v1.15.2

Fixed

• Tool descriptions and the runtime low-saving nudge no longer reference tools removed in
  the 3-tool strip. MinifyFile, OutlineCSharpFile, and the nudge pointed the model at
  focus_method, and TraceDiRegistrations referenced trace_implementors — none of which
  are registered tools, so a model following that guidance reached for a tool that doesn't
  exist and fell back to reading the whole file. They now point at outline_c_sharp_file
  plus a narrow Read. OutlineCSharpFile and MinifyFile also lead with an imperative
  trigger ("call this first before reading a C#/VB file over ~50 lines") so the host model
  reaches for them earlier. Descriptions only — no behaviour change (#103).

v1.15.1

Fixed

• OutlineCSharpFile on .razor files now reports correct // L.. line ranges.
  The Razor preprocessor previously rebuilt the @code/@functions C# into a fresh
  synthetic class, so every member's line number was offset by however far the @code
  block sat in the file — a narrow Read of a printed range landed on the wrong lines.
  Extraction is now line-aligned: each extracted line keeps its original position, so
  the ranges map straight back to the real .razor file. (Surfaced once the warm-start
  flow — outline then Read the line-range — became the primary path.)

v1.15.0

Changed — stripped to a warm-start tool set

• Reduced the advertised MCP surface from 11 tools to 3: OutlineCSharpFile
  (structural warm start), MinifyFile (whole-file, all languages), and
  TraceDiRegistrations (DI wiring — the one discovery grep can't do cleanly).
  The focus/trace family (FocusMethod, FocusMultipleMethods,
  FocusMethodsAcrossFiles, FocusType, FocusCallers, TraceCallers,
  TraceImplementors, MapProject) is no longer registered as an MCP tool (#99).
• Fixed per-session overhead dropped 4,134 → 1,415 tokens (−66%):
  ServerInstructions rewritten for the 3-tool set, 8 tool schemas removed, and a
  legacy duplicate focus_method registration (FocusedEmissionTool) deleted.
  Rationale: a controlled A/B (outline-only vs full toolset vs no-MCP) showed the
  focus/trace tools added no measured token value on scoped tasks; the durable win
  is a cheap structural warm start, after which Grep + a narrow Read of the
  outline's line-ranges does the rest. On a real cheaper-model task this flips the
  session ledger from net-negative to net-positive.
• Removed the Markdown emitter.

Documentation

• Reframed the README, NuGet readme, viewer pages, llms.txt, and registry
  description around the warm start: it curbs wasteful reading, with the biggest win
  on smaller/cheaper models (~8%) and large files. Dropped the "30–95% for everyone"
  framing — the win scales inversely with model capability (~1% top-tier, ~5–7%
  mid-tier, ~8% cheaper).

v1.14.1

Improvements

• Fixed MCP context overhead trimmed: ServerInstructions compacted from 2,207 to 1,583 tokens with no rule removed — selection rules now precede the supported-types table so clients that truncate long instructions keep the behavior-critical part — and emitter result banners collapsed from 2–5 lines to a single line, with a shorter session: header line (#94)

Fixes

• Background self-update no longer downgrades a config pinned to a newer version than the discovered latest (e.g. a local dev build) — SetPinnedVersion refuses to lower an existing pin (#96)
• Markdown minify banner note collapsed to a single line, matching the other emitters (#96)

v1.14.0

New Tools

• trace_di_registrations — discover every Add/TryAdd/AddKeyed registration for a type across the project; detects generic, typeof(), and factory-lambda forms (#86)
• map_project — project-wide type-to-file map with kind and base list; opt-in via TOKENSAVER_ENABLE_MAP_PROJECT=1 (#87, #89)

Improvements

• Agentic-mode optimizations: parse cache avoids re-parsing the same syntax tree across consecutive tool calls (#87)
• trace_di_registrations now recognizes TryAddKeyed* registrations (#88)
• docs/tools.json is now the single source of truth for tool metadata; README and viewer tool grid are generated from it (#90)

Fixes

• Removed duplicate tool cards in Capabilities.razor that caused a build failure (#91)

