All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Develop-line release focused on broader AI maintainer workflows: codebase graph analysis, reversible context compression, structured mail tools, CAD read/edit tools, and long-running agent-output evaluation.

• anvil-defs codebase graph (Doc 57) — expands the SQLite defs index into call-graph analysis for Elisp, Python, JavaScript, and TypeScript. Adds trace-path, change blast-radius, caller/callee ranking, dead-code hints, file-level clustering, architecture summaries, and default exclusions for .worktrees/ and target/.
• anvil-semantic (Doc 18) — lands the local-first semantic search module on develop/master. Search combines SQLite FTS5, CJK LIKE fallback, lexical overlap, and optional vector cosine retrieval. The embedding provider defaults to ollama-local, with Gemini/OpenAI opt-in.
• anvil-context — Headroom-inspired reversible context compression for arbitrary JSON, diffs, logs, code excerpts, and prose. Adds context-compress, context-retrieve, and context-stats; raw context is retained by ccr-id through anvil-state.
• anvil-mu4e and anvil-wl (Docs 53/56) — structured mail tools over local mu4e and Wanderlust/Maildir stores. Read/search/compose paths are structured and CJK-aware; send paths are gated.
• anvil-cad — ASCII DXF/SVG read/edit/generate tools through a shared entity IR: outline, extract, annotate, batch-update, and generate.
• Fusion modules (Docs 54/55) — panel, async, long-run store, and MCP surfaces for comparing agent outputs and preserving compact review summaries.
• Client documentation — README now documents both Claude Code and Codex CLI MCP registration, plus AGENTS.md / CLAUDE.md guidance for shell-run, shell-tee-get, context-compress, context-retrieve, and context-stats.

• Use #!/usr/bin/env bash in the stdio bridge for NixOS-friendly execution.
• Decode literal newline escapes in the mu4e compose body path.
• Improve test isolation around orchestrator task-table state.
• Skip org-habit assertions on older org-habit builds where the habit summary format differs.

First stable release. Two landmark capabilities ship together:

1. anvil.el now runs without an Emacs install — thanks to NeLisp's v1.0 standalone Rust runtime (anvil-runtime). bin/anvil mcp serve --no-emacs spawns zero Emacs processes; the Rust binary itself reads + evaluates the loaded anvil modules and serves MCP tools over stdio.
2. anvil-ide split — Emacs-only IDE features (treesit-based code navigation, the worker dashboard, Info/Help lookup) moved out of anvil.el into the dedicated anvil-ide.el repo. The remaining anvil.el is the AI-side workbench: pure stdio MCP server + tool registry + library helpers, no human IDE surface.

This is a breaking release (anvil-ide tools are no longer in this repo) and a major architecture milestone (anvil + NeLisp standalone proves the AI workbench can run on a substrate other than the Emacs binary). Hence the jump 0.4.x → 1.0.0.

• Doc 38 §3 — full split pipeline shipped 2026-04-26 → 04-27. Phase A (audit + classification of all 60+ modules) → Phase B (5-wave adapter migration: anvil-buffer / anvil-elisp / anvil-org-index / anvil-org / anvil-browser) → Phase C (rename anvil-treesit / anvil-worker-ui → anvil-ide-*, split info-lookup-symbol into anvil-ide-elisp.el) → Phase D (git filter-repo extract preserving 12-commit history) → Phase E (cleanup + soft (require 'anvil-ide-* nil 'noerror)) → Phase F (treesit backend abstraction so anvil-ts / anvil-js / anvil-py stay AI-callable on NeLisp substrate via subprocess fallback) → Phase G (subprocess backend impl: Python full ast / JavaScript acorn / TypeScript degraded acorn). All phases documented in docs/design/38-anvil-ide-split.org.
• Architecture α now active — anvil-defs, anvil-state, anvil-http internals delegate to NeLisp side via fboundp guard + fallback, letting the Rust runtime evaluate them without a host Emacs.

• Stage D v2.0 (bundled-Emacs path) and v3.0 (Rust-only path via bin/anvil mcp serve --no-emacs) ship as part of NeLisp v1.0; both expose the same MCP tools so existing Claude Code .mcp.json entries keep working.

