dcondrey · dcondrey · Jul 27, 2026 · Jul 14, 2026 · Jul 14, 2026 · Jul 14, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -9,6 +9,71 @@ type.
 
 ## [Unreleased]
 
+## [0.5.0] - 2026-07-23
+
+A large hardening + feature run: all planned tiers, the critical/high/medium audit
+findings, and the async-job lifecycle over MCP.
+
+### Added
+
+- **Async job lifecycle over MCP.** `build_async`/`run_async` jobs now persist
+  across a server restart (a job still `running` when the process died reloads as
+  `interrupted`) and expose task-level progress — `done`/`total`/`phase`/`message` —
+  surfaced by `job_status`. Persistent by default (`~/.misterdev/jobs.json`,
+  `MISTERDEV_STATE_DIR` override).
+- **`full_rewrite` escalation rung** — a structurally different retry (rewrite the
+  whole target region) between context-widen and model-swap.
+- **Escalation sub-steps execute as real gated child tasks** (depth-guarded);
+  **proactive keystone splitting** of high-fan-in tasks; a **size/verifiability
+  invariant** flags over-large or unverifiable decomposed tasks.
+- **Runtime FailureLog read-back** (a task sees its own prior failures within the
+  run) and a **schedulable, benchmark-gated evolution pass**.
+- **compile_view per-language adapter registry** — real rust/typescript/go/swift/
+  csharp diagnostic parsers — and **named-symbol definitions surfaced** on
+  "cannot find X" / type-mismatch diagnostics.
+- **Documentation tool (context7/fetch) mounted by default** (isolation-gated,
+  opt-out via `mcp.docs_tool`).
+
+### Changed
+
+- **spec-as-tests is default-on** (advisory). The **mutation gate scores every
+  changed source file**, not just the largest. The **edit region is never
+  truncated** by the context budget. The **independent judge** detects a
+  same-model config as non-independent instead of silently claiming independence.
+
+### Fixed
+
+- **CRITICAL — the integration gate reverts a wave that breaks the build.** A
+  post-wave suite that no longer compiles/collects was waved through as unparseable
+  (a false-GREEN); it is now reverted while genuinely ambiguous failures are still
+  left alone.
+- A **zero-test gate** and a **MANIFEST acceptance pass-through with no prior
+  objective gate** are hard-rejected. **dotnet/VSTest counts sum across all
+  projects**; the **C# classifier anchors on the Roslyn `error CSxxxx:` prefix**.
+  The decomposer **prunes dependencies pointing at trimmed tasks**; model selection
+  **widens instead of silently defaulting** when the top tier is empty;
+  `RealTimeAligner` tolerates a malformed `consensus.json`; an unchanged file is no
+  longer treated as fully mutated; deferral task ids containing ` - ` are preserved.
+
+### Security
+
+- **Reject an untrusted MCP `runtimeHint`** outside the npx/uvx allowlist, closing
+  arbitrary-binary execution via a discovered server.
+
+### Performance
+
+- **SymbolGraph per-file index** (file-scoped queries are no longer
+  O(all-symbols)); the **Gatekeeper git-diff is memoized** within one gate run
+  (3 subprocess passes → 1).
+
+### Removed / robustness
+
+- **web-verify no longer leaks a dev-server** on a silent readiness wait;
+  **worktrees are torn down** if batch prep raises; **venv setup is
+  timeout-bounded**.
+- Orphaned dead modules `held_out_oracle` and `early_abort`, and the unused
+  `ModelSelector.is_ready`.
+
 ### Changed
 
 - **Integration gate: identity-based regression detection on a red baseline.** The

