v0.7.0

Claude Desktop test follow-up (2026-10-21)

• Declare the process_thought stage parameter as a string listing the five
  stages, so strict clients accept the documented case-insensitive spelling
  ("research"); unknown stages still return INVALID_INPUT.
• Stage-jump warnings in process_thought and generate_summary.stage_transitions
  compare with the previous mainline or revision note; a stage reached in a
  revision is no longer reported as skipped.
• Add deprecation_notice to generate_summary, clear_history and legacy-session
  export_session/import_session results (null for explicit sessions).
• Name the involved fields in INVALID_INPUT hints for cross-field validation
  errors (e.g. Fields: branch_from_thought, branch_id).
• read_session steps view: a first note larger than max_chars degrades to its
  first content chunk (content_truncated, next_content_offset) instead of
  returning RESPONSE_LIMIT.
• Kept by decision: truncated stays one flag (next_cursor signals more pages),
  cursors stay opaque per view, and a create_session retry returns the original
  snapshot (use read_session for current state).

Cleanup (2026-10-21)

• Remove run_server.py; use mcp-sequential-thinking or
  python -m mcp_sequential_thinking.server.
• Remove debug_mcp_connection.py; scripts/stdio_smoke.py covers the stdio check
  with the real MCP client.
• Remove the unused utils.to_camel_case and ThoughtStage.from_string;
  ThoughtStage(value) accepts the same case-insensitive spelling.
• Drop the unused vis, web and all extras (matplotlib, numpy, fastapi, uvicorn).
• ThoughtData serializes through Pydantic camelCase aliases; the stored format and
  key order are unchanged, unknown keyword arguments are now rejected.
• Rely on the MCP SDK's UTF-8 stdio wrapping; drop the duplicate re-wrap and the
  script-mode import fallback in server.py.

Second manual MCP audit follow-up (2026-09-08)

• Return PATH_OUTSIDE_EXPORTS for rejected import/export paths with a useful,
  sanitized correction hint.
• Include migration guidance in process_thought.deprecation_notice, and explain
  the basis of mainline progress and the database-wide sequence cursor in results.
• Add MCP regression coverage for the quoted German lexical overlap (score 0.6),
  tag matching, and atomic rejection of nonexistent/cross-session completion evidence.
  Evidence validation already existed; the reported bypass was not reproduced.
  Empty related matches for the full manual inputs remain unverified without those inputs.

Manual MCP audit follow-up (2026-09-08)

• Preserve decimal/version numbers in summary excerpts.
• Add explicit session_id export/import with a validated, atomic single-session
  archive preserving branches, IDs, sources and completion. Existing sessions
  conflict on import; legacy v1/v2 behavior stays available.
• Distinguish SESSION_FINALIZED from retryable version conflicts and expose
  current_version through sanitized MCP errors.
• Qualify legacy analysis references and revision chains by branch.
• Paginate the prioritized resume view without repeating page zero; document its
  ordering and omit completion from later pages unless explicitly requested.
• Deprecate the legacy workflow without removing it; emphasize working notes,
  auditability and resumption rather than improved reasoning.

Reliability upgrade (2026-09-05)

• Serialize JSONL recovery and mutations; repair incomplete UTF-8/JSON tails with
  exact backups; reject complete corruption, future schemas and identity conflicts.
• Add transactional SQLite sessions, ID references, request-ID retries, versioned
  finalization, bounded resume/full reads and explicit typed completion records.
• Migrate legacy UUIDs and contents without deduplication; keep the five tools on
  the explicit legacy session, with guarded old-writer exclusion and snapshot restore.
• Add memory-only operation, controlled diagnostics, schema limits and sanitized
  correctable tool errors. Legacy stage spelling remains case insensitive.
• Extend process/crash/fault/stdio tests and complete Python/Windows/SDK gates;
  test installed artifacts outside the checkout and publish the same tested files.
• Prepare a bilingual evaluation collection and registry metadata. A completed
  model usefulness comparison is still required; no general quality claim is made.