• anvil-treesit.el, anvil-worker-ui.el, anvil-ide.el, the elisp-info-lookup-symbol MCP tool, and the org-element / org-edit-body tree-walk surface are no longer in this repo. Install anvil-ide.el separately if you want the IDE layer. Cross-package wires use (require 'anvil-ide-* nil 'noerror) so anvil.el continues to load even when anvil-ide is absent.

• CI subprocess test gating — 49 slow subprocess ERTs gated on ANVIL_SLOW_TESTS=1 so the default GitHub Actions matrix exits 0 on every push. Local make test-all still runs everything.
• Test count: 1198 (v0.4.0) → 1610+ (v1.0.0), with all of Phase F backend abstraction + Phase G subprocess paths under ERT.

• 3 audit-correction notes appended to Doc 38: anvil-buffer is namespace-disjoint, not drop-in compatible; the org-capture-templates reference in anvil-browser is docstring-only; anvil-ts / -js / -py wrap treesit-* and were reclassified PURE-via-backend in Phase F.

Two small fixes released same-day as v0.4.0, both reached the repo after the v0.4.0 tag was pushed.

• Release-audit false positive on anvil-discovery.el — four private accessors (--tool-intent / --tool-layer / --tool-stability / --tool-summary) triggered the MCP-tool wrapper regex and failed the release-audit gate on master, leaving the v0.4.0 CI build red. Renamed to -of suffixes (--intent-of / --layer-of / --stability-of / --summary-of) so the audit no longer treats them as MCP tools. No public API change; the actual anvil-tools-by-intent / anvil-tools- usage-report surfaces are unchanged.
• treesit_info parser initialisation (@yours57) — the tool ran through the TreeSitter node walker without creating a parser first, so callers got a stale / empty info payload on buffers where treesit hadn't been warmed. Added treesit-parser-create
  • integer-param parse helper + autoload for anvil-treesit-language-for-file / -ensure-grammar.

AI agent workbench release. Fourteen new design documents (Doc 21-34) worth of primitives turn anvil from a tool-per-task module collection into a cohesive agent development platform: memory engine, intent-based tool discovery + profile filter, structural edits for Python / TS / JS / Elisp-CST, shell output compression, session snapshot + Claude Code lifecycle hooks.

Test suite: 640 → 1198 passing (+558, +87%). 20 new anvil-*.el modules. +39,343 / -408 lines across 145 files.

• memory module (Doc 29) — Bayesian + TTL + FTS5 auto-memory engine with 16 MCP tools (memory-scan / memory-audit / memory-access / memory-list / memory-search / memory-duplicates / memory-promote / memory-serve-start etc.). Indexes ~/.claude/projects/*/memory/*.md into SQLite, tracks contradictions and URL liveness, surfaces stale rows to memory-pruner. +115 ERT.
• session module (Doc 17) — session snapshot / resume plus 5 Claude Code lifecycle hooks (PreCompact / SessionStart / PostToolUse / UserPromptSubmit / SessionEnd). session-snapshot captures branch
  • task summary into anvil-state ns=session (14-day TTL), returns a preamble-suggested resume block; anvil-hook install wires the hook set into ~/.claude/settings.json. +19 ERT.
• shell-filter module (Doc 27) — shell output compression with 20 bundled filters (git status / git log / git diff / rg / find / pytest / ert-batch / emacs-batch / make / docker-logs / …) + tee + gain statistics. Raw output stashed under shell-tee namespace with TTL so callers can recover on demand. Depends on state.
• disclosure module (Doc 28) — 3-layer read contract and citation URI scheme. file-outline / org-index-index / defs-index / journal-index / http-cache-index (Layer 1), file-read-snippet (Layer 2), file-read / org-read-headline / org-read-by-id / elisp-get-function-definition / http-cache-get (Layer 3). disclosure-help tool prints the contract; anvil-uri-fetch is a cross-layer resolver.
• discovery module (Doc 34) — intent-based MCP tool discovery (Phase A) + agent / edit intent-based profiles (Phase B) + orchestrator auto-injection + per-tool usage counter + release-audit :unused-since scanner (Phase C). anvil-tools-by-intent returns tools matching intent CSV / layer / query regex, sorted by layer rank and intent overlap. All 198 registered tools tagged with :intent / :layer / :stability metadata. +27 ERT across discovery / manifest / orchestrator / dev.