diff --git a/GAP_ANALYSIS.md b/GAP_ANALYSIS.md
@@ -0,0 +1,86 @@
+# GAP_ANALYSIS — project-completer (donor) → misterdev (canonical)
+
+Phase 1, read-only. Verdict up front, evidence below. No code changed.
+
+**Scope:** project-completer is a ~10k-LOC JavaScript MCP server (`/Volumes/A/autocoder/project-completer`) covering the same domain as misterdev (autonomously audit/fix/complete/build projects). misterdev is the maintained, published product (PyPI + MCP registry, `dcondrey`). Goal: reimplement the donor's *good ideas* as idiomatic Python, not merge the JS.
+
+---
+
+## TL;DR recommendation
+
+**Port #1 — Reference-guided build (`reference_dir` digest).** Highest value-per-unit-effort. misterdev already solves the hard part (multi-language symbol extraction via tree-sitter topography across 9 languages); the donor's `reference-analyzer` is a thin digest layer on top of capability misterdev *already has*. It's additive, offline-testable, self-contained, fits "one reviewable commit + tests," and is directly the workflow you're using right now (porting a reference implementation into a target). Effort **S–M**, risk **MED**.
+
+**Port #2 (strategic, not a quick port) — Async job lifecycle over MCP.** The donor's single most differentiated architecture: `start → run_id → status / pause / resume / stop`. misterdev's MCP `build`/`run` are **fully synchronous** — they block until the entire build finishes and return a string. For any non-trivial autonomous build over MCP, the client can't monitor, pause, or survive a timeout. Highest raw value, but effort **L** and risk **HIGH** (deep orchestrator rework, hard to test offline). Deserves its own dedicated effort, not folded into a quick port. **Critically: do NOT copy the donor's concurrency model — that's exactly where its 5 critical bugs live** (see "Bugs not to port"). misterdev's existing worktree/parallel scheduler is already race-safe; reuse it.
+
+Everything else the donor does, misterdev already does at equal-or-better quality. Details below.
+
+---
+
+## Inventories (verified against source)
+
+### project-completer — 18 MCP tools, 3 workflows
+
+| Workflow | Tools |
+|---|---|
+| **Completion** (audit→fix→verify) | `start_completion`, `get_completion_status`, `get_completion_findings`, `get_completion_report`, `pause_completion`, `resume_completion`, `stop_completion` |
+| **Evolution** (analyze→plan→approve→execute) | `analyze_project`, `create_evolution_plan`, `get_evolution_plan`, `approve_plan`, `execute_plan`, `get_execution_status` |
+| **Spec build** (decompose→scaffold→implement→converge) | `build_from_spec` (supports `reference_dir` porting), `get_spec_build_status`, `stop_spec_build` |
+
+Key `lib/*.js`: `orchestrator`, `evolution-orchestrator`, `spec-build-orchestrator`, `scheduler` (concurrency + file-lock + budget), `model-router` (complexity→model), `convergence` (diminishing-returns detector), `strategies` (per-stack audit prompts), `knowledge` (approve/reject learning), `library-intel` (ecosystem detect + `npm/pip audit`), `reference-analyzer` (extract interfaces/data-models/module-graph from a donor impl), `sandbox` (native/Docker), `write-queue` (serialize SQLite). Persistence: `better-sqlite3`.
+
+### misterdev — 5 MCP tools, 8 CLI subcommands, deep subsystems
+
+- **MCP tools:** `list_projects`, `status`, `scan`, `build` (plan+execute a goal, synchronous), `run` (execute planned tasks, synchronous).
+- **CLI:** `scan`, `list`, `status`, `report`, `run`, `build`, `plan`, `interactive` (+ `mcp`), plus natural-language routing.
+- **Subsystems:** planning (`decompose_spec`, `advisor.recommend_work`, `sovereign` strategies), verification (build/test/lint/typecheck/mutation/smoke/web/held-out-oracle gates), learning (`lesson_store`, `failure_taxonomy`, `warm_start`), evolution (`EvolutionLoop`, tool invention), economics (ledger-based UCB `model_selector`), context (tree-sitter topography for 9 languages), integration (`MCPManager`).
+
+---
+
+## Gap table
+
+Legend — **Have**: equivalent / partial / missing. **Value/Effort/Risk** apply only to *missing* or *partial* rows worth porting.
+
+| # | Donor capability | misterdev today | Have | Port value | Effort | Risk | Verdict |
+|---|---|---|---|---|---|---|---|
+| 1 | **Reference-guided build** (`reference-analyzer`: extract interfaces / data-models / module graph from a donor impl, feed the builder) | `decompose_spec` + tree-sitter topography (9 langs) exist, but **no reference-porting layer**; `build` takes a goal string only | **missing** (spec decomposition partial) | **HIGH** | **S–M** | MED | **PORT FIRST.** Topography already does the hard multi-language extraction; this is a digest + prompt-wiring layer. |
