manojmallick · manojmallick · Jun 5, 2026 · Jun 5, 2026
@@ -10,6 +10,24 @@ Format: [Semantic Versioning](https://semver.org/)
 
 ---
 
+## [6.12.0] — 2026-06-05
+
+### Added
+
+- **Surgical Context Phase 2 — demand-driven retrieval (#219, PR #220):**
+  - **`get_lines` MCP tool** (10th tool) — fetch an exact `{ file, start, end }` line range on demand. Lines are clamped to the file bounds, secret-scanned via the existing redactor, and sandboxed to the project root. This is the demand-driven workhorse: agents read the lines behind a `:start-end` anchor instead of re-opening whole files.
+  - **`sigmap ask --mode index`** — two-tier output that emits only symbol-header pointers (`symbol  :start-end`), dropping parameter lists, return types, and bodies. Agents re-fetch bodies via `get_lines`.
+  - **`sigmap ask --since <ref>`** — delta context that restricts ranked output to files changed since a git ref.
+- **Token Reduction dashboard panel (Surface A)** — `sigmap --report` now renders a "Token Reduction" panel (whole-file baseline vs ranked signatures vs surgical, with per-repo rows), sourced from `benchmarks/reports/token-reduction.json` — numbers are never hand-typed.
+- New **Surgical Context** guide (`docs-vp/guide/surgical-context.md`) covering line anchors, `--mode index`, `--since`, and the `get_lines` MCP tool.
+
+### Changed
+
+- **Budget-aware progressive disclosure** — when generated context exceeds `maxTokens`, the token budget now collapses signature bodies to their line anchors (keeping `symbol  :start-end`) *before* dropping whole files, degrading gracefully.
+- **CI** — added `workflow_dispatch` to the develop→main sync workflow so it can be run on demand (PR #218).
+
+---
+
 ## [6.11.1] — 2026-06-04
 
 ### Fixed

@@ -35,6 +35,10 @@ To ensure proper attribution:
 
 We welcome contributions! See [Contributing](./docs/CONTRIBUTING.md) for guidelines.
 
+### Recent Contributors (v6.12.0)
+- **@manojmallick** — feat: Surgical Context Phase 2 — `get_lines` MCP tool, `ask --mode index`, `ask --since`, budget-aware body collapse, Token Reduction dashboard panel (#219, PR #220)
+- **@manojmallick** — ci: add `workflow_dispatch` to the develop→main sync workflow (PR #218)
+
 ### Recent Contributors (v6.11.1)
 - **@rudi193-cmd** — fix: include hot-cold cold signatures in the bundled MCP server (#201, PR #216)
 

@@ -87,8 +87,8 @@ Ask → Rank → Context → Validate → Judge → Learn
 ## Benchmark
 
 ```
-Benchmark : sigmap-v6.11-main (21 repositories, including R language)
-Date      : 2026-06-04
+Benchmark : sigmap-v6.12-main (21 repositories, including R language)
+Date      : 2026-06-05
 
 Hit@5          : 81.1%   (baseline 13.6%  — 6.0× lift)
 Token reduction: 96.5%   (across 21 repos)
@@ -183,7 +183,7 @@ Use SigMap with open-source tools and fully self-hosted setups:
 | **JetBrains** | [Marketplace](https://plugins.jetbrains.com/plugin/31109-sigmap--ai-context-engine/) | [github.com/manojmallick/sigmap-jetbrains](https://github.com/manojmallick/sigmap-jetbrains) | IntelliJ IDEA, WebStorm, PyCharm, GoLand — tool window + actions |
 | **Neovim** | lazy.nvim / packer / vim-plug | [github.com/manojmallick/sigmap.nvim](https://github.com/manojmallick/sigmap.nvim) | `:SigMap`, `:SigMapQuery` float window, statusline widget |
 
-**MCP server** — 9 on-demand tools for Claude Code and Cursor:
+**MCP server** — 10 on-demand tools for Claude Code and Cursor:
 
 ```bash
 sigmap --mcp

@@ -1,22 +1,22 @@
 ---
 title: Benchmark overview
-description: Official v6.11.1 benchmark snapshot. 96.5% average token reduction across 21 repos, 81% retrieval hit@5, 41.8% fewer prompts, and R language support verified.
+description: Official v6.12.0 benchmark snapshot. 96.5% average token reduction across 21 repos, 81% retrieval hit@5, 41.8% fewer prompts, and R language support verified.
 head:
   - - meta
     - property: og:title
-      content: "SigMap benchmark overview — v6.11.1 snapshot with R language"
+      content: "SigMap benchmark overview — v6.12.0 snapshot with R language"
   - - meta
     - property: og:description
-      content: "Token, retrieval, quality, and task metrics from latest v6.11.1 benchmark run (2026-06-04) with 21 repositories including R language support."
+      content: "Token, retrieval, quality, and task metrics from latest v6.12.0 benchmark run (2026-06-05) with 21 repositories including R language support."
   - - meta
     - property: og:url
       content: "https://manojmallick.github.io/sigmap/guide/benchmark"
 ---
 
