release: prepare v0.2.0

Update CHANGES.md, README.md (introspect docs, version refs, pipeline
diagram, command table), fix report contradictions and stale stats,
bump version.edn to 0.2.0.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: leifericf

CHANGES.md

@@ -2,7 +2,58 @@
 
 All notable changes to Noumenon are documented in this file.
 
-## 0.1.0 — 2026-03-27 — Initial Public Release
+## 0.2.0
+
+### Features
+
+- **Introspect** — Autonomous self-improvement loop. An LLM optimizer
+  analyzes benchmark gaps and proposes changes to prompts, example queries,
+  Datalog rules, source code, and ML model hyperparameters. Improvements are
+  kept; regressions are reverted automatically. Inspired by Karpathy's
+  autoresearch. See [reports/introspect-development-2026-03-28.md](reports/introspect-development-2026-03-28.md).
+- **ML query routing model** — On-device feedforward network predicts which
+  Datalog queries to try for a given question. Trained via introspect, runs at
+  zero token cost. Integrated into the ask agent as optional hints.
+- **Multi-repo evaluation** — Introspect evaluates across multiple repos to
+  prevent overfitting prompts to a single codebase.
+- **Async MCP introspect** — `noumenon_introspect_start`, `_status`, `_stop`
+  tools for background optimization runs.
+- **Internal meta database** — Dedicated Datomic database (`noumenon-internal`)
+  for cross-repo introspect history, with 5 named Datalog queries.
+- **Ask agent improvements** — Prompt caching (Anthropic API), parallel tool
+  execution, column headers in query results, iteration-aware budget warnings,
+  partial answers on budget exhaustion with resumable sessions.
+- **Issue reference extraction** — Commits now extract issue refs (`#123`,
+  `PROJ-456`, URLs) into `:commit/issue-refs`.
+- **Scoped re-analysis** — `--reanalyze` flag for `analyze` command supports
+  `all`, `prompt-changed`, `model-changed`, and `stale` modes.
+
+### Fixes
+
+- Pin Clojure 1.12.3 in deps.edn to fix uberjar compilation (tools.build
+  resolves 1.11.2 transitively, breaking AOT with neanderthal).
+- Python relative import handling in `enrich`.
+- Rust `use` statement capture alongside `mod` declarations.
+- Analysis prompt improvements: tighter confidence scoring, segment naming,
+  call-names accuracy, safety concern calibration.
+- Ask agent prompt hardening against hedge paragraphs and answer quality issues.
+- Numerous introspect robustness fixes (path traversal blocking, TOCTOU
+  mitigation, exception recovery, formatting-preserving revert).
+
+### CLI Commands
+
+Added `introspect` subcommand with `--max-iterations`, `--max-hours`,
+`--max-cost`, `--target`, `--eval-runs`, `--git-commit`, and `--verbose` flags.
+
+### Dependencies
+
+- Added `uncomplicate/deep-diamond` 0.43.0 and `uncomplicate/neanderthal` 0.61.0
+  (ML model training).
+- Pinned `org.clojure/clojure` 1.12.3.
+
+---
+
+## 0.1.0
 
 First public release of Noumenon, a Datomic-backed knowledge graph for
 AI-assisted codebase understanding.
@@ -42,10 +93,3 @@ AI-assisted codebase understanding.
 `digest`, `import`, `analyze`, `enrich`, `update`, `watch`, `query`, `ask`,
 `show-schema`, `status`, `list-databases`, `serve`, `benchmark`
 
-### Supported Languages
-
-| Tier | Languages |
-|------|-----------|
-| Full (deterministic imports) | Clojure |
-| Import extraction | Elixir, Erlang, Python, JavaScript/TypeScript, Rust, Java |
-| Analysis only | All languages with text files |

README.md

@@ -36,22 +36,22 @@ clj -M:run --help
 Download the latest JAR from [GitHub Releases](https://github.com/leifericf/noumenon/releases):
 
 ```bash
-java -jar noumenon-0.1.0.jar --help
+java -jar noumenon-0.2.0.jar --help
 ```
 
