Tags give the ability to mark specific points in history as being important
-
v0.1.52
Release: v0.1.52f8547d59 · ·v0.1.52 — attempt-spine observability + token accuracy + fix-brackets Observability (retries & re-reviews are first-class now): - Node execution SPINE: watch/turns rebuilt around (session-id, attempt) — one block per attempt, review rounds inline, camera follows the live node. node-attempts RPC + graph node-attempts reads. - Per-round contract reviews surfaced (fail reason = the rework spec), no longer overwritten by the latest verdict. - HTML in tool outputs escaped (was destroying the watch layout). - Stable graph-tree node order (was reshuffling every refresh). Correctness: - Token double-count fixed: OpenAI-family prompt_tokens is cache-inclusive; only Anthropic input needs cache added. A 57.6M run had displayed as ~103M. - Run-auditor brief now hands over the contract-reviewer sessions, so it verifies a verdict's claims against the session that MADE them (was filing false findings against clean nodes). - Declared-failure exit (complete-task outcome=fail always accepted) replaces the open-items completion deadlock; loop runtime droppings out of workspace. Tooling / safety: - lisp-editor fix-brackets: mechanical bracket repair (.asd included) — ends the paren-hand-counting thrash. - naive-sup guard: refuses to restart/stop the orchestrator while agents are live on the hub (--force override). Smoke gate: native + claude + codex graphs all green before cut.
-
v0.1.51
Release: v0.1.51d3498cd2 · ·v0.1.51 — knowledge graph traversal via cl-naive-graph cl-naive-knowledge routes traversal through the new cl-naive-graph library (persistence to cl-naive-documents, 1M-edge scale-smoked). Verified: graph 113/0, knowledge 66/0, orchestrator loads, three-path release smoke green, and a live knowledge query through the stdio MCP on freshly-compiled fasls.
-
v0.1.50
Release: v0.1.50d3498cd2 · ·v0.1.50 — sleep-inhibited boot, synchronous shutdown, docs restructure - naive-sup boot holds a systemd sleep inhibitor for the hub's lifetime (idle host suspend was killing in-flight agent runs); shutdown is synchronous (waits for the port; SIGKILL escalation), so shutdown && boot is safe. - Docs restructure: load-bearing-only top level, sundry/archived for dated material, known-issues backlog split open/resolved, observer-fixer promoted to a future spec; published paper (Zenodo DOI 10.5281/zenodo.21139628) cited. - Shipped offline docs track the load-bearing set.
-
v0.1.49
Release: v0.1.4977c3c410 · ·v0.1.49 — shared MCP client (cl-naive-mcp-client) + skill-registry write side - cl-naive-mcp-client: MCP protocol/transports/lifecycle extracted from agent-loop; connect-tools consumer surface for in-process hosts (canvas). - skill-registry: validate-proposal / install-proposal / provenance — the three-layer CI gates codified as one install gate. - All three execution paths graph-verified on this exact tree (native+MCP, claude, codex).
-
v0.1.48
Release: v0.1.480e6408d6 · ·v0.1.48 — codex worker path fixed; beta notice + 1.0 readiness gate - codex cli-adapter workers: disable codex's own sandbox when already bwrap-confined (nested Bubblewrap failed -> 0-turn nodes). Verified by a live codex worker graph completing through the verification gate. - README notice softened to Beta (use tagged releases); the remaining gate to 1.0 tracked in docs/production-readiness.md.
-
v0.1.47
Release: v0.1.4776b27f67 · ·v0.1.47 — skill-registry compose() + folder-sourced roles - compose(naive-home domain role): one-call composition -> :domains (requires- expanded) :roles (roles/ folder scan) :skills :tools. Public entry point for consumers (canvas, launchers, UIs) — no agent-loop dependency needed. - available-roles exported — roles discovered from the roles/ folder (single source of truth), not per-domain config lists that drift. - domain/role specs accept a comma-string OR a list uniformly across the API.
-
v0.1.46
Release: v0.1.4676b27f67 · ·v0.1.46 — worker-image spawn, domain-resolution consolidation, agent guidance - NAIVE_SPAWN_FROM_IMAGE: opt-in dev worker spawn from the prebuilt image (shared core, big per-worker memory saving). - Domain resolution consolidated in cl-naive-skill-registry: resolve-skill-names now expands domain :requires internally (correct standalone); agent-loop delegates. Verified live through native-loop AND cli-adapter graph runs. - agent-baseline + generic-floor: 'spill large output to a scratch file and search it' data-access rule; complete-task clarification propagated.
-
v0.1.45
Release: v0.1.4598c38d6e · ·v0.1.45 — dashboard perf + always-live token totals - session-query / graph-list token totals batched into single-pass indexes (was O(sessions x transcript) / per-graph scans): dev hub graphs 1.5s->0.05s, sessions 1.5s->0.11s, graph-detail 3s->0.15s. - Node token totals now always-live (worker+recovery+judge) via a per-request transcript index — no staleness, no per-node scan. - Pagination on the graphs + sessions lists with an on-page rows-per-page field. - Node token-count-total no longer stamped-then-stale after the implementer.
-
v0.1.40
cd265f00 · ·naive binaries v0.1.40 — naive-sup config-verify routing fix (+ v0.1.39 agent-tools MCP)
-
v0.1.39
71a4bd8b · ·naive binaries v0.1.39 — agent-tools MCP verdict/completion-via-call + config-verify
-
-
-
-
-
v0.1.34
7734dbaa · ·v0.1.34 — confined dep-resolution (source-registry live trees) + analyzer code-index framing
-
v0.1.33
d6facc03 · ·v0.1.33 — bash-tool two-shell fix + mandatory/cheap analyzer skills (analyze-file refresh)