 # Benchmark overview
 
-::: info Official v6.11.1 benchmark snapshot (21 repos, including R language)
-**Benchmark ID:** sigmap-v6.11-main &nbsp;·&nbsp; **Date:** 2026-06-04
+::: info Official v6.12.0 benchmark snapshot (21 repos, including R language)
+**Benchmark ID:** sigmap-v6.12-main &nbsp;·&nbsp; **Date:** 2026-06-05
 
 | Metric | Value |
 |---|---:|
@@ -37,9 +37,9 @@ This is the landing page for the public benchmark story. It answers four differe
 | SigMap reduces retries and wrong-context answers | [Task benchmark](/guide/task-benchmark) |
 | SigMap keeps large repos inside model limits | [Quality benchmark](/guide/quality-benchmark) |
 
-## Official v6.11.1 snapshot (with R language support)
+## Official v6.12.0 snapshot (with R language support)
 
-Latest saved benchmark run: **2026-06-04 (v6.11.1)**
+Latest saved benchmark run: **2026-06-05 (v6.12.0)**
 
 | Metric | Result |
 |---|---:|
@@ -65,6 +65,7 @@ Latest saved benchmark run: **2026-06-04 (v6.11.1)**
   - ggplot2: 94.3% reduction (381.5K → 21.7K tokens)
   - dplyr: 93.4% reduction (145.1K → 9.5K tokens)
   - shiny: 96.2% reduction (264.6K → 10.0K tokens)
+- **New in v6.12.0:** demand-driven *Surgical Context* (`ask --mode index` + the `get_lines` MCP tool) cuts upfront `ask` context further on top of the figures above by emitting symbol pointers instead of bodies — see the [Surgical Context guide](/guide/surgical-context).
 
 ### 2. Retrieval quality
 

@@ -7,7 +7,7 @@ head:
       content: "SigMap CLI Reference — every command and flag with examples"
   - - meta
     - property: og:description
-      content: "All 37 SigMap commands and flags documented with examples. ask, plan, bench, judge, validate, roots, history, --ci, --cost, --coverage, --watch, --diff, --mcp, --report, --health, weights --export/--import and more."
+      content: "All 39 SigMap commands and flags documented with examples. ask, plan, bench, judge, validate, roots, history, --ci, --cost, --coverage, --watch, --diff, --mcp, --report, --health, weights --export/--import and more."
   - - meta
     - property: og:url
       content: "https://manojmallick.github.io/sigmap/guide/cli"
@@ -19,7 +19,7 @@ head:
       content: "SigMap CLI Reference — every command and flag with examples"
   - - meta
     - name: twitter:description
-      content: "All 37 SigMap commands and flags documented with examples. ask, plan, bench, judge, validate, history, --ci, --cost, --coverage, --watch, --diff, --mcp, --report, --health, weights --export/--import and more."
+      content: "All 39 SigMap commands and flags documented with examples. ask, plan, bench, judge, validate, history, --ci, --cost, --coverage, --watch, --diff, --mcp, --report, --health, weights --export/--import and more."
   - - meta
     - name: twitter:image:alt
       content: "SigMap CLI Reference"
@@ -47,6 +47,8 @@ If you are new to the product, start with the workflow pages first:
 | `ask "<query>" --followup` | Reuse previous session context for follow-up queries (session carry-forward) |
 | `ask "<query>" --package <name>` | Scope retrieval to a specific monorepo workspace package |
 | `ask "<query>" --global` | Disable package scoping; search entire repo (monorepo override) |
+| `ask "<query>" --mode index` | Surgical Context: emit symbol-header pointers (`symbol :start-end`) only — no bodies; fetch on demand via `get_lines` |
+| `ask "<query>" --since <ref>` | Delta context: restrict ranked output to files changed since a git ref |
 | `plan "<goal>"` | Analyze change impact and plan modifications — returns files grouped by confidence |
 | `judge --response <f> --context <f>` | Rule-based groundedness scoring for LLM responses |
 | `validate` | Validate config and coverage; optional query symbol check |
@@ -180,6 +182,42 @@ sigmap ask "find all auth handlers" --global --json
 |--------|-------------|
 | `--global` | Disable automatic package inference; search entire repo |
 