 Build from source if needed:
 
 ```bash
 clj -T:build uber
-java -jar target/noumenon-0.1.0.jar --version
+java -jar target/noumenon-0.2.0.jar --version
 ```
 
 ### Option 3: Use as a Clojure dependency
 
 ```clojure
 {:aliases
  {:noumenon
-  {:extra-deps {io.github.leifericf/noumenon {:git/tag "v0.1.0" :git/sha "284ee79"}}
+  {:extra-deps {io.github.leifericf/noumenon {:git/tag "v0.2.0" :git/sha "TBD"}}
    :main-opts ["-m" "noumenon.main"]}}}
 ```
 
@@ -130,9 +130,11 @@ flowchart LR
   F --> G
 
   C --> H[Benchmark\nEvaluation workflows]
+  H --> I[Introspect\nAutonomous self-improvement]
+  I -.->|optimizes| E
 ```
 
-`enrich` is optional but recommended for deterministic dependency and test-impact analysis. `update` replaces the manual `import` + `enrich` workflow.
+`enrich` is optional but recommended for deterministic dependency and test-impact analysis. `update` replaces the manual `import` + `enrich` workflow. `introspect` uses benchmark results to autonomously improve the ask agent's prompts, examples, rules, and code.
 
 ## Command Reference
 
@@ -141,7 +143,7 @@ clj -M:run <command> [options]
 clj -M:run <command> --help
 ```
 
-The CLI and [MCP](https://modelcontextprotocol.io) server expose the same capabilities.
+The CLI and [MCP](https://modelcontextprotocol.io) server expose the same capabilities. Use `--help` on any subcommand for details.
 
 | Command | CLI | MCP tool | Description |
 |---|---|---|---|
@@ -159,12 +161,15 @@ The CLI and [MCP](https://modelcontextprotocol.io) server expose the same capabi
 | Benchmark | `benchmark <path>` | `noumenon_benchmark_run` | Evaluate knowledge graph efficacy |
 | Benchmark results | -- | `noumenon_benchmark_results` | Get benchmark results (latest or by ID) |
 | Benchmark compare | -- | `noumenon_benchmark_compare` | Compare two benchmark runs by score differences |
+| Introspect | `introspect <path>` | `noumenon_introspect_start` | Autonomous self-improvement loop |
+| Introspect status | -- | `noumenon_introspect_status` | Check running introspect session |
+| Introspect history | -- | `noumenon_introspect_history` | Query introspect runs and iterations |
 | Watch | `watch <path>` | -- | Auto-sync on new commits (CLI-only) |
 | Serve | `serve` | -- | Start MCP server (CLI-only) |
 
 ## Named Queries
 
-48 named Datalog queries live in `resources/queries/` (EDN), covering hotspots, ownership, dependencies, complexity, churn, impact analysis, issue tracking, LLM cost tracking, and benchmarks. Run `clj -M:run query list` to see them all.
+56 named Datalog queries live in `resources/queries/` (EDN), covering hotspots, ownership, dependencies, complexity, churn, impact analysis, issue tracking, LLM cost tracking, benchmarks, and introspect history. Run `clj -M:run query list` to see them all.
 
 ## Data Model
 
@@ -253,7 +258,7 @@ Run Noumenon as an [MCP](https://modelcontextprotocol.io) server so agents can c
 
 ```bash
 clj -M:run serve
-# or java -jar noumenon-0.1.0.jar serve
+# or java -jar noumenon-0.2.0.jar serve
 ```
 
 ### [Claude Desktop](https://claude.ai/download) config
@@ -265,7 +270,7 @@ Add to `~/Library/Application Support/Claude/claude_desktop_config.json`:
   "mcpServers": {
     "noumenon": {
       "command": "java",
-      "args": ["-jar", "/path/to/noumenon-0.1.0.jar", "serve"]
+      "args": ["-jar", "/path/to/noumenon-0.2.0.jar", "serve"]
     }
   }
 }