v0.5.0

What's new

Bundled stdio MCP transport (Phase A)

• Zero-dependency McpStdioTransport speaking LSP-style Content-Length framed JSON-RPC to a child MCP server process, with bounded frames, per-request timeouts, error-code mapping, and spawn/exit failure propagation.
• Checked-in conformance server and pnpm test:mcp-live, including a tampered-digest fail-closed run. SSE/HTTP transports remain consumer-provided.

Federated ecosystem indexes (Phase B, experimental)

• RegistryIndexClient validates pinned entries (40-hex commit, skill name, owner/repo, advisory tier) and bounds listings.
• Index seeds honor their own pinned commit and join the existing repository enrichment, evidence, trust, and approval pipeline.
• official | verified | community | unreviewed tiers are advisory evidence only: they surface as quality signals and warnings and never grant trust or relax any boundary.
• Off by default (registryDiscovery); register with ctx.skillFlux.registerRegistryIndex(label, transport).

Token value governance (Phase C)

• Persisted running body-token totals join installed-cache pruning evidence. Among versions tied on recency, uses, and mounts, the higher token-cost version is evicted first; deterministic and backward compatible.

Scope notes

• No new runtime dependencies were added across all three phases.
• Deferred work is unchanged: LLM Router, GUI marketplace, heavy malware scanning.

No breaking changes

• Existing configuration and behavior carry over from v0.4. The peer range >=0.1.1-rc.2 <0.2.0 remains verified against DeepSeek Harness 0.1.6-alpha.2.

Evidence

• pnpm check: typecheck, oxlint zero warnings, 244 tests, build.
• pnpm eval: 36 corpus cases including the 12-case MCP and 10-case registry entry contracts.
• pnpm test:discovery-live, pnpm test:runtime-live, and pnpm test:mcp-live: real online discovery, a real remote lazy install, and the stdio conformance round trip with a tampered-digest fail-closed run.
• Live demo against DeepSeek Harness 0.1.6-alpha.2: native token-meter and portable-fallback paths, plus a full dsh CLI headless boot with a real DeepSeek turn.

v0.4.0

What's new

MCP Skills sources (Phase A)

• Transport-agnostic McpSkillsClient speaking the ratified MCP Skills extension (SEP-2640): skills/list, skills/get, resources/read, with pagination and strict entry validation.
• Digest and size verification on every read, field-by-field frontmatter verification, duplicate-name disambiguation, and refusal of "dynamic" skills.
• Content-bound approvals and cache installs keyed by host-assigned server label, SKILL.md URI, and the digest set; changed content lands at a new cache id and cached files are re-digested on every load.
• MCP candidates reuse the existing trust/blocked-server/approval/catalog-budget and fail-open discovery boundaries, including restart-safe cache restoration.
• Register a source with ctx.skillFlux.registerMcpSource(label, client). New config keys: mcpDiscovery, mcpTrustedServers, mcpBlockedServers. No MCP dispatch sandbox and no skill execution: discovery and loading only.

Token telemetry (Phase B)

• Optional integration with the DSH token-meter service, resolved at runtime with no package dependency; missing or failing meters silently fall back to the portable catalog-budget estimate.
• Per-skill catalog-footprint and loaded-body token counts with an estimator marker, persisted as backward-compatible usage records and shown by /skillflux status and /skillflux usage.
• Only token counts are stored; task text and skill bodies never reach usage.json.

Fixes

• Accept mcp origin in persisted usage records so MCP mounts accumulate usage evidence.

No breaking changes

• Existing configuration and behavior carry over from v0.3 unchanged. The peer range >=0.1.1-rc.2 <0.2.0 remains verified against the targeted DeepSeek Harness 0.1.6-alpha.2 line.

Deferred

• LLM Router: no meaningful community traction in the last 30 days.
• GUI marketplace and heavy malware scanning: intentionally out of scope.

Evidence

• pnpm check: typecheck, oxlint with zero warnings, 227 tests, build.
• pnpm eval: 35 corpus cases including the 12-case MCP entry contract.
• pnpm pack --dry-run: exactly the intended package files.
• pnpm test:discovery-live and pnpm test:runtime-live: real online discovery and a real remote lazy install end to end.
• Live demo against DeepSeek Harness 0.1.6-alpha.2: real stdio MCP server with content-bound install, native token-meter and portable-fallback paths, and a full dsh CLI headless boot with a real DeepSeek turn.

dsh-skillflux v0.3.0-rc.0

v0.3.0-rc.0

Provider-native lazy runtime for DeepSeek Harness. All six v0.3 phases are merged into main and this is the first release candidate.

What's new

• Provider-native lazy runtime - routing publishes metadata-only summaries; cached and remote bodies download, SHA-256 verify, and load only when the model calls the skill tool. Registry skills still mount eagerly. Same-name candidates form a fallback chain under one shared install deadline, and the provider catalog releases at urn/end.
• Approval on lazy activation - the configured �lways / session / �utomatic policy gates remote downloads at the skill boundary, failing closed without an approval channel.
• Adaptive discovery - a confident local shortlist skips the network; partial hits fill remaining slots. Per-source health cooldowns degrade failing providers, and non-authoritative observations keep last-good catalogs during outages.
• DeepSeek Harness 0.1.6 compatibility - catalog history reads work across the 0.1.1 �vents array and the 0.1.6 seq/�ventAt accessors; the provider registers on the host registry with per-agent scope routing.
• Better Chinese routing - short CJK queries now score verbatim fragments in Chinese descriptions.
• Evaluations - 189 unit tests and 34 evaluation cases, including provider-native lazy-runtime, remote fallback, approval, concurrency, cancellation, and turn-release scenarios.

Verified against

• DeepSeek Harness