Skip to content

armgit / mso

Public Imported from GitHub
0

Browser-based graphical shell and control plane for a Linux server you own — real terminal, file manager, live metrics and BYOK AI in one mobile-first workspace. Next.js 16 · React 19 · node-pty · Tailwind v4 · shadcn/ui. Public alpha, self-hosted.

docs: publish catalog-first RASMIC runtime

Rahman1715ee5e3927
diff --git a/docs/RASMIC.md b/docs/RASMIC.md
index ba31dc00..be44f046 100644
--- a/docs/RASMIC.md
+++ b/docs/RASMIC.md
@@ -66,7 +66,7 @@ Memory is compact structured JSON, not raw chat history. It distills what happen
Lifecycle: `active → confirmed → superseded → archived`.
-Retrieval combines lexical relevance, confidence, importance, freshness, lifecycle, and manual-user authority. Superseded/archived records are excluded by default, stale records lose rank, and a current failed manual user test outranks a conflicting automated healthy assumption. `project_memory_search` is read-only and does not create `.agent/` merely because an agent searched. `project_memory_upsert` writes one task/debug/test/decision/failure record. LOW trivial successful tasks remain memory-light unless explicit evidence or relevant memory warrants persistence.
+Retrieval combines lexical relevance, confidence, importance, freshness, lifecycle, and manual-user authority. Superseded/archived records are excluded by default, stale records lose rank, and a current failed manual user test outranks a conflicting automated healthy assumption. `project_memory_search` is one token-efficient read surface with `search`, `related`, and `timeline` views; relation edges are derived deterministically from supersession, scope/tags, result conflicts, and lexical overlap rather than an LLM call. It does not create `.agent/` merely because an agent searched. `project_memory_upsert` writes one task/debug/test/decision/failure record. LOW trivial successful tasks remain memory-light unless explicit evidence or relevant memory warrants persistence.
## Manual user tests
@@ -103,9 +103,15 @@ A one-off trace may remain searchable history but is not returned as a reusable
Stop on `maxIterations`, repeated no-progress, new regressions, destructive uncertainty, or required human decision. Static analysis, tests, security scans, architecture rules, LLM reviewers, and external reviewers are optional backends; none is mandatory.
+## Catalog-first token budget
+
+The terminal runtime routes the latest intent through a shared deterministic capability catalog **before** projecting model history. Known intents do not pay for `skills_search` or an always-on tool core. Capability packs are phase-aware, dependency-complete, and bilingual for the common Indonesian/English operations used by MSO. The model sees only the small selected schema pack and a route-requested history slice; broader semantic discovery is the fallback for genuinely unknown intent.
+
+The benchmark currently gates 100% required-tool recall and deterministic selection while averaging 2.7 active tools / 2,252 schema bytes per turn, a 95.8% reduction from the 70-tool catalog. Routing text averages 76 bytes and the route-requested history budget 10,571 tokens in the checked corpus. These are reproducible harness metrics, not a claim about total provider billing. Runtime status/one-shot JSON expose routing metadata outside the prompt so real token-per-success can be measured without spending model context on telemetry.
+
## Metrics
-`workflow_finish` reports tool calls, execution duration, retries, memory hits, approximate context tokens, recipe reuse/stage, and whether a script candidate was created. The objective is repeated reasoning ↓, tool calls ↓, context usage ↓, consistency ↑.
+`workflow_finish` reports tool calls, execution duration, retries, memory hits, approximate context tokens, recipe reuse/stage, and whether a script candidate was created. It also reports reservation/worktree cleanup guidance. The objective is repeated reasoning ↓, tool calls ↓, context usage ↓, consistency ↑.
## Security
@@ -115,15 +121,15 @@ RASMIC redacts common token/password/API-key/Authorization/private-key/credentia
RASMIC is additive. Existing workflow ids and learned recipe storage remain valid; old in-flight workflows without RASMIC metadata can close under the old finish contract; new start fields and finish evidence are optional additions; `.agent` is portable/optional; no new package dependency is required.
-## Future optional API
+## Optional external memory transport
-Repo-local memory stays canonical. A future service can map to `memory.search`, `memory.get`, `memory.related`, `memory.timeline`, `memory.upsert`, `memory.supersede`, `recipe.find`, and `script.find` without becoming a core dependency.
+Repo-local memory stays canonical. MSO now has internal bounded export/import adapters for a portable redacted memory bundle, so a future external service can synchronize records without becoming a core dependency. `related` and `timeline` already exist locally through `project_memory_search`; no extra model tool names were added for them.
