flowseeker 0.1.7 → 0.1.8

package/.env.example

@@ -0,0 +1,7 @@
+# FlowSeeker semantic retrieval API key.
+# Copy this file to .env in the repo you want to scan.
+# Semantic search is optional and OFF unless enabled in .flowseeker/config.json.
+# Run: flowseeker semantic init --jina
+# Then fill the key below.
+
+FLOWSEEKER_EMBEDDING_API_KEY=

package/CHANGELOG.md

@@ -1,111 +1,134 @@
 # FlowSeeker Changelog
 
-## 0.1.7 - Update Notice and Deterministic Accuracy
-
-- **Update notice:** Extension checks npm for newer FlowSeeker versions at most once per day, shows a non-modal update notice, supports dismissing a specific version, and fails silently when offline.
-- **MCP CLI version checks:** `flowseeker-mcp --version` prints the package version, and `flowseeker-mcp --check-updates` checks npm explicitly without running the MCP stdio server.
-- **MCP protocol safety:** Normal MCP server startup does not perform passive network update checks and keeps stdout reserved for JSON-RPC only.
-- **Caller/callee candidate expansion:** Workspace index now records extracted call names and uses symbol/call relationships to expand candidate files beyond imports.
-- **Generated/noise downranking:** Generated folders, snapshots, mocks, protobuf/gRPC outputs, bundles, framework build folders, and similar noise paths are downranked without being excluded.
-- **High-precision task terms:** Quoted literals and route-like paths are preserved as literal retrieval terms before broad keyword matching.
-
-### 0.1.7 Files changed
-`src/utils/updateChecker.ts` (new), `src/extension.ts`, `src/cli/mcpServer.ts`, `src/index/workspaceIndex.ts`, `src/pipeline/taskUnderstanding.ts`, `package.json`, `package-lock.json`, `CHANGELOG.md`, `README.md`, `COMPATIBILITY.md`
-
-## 0.1.6 - Accuracy, MCP Native Chat, and Package Readiness
-
-- **Quality Score:** Added single composite metric `computeQualityScore()` combining hit@10, context coverage, avoid-hit rate, MRR, and token reduction into one comparable number (Quality = 0.35·Hit@10 + 0.25·Coverage + 0.20·(1−AvoidHitRate) + 0.10·MRR + 0.10·TokenReduction). Printed in every headless evaluation summary line.
-- **Benchmark expansion:** Wired up 22 fixture repos with 697 labeled cases (including 2 negative cases for AvoidHitRate) via `.flowseeker/eval-benchmark-mega.json`. Added `scripts/validate-eval-paths.js` — deterministic ground-truth checker that mirrors runtime path matching, with `--fix` mode to prune dead paths. Recorded baseline Quality Score at 0.741.
-- **Negative term utilization:** Extended `detectNegativeTerms()` with `except` and `ngoại trừ` (Vietnamese exclusion) cues. Existing ranker penalty (-24 score + contrastContext) already wired via `applyTaskMetadata()`. Tested against the full benchmark; zero regression at the conservative pattern set.
-- **Active file context boost:** Ranker now accepts optional `editorContext` (active file, directory siblings, recently opened tabs) and applies small score boosts (+15 active file, +5 sibling dir, +8 recent tabs). Editor context is purely optional — headless evaluation is completely unaffected.
-- **Deterministic eval validator:** New `scripts/validate-eval-paths.js` validates every expected/avoid/group path resolves to a real fixture file. Mirrors runtime `pathMatches()` semantics (exact, suffix, substring). `--fix` mode prunes dead paths and drops unmeasurable cases.
-- **MCP prompt discovery:** Standalone MCP server now advertises `prompts` capability and supports `prompts/list` + `prompts/get` for `retrieve`, `guide`, and `files`. These prompts map to the stable tools `flowseeker_retrieve`, `flowseeker_guide`, and `flowseeker_files` so Native Chat hosts can show short suggestions when their UI supports MCP prompt/tool autocomplete.
-- **MCP metadata polish:** Tightened tool descriptions and input schema wording for truncated IDE suggestion UIs while keeping the stable tool names unchanged.
-- **Cross-IDE config correctness:** Fixed MCP help/docs to use `mcpServers` for Cursor, Antigravity, Windsurf, Claude Desktop, and `servers` only for VS Code/Copilot `.vscode/mcp.json`.
