Midas 1.0.0 — control-plane, trust, and adoption power-up - #17
Merged
Merged
Conversation
…doctor --json - `midas init` now wires VS Code (user mcp.json, `servers` key), Gemini CLI, Cline, and Zed (`context_servers`), each with its own config schema; status/uninstall understand the per-client server-map keys too. - `midas serve --http` accepts MIDAS_MCP_TOKEN / --token: a constant-time bearer-token ASGI gate, since the HTTP transport otherwise exposes the whole store to any local process. - CI runs the core+MCP suite on Windows and macOS (the client-wiring code is full of per-OS paths that only Linux exercised). - `midas doctor --json` emits the same machine-readable envelope as the wiring receipt. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
…ention
- midas/continuity.py — three primitives similarity search can't express:
* memory_conflicts: live beliefs that contradict each other with neither
superseding the other (the multi-agent shared-memory failure mode). NLI-
scored when available; else a same-slot heuristic (value swap on a shared
frame, numbers disagree, one side negates). Ranked candidates only —
nothing is resolved silently.
* open loops: kind="commitment" records a promise, close_loop supersedes it
with its resolution, open_loops lists what's still unclosed (oldest first).
* resume: the one-call session-onboarding pack — pinned directives,
forbidden rules, what changed, current state, open loops, conflicts —
token-budgeted and prompt-ready.
- Memory.forget_expired(ttl_by_kind) + MIDAS_MCP_TTL / maintain(ttl=...):
age-based per-kind retention; user-confirmed, standing, and supersession-
chain records never expire silently.
- MCP tools: resume, memory_conflicts, open_loops, remember_commitment,
close_loop; agent instructions now start sessions with resume and keep
promises via open loops.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
Every SQLite mutation (put / delete / clear — so remember, supersede, and forget) appends an entry to an append-only audit_log table: seq, timestamp, op, record id, a sha256 of the record's audited state, and a hash chained over the previous entry. Editing, removing, or reordering ANY past entry breaks every hash after it. Entries carry hashes only — never memory content, preserving the receipt/audit boundary. - SQLiteStore.audit_log() / verify_audit_log(); audit=False opt-out for perf-sensitive paths. Additive table: older Midas versions still open the file. - `midas audit [--limit N] [--json]` shows the tail and verifies the chain (exit 1 + first_invalid_seq when broken). - `midas doctor` now checks chain integrity. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
…ntic embeddings - midas/importers.py + `midas import --from …`: file-based agent memory (CLAUDE.md, .cursorrules, exported JSONL) becomes first-class Midas records — bullets/paragraphs parsed per item with heading context, fenced code skipped, idempotent re-runs, and governance-safe defaults (observation provenance; imported rules can't authorize guarded actions unless --confirmed). - TypeScript port: LocalEmbedder — local ONNX semantic embeddings (bge-small, 384d) via the optional @huggingface/transformers behind MIDAS_MCP_EMBEDDER=local, with an announced fallback to the byte-parity hashing embedder when the package is absent. The TS Memory API is now async (remember/capture/recall/buildContext/forgetMatching), verified end-to-end (paraphrase recall 0.731 vs 0.491 noise). 14/14 TS tests pass. - CHANGELOG for the whole batch; README updates. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
…tamper test The new macOS/Windows jobs caught exactly what they exist for — in the tests: - VS Code/Zed/Claude Desktop resolve OS-specific roots (macOS Library/, Windows %APPDATA%) that ignore a patched Path.home(), so the client-wiring tests looked in the wrong place off-Linux (and could have touched a runner's real configs). A _sandbox_client_paths helper now pins every lookup into tmp_path on all three OSes. - test_tampering_breaks_the_chain rewrote the sha's first hex char to 'f' — a 1/16 no-op (hit on the Windows runner). Tamper the op field instead, which is always a real change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