+| 2 | **Async job lifecycle over MCP** (`start`→`run_id`→`status`/`pause`/`resume`/`stop`) | MCP `build`/`run` are **synchronous, blocking, return a string** (verified: no `run_id`/job/thread/asyncio in `mcp_server.py`) | **missing** | **HIGH** | **L** | HIGH | **PORT SECOND, standalone.** Real defect for MCP use. Big lift; do NOT copy donor concurrency (buggy). |
+| 3 | **Human-in-loop plan approval over MCP** (`create_plan`→`approve_items`/`reject_items`→`execute`) | Interactive `plan` CLI + `advisor.recommend_work`; **no MCP approve/reject surface** | **partial** | MED | M | MED | Defer. Overlaps existing planning; mostly an MCP-surface exposure of what CLI already does. |
+| 4 | **Structured findings query** (`get_completion_findings` by severity/file/status) | Produces build *reports* + audit trail; **no queryable findings store** | **partial** | LOW–MED | M | LOW | Defer. Requires a findings store misterdev's model doesn't currently keep. |
+| 5 | **Ecosystem security audit** (`library-intel`: `npm audit` / `pip-audit` / `cargo audit`) | `detection.py:139-148` **already** resolves `cargo audit` / `npm audit --omit=dev` / `pip-audit` as an advisory gate | **equivalent** | — | — | — | **SKIP — would duplicate.** |
+| 6 | **Preference learning** (`knowledge`: approve/reject/success/failure) | `lesson_store`, `failure_taxonomy`, `warm_start`, metacognition | **equivalent** | — | — | — | Skip. |
+| 7 | **Model routing** (`model-router`: complexity→opus/sonnet/haiku) | `model_selector` (ledger-driven UCB, quality-per-dollar under a validation floor) | **equivalent (better)** | — | — | — | Skip — misterdev's is superior. |
+| 8 | **Convergence detector** (`convergence`) | `metacognition` / `advisor` diminishing-returns handling | **equivalent** | — | — | — | Skip (donor's has div-by-zero bugs anyway). |
+| 9 | **Concurrency scheduler + file lock** (`scheduler`) | Parallel isolated git worktrees + integration gate + auto-bisect revert | **equivalent (safer)** | — | — | — | Skip. |
+| 10 | **Sandbox** (native/Docker) | `container_env` + `evolution/sandbox` | **equivalent** | — | — | — | Skip. |
+| 11 | **Per-stack audit prompts** (`strategies`) | `guidance/*` per-language conventions + analyzer prompts | **equivalent** | — | — | — | Skip. |
+| 12 | **SQLite persistence + write-queue** | YAML/project registry + `.orchestrator/` progress (content-hash replay) | **equivalent (different model)** | — | — | — | Skip — no reason to adopt SQLite. |
+
+**Net:** of 12 donor capability clusters, 8 are already equal-or-better in misterdev (skip), 2 are partial/low-priority (defer), and **2 are genuine high-value gaps** (rows 1 and 2).
+
+---
+
+## Bugs not to port (from donor `todo.md`: 5 critical / 14 high)
+
+The donor's open criticals cluster almost entirely in its **async/concurrency machinery** — the exact subsystem behind Port #2. Reimplementing that surface on misterdev's existing race-safe worktree scheduler *avoids* these by construction; copying the JS structure would import them:
+
+- **C-006 / C-009** scheduler: file lock not atomic with CLI exec; budget check reads stale value under concurrent dispatch → overruns.
+- **C-007 / C-008** CLI wrapper: double-resolve on exit+error race; SIGKILL timer not cleared on normal exit (leaks event loop).
+- **C-010 / H-020** server: no guard against concurrent `start_completion` on same dir (DB/file corruption); `project_dir` unvalidated (path traversal).
+- **CLU-003 / CLU-004** systemic: concurrent-run corruption; resource exhaustion (DB conns / orchestrators / AbortControllers never released).
+- **H-021 / H-022** convergence: div-by-zero and off-by-one → premature convergence (another reason to skip row 8).
+- **SYS-001** silent error swallow across `git.js` / `fix.js` / `report.js` (catch-return-empty) — violates misterdev's "no silent swallow" baseline; do not reproduce.
+
+---
+
+## Recommended port order
+
+1. **Reference-guided build** — `reference_dir` digest built on existing topography; new optional param on `build` (+ optional CLI flag), offline-testable against a fixture reference tree. **← start here.**
+2. **Async MCP job lifecycle** — dedicated effort; new `run_id`-based tools alongside (not replacing) the synchronous ones; reuse misterdev's worktree scheduler; never copy the donor's concurrency code.
+3. *(defer)* MCP plan-approval surface (row 3).
+4. *(defer)* Findings query store (row 4).
+
+Rows 5–12: **do not port** (duplicate existing, equal-or-better misterdev capability).
+
+---
+
+*Phase 2 will implement exactly one capability — your pick — as a single reviewable commit with tests, `ruff` + `pytest` green, no heavy deps, public surface intact.*
diff --git a/misterdev/__init__.py b/misterdev/__init__.py
@@ -1,3 +1,3 @@
 """misterdev — an autonomous, extensible LLM build orchestrator."""
 
-__version__ = "0.4.1"
+__version__ = "0.5.0"