context Alpha 0.9.4
Any agent host can now use RSO on any bounded project folder through MCP. Host entries used to pin the folders named at setup time, so an agent working anywhere else was refused; that was the usual reason agents reported they could not run RSO.
Install
Download RSOContextAlpha-0.9.4.zip, extract it, and run python -B install.py from the extracted folder. rso-context --version should report 0.9.4 and doctor should report ready: true. Then run rso-context mcp --setup --client all and restart your agent apps. INSTALL.md covers Windows, Linux and macOS paths; QUICKSTART.md is the beginner walkthrough.
Requires Python 3.11 or newer, Git, and Python's standard-library SQLite with FTS5. The core CLI needs no pip packages, API key or network service.
sha256 of the attached ZIP: 2ce25c2e72a42358f4a74f416f4a65b16a524391eb5e32a9beec4f02ec9826a3 (62 files, 209,170 bytes). The same value is in the attached .sha256 file.
Changed
- MCP access is open by default. A tool call may name any bounded project folder, whether or not it sits under a launch
--root.--strict-rootsrestores launch-only access. - One bounded-folder rule covers every MCP path, launch roots included. It refuses filesystem roots, user profiles and folders that contain one, other accounts' profiles, Desktop, Documents, Downloads, OneDrive roots, hidden profile folders such as
.ssh, operating-system folders, and the ledger home. - The server starts however a host launches it. A desktop app that starts servers from its install folder or the user profile used to get a startup error; that directory is now simply not used as a root.
mcp --setuppins no--rootunless one is given. A rewrite keeps keys you added to a JSON entry and Codex tool-approval tables.
Added
mcp --setup --client allconfigures every installed host and lists the ones it skipped. New hosts:claude-desktop,cursor,windsurf.- The CLI accepts
--client opencode; the parser used to reject it even though setup supported it.doctorreports every supported host.
Fixed
rso_query,rso_checkandrso_expandcould return an emptyinsufficient_budgetpacket at the default byte budget. JSON escaping made the wire result a few bytes larger than the packet, and the shrink loop fell back to 256 bytes before getting under. The budget is now found by binary search.- A host that replaces the child environment left Git off
PATH. Ingest stopped with a generic error, and project identity fell back to a non-Git key, so one checkout could register as two projects. RSO now finds Git throughRSO_GIT,PATH, or the standard Windows install folders, and setup writesRSO_GITinto the entry. - The user-profile refusal no longer depends on
USERPROFILEorHOME; it asks the operating system for the profile folder. - Every tool failure now reports its reason instead of "Error executing tool".
Verified, and what is not
176 regression tests pass with one POSIX-only skip. GitHub Actions passed on Windows, Ubuntu and macOS with Python 3.11 and 3.12 for both merges in this release. The attached ZIP was rebuilt from the tagged commit and matches, byte for byte, the archive that was installed into a temporary prefix and checked with a scratch ledger (--version 0.9.4, doctor ready).
Open access was checked live over stdio the way a desktop app launches servers: working directory set to the user profile, no --root, environment reduced to SystemRoot and a System32 PATH. On a folder outside every previously configured root, rso_use ingested 35 files, rso_query returned six evidence spans at the default budget, and rso_use on the profile was refused. The entries written by --setup --client all for Codex, Claude Code, Claude Desktop, Cursor, Gemini CLI, Antigravity and OpenCode were each started with their own command and environment and gave the same results.
That checks the entries, not the host applications. Acceptance inside each app after a restart is unverified, as are Windsurf and native macOS live-agent use. See VERIFICATION.md for the recorded scope.
This is an alpha pre-release.
RSO Context Alpha 0.9.3
An MCP host no longer needs a config edit to give RSO a new project, and two startup failures that a host could only see as a timeout now report themselves.
Install
Download RSOContextAlpha-0.9.3.zip, extract it, and run python -B install.py from the extracted folder. rso-context --version should report 0.9.3 and doctor should report ready: true. INSTALL.md covers Windows, Linux and macOS paths; QUICKSTART.md is the beginner walkthrough.