Other

• Added MIT License
• Startup log lines now show both UTC and local time (#85)

v1.13.7

Fixed

• Telemetry uploads are now idempotent, ending duplicate dashboard rows.
  The durable startup resend re-POSTs every pending row, and a client can spawn
  several MCP server processes at once, so the same logical event could be
  POSTed more than once — the payload carried no identifier and the dashboard
  inserted every POST as a new row. Each recorded event now carries a stable
  EventId (GUID) that is sent with the upload; the dashboard dedupes on it via
  a unique index plus an idempotent insert, so a re-send collapses to a no-op.
  Older clients that send no EventId, and older dashboards that ignore the
  field, keep working unchanged.

v1.13.6

Added

• Telemetry uploads are now durable. The dashboard upload was fire-and-forget
  with no retry and an ignored HTTP status, so a transient failure or a process exit
  mid-flight silently dropped a row that was recorded locally. Each report.json row
  now carries an upload-tracking flag: a confirmed 2xx — or a permanent 4xx
  rejection that retrying can't fix — settles the row, while a transient failure
  (5xx / 429 / network) leaves it pending. On startup the server resends any
  still-pending row. Rows written before this change are left untouched (never
  resent), so there is no mass re-upload of history. Rows the dashboard would
  reject — chiefly an honest negative saving where a focused view costs more than
  the bare relevant-code baseline (TokensWithTool > TokensWithoutTool) — are
  skipped client-side and settled locally rather than sent for a guaranteed 400.

Fixed

• NOT FOUND rows no longer log a bogus 0% saving. A focus miss returns a
  small members outline (plus the partial/inherited hint), not the whole file,
  yet telemetry logged whole-file → whole-file (0% saved). It now logs
  whole-file → actual response, crediting the real saving versus the model
  reading the whole file to discover the member isn't there. Applies to the
  NOT FOUND paths of focus_method, focus_multiple_methods, focus_type,
  and focus_callers.

v1.13.5

Added

• Focus misses now hint where to look. When focus_method or
  focus_multiple_methods returns NOT FOUND and the file's type is partial,
  the response notes the member may be in a sibling file in the same
  namespace/folder; when the type has a base list, it notes the member may be
  inherited and points at the base type's file. Turns a bare NOT FOUND into an
  actionable next step instead of a dead end.

v1.13.4

Fixed

• Cache hits no longer overstate savings against the whole-file baseline. A
  cached re-serve logged the raw whole-file token count as its baseline (e.g.
  8,201 → 246, 97%) while the original call recorded the conservative
  relevant-code baseline (303 → 246, 19%) — re-crediting a whole-file saving the
  session ledger and in-chat header never re-credit, and inflating the
  dashboard's aggregate totals on every repeat view. The cache now stores the
  same conservative baseline the first call logged, so a cache-hit row is
  identical to the original. Applies to focus_method, focus_multiple_methods,
  focus_type, and focus_callers.

Changed

• Cache-hit telemetry rows are tagged with the originating tool. A re-served
  result logged a bare Cache tool name, hiding which tool produced it; rows are
  now labelled <tool> Cache (e.g. Focused Emitter (multi) Cache).

v1.13.3

Added

• Reports now record the TokenSaver version that produced them. Each
  uploaded report carries an McpVersion field resolved from the running
  build, so the dashboard can attribute savings to a specific release and tell
  old clients from new ones. The viewer stores it in a new nullable column and
  surfaces it in the admin log. Existing databases are migrated in place by a
  guarded ALTER TABLE ... ADD COLUMN on startup — additive and idempotent, so
  rows ingested before this field keep a null version and no data is touched.

Changed

• Session savings are honest about repeat views of the same file. The
  session ledger added each call's whole-file baseline every time, so viewing
  one file several ways (a different method, or outline-then-minify) credited the
  whole-file saving once per view — inflating a real ~50% into an apparent ~90%.
  The ledger now tracks which sources have already been counted and adds the
  whole-file baseline only on first sighting; later distinct views add only their
  own output. Repeat views drop the "% saved" headline and state the baseline was
  already counted, so per-call lines can no longer be summed into an overstated
  total. (Identical repeat calls were already served from cache without touching
  the ledger.)

Fixed

• A comma-containing focus_method name now auto-routes to
  focus_multiple_methods. Passing "A,B" as methodName was treated as a
  single missing method, dumping the whole outline as a "not found" reply; it is
  now split and routed to the multi-method tool.