+### ask --mode index (Surgical Context)
+
+Emit a two-tier **symbol index** instead of full signature blocks. Each ranked file is reduced to its declaration heads plus line anchors (`symbol  :start-end`) — parameter lists, return types, and bodies are dropped. The agent reads this minimal map, then fetches the exact lines it needs on demand via the [`get_lines` MCP tool](/guide/mcp). This is the demand-driven half of *Surgical Context* (line anchors are the first half — see above).
+
+```bash
+sigmap ask "where is config loaded" --mode index
+```
+
+```
+# SigMap Query Context (index mode)
+> Symbol index only — fetch exact lines on demand via the `get_lines` MCP tool.
+
+## src/config/loader.js
+function loadConfig  :42-58
+function detectAutoSrcDirs  :12-39
+```
+
+| Option | Description |
+|--------|-------------|
+| `--mode index` | Emit symbol-header pointers only; bodies fetched on demand via `get_lines` |
+
+When over `maxTokens`, the regular (non-index) generate path now degrades the same way automatically: it collapses bodies to anchors before dropping whole files. See the [Surgical Context guide](/guide/surgical-context).
+
+### ask --since (delta context)
+
+Restrict ranked output to files changed since a git ref, so a steady-state turn carries near-zero context. Combine with `--mode index` for the leanest possible turn.
+
+```bash
+sigmap ask "finish the refactor" --since main
+sigmap ask "what did I touch" --since HEAD~3 --mode index
+```
+
+| Option | Description |
+|--------|-------------|
+| `--since <ref>` | Keep only ranked files changed since `<ref>` (any git ref: branch, tag, or SHA) |
+
 ---
 
 ## plan

@@ -1,6 +1,6 @@
 ---
 title: Generalization — SigMap across languages, domains & repo sizes
-description: SigMap generalizes across 21 repos, 31 languages, and multiple domains with 81.1% hit@5 in the latest saved v6.11.1 retrieval run.
+description: SigMap generalizes across 21 repos, 31 languages, and multiple domains with 81.1% hit@5 in the latest saved v6.12.0 retrieval run.
 head:
   - - meta
     - property: og:title
@@ -19,8 +19,8 @@ head:
 SigMap was not tuned for one repo. This benchmark matters because it shows the same workflow transfers across different languages, repo sizes, and architectures without manual tuning.
 :::
 
-::: info Official v6.11.1 benchmark snapshot
-**Benchmark ID:** sigmap-v6.11-main &nbsp;·&nbsp; **Date:** 2026-06-04 (with R language)
+::: info Official v6.12.0 benchmark snapshot
+**Benchmark ID:** sigmap-v6.12-main &nbsp;·&nbsp; **Date:** 2026-06-05 (with R language)
 
 | Metric | Value |
 |---|---:|
@@ -37,7 +37,7 @@ The important part of SigMap's benchmark story is not just the topline score. It
 ::: info What "generalization" means here
 SigMap's signature extractors are hand-written regex patterns, not ML models. Generalization
 means: *do the patterns hold up on codebases the authors never inspected?* The answer across
-these 90 tasks is yes — 81% hit@5 with no per-repo tuning in the latest saved v6.11.1 run.
+these 90 tasks is yes — 81% hit@5 with no per-repo tuning in the latest saved v6.12.0 run.
 :::
 
 - **21 repos** (including 3 R language repos)

@@ -1,13 +1,13 @@
 ---
 title: MCP server setup
-description: Set up the SigMap MCP server for Claude Code, Cursor, and Windsurf. On-demand codebase access with 9 tools over stdio. Zero npm install.
+description: Set up the SigMap MCP server for Claude Code, Cursor, and Windsurf. On-demand codebase access with 10 tools over stdio. Zero npm install.
 head:
   - - meta
     - property: og:title
       content: "SigMap MCP Server — on-demand codebase context"
   - - meta
     - property: og:description
-      content: "Give Claude Code, Cursor, and Windsurf on-demand access to your codebase signatures. 9 MCP tools over stdio."
+      content: "Give Claude Code, Cursor, and Windsurf on-demand access to your codebase signatures. 10 MCP tools over stdio."
   - - meta
     - property: og:url
       content: "https://manojmallick.github.io/sigmap/guide/mcp"
@@ -22,7 +22,7 @@ head:
 
 Give Claude Code, Cursor, and Windsurf on-demand access to your codebase signatures. Zero npm install.
 
-The SigMap MCP server exposes 9 tools over the stdio Model Context Protocol. Your AI agent calls only what it needs — keeping token costs low.
+The SigMap MCP server exposes 10 tools over the stdio Model Context Protocol. Your AI agent calls only what it needs — keeping token costs low.
 
 > **Setup time: under 2 minutes.** Use `sigmap --setup` for automatic configuration.
 