Requires Python 3.11 or newer, Git on PATH, and Python's standard-library SQLite with FTS5. The core CLI needs no pip packages, API key or network service.
sha256 of the attached ZIP: 7e6cd1c123febd0cf2f15b94e7094b8dde8769f5551cb5c1431acd289fbdb86a (62 files, 199,271 bytes).
Added
- An MCP launch with no
--rootbinds the directory the host started the server in, which is the project the user opened. One configured entry serves every project. This is a launch-time bound rather than run-time widening: a user profile or an unbounded home child is still refused, and--rootstill wins when given. mcp --setup --client opencodewrites that host's entry. OpenCode keys servers undermcpwith a command list rather thanmcpServers, and its config may be.jsonc, so the writer tolerates comments while leaving unrelated settings alone.--setupmay now omit--rootand use the current directory.- Generated host entries carry
RSO_CONTEXT_HOMEand, on Windows,SystemRootwith a two-entry systemPATH. A host that replaces rather than augments the child environment otherwise spawns a server that can neither locate a ledger nor load winsock. The user's ownPATHis deliberately not copied into a config file. rso-context replayre-runs the queries and checks already recorded in the ledger against the current build and reports which packets changed, separating a moved corpus and edited sources from a real regression. It writes nothing.benchmarks/budget_sweep.pysweeps--limitand--token-budgetover recorded runs and prints a coverage-against-cost table. Advisory only.
Fixed
- The stdio adapter no longer dies before answering
initializewhen the spawn environment carries no home directory.Path.home()raisesRuntimeError, notOSError, so the existing handler never fired. A host that replaces the child environment produced exactly that and reported the silent exit as-32001 Request timed out. - Startup failures now emit a JSON-RPC error frame on stdout as well as the JSON on stderr: code
-32099, null id, cause in the message. A client watching only stdout previously could not tell a crash from a hang. - The
--dbdefault no longer needs a home directory to build the argument parser. In a home-less environment that raised before the CLI could report it, giving a bare traceback and exit 1 even when--dbwas passed explicitly. default_home()namesRSO_CONTEXT_HOMEand--dbinstead of surfacing pathlib'sRuntimeError, and an unreadable working directory is reported as a launch-root failure rather than an uncaughtOSError.- The user-profile root refusal no longer disappears when
Path.home()is unavailable. It also readsUSERPROFILE,HOMEandHOMEDRIVE+HOMEPATH, so a partial environment keeps the guard.
Verified, and what is not
161 regression tests pass with one POSIX-only skip. GitHub Actions passed on Windows, Ubuntu and macOS with Python 3.11 and 3.12 for every merge in this release.
The startup changes were also checked live over stdio against the isolated runtime with a deliberately stripped environment, not only in unit tests: no home with --db starts; no home without --db exits 2 with the frame on stdout; a working directory set to the user profile still refuses.
Two limits stated plainly. The SystemRoot half of the original failure is environmental — a Windows child process without it cannot load winsock, and no code here changes that, which is why setup writes the variable into the host entry. Live OpenCode acceptance is unverified: the generated entry was checked by round-tripping the config file and starting the server with that entry's own command and environment, not by driving OpenCode. Native macOS live-agent acceptance also remains unverified. See VERIFICATION.md for the recorded scope.
This is an alpha pre-release.
RSO Context Alpha 0.9.2
RSO Context Alpha 0.9.2 stops rso_check from reading an evaluated or tested option as a chosen one.
Download and install
Download RSOContextAlpha-0.9.2.zip, extract it, and follow QUICKSTART.md or INSTALL.md. Run python -B install.py from the extracted folder. Check that the installed command reports 0.9.2 and doctor reports ready: true. Python 3.11+, Git, and SQLite FTS5 are required. Verify the download against SHA256SUMS.txt.