-- **Antigravity Native Chat note:** Documented current Antigravity behavior: selecting FlowSeeker may insert `@mcp:flowseeker:` without showing second-level command suggestions. FlowSeeker still works through natural language or direct tool names; autocomplete visibility is host-controlled.
-- **npm MCP readiness:** Verified `flowseeker-mcp` builds from `dist/cli/mcpServer.js`, responds to `tools/list` and `prompts/list`, and is included by `npm pack --dry-run` for users who want MCP-only integration without installing the extension.
-
-### 0.1.6 Files changed (vs 0.1.5 baseline)
-`src/pipeline/evaluationMetrics.ts`, `src/cli/runEvaluation.ts`, `scripts/validate-eval-paths.js` (new), `scripts/render-icon.js` (new), `TODO-0.1.6.md` (new), `src/pipeline/taskUnderstanding.ts`, `src/pipeline/ranker.ts`, `src/pipeline/runPipeline.ts`, `src/cli/mcpServer.ts`, `src/mcp/mcpTools.ts`, `notes/Req5.txt`, `notes/mcp-config-paths.md`, `.gitignore`, `.vscodeignore`, `.npmignore` (new), `README.md`, `ARCHITECTURE.md`, `COMPATIBILITY.md`, `MARKETPLACE_DEMO.md`, `UPDATE_GUIDE.md`, `plan.md`
-
-## 0.1.5 - Agent UX Stability, Cross-Host Compatibility, and MCP Server
-
-- **MCP Server:** Added standalone MCP server (`flowseeker-mcp`) exposing FlowSeeker retrieval as MCP tools. Works with Cursor, Antigravity, Windsurf, Claude Desktop, and any MCP-compliant AI client.
-- **MCP tools:** `flowseeker_retrieve` (full Solve Packet), `flowseeker_guide` (agent prompt), `flowseeker_files` (lightweight file list).
-- **Cross-IDE notification:** FlowSeeker now detects Cursor, Antigravity, and Windsurf hosts at startup and shows a one-time guidance notification pointing users to the sidebar chat or MCP setup.
-- **Headless agent prompt:** Added `agentPromptHeadless.ts` — agent solve prompt generator without VS Code dependency, used by MCP server and CLI.
-- **Standalone CLI binary:** `flowseeker-mcp` bin for running the FlowSeeker MCP server outside of any IDE.
-- **Host detection:** FlowSeeker now identifies the editor host at startup (VS Code, Cursor, Antigravity, Windsurf, VSCodium, Theia) and logs the identity + capabilities to the Output channel.
-- **Runtime probing:** Beyond `typeof` checks, FlowSeeker now calls `vscode.lm.selectChatModels()` to count actual available models and vendors — a reliable signal of whether the host exposes usable AI models.
-- **Tier table:** Single source of truth for host capabilities — `native-full` (VS Code/Insiders), `native-unverified` (Cursor, Antigravity, Windsurf — API exists but not confirmed to render), `sidebar-primary` (VSCodium, Theia — no API).
-- **Shared command runner:** Core retrieve→ask flow extracted into `flowCommandRunner.ts` so both native chat and sidebar use the same logic.
-- **Enhanced compatibility check:** `FlowSeeker: Check Host Compatibility` now shows host identity, tier, live model probe results, and offers human-in-the-loop verification for unverified hosts.
-- **Status bar indicator:** Shows current tier mode (`FlowSeeker` for native, `FlowSeeker: Sidebar` for sidebar-only) with tooltip explaining why. Click opens compatibility report.
-- **JSON prompt hardening:** Edit generation instruction moved to END of prompt (after context pack), strengthened wording, added `response_format: { type: "json_object" }` for OpenAI-compatible providers, improved JSON extraction from prose responses.
-- **Per-file apply status:** Edits applied file-by-file with individual success/failure reporting including file name and reason.
-- **Graded provider diagnosis:** `testProvider` classifies errors (401/403 credentials, 429 quota vs rate-limit, 5xx server, network, missing CLI) with actionable fix steps.
-- **MCP server visibility:** Agent policy tooltip shows each MCP server name, type, and status (configured/disabled/misconfigured).
