rubik 0.1.6 → 0.3.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: f281eb015bdfadb67a18d2a61972c2559f448a0458915393443a39c042b8a1d9
-  data.tar.gz: dedd94f98ab0785e0688e256b8be35244d764c88ccbe87c9367836715c3c788d
+  metadata.gz: 14020218aa50674cdf508953465d4f31c285420979b8a9c518ef4d5e6f6ef300
+  data.tar.gz: caa28cd931b29aecf1de25e66d794a304f04ea0331e9b58f2b198c985488a8e2
 SHA512:
-  metadata.gz: bcaf73fa4a12d8084af7b7e8ae5f1886e14c0f55f7e8817112addc073d17533fce0bc0959971fc18654d217efb24cadd0cdb4fa128c9c6980f49c0ff7c3255d3
-  data.tar.gz: 7d5dfa25a399e9e2710d9cbd16ec7c5cb0cbc7b373c62dfd9b3eae76bc4505df08f1bb51f9d893715ca4045da61b9b25c0b8b9c835c4960549788a98fb256f0b
+  metadata.gz: 27fdbf68f5e00c88e87bdfe3ba33e837783b3b4a573ae3ccbd8db707fd7b320af8256c5a875811d33feaaf2118f18c701579266bdb2795d27815d0d0199d30b7
+  data.tar.gz: d050ff897cbe97a68399a31a28ca68e53f7edf7e894e097bbba338f722406975f397cd5fcd3fe21946c98d5c6a8a158f7193c04979d7878dcd985cdc524083b0

data/ATTRIBUTION.md

@@ -121,7 +121,7 @@ SOFTWARE.
 | **`strings-ansi`** | Piotr Murach | MIT | ANSI-aware string truncation and alignment for REPL formatters. |
 | **`reline`** | Hiroshi SHIBATA & Ruby Team | Ruby License / BSD-2-Clause | Line-editing, history navigation, autocompletion hooks, and headless terminal adaptation. |
 | **`prism`** | Kevin Newton & Ruby Core Team | MIT | High-performance Ruby parsing and AST extraction ([`Rubik::AST::RubyExtractor`](lib/rubik/ast/ruby_extractor.rb)). |
-| **`minitest`** | Ryan Davis & seattlerb | MIT | Fast, modular test suite powering Rubik's 3,600+ unit, integration, and contract tests. |
+| **`minitest`** | Ryan Davis & seattlerb | MIT | Fast, modular test suite powering Rubik's 3,700+ unit, integration, and contract tests. |
 | **`rubocop`** | Bozhidar Batsov & contributors | MIT | Static code analysis, style enforcement, and architectural hygiene gates. |
 | **`benchmark`, `logger`, `pstore`, `reline`, `fiddle`** | Ruby core / rubygems teams | Ruby License / BSD-2-Clause | Ruby stdlib gems pinned explicitly in `rubik.gemspec` as they are unbundled from the default set in Ruby 3.4/3.5. |
 
@@ -146,3 +146,7 @@ SOFTWARE.
 
 Rubik itself is licensed under the **MIT License**.  
 See the [`LICENSE`](LICENSE) file for the full license text.
+
+---
+
+See the [documentation index](docs/index.md) or return to the [README](README.md).

data/CHANGELOG.md

@@ -10,6 +10,175 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 Nothing yet.
 