The package contains 60 allowlisted files and no personal ledger or host configuration.
Fix
| Source sentence | 0.9.1 | 0.9.2 |
|---|---|---|
| We evaluated the Arnold renderer for finals last year. | selected: Arnold |
unknown |
| For finals we evaluated the Arnold renderer and chose Cycles. | disagreement |
selected: Cycles |
| After testing, we use Arnold to render finals. | selected: Arnold |
selected: Arnold |
| We tested exporting the preview at 30 fps. | claim supported, value 30 |
unknown |
A sentence or clause that records an experiment ("evaluated", "tested", "compared", "tried", "benchmarked", "looked at") without a decision word ("chose", "use", "decided", "must") is no longer evidence. Choice sentences also split just before a decision verb joined by "and" or "then", so each half is judged on its own. Every 0.9.1 cell above was confirmed by running the released code.
Verification and limits
- 145 regression tests passed on Windows with Python 3.12.5 (one POSIX-only test skipped). GitHub Actions passed on Windows, Ubuntu, and macOS with Python 3.11 and 3.12 for the release commit.
- The ZIP was built from the tagged commit, installed into a temporary prefix, and checked with
--version,doctor, and scratch-ledger checks for an evaluated-only sentence (unknown) and an evaluate-then-choose sentence (selected: Cycles). - Twelve sample questions over this repository gave identical answers and timing under 0.9.1 and 0.9.2 on the same corpus.
- The rule is word-based. "We tried to use Arnold" still counts as a decision because it contains "use", and experiment wording outside the cue list gets through. A choice sentence must also share two topic words with the question to count at all, a rule unchanged since 0.9.0. See VERIFICATION.md.
RSO Context Alpha 0.9.1
RSO Context Alpha 0.9.1 fixes three accuracy problems in how RSO reads source sentences, and makes rso_check faster.
Download and install
Download RSOContextAlpha-0.9.1.zip, extract it, and follow QUICKSTART.md or INSTALL.md. Run python -B install.py from the extracted folder. Check that the installed command reports 0.9.1 and doctor reports ready: true. Python 3.11+, Git, and SQLite FTS5 are required. Verify the download against SHA256SUMS.txt.
The package contains 60 allowlisted files and no personal ledger or host configuration.
Fixes
| Source text | 0.9.0 | 0.9.1 |
|---|---|---|
| A sentence hard-wrapped across two lines | unknown |
read as one sentence; evidence spans both lines |
| "Someone asked whether preview export is 24 fps." | supported |
not evidence either way |
| A plain "not" | the conflict detector ignored it; rso_check counted it |
one shared negation rule |
- Wrapped sentences. A line continues into the next only when it does not end in sentence punctuation and neither line is structural (heading, list item, table row, fence, indented code). List items and table rows stay separate.
- Recorded questions. A question mark or a cue such as "asked", "whether", "proposed", "TBD" or "should we" means the sentence records a question or suggestion, not a decision.
- Negation. Measured before release: across 15 queries over this repository's documentation, the wider conflict rule changed no query result.
Faster checks
One claims query and one validations query per check instead of two per answer, cached topic-word sets, and no result copy when nothing needs trimming. A twelve-question check over this repository went from a median of 163 ms to 138 ms with identical answers.
How these were found
Jev (TypeSafe System One) was used as an advisory static reviewer over the source: yes/no questions about specific functions, then choices between alternative fixes. Every answer was confirmed by running the code before anything changed; two of its 30 audit answers were wrong. No model is part of RSO.
Verification and limits
- 141 regression tests passed on Windows with Python 3.12.5 (one POSIX-only test skipped). GitHub Actions passed on Windows, Ubuntu, and macOS with Python 3.11 and 3.12 for the release commit.
- The ZIP was built from the tagged commit, installed into a temporary prefix, and checked with
--version,doctor, and a scratch-ledgercheckcovering a wrapped sentence and a recorded question. - Still open: a sentence that mentions an option in the question's own words ("We evaluated the Arnold renderer for finals") counts as selecting it.