-- **Feedback hardening:** Save confirmation includes file path and replay hint. CLI replay shows original top files for drift detection.
-- Added `COMPATIBILITY.md` with detailed host compatibility matrix and model sharing explanation.
-
-### 0.1.5 Files changed
-`src/runtime/hostDetect.ts` (new), `src/runtime/hostTier.ts` (new), `src/runtime/statusBar.ts` (new), `src/agent/flowCommandRunner.ts` (new), `src/mcp/mcpTools.ts` (new), `src/pipeline/agentPromptHeadless.ts` (new), `src/cli/mcpServer.ts` (new), `src/commands/checkHostCompatibility.ts`, `src/extension.ts`, `src/ui/chatViewProvider.ts`, `src/gateway/aiGateway.ts`, `src/gateway/aiProviders.ts`, `src/gateway/agentPrompts.ts`, `src/gateway/editProposalStore.ts`, `src/cli/runReplay.ts`, `src/pipeline/tokenSavings.ts`, `package.json`, `COMPATIBILITY.md` (new)
-
-## 0.1.4 - Token Proof Panel and Feedback Hardening
+## 0.1.8 - Local Semantic, One-Command MCP, And Solve Packet UX
+
+- Added full CLI entry points: `flowseeker`, `fs-guide`, `fs-auto`, `fs-files`, `fs-retrieve`, and `fs-mcp`.
+- Added one-command MCP setup UX for Claude Code, Codex CLI, Cursor, and VS Code, with `flowseeker doctor --mcp` self-checks.
+- Added optional local semantic embedding support through the lightweight core package plus the separate `flowseeker-local` model/runtime package path.
+- Added provider-neutral hosted semantic setup with `flowseeker semantic init --openai-compatible` and `flowseeker semantic check --test`.
+- Improved Solve Packet output with Retrieval Mode, First 3 Files To Open, Why These Files, Candidate Flow, Do Not Edit Yet, Missing Links, and Verification Command Hints.
+- Hardened read-only/investigation Solve Packets so they use context labels instead of edit-surface labels.
+- Added release-candidate proof gates for CLI/MCP output, golden precision, side-effect safety, package hygiene, local package hygiene, and semantic diagnostics.
+- Accepted golden precision gates for 0.1.8: Top-1 6, Top-3 12, Top-5 14, Top-10 19, Expected edit primary hits 12, Expected read hits 2, must-not violations 0, semantic active 8/8 UMS tasks.
+
+## 0.1.8 Public Proof, README, And MCP Onboarding
+
+- Reworked README wording around evidence-backed capabilities and known limits.
+- Added MCP onboarding snippets for Cursor, Antigravity, Windsurf, Claude Desktop, and VS Code/Copilot in `docs/mcp-onboarding.md`.
+- Added demo/screenshot review checklist in `docs/demo-screenshot-checklist.md` so public assets stay aligned with current Solve Packet 3.0 output.
+- Added marketplace/OpenVSX/NPM copy source in `docs/marketplace-copy.md`.
+- Public proof commands:
+  - `npm.cmd run compile --silent`
+  - `npm.cmd run test:manifest`
+  - `node scripts/analyze-phase08-public-proof.js --out-md .flowseeker/reports/phase08-8a-public-proof.md --out-json .flowseeker/reports/phase08-8a-public-proof.json`
+- Current public limits remain explicit: no sub-second guarantee, no default AI/LLM ranking, semantic search disabled by default, tree-sitter disabled by default for public benchmark claims, graph expansion disabled in production, and no monorepo/LSP/git-history guarantee.
 
-- Added sidebar proof panel showing token proof status (complete/partial/unavailable) with differentiated visual states per retrieval card.
-- Added context coverage display in the proof panel showing found/missing required slots.
-- Added proof warnings in the metric card for partial scans, capped evidence, and missing slots.
-- Added one-click "Copy token proof" action that copies a formatted token savings summary including proof status, coverage, and warnings.
-- Feedback save confirmation now includes the saved file path and a replay hint.
-- CLI replay reports now show original top files from feedback snapshots alongside current top files for drift visibility.
-
-## 0.1.3 - Token Savings Proof
-
-- Added measured token savings to FlowSeeker results, comparing Solve Packet tokens against indexed workspace text tokens.