…licts, loops, audit chain - Agent Continuity Bench grows three axes, all deterministic and $0: resume_fidelity (the one-call pack contains live state / forbidden rules / open loops and never presents superseded values or closed loops as live), conflict_detection (every planted live-live contradiction found), and conflict_precision (no benign look-alike flagged). All → 1.00; wired into `midas bench` / eval.benches with the same all-green verdict rule. - Inspector: three new glass-box views + pure API endpoints — Conflicts (ranked pairs with per-side forget), Open loops (close with a recorded resolution), Audit log (chain verification + hash-only tail). Conflict similarity is cast to float at the source so every JSON consumer works. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
…ind TTL - packages/midas-ts/src/continuity.ts — resume (one-call onboarding pack), memoryConflicts (heuristic tier: value swap / numeric drift / negation), and open loops (rememberCommitment / openLoops / closeLoop), exposed as MCP tools and SDK exports; "commitment" added to the TS MemoryKind. - SQLiteStore writes the same hash-chained audit_log as Python. The hash formula is canonicalised to integer microseconds on BOTH sides (float FORMATTING differs across runtimes; IEEE multiply+floor doesn't), verified bidirectionally: Python validates TS-written chains and vice versa. - Memory.forgetExpired + parseTtlSpec + maintain(ttl=…) / MIDAS_MCP_TTL. - 20/20 TS tests (6 new: conflicts, loop lifecycle, resume budget, chain tamper detection, audit opt-out). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
…r at rest; release 0.2.0 - `midas init --claude-hook` installs a Claude Code SessionEnd hook (`midas hook capture-session`): each finished session's user turns are OFFERED to capture — the no-LLM policy still decides what's kept. Fail-open (a memory hook must never break a session), idempotent install/uninstall into ~/.claude/settings.json with backup; removed by `midas uninstall`. - `midas import --from mem0|zep`: Mem0 distilled memories → facts (original ids/user/agent attribution preserved in metadata); Zep facts/edges → facts, messages → chat turns. Tolerant of the wrapped and bare-list export shapes. - Encryption at rest (opt-in): SQLiteStore(key=…) / MIDAS_MCP_KEY via the new [encrypted] extra (SQLCipher). Ciphertext on disk (verified: no plaintext, no SQLite header), wrong/no key can't read, audit chain works under encryption, and a set key without the extra FAILS CLOSED. - Release prep 0.2.0: version bump (pyproject, __init__, npm package + MCP handshake), dated CHANGELOG covering the whole batch; Linux CI now installs [encrypted] so those tests run (they skip where wheels are absent). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
vornicx
changed the title
…ystem
The inspector had one problem underneath its polish: every kind, every
provenance, every signal rendered in the same monochrome gold accent, so a
glass-box tool for governed memory couldn't be visually parsed at a glance.
And it was dark-only, had no real charts, used native confirm()/prompt(),
and its hash router never listened for hashchange (back/forward and direct
links silently did nothing).
- Fixed categorical color system: kind and provenance each get a
consistent, validated color (colorblind-safe contrast/separation checked
in both themes) used identically across every view — Overview bars,
Browse tags, Project governance, Conflicts. Status color (good/warning/
critical) reserved separately for verdicts/chain-integrity/conflicts, so
a revision reads as "normal lifecycle" (amber) not "error" (red).
- A real second theme: light, deliberately designed (not an inverted
dark), toggle + system-preference default + persisted choice.
- Real charts: a 30-day activity line/area chart with a hover crosshair +
tooltip, and a recency (short/medium/long) stacked bar — backed by two
new pure API functions, api_timeseries and api_meta (plus by_tier added
to api_overview).
- Native confirm()/prompt() replaced with in-app modal + toast components;
grouped nav (Memory/Coding/Governance) with live badge counts; a command
palette (Cmd/Ctrl-K) and "/" search shortcut; responsive down to phone
width with a horizontal pill nav.
- Fixed the hashchange bug (the router only ran once, at load) and two
real API-shape bugs the fixes surfaced under interaction testing:
api_conflicts/api_loops return bare lists (the code assumed a
{count,...} envelope) and the forget receipt's key is content_sha256
(the code read content_sha).