- The timing and negation measurements are single runs on one corpus and one machine. This is alpha software; see VERIFICATION.md.
RSO Context Alpha 0.9.0
RSO Context Alpha 0.9.0 adds typed evidence questions. rso-context check and MCP rso_check answer up to 12 questions per call from current source sentences and return rso-check/v1, with the spans behind every answer.
Download and install
Download RSOContextAlpha-0.9.0.zip, extract it, and follow QUICKSTART.md or INSTALL.md. Run python -B install.py from the extracted folder. Check that the installed command reports 0.9.0 and doctor reports ready: true. Python 3.11+, Git, and SQLite FTS5 are required. Verify the download against SHA256SUMS.txt.
The package contains 60 allowlisted files: LICENSE, NOTICE, the installer, user documentation, runtime, and regression tests. It contains no personal ledger or host configuration.
Changes
- Typed questions. A
claimanswers supported, contradicted, disagreement, or unknown, and a number in the claim must match the source. Achoiceanswers selected, excluded, disagreement, or unknown; options take aliases, and a tie is never broken by search rank. Avalueanswers found, disagreement, or unknown, reading a number before a unit or the first capture group of a caller pattern. - No probabilities. Answers are evidence statuses. A check never changes trust state, and
supportedis not verification. A named validation is shown beside an answer but does not override it. - Budgets and recovery. One call uses one run-budget unit. Results fit a byte budget, keeping one span from each side of a disagreement, then reporting
insufficient_budgetwith expand references. explainaccepts a check hash and marks evidence whose source file changed since the check ran.- Launch roots. MCP checks skip sources outside the roots and no longer list out-of-root projects by name or identifier.
DESIGN-RSO-CHECK.mdships with the package and carries the rules and the v1 limits.
Verification and limits
- 137 regression tests passed on Windows with Python 3.12.5 (one POSIX-only test skipped), including a stdio MCP
rso_checkround trip inside its byte budget and a root-isolation test that fails without the fix it covers. - GitHub Actions passed on Windows, Ubuntu, and macOS with Python 3.11 and 3.12 for the release commit.
- The ZIP was built from the tagged commit, installed into a temporary prefix, and checked with
--version,doctor, and a scratch-ledgeruseandcheck. - Answers were spot-checked against this repository, not measured for accuracy on a labeled set. The lexical rules have known gaps: a sentence that merely mentions an option counts as selecting it, and conflicts phrased in different words are missed. Read the evidence before acting on an answer.
- Gemini CLI and Antigravity setup remain covered by configuration tests only, with no live session recorded.
- This is alpha software. Checks on one host do not establish behavior on untested hosts. See VERIFICATION.md.
RSO Context Alpha 0.8.1
RSO Context Alpha 0.8.1 fixes false disagreements and empty compact packets, adds MCP setup for Gemini CLI and Antigravity, and packages the Linux and launcher fixes that reached main after 0.8.0.
Download and install
Download RSOContextAlpha-0.8.1.zip, extract it, and follow QUICKSTART.md or INSTALL.md. Run python -B install.py from the extracted folder. Check that the installed command reports 0.8.1 and doctor reports ready: true. Python 3.11+, Git, and SQLite FTS5 are required. Verify the download against SHA256SUMS.txt.
The package contains 57 allowlisted files: LICENSE, NOTICE, the installer, user documentation, runtime, and regression tests. It contains no personal ledger or host configuration.