• manifest module (Doc 26) — per-session tools/list profile filter driven by ANVIL_PROFILE. Five legacy profiles (ultra / nav / core / lean / full) advertise curated tool subsets to shrink the manifest token cost; handlers of hidden tools stay callable via explicit tools/call. Phase 1b auto-injects --mcp-config into orchestrator child sessions.
• Intent-based profiles (Doc 34 Phase B) — agent (orchestrator / session / memory / browser + edit tools, layer=core+workflow) and edit (file / org / code / json / db only, layer=core). Filter by metadata instead of hand-curated ID lists.

• sexp module (Doc 12 Phase 1+2) — reader-based edits for Elisp: sexp-read-file / sexp-surrounding-form / sexp-rename-symbol / sexp-replace-call / sexp-replace-defun / sexp-wrap-form / sexp-macroexpand / sexp-verify. +48 ERT.
• sexp-cst module (Doc 31) — tree-sitter-elisp CST + runtime inspect-object tool for any live Lisp value. sexp-cst-read (comment-preserving CST), sexp-cst-edit + -write (point-offset replacement with re-parse validation), sexp-cst-repair (paren + unterminated-string balancing). +47 ERT.
• py module (Doc 21 Phase 1) — Python structural locators and edits via treesit: py-list-imports / py-list-functions / py-list-classes / py-list-methods / py-list-decorators / py-find-definition / py-surrounding-form plus edit tools py-add-import / py-remove-import / py-rename-import / py-replace-function / py-wrap-expr. +55 ERT.
• ts + js modules (Doc 21 Phase 2) — TS/TSX + JS/JSX locators: ts-list-imports/exports/functions/classes/methods/interfaces/ type-aliases/find-definition/surrounding-form and the js-* mirror. +25 ERT (ts) + ts-test fixtures.
• defs module (Doc 11) — SQLite-backed Elisp symbol index. defs-search / defs-references / defs-signature / defs-who-requires / defs-index-rebuild / defs-index-status. +52 ERT.

• bench module (Doc 14) — bench-compare / bench-profile-expr / bench-last.
• bisect module (Doc 13) — test-driven git bisect via worktree- isolated emacs --batch. bisect-test pins a failing ERT to the introducing commit. +12 ERT.
• git-msg module (Doc 15) — git-commit-message (from staged diff) and git-pr-body (from branch log). +17 ERT.
• lint module (Doc 16) — repo hygiene scanner with pluggable registry: conflict-markers (error), orphan-ids (info), broken-scheduled (warning). +13 ERT.
• data module (Doc 33) — JSON path-based edits with preview-by- default: data-get-path / data-set-path / data-delete-path / data-list-keys for ~/.claude.json, package.json, MCP configs. +28 ERT.

• Provider latency routing (Doc 22) — orchestrator-routing-select for per-provider latency-aware dispatch. +13 ERT.
• Consensus presets (Doc 23) — named provider combinations ship with orchestrator-consensus-* family. +16 ERT.
• Orchestrator submit-and-collect — one-shot dispatch + wait combinator for programmatic callers.
• Preamble management — orchestrator-preamble-set/-get/-list/ -delete/-set-from-file for reusable system prompts.
• Live streaming — orchestrator-stream + orchestrator-tail forward provider stdout as MCP events while the task runs.
• Cross-session stats — orchestrator-stats aggregates batch history across daemon restarts via anvil-state.

• anvil-tools-by-intent — intent CSV / layer / query regex / stability filter, deprecated always hidden, experimental opt-in, layer-ranked output.
• anvil-tools-usage-report — per-tool counter summary (days threshold, never-called bucket, unused-since bucket, recency sort).
• anvil-dev-release-audit gains :unused-since N scanner — advisory-only Phase C hazard: does not flip :clean-p.