+## [0.3.0] - 2026-10-18
+
+The subagent-governance minor: runtime-defined archetypes become
+context-scoped and reversible, spawn turns transactional, outcomes
+are measured, and the first interactive launch asks before trusting
+a workspace. Prompt teachings and tool schemas are aligned against
+the declared contracts.
+
+### Added
+
+- **Context-scoped, reversible archetype registration (#175, closes
+  #171):** runtime `define_subagent` profiles are owned by an explicit
+  runtime context with LIFO revision stacks and context-bound disposal
+  inverses — disposing the owner pops the registration and restores
+  the prior profile; stale out-of-order cleanups remove only what they
+  own; built-ins stay reserved; re-registering a live archetype
+  requires `replace: true`; and nested profiles are owned and cleaned
+  up by the child context.
+- **Workspace trust prompt (#183, closes #182):** interactive launches
+  in an untrusted workspace show the resolved path, explain Rubik's
+  file and command capabilities and the `buffer.model_can_apply:
+  false` limitation, and offer a per-workspace Trust and continue or
+  Exit choice (exit is the default); script and non-interactive
+  launches do not block on the prompt.
+- **Validated subagent profile contracts (#177, closes #172):** custom
+  archetypes declare required coeffects and allowlisted capability
+  grants; profiles default to no write or nested-spawn grants; invalid
+  names, dependencies, grants and field limits return structured
+  validation errors before registration; custom tool filters derive
+  only from allowlisted grants.
+- **Subagent profile outcomes and bounded prompt visibility (#179,
+  closes #174):** root prompts keep every built-in and show only the
+  eight newest active custom profiles; run journals record profile
+  identity and version, outcome, bounded redacted error summaries,
+  duration, step and tool-call counts, and provider-reported token
+  usage — missing token and cost data are marked unavailable, never
+  estimated.
+
+### Fixed
+
+- **Subagent spawn is transactional (#178, closes #173):** required
+  coeffects are validated before the worker starts; a startup failure
+  rolls back the child context, scoped registry, kernel registration,
+  instance and generated ID; each instance pins an immutable profile
+  snapshot (redeploys keep the original); and honest spawn failures
+  reach the maestro.
+- **The system prompt follows active tools and mode (#170, closes
+  #166):** teachings for gated tools are filtered when their
+  coeffects are unsatisfied, plan-only mode directs no tool calls or
+  synthesis, and read-only orientation vs delegated execution is
+  harmonized across rules and role headers.
+- **Storage and Vakanüvis write rules are aligned (#169, closes
+  #165):** direct `storage put` targets work artifacts
+  (goal/plan/review/note) while durable `memory` cards and `wiki`
+  pages route through `vakanuvis` record/digest when available; the
+  tool declaration documents the contract and the coder prompt
+  surfaces durable lessons to the maestro.
+- **Tool teachings match declared schemas (#168, closes #167):**
+  `replace_file_content` guidance drops the unadvertised aliases
+  (runtime tolerance preserved) and `synthesize_tool` requires
+  `params_schema` when the synthesized tool accepts arguments.
+
+### Changed
+
+- **Documentation navigation:** user and developer guides reorganized
+  under `docs/`, README quickstart and feature summary refreshed, and
+  GitHub issue and pull request templates plus a code of conduct
+  added; provider and local model IDs refreshed across README,
+  configuration, command, API and first-session examples.
+- **Process stream capture is consolidated (#180):**
+  `Rubik::Platform.popen_capture` centralizes Open3 startup,
+  concurrent stdout/stderr capture, exit-status mapping and timeout
+  process-tree termination; ProcessDriver, Bubblewrap, Docker and
+  Podman drivers share it.
+- **Hot paths get cheaper:** `TokenEstimator.estimate_text` switches
+  to exact integer math with zero allocations (~7.8% on history
+  estimation, #181); `Store#search_ranked` ranking cuts allocations
+  by ~43% (#176).
+- **Deps:** `tree_sitter_language_pack` bumps to 1.22.1 (#162); dev
+  rubocop to 1.92.0 (#163).
+
+## [0.2.0] - 2026-10-16
+
+The first minor of the 0.2 era is a hardening minor: a reproducible
+runtime validation harness with capacity proofs, a fail-closed
+security posture across the command guard, transactional plugin
+reload, and failure reporting that names what actually happened.
+
+### Added
+