Changes
- Fewer false conflicts. Disagreement detection compares sentences. A negated sentence and an affirmative one conflict only when they share at least two topic words and come from different files. Before, any "do not" in one file plus any "must" or "required" in another set
clarification_needed. - No more empty compact packets. When conflicting evidence exceeds the byte budget, the packet keeps the smallest two-sided pair of spans and lists the rest as expand refs. If even that can't fit, it reports
insufficient_budgetand keeps the refs sorso_expandcan recover the evidence. - Gemini CLI and Antigravity MCP setup.
rso-context mcp --setup --client geminiwrites~/.gemini/settings.json;--client antigravitywrites~/.gemini/config/mcp_config.json. Each has an environment override, anddoctorreports both. RSO_MCP_RUNTIMEin the Windows launchers. Thecmdand PowerShell launchers use it when set and stop with a clear message when it has nopython.exe.- Linux and Git Bash fixes from
mainafter the 0.8.0 package are now in the ZIP. DESIGN-RSO-CHECK.mdproposes typed evidence questions. It is a design note only; nothing in this release implements it.
Verification and limits
- 114 regression tests passed on Windows with Python 3.12.5 (one POSIX-only test skipped).
- GitHub Actions passed on Windows, Ubuntu, and macOS with Python 3.11 and 3.12 for the release commit. One Windows 3.11 job first failed a 2-second timing assertion on the runner (12.3 s; 0.4 s locally) and passed on rerun.
- The ZIP was built from the tagged commit, installed into a temporary prefix, and checked with
--version,doctor, and a scratch-ledgeruseandquery. - Gemini CLI and Antigravity setup are covered by isolated configuration tests only. No live MCP session with either host has been recorded.
- This is alpha software. Checks on one host don't establish behavior on untested hosts. See VERIFICATION.md.
RSO Context Alpha 0.8.0
RSO Context Alpha 0.8.0 adds an optional local stdio MCP adapter to the shared evidence ledger. The ordinary CLI remains available without third-party Python dependencies.
Download and install
Download RSOContextAlpha-0.8.0.zip, extract it, and follow QUICKSTART.md or INSTALL.md. Run python -B install.py from the extracted folder. Check that the installed command reports 0.8.0 and doctor reports ready: true. Python 3.11+, Git, and SQLite FTS5 are required.
The attached package contains 56 explicitly allowlisted files, including LICENSE, NOTICE, the installer, user documentation, runtime, and regression tests. Each person keeps a separate ledger; the archive contains no personal index or host configuration.
Changes
- MCP tools: use, resume, query, explain, and evidence expansion.
- Compact evidence packets with byte budgets and expansion references.
- Bounded ingestion coverage reports and explicit empty-result diagnostics.
- Codex and Claude Code setup/removal helpers; MCP runtime pinned to
mcp==2.2.0in an isolated environment. - Fixes for commented TOML header preservation, launch-root filtering, and serialized MCP response budgets.
- Expanded README, beginner download instructions, installation guide, manual command/tool references, changelog, and publication checklist.
Verification and limits
- All 103 combined local regression tests passed on Windows.
- The archive matches the exact Git blobs in source commit
2a91034bbb36818ece2594715a15e8b01ba45302. - Installation from the final ZIP into a temporary prefix reported 0.8.0, passed doctor, and returned current synthetic evidence using a scratch database.
- Live Codex MCP use/resume/query/explain/expand was previously checked on Windows. Claude Code and Antigravity have completed the first two phases of the ongoing shared-ledger CLI trial; that trial is not a completed performance comparison or live MCP acceptance for those hosts.
- Native macOS execution remains unverified. This is alpha software and an evidence-ledger subset of the RSO research proposal.
SHA-256 for RSOContextAlpha-0.8.0.zip:
6354078bf595cf74072521271866e6e1e822a9610f479751a5333995cf4f28c2
See the packaged CHANGELOG.md and VERIFICATION.md for upgrade guidance and the scope of recorded checks. Repository visibility remains private.
The Kinestrand repository preserves both source histories and its original 25 tests alongside the 78 focused tests. All source and download links now point to Kinestrand/RSOContextAlpha.
GitHub CI passed on Windows with Python 3.11 and 3.12: https://github.com/Kinestrand/RSOContextAlpha/actions/runs/34562116750