Compatibility changes: corrupt storage fails closed rather than starting empty;
correctable legacy errors are now isError tool results; inputs/outputs have explicit
limits; source imports no longer initialize storage. SQLite snapshots are required
for complete restore after new writes; legacy JSONL alone cannot downgrade all
session semantics. See the README for details.

Earlier 0.7 development work (2026-08-24; not a published release)

Breaking

• mcp dependency moves to the 2.x line: mcp>=2,<3 (was
  >=1.2.0,<2.0.0). mcp.server.fastmcp.FastMCP is gone upstream; the
  server is rewritten against mcp.server.mcpserver.MCPServer. If you
  vendor or subclass anything from this package's server module, the SDK
  types you get back changed from camelCase to snake_case (isError →
  is_error, structuredContent → structured_content, etc.) — this
  package's own request/response shapes below are affected the same way.
• Every tool now returns structured, typed output (a Pydantic model,
  surfaced as structured_content) instead of an untyped dict serialized
  to JSON text. Field names are snake_case throughout, replacing the old
  camelCase dict shape end to end — not just the fields called out below.
  tools/list now advertises an output_schema for all five tools.
• process_thought analysis output: the old single progress float
  (which froze/misreported once revisions or branches entered the
  picture) is replaced by four explicit fields: main_line_progress,
  main_line_position, total_thoughts_recorded, branch_count,
  revision_count. relatedThoughtsCount/relatedThoughtSummaries —
  which matched on stage equality or tag overlap and called that
  "related" — are replaced by related_thoughts (lexical similarity
  across the actual thought text, stage-independent) and
  same_category_thoughts (the honest name for the old categorical
  match, now requiring an actual shared tag rather than stage alone).
  Every response also carries a warnings list (stage-order issues; see
  below).
• thought_number is now optional on process_thought — omit it to
  have the server assign the next free number on the current line
  (mainline, or the given branch_id). Sending a thought_number that's
  already used on that line is now rejected (previously accepted
  silently, corrupting progress accounting).
• generate_summary output restructured: a new content section
  (per-stage thought excerpts, aggregated assumptions_challenged,
  open_branches, revision_chains) sits alongside the old counters,
  now under structure. completionStatus.percentComplete is replaced
  by structure.completion.stages_used_percent, whose denominator is
  always len(ThoughtStage) (5), never hardcoded.
• Export/import path errors, invalid stage names, and duplicate
  thought_number now fail the tool call as a protocol-level error
  (MCPError, JSON-RPC INVALID_PARAMS/REQUEST_TIMEOUT) instead of
  being swallowed into a {"status": "failed"} response. Execution
  failures the model can retry past (a missing import file, an
  unexpected storage error) instead come back as is_error=True tool
  results.
• Session/export files on disk are unaffected — no schema migration
  needed, existing current_session.jsonl and v1/v2 exports keep
  working (see docs/MIGRATION_PLAN.md §6).

Added

• Stage-order warnings: every process_thought response includes a
  warnings list noting a skipped or backward stage transition on the
  mainline. Permissive by default. New --strict-stages CLI flag (and
  ThoughtStorage-adjacent server state) rejects such transitions with
  MCPError instead of just warning.
• --health CLI subcommand: checks the storage directory exists, is
  writable, and that the session lock can be acquired; prints a report
  and exits 0/1.
• --transport {stdio,sse,streamable-http} CLI flag (plus
  --host/--port); stdio remains the default.
• Structured stderr logging: each tool call logs start/completion (or
  failure) with elapsed time in milliseconds.
• Tool annotations (readOnlyHint, destructiveHint, idempotentHint)
  are now set per tool — clear_history and import_session are marked
  destructive.

Fixed

• A handful of concurrent process_thought calls against the same
  session could stall for minutes: blocking file I/O (a file lock
  acquire plus fsync) ran synchronously inside an async def handler,
  serializing every concurrent request behind it on the single event-loop
  thread. Storage calls now run in worker threads
  (anyio.to_thread.run_sync); a lock that can't be acquired within its
  timeout now fails fast with a clear error instead of blocking.
• generate_summary previously reported only structural statistics
  (stage/branch/tag counts) with no actual thinking content — it now
  includes per-stage excerpts of the recorded thoughts, aggregated
  challenged assumptions, open branches, and revision chains.