+- **Reproducible runtime validation harness (#160, closes #151):**
+  `bundle exec ruby dev/validate_runtime_capacity.rb` runs without
+  provider credentials and writes a dated report — environment, seed,
+  fixture sizes, raw metrics, resource checkpoints, thresholds and
+  verdicts. Nine lifecycle families soak for the configured 1,000
+  epochs (real subprocess shutdown, failed JSON-RPC handshake,
+  crash/recovery); two single-run probes cover in-flight
+  disable/re-enable and known-descendant cleanup. Missing
+  capabilities report INCONCLUSIVE (exit 2), failed thresholds FAIL
+  (exit 1). The soak also surfaced production fixes that ship here:
+  nonblocking IO hooks keep their return codes and minimum lengths,
+  select wakes only fibers interested in each readiness direction,
+  queued wakeups carry their park generation so stale events cannot
+  resume a later operation, and a crashed MCP pipe closes with the
+  process tree terminated before stdin (recursive descendant
+  discovery on Unix).
+
+### Fixed
+
+- **Command classifier fails closed (#143):** a classifier crash or a
+  partial tree-sitter parse no longer downgrades to the catastrophic
+  regex gate — both raise structured blocks (`:classifier_error` /
+  `:parse_ambiguous`) with the crash cause audited; the catastrophic
+  regex keeps first refusal for precise destructive-shape reasons and
+  the operator allowlist stays the only valve. SECURITY.md's
+  fallback residual is updated to the fail-closed policy.
+- **Plugins verify whole packages and reload transactionally
+  (#144):** directory-plugin manifests hash every regular file under
+  the package root (additions, deletions and modifications audited;
+  symlinks and special files are containment violations); legacy
+  entry-file-only manifests require explicit re-consent; and
+  `reload_plugin` discovers, verifies and prepares the replacement
+  before touching the running version — a failed reload leaves the
+  old version running instead of claiming success.
+- **Runtime repair verifies before promoting (#145):** LSP candidates
+  must complete an isolated initialize handshake and other targets
+  run bounded argv probes that preserve the proposed arguments; live
+  promotion records inverses and unwinds the transaction on failure,
+  rolling back to the previously configured client (with an honest
+  report if restoration itself fails).
+- **Buffer apply names its authority and its partial truth (#146):**
+  `buffer.model_can_apply` is honored in every registry rebuild
+  (branch workspaces included); a multi-file `apply!` failure raises
+  a structured `ApplyFailedError` carrying what reached disk, what
+  failed and what a retry still writes; retry is incremental and the
+  tool and `!apply` render one shared failure vocabulary — a partial
+  apply is never reported as "Successfully applied".
+- **C++ extraction sees classes and reports parse status (#147):**
+  the shared C/C++ walker no longer crashes on unknown declarator
+  shapes — class/struct/union symbols, namespaces, in-class members
+  (definitions, declarations, ctors, operators) and templates are
+  recorded; the extractor returns a structured status
+  (`:ok`/`:partial`/`:failed`/`:unavailable`) and a failed reindex
+  keeps the file's previously indexed symbols instead of silently
+  deleting them.
+- **Plugin reinstallation survives symlinked ancestors (#158):**
+  resolved-source × lexical-destination comparison could delete the
+  source before copying it; file identity now detects an
+  already-installed directory.
+- **Unit tests are isolated from external network and terminal env
+  (#148):** handshake-deadline assertions no longer depend on cleanup
+  timing.
+
+### Changed
+
+- **CI measures the platform matrix (#158):** four jobs — Linux Ruby
+  3.4/4.0, macOS Ruby 4.0, Windows Ruby 4.0 — each install the gem
+  into an isolated GEM_HOME, run CLI/require smoke and all three
+  suite loaders, and publish evidence artifacts; README and TESTING
+  replace the universal green-run guarantee with commit-specific
+  evidence.
+- **Docs calibrate absolute guarantees (#159):** README, SECURITY,
+  CONFIG, PLUGINS and SELF_EVOLUTION swap "zero-risk"-style claims
+  for concrete preconditions and bounded behavior (the full buffer
+  review gate is `auto_apply: manual` + `model_can_apply: false`).
+- **Cold scan gets cheaper (#161):** integer-micros mtime math, a
+  backslash translation that runs only when backslashes exist, and
+  in-place dedup; the fast scan keeps bracket-class ignores intact.
+
 ## [0.1.6] - 2026-10-15
 
 The execution-contract release: subagent runs become first-class,