Verified with Playwright end-to-end (not just screenshots): forget with
modal confirm, close-loop with a resolution textarea, theme toggle,
command-palette keyboard nav, "/" focus, and browser back/forward through
the hash router — all against a seeded demo store, in both themes and a
mobile viewport, zero console/page errors. 336 tests pass, ruff clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
main jumped straight to a 1.0.0 semver-freeze release while this branch was open, so the branch (cut from the pre-1.0 base) conflicted. Reconcile by merging main in — code auto-merged cleanly; the only conflicts were the version/changelog files — and renumbering the batch from the pre-release 0.2.0 to 1.1.0, the correct minor on top of 1.0.0 (all additive, no breaks). - pyproject / __init__ / midas-ts package.json + lock / mcp.ts handshake / mcpb manifest / server.json: all 1.1.0 (the `forwarded` dep's own 0.2.0 in the lockfile is left untouched). - CHANGELOG: the batch + inspector redesign become [1.1.0] — 2026-07-06, above main's [1.0.0]; nothing dropped or duplicated. Merged tree: 336 passed, 5 skipped; ruff clean; midas.__version__ == 1.1.0. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
vornicx
changed the title
…ease) 1.0.0 hasn't been published yet, so instead of shipping this batch as a separate 1.1.0 on top, fold it into the one comprehensive 1.0.0 release. - All own-version references back to 1.0.0 (pyproject, __init__, midas-ts package.json + lock own-version lines, mcp.ts handshake, mcpb manifest, server.json); the lockfile's `forwarded` dependency version is untouched. - CHANGELOG: [1.1.0] folded into a single [1.0.0] — 2026-07-06; the contract paragraph reworded (this release now genuinely adds the control-plane, audit chain, and tooling, not just consolidates), Added/Changed/Security merged with nothing dropped or duplicated. Merged tree: 336 passed, 5 skipped; ruff clean; midas.__version__ == 1.0.0. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4
vornicx
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The full power-up batch, in the order agreed: quick wins → the differentiating control-plane → trust → adoption — plus the follow-up backlog (bench the new features, inspector views, TS parity, Claude Code hook, Mem0/Zep import, encryption at rest, the inspector redesign).
Highlights (each phase self-contained)
midas initclients (VS Codeserverskey, Gemini CLI, Cline, Zedcontext_servers);midas serve --http --token/MIDAS_MCP_TOKENbearer auth (constant-time; the HTTP transport was previously open to any local process); CI on Windows + macOS;midas doctor --json.resume(one-call session-onboarding pack, token-budgeted),memory_conflicts(live contradictions with neither side superseded — NLI when available, same-slot heuristic otherwise; ranked candidates, never auto-resolved), open loops (remember_commitment/open_loops/close_loop), per-kind TTL retention (MIDAS_MCP_TTL). Agent policy now starts sessions withresume.audit_login SQLiteStore (hashes only, never content);midas audit [--json]; doctor checks chain integrity.midas import --from claude-md|cursorrules|jsonl|mem0|zep(governance-safe provenance defaults); TS port gains local ONNX semantic embeddings (LocalEmbedder, optional@huggingface/transformers, announced fallback; asyncMemoryAPI).resume_fidelity,conflict_detection,conflict_precision(all → 1.00, in themidas benchverdict); inspector gains Conflicts / Open loops / Audit log views.midas init --claude-hook(SessionEnd auto-capture, fail-open, policy-gated); encryption at rest via SQLCipher (MIDAS_MCP_KEY+[encrypted]extra, fails closed; ciphertext verified on disk).midas inspectredesigned end to end — a real light theme, a fixed/validated categorical color system consistent across every view, real charts (30-day activity + recency), a command palette (⌘K), in-app modals/toasts, responsive to phone width; fixed a realhashchangerouting bug in the process. Verified with Playwright across views/themes/mobile.Tests
midas bench: all green ✓ including the three new continuity axes.Publishing note
Versions are 1.0.0 (PyPI
midas-memory, npmmidas-memory-mcp, server.json / mcpb manifest) but publishing needs maintainer credentials — after merge:python -m build && twine uploadandnpm publishfrompackages/midas-ts.🤖 Generated with Claude Code
https://claude.ai/code/session_01Deray2qcPRy1hZVQnboHp4