• anvil-server-encode-handler internals — replaced eval + constructed lambda with make-symbol + fset + apply-partially + symbol-property (anvil-server-raw-handler / anvil-server-encode-result). anvil-server-register-tool normalizes wrappers back to the raw handler for schema generation and docstring validation. Fixes silent enable skip for sexp / py / bench / git-msg optional modules (PR #12, @yours57).
• git MCP handlers — sentinels :null / :empty-array → literal strings "null" / "[]" (visible wire-format change). All 8 git tools now wrapped via anvil-server-encode-handler at registration; 6 dedicated ERT.
• Worker server files — path moved from user-emacs-directory/server/ to server-auth-dir / server-socket-dir; liveness check uses server-running-p for local sockets, PID parsing for TCP auth files.
• anvil-server-tool-filter-function — now receives (TOOL-ID TOOL-PLIST SERVER-ID) so filters can branch on metadata. anvil-manifest uses the full signature for intent-based filtering.
• anvil-server-tool-dispatch-hook (new abnormal hook) — runs after each successful handler return with (TOOL-ID SERVER-ID); used by anvil-discovery to maintain usage counters without coupling anvil-server to anvil-state.

• Encoded handler registration bug — anvil-server-encode-handler previously returned (lambda (&rest args) ...) which anvil-server--generate-schema-from-function rejected; silently skipped enable for sexp / py / bench / git-msg and made every re-register via unload-feature break. Fixed via the symbol-backed wrapper path above (PR #12).
• JSON encoder on dotted pairs — anvil-server--to-json-value crashed with listp error on alist entries like (cons "k" 1.0). New anvil-server--list-to-json-array emits [car, cdr] for improper lists so mapcar can't trip.
• xref_find_apropos — missing (require 'apropos) made the tool void-function on fresh Emacs installs.
• Worker liveness — Unix-socket workers wrongly reported alive via file existence alone; now goes through server-running-p.

• Fourteen new design docs (Doc 21-34). Six docs (18 / 19 / 20 / 24 / 25 / 30) explicitly DEFERRED — out of scope for this release. Doc 32 is an audit memo (rhblind/emacs-mcp-server positioning, informational only).
• CLAUDE.md selection flowchart trimmed (11 rows → 5) — discoverable entries moved to anvil-tools-by-intent runtime query; only size/shape heuristics kept.
• Platform files (linux.md / windows.md) gained a tool 探索 section documenting ANVIL_PROFILE=agent / edit usage.

Hotfix surfaced by the v0.3.0 orchestrator benchmark (develop: benchmarks/results/report-2026-04-19.org). The bench's first programmatic orchestrator-* call hit the cons-return bug; every fix here fell out of the investigation.

• orchestrator MCP wrappers returned plists instead of strings, tripping the anvil-server contract and breaking every programmatic orchestrator-* call from Claude Code / OpenCode / other MCP clients with "Tool handler must return string or nil, got: cons". Added anvil-orchestrator--encode-handler which wraps each registered tool at registration time and JSON-encodes the plist result. Tool bodies stay unchanged (and still return rich plists for direct Elisp / ERT callers).
• Added anvil-orchestrator--batch-task-ids accessor + docstring note on the anvil-orchestrator--batches / anvil-orchestrator--consensus-groups shape asymmetry, so future callers stay decoupled from the storage representation.

• Raised the codex per-provider concurrency default from 3 → 6. The v0.3.0 ramp benchmark reached 8 concurrent codex jobs on ChatGPT Plus OAuth with zero 429s; the old default was defensive past its evidence. Users on throttled accounts can set it back via anvil-orchestrator-per-provider-concurrency.

Major release covering 10 design documents worth of new primitives: orchestrator (Doc 10), browser framework (Doc 07), disk-first helpers (Doc 05), code-transform tools (Doc 06), state KV store (Doc 08), HTTP client (Doc 09), pty-broker (Doc 04), worker pool v2 (Doc 01), org index (Doc 02), offload framework (Doc 03). Test suite grew from 292 to 508 passing.