## MCP surfaces
- `workflow_start`: classify, collision check, minimal memory/recipe/script retrieval.
- `workflow_finish`: Evidence Receipt, memory, recipe quality, metrics, script-candidate promotion.
-- `project_memory_search`: ranked repo-local retrieval.
+- `project_memory_search`: ranked `search`, deterministic `related`, or compact `timeline` projection through one read tool.
- `project_memory_upsert`: structured memory and manual-test ingestion.
- `project_script_run`: bounded read-only replay and candidate→tested promotion.
diff --git a/docs/SECURITY-ASSURANCE.md b/docs/SECURITY-ASSURANCE.md
index 00da4310..924b2976 100644
--- a/docs/SECURITY-ASSURANCE.md
+++ b/docs/SECURITY-ASSURANCE.md
@@ -47,7 +47,7 @@ From a clean development checkout with Docker and the normal MSO toolchain:
bun run security:ultimate
```
-The command creates a `0700` private run directory under `~/.mso/security-assurance`, snapshots **tracked files only**, and runs the repository verification plus Trivy, OSV-Scanner, Gitleaks, Semgrep, ShellCheck, the official Codex Security component scan across the full tracked repository and the passive ZAP baseline. Detailed scanner output stays outside the repository. A failing scanner produces a failing command; the script does not convert scanner outages into security passes.
+The command creates a `0700` private run directory under `~/.mso/security-assurance`, snapshots **tracked files only**, and runs the repository verification plus Trivy, OSV-Scanner, Gitleaks, Semgrep, ShellCheck, the official Codex Security component scan across the full tracked repository and the passive ZAP baseline. OSV-Scanner is a version-pinned official release binary cached under owner-local MSO state and verified against its pinned SHA-256 before use; the gate does not depend on a mutable/denied GHCR image. Detailed scanner output stays outside the repository. A failing scanner produces a failing command; the script does not convert scanner outages into security passes.
Codex Security is intentionally not an automatic per-commit hosted gate because it consumes model/API budget. MSO uses OpenAI's official `scan-components` command because repeated monolithic Standard scans of this repository reached bounded spend limits without producing a terminal coverage result. The committed [`security/codex-components.json`](../security/codex-components.json) plan was generated by the official planner and covers all tracked files through twenty non-overlapping components. The official auto-plan was refined along subsystem boundaries after bounded trial scans proved that its two largest groups were too broad to reach a terminal result within the per-component guard. Each component receives a Standard scan, then Codex Security performs cross-component root-cause matching; incomplete component coverage or incomplete matching fails closed. The default guard is $8 **per component**, with two concurrent workers. MSO pins the component lane to `gpt-5.6-terra` with `high` reasoning because calibration on the security-critical `lib/host` component reached complete terminal coverage under that bounded budget, while `gpt-5.6-sol` at xhigh/medium repeatedly exhausted bounded spend without a terminal coverage result. Model, effort, budget and workers remain explicit overrides through `CODEX_SECURITY_MODEL`, `CODEX_SECURITY_EFFORT`, `CODEX_SECURITY_COMPONENT_MAX_COST_USD` and `CODEX_SECURITY_COMPONENT_WORKERS`. The result gate independently validates `summary.json` and the combined `findings.json`; High/Critical groups fail the run. A maintainer can explicitly skip the lane with `MSO_SECURITY_SKIP_CODEX=1`, but a run with a skipped lane must not be advertised as the full ultimate gate. ZAP is likewise skippable for offline development, but not for a published assurance run.
diff --git a/lib/mcp/toolset.ts b/lib/mcp/toolset.ts
index 69a41120..f7d55447 100644
--- a/lib/mcp/toolset.ts
+++ b/lib/mcp/toolset.ts
@@ -3,8 +3,8 @@ import type { Scope } from "./scope";
import type { McpTool } from "./tool-kit";
export const MCP_SERVER_VERSION = "1.6.0";
-export const MCP_TOOLSET_VERSION = "2026.09.02.8";
-export const MCP_TOOLSET_CHANGED_AT = "2026-09-02T17:00:00Z";
+export const MCP_TOOLSET_VERSION = "2026.09.02.9";
+export const MCP_TOOLSET_CHANGED_AT = "2026-09-02T18:30:00Z";
export type McpToolsetInfo = {
serverVersion: string;
diff --git a/scripts/check-docs.mjs b/scripts/check-docs.mjs
index 2bb585e1..92422b68 100755
--- a/scripts/check-docs.mjs
+++ b/scripts/check-docs.mjs
@@ -19,6 +19,9 @@ const toolFiles = [
"lib/mcp/tools-read.ts",
"lib/mcp/tools-discovery.ts",
"lib/mcp/tools-learning.ts",
+ "lib/mcp/tools-workflow-start.ts",
+ "lib/mcp/tools-workflow-lifecycle.ts",
+ "lib/mcp/tools-project-memory.ts",
"lib/mcp/tools-infra.ts",
"lib/mcp/tools-power.ts",
"lib/mcp/tools-agent.ts",