@@ -96,10 +96,10 @@ Stack both MCP servers for the two-layer context strategy — SigMap for always-
 }
 ```
 
-## 9 available tools
+## 10 available tools
 
 ::: tip New in v6.3.0 — native tool registration
-Claude Code and Codex now receive the full tool list at MCP startup without a discovery round-trip. The server declares all 9 tools in the `initialize` response, so your AI sees them immediately. No config change needed — upgrade via `npm install -g sigmap@latest`.
+Claude Code and Codex now receive the full tool list at MCP startup without a discovery round-trip. The server declares all 10 tools in the `initialize` response, so your AI sees them immediately. No config change needed — upgrade via `npm install -g sigmap@latest`.
 :::
 
 All tools are available on-demand — your AI agent calls only what it needs.
@@ -114,6 +114,8 @@ All tools are available on-demand — your AI agent calls only what it needs.
 | `create_checkpoint` | Records session progress with a git state snapshot. | `summary` (required string) | `create_checkpoint(summary="Added rate limiting")` |
 | `get_routing` | Returns the model tier hints table (fast / balanced / powerful per file) based on complexity scores. | none | `get_routing()` |
 | `query_context` | Ranks all files by relevance to a free-text query using TF-IDF scoring. Returns top-K files. New in v2.3. | `query` (required string), `topK` (optional number, default 10) | `query_context(query="authentication flow")` |
+| `get_impact` | Returns the blast radius of a file — direct importers, transitive importers, affected tests and routes. | `file` (required string), `depth` (optional number, default 3) | `get_impact(file="src/auth/service.ts")` |
+| `get_lines` | **Surgical Context** demand-driven fetch: returns an exact line range from a file behind a `:start-end` anchor. Clamped to file bounds, secret-scanned, sandboxed to the project root. New in v6.12.0. | `file` (required string), `start` (required number), `end` (required number) | `get_lines(file="src/config/loader.js", start=42, end=58)` |
 
 ## Token cost per tool call
 
@@ -165,7 +167,9 @@ Expected output:
       { "name": "list_modules" },
       { "name": "create_checkpoint" },
       { "name": "get_routing" },
-      { "name": "query_context" }
+      { "name": "query_context" },
+      { "name": "get_impact" },
+      { "name": "get_lines" }
     ]
   }
 }

@@ -1,6 +1,6 @@
 ---
 title: Quality benchmark
-description: What token reduction means operationally in v6.11.1. 16/21 repos overflow GPT-4o without SigMap, 5,200+ files would be hidden, and GPT-4o input savings reach $10,500+/month at 10 calls/day.
+description: What token reduction means operationally in v6.12.0. 16/21 repos overflow GPT-4o without SigMap, 5,200+ files would be hidden, and GPT-4o input savings reach $10,500+/month at 10 calls/day.
 head:
   - - meta
     - property: og:title
@@ -15,8 +15,8 @@ head:
 
 # Quality benchmark
 
-::: info Official v6.11.1 benchmark snapshot
-**Benchmark ID:** sigmap-v6.11-main &nbsp;·&nbsp; **Date:** 2026-06-04 (with R language)
+::: info Official v6.12.0 benchmark snapshot
+**Benchmark ID:** sigmap-v6.12-main &nbsp;·&nbsp; **Date:** 2026-06-05 (with R language)
 
 | Metric | Value |
 |---|---:|
@@ -34,7 +34,7 @@ Token reduction is the mechanism. This benchmark shows the operational consequen
 - how much code would be hidden without SigMap?
 - what does that mean for API cost?
 
-Latest saved run: **2026-06-04 (v6.11.1)**
+Latest saved run: **2026-06-05 (v6.12.0)**
 
 ## Headline numbers
 

@@ -1,6 +1,6 @@
 ---
 title: Retrieval benchmark
-description: Latest saved retrieval benchmark for SigMap v6.11.1. 81% hit@5 vs 13.6% random baseline across 90 tasks on 18 repos, with R language support.
+description: Latest saved retrieval benchmark for SigMap v6.12.0. 81% hit@5 vs 13.6% random baseline across 90 tasks on 18 repos, with R language support.
 head:
   - - meta
     - property: og:title
@@ -15,8 +15,8 @@ head:
 
 # Retrieval benchmark
 
-::: info Official v6.11.1 benchmark snapshot
-**Benchmark ID:** sigmap-v6.11-main &nbsp;·&nbsp; **Date:** 2026-06-04 (with R language)
+::: info Official v6.12.0 benchmark snapshot
+**Benchmark ID:** sigmap-v6.12-main &nbsp;·&nbsp; **Date:** 2026-06-05 (with R language)
 
 | Metric | Value |
 |---|---:|
@@ -29,7 +29,7 @@ head:
 | GPT-4o overflow (without → with) | **16/21 → 0/21** |
 :::
 
-Latest saved run: **2026-06-04 (v6.11.1)**
+Latest saved run: **2026-06-05 (v6.12.0)**
 
 **Result:** SigMap finds the right file in the top 5 far more often than chance — **81% hit@5** vs **13.6%** random baseline across 90 tasks on 18 real repos.