• orchestrator module (Doc 10) — parallel AI CLI dispatcher.
  • 5 native providers: claude, aider, gemini, ollama, codex (OpenAI ChatGPT Plus OAuth).
  • Pool with per-provider concurrency caps, state-persisted queue, tabulated-list dashboard, git worktree isolation.
  • :depends-on DAG with cycle / unknown-name detection and propagating failure semantics. DAG resume across daemon restart via orchestrator-resume-interrupted.
  • Cross-model consensus (Jaccard verdict) and meta-LLM judge (synthesized answer via 3rd provider).
  • Auto-retry with exponential backoff + jitter, Retry-After honouring, per-provider backoff overrides.
  • Live streaming events (Phase 7c) — per-event callbacks without SQLite writes, overflow compression.
  • Context preamble registry (Phase 7b) — reusable prompt preludes via :preamble-ref, auto-resolved at submit.
  • submit-and-collect convenience for synchronous fan-out, submit-one / extract-result / tail glue helpers.
  • Observability: orchestrator-stats with percentile / provider filter, anvil-orchestrator-stats-dashboard command.
  • anvil-cron integration (Phase 2b) via anvil-cron-register :fn
    • examples/nightly-orchestrator.org.
  • gemma4 thinking-mode guard (XML <think> and plain-text delimiter).
• browser module (Doc 07) — agent-browser wrapper.
  • Phase A: browser-fetch / -interact / -capture / -screenshot / -close MCP tools with in-memory cache + metrics + M-x anvil-browser-status.
  • Phase A': auth primitives (profile, user-agent, auto-connect, session-presets) for login-walled / anti-bot sites.
• pty-broker module (Doc 04 Phase 1) — Node-pty TCP broker + Elisp client + 5 MCP tools, structurally avoids filter-starvation and ConPTY headless-stdin limitations.
• state module (Doc 08) — SQLite-backed KV store with namespaces, TTL, and prin1 value serialization. anvil-browser cache now persists across daemon restarts.
• http module (Doc 09 Phase 1a) — http-fetch / -head / -cache-purge with ETag/TTL cache via anvil-state.
• org-index (Doc 02 Phase 1-4 + 5a) — SQLite org index, filenotify watcher + periodic scan, 177× speed-up on 1500 files / 71k headlines; org-read-* routed through index.
• Disk-first helpers (Doc 05) — :warnings on every mutating file-* tool, plus buffer-read / buffer-save / buffer-list-modified MCP tools.
• Code-transform tools (Doc 06) — code-extract-pattern (read-only structured block extraction) and code-add-field-by-map (TS/JS object literal bulk field add for i18n).
• Worker pool v2 (Doc 01) — 3-lane split with :kind classifier, batch warmup, latency metrics, optional server-eval-at transport.
• Offload framework (Doc 03) — pipe REPL + pool, :offload t, hard-kill, handler-side checkpoint for :value / :cursor persistence.
• anvil-dev tools — anvil-dev-release-audit (3 scanners: Emacs 30 _args arglist-strip hazard, missing MCP Parameters: docstring sections, non-SHIPPED design docs), anvil-dev-test-run-all :minimal, anvil-dev-self-sync-check, anvil-dev-journal-memo-append, anvil-scaffold-module.
• Efficiency bundle — anvil-server-register-tools bulk registration, anvil-orchestrator-preamble-set-from-file, anvil-orchestrator-dashboard-autofollow, anvil-dev-release-audit :scope.
• anvil-git extensions — worktree helpers + git-branch-current / git-head-sha / git-repo-root / git-worktree-list MCP tools.

• anvil-orchestrator-summary-max-chars default raised 300 → 4000 to eliminate the common re-parse round-trip; pass :full t to orchestrator-extract-result for the unbounded form.
• No-arg MCP wrapper convention: write () not (_args). Emacs 30 help-function-arglist strips _-prefix, which broke MCP schema registration for worker, cron, orchestrator, browser, and other modules — all audited and fixed.

• orchestrator: close stdin after spawn to prevent ollama CLI hang (31d2f3f).
• orchestrator: deep-copy task plist in submit so consensus fan-out does not share tails — previously caused :cost-usd / :elapsed-ms clobber across providers (5a2610f).
• orchestrator: gemma4 thinking-strip now handles the plain-text Thinking... ...done thinking. delimiter in addition to XML <think> blocks (86a0e7c).
• Windows stdio CR robustness — tools/list decode reliably on MSYS2 + gawk 5.0.0 + coreutils 8.32 (issue #6).
• Emacs 30 help-function-arglist underscore-strip regressions across worker, cron, orchestrator (issue #9, v0.2.1) + remaining no-arg wrappers audited via anvil-release-audit.

• Worker + cron module load failure on Emacs 30 due to help-function-arglist stripping _args prefix (issue #9).

Initial tagged release — see git history for details.

Project inception.