-- Uses local BPE token counting with `cl100k_base` for headless, sidebar, and report metrics, with heuristic fallback only when tokenization fails.
-- Counts the actual prompt through VS Code `LanguageModelChat.countTokens()` when the selected native chat model exposes token counting.
-- Added generic Context Blueprint coverage so FlowSeeker checks required flow slots for each task shape instead of ranking only by keyword score.
-- Updated Solve Packet generation to preserve required context slots, report missing links, and mark likely noise.
-- Improved plan approval safety: responses that only ask to read more files are no longer treated as approvable plans.
-- Added bounded context expansion and retry when the AI explains that the first Solve Packet is insufficient.
-- Added structured-edit handling with diff preview, scope lock, apply actions, and verification command proposals.
-- Added sidebar chat history, provider profiles, model switching, provider model fetch, and compact quota/usage status.
-- Added agent controls for workspace edit approval, safe command proposals, MCP/tool proposals, and user-defined prompt rules.
-- Added host compatibility reporting for Native Chat, Language Model, and MCP support across VS Code-compatible editors.
-- Added sidebar feedback capture and real-world replay support so retrieval misses can become regression cases.
-- Normalized feedback kinds from the sidebar and legacy records so replay metrics classify missing context consistently.
-- Shows compact token reduction in native VS Code Chat after retrieval.
-- Added a sidebar chat fallback for VS Code-compatible editors that do not expose the native VS Code Chat participant API.
-- Added token savings and context coverage metrics to headless and VS Code evaluation reports alongside hit@K, MRR, and scan duration.
-- Keeps token reduction wording technical, with retrieval quality measured separately.
-
-## 0.1.2 - Scan UX and Logo Polish
-
-- Refined the FlowSeeker logo assets for extension icon, activity bar, and native chat usage.
-- Kept chat scan progress compact by updating status lines instead of flooding the conversation.
-- Improved AI plan language behavior so generated plans follow the user's request language when set to auto.
-- Tightened package excludes so Marketplace builds do not ship local caches, source maps, or reference raster assets.
-
-## 0.1.1 - Large Repo Stabilization
-
-- Added persistent workspace indexing and dependency expansion to reduce full-codebase scans.
-- Added scan budgets for max files, bytes, duration, raw evidence, and indexed candidates.
-- Improved Solve Packet output with clearer trace, context buckets, uncertainty notes, and agent guidance.
-- Added headless evaluation support for repeatable retrieval tests across sample workspaces.
-
-## 0.1.0 - Initial Release
-
-- **Core retrieval pipeline**: task understanding, workspace indexing, universal text scan, evidence ranking, evidence graph with framework edges
-- **Solve Packet v1**: structured reasoning context with flow buckets (entry/handler/service/schema/tests), problem hypothesis, noise risk detection, missing link identification, agent instructions
-- **Native VS Code Chat participant**: `@flowseeker /guide`, `/auto`, `/apply` commands with plan-first editing guardrail
-- **Framework adapters**: Laravel (PHP), NestJS (TypeScript), Express (JavaScript/TypeScript) - route->controller, event->listener, controller->service, view rendering, DI injection edges
-- **Structured extraction**: regex-based code structure extraction for 11 languages (TS, JS, Python, PHP, Java, Go, C#, Ruby, Rust, Kotlin) - declarations, calls, string literals, comments
-- **Scope lock**: edits are validated against approved plan scope before applying
-- **Diff preview**: file-level change preview before applying structured edits
-- **Verification loop**: auto-detection of test/lint/typecheck commands after edits, with pass/fail reporting
-- **Headless evaluation**: repeatable retrieval benchmarking with hit@K, MRR, expected-path coverage
-- **VS Code Language Model API integration**: works with the selected VS Code Chat model; AI Gateway fallback
-- **Result sidebar**: ranked evidence tree with checkbox selection, context pack copy, agent prompt copy
-- **Configuration**: `.flowseeker/config.json` workspace config + VS Code settings overrides
-- **Scan budget safety**: max files, max bytes, max time limits prevent runaway scans on large repos
+## 0.1.7 - Update Notice and Deterministic Accuracy
+
+- **Update notice:** Extension checks npm for newer FlowSeeker versions at most once per day, shows a non-modal update notice, supports dismissing a specific version, and fails silently when offline.
+- **MCP CLI version checks:** `flowseeker-mcp --version` prints the package version, and `flowseeker-mcp --check-updates` checks npm explicitly without running the MCP stdio server.
+- **MCP protocol safety:** Normal MCP server startup does not perform passive network update checks and keeps stdout reserved for JSON-RPC only.
+- **Caller/callee candidate expansion:** Workspace index now records extracted call names and uses symbol/call relationships to expand candidate files beyond imports.
+- **Generated/noise downranking:** Generated folders, snapshots, mocks, protobuf/gRPC outputs, bundles, framework build folders, and similar noise paths are downranked without being excluded.
+- **High-precision task terms:** Quoted literals and route-like paths are preserved as literal retrieval terms before broad keyword matching.
+
+### 0.1.7 Files changed
+`src/utils/updateChecker.ts` (new), `src/extension.ts`, `src/cli/mcpServer.ts`, `src/index/workspaceIndex.ts`, `src/pipeline/taskUnderstanding.ts`, `package.json`, `package-lock.json`, `CHANGELOG.md`, `README.md`, `COMPATIBILITY.md`
+
+## 0.1.6 - Accuracy, MCP Native Chat, and Package Readiness
+
+- **Quality Score:** Added single composite metric `computeQualityScore()` combining hit@10, context coverage, avoid-hit rate, MRR, and token reduction into one comparable number (Quality = 0.35·Hit@10 + 0.25·Coverage + 0.20·(1−AvoidHitRate) + 0.10·MRR + 0.10·TokenReduction). Printed in every headless evaluation summary line.
+- **Benchmark expansion:** Wired up 22 fixture repos with 697 labeled cases (including 2 negative cases for AvoidHitRate) via `.flowseeker/eval-benchmark-mega.json`. Added `scripts/validate-eval-paths.js` — deterministic ground-truth checker that mirrors runtime path matching, with `--fix` mode to prune dead paths. Recorded baseline Quality Score at 0.741.
+- **Negative term utilization:** Extended `detectNegativeTerms()` with `except` and `ngoại trừ` (Vietnamese exclusion) cues. Existing ranker penalty (-24 score + contrastContext) already wired via `applyTaskMetadata()`. Tested against the full benchmark; zero regression at the conservative pattern set.
+- **Active file context boost:** Ranker now accepts optional `editorContext` (active file, directory siblings, recently opened tabs) and applies small score boosts (+15 active file, +5 sibling dir, +8 recent tabs). Editor context is purely optional — headless evaluation is completely unaffected.
+- **Deterministic eval validator:** New `scripts/validate-eval-paths.js` validates every expected/avoid/group path resolves to a real fixture file. Mirrors runtime `pathMatches()` semantics (exact, suffix, substring). `--fix` mode prunes dead paths and drops unmeasurable cases.
+- **MCP prompt discovery:** Standalone MCP server now advertises `prompts` capability and supports `prompts/list` + `prompts/get` for `retrieve`, `guide`, and `files`. These prompts map to the stable tools `flowseeker_retrieve`, `flowseeker_guide`, and `flowseeker_files` so Native Chat hosts can show short suggestions when their UI supports MCP prompt/tool autocomplete.
+- **MCP metadata polish:** Tightened tool descriptions and input schema wording for truncated IDE suggestion UIs while keeping the stable tool names unchanged.
+- **Cross-IDE config correctness:** Fixed MCP help/docs to use `mcpServers` for Cursor, Antigravity, Windsurf, Claude Desktop, and `servers` only for VS Code/Copilot `.vscode/mcp.json`.
+- **Antigravity Native Chat note:** Documented current Antigravity behavior: selecting FlowSeeker may insert `@mcp:flowseeker:` without showing second-level command suggestions. FlowSeeker still works through natural language or direct tool names; autocomplete visibility is host-controlled.
+- **npm MCP readiness:** Verified `flowseeker-mcp` builds from `dist/cli/mcpServer.js`, responds to `tools/list` and `prompts/list`, and is included by `npm pack --dry-run` for users who want MCP-only integration without installing the extension.
+
+### 0.1.6 Files changed (vs 0.1.5 baseline)
+`src/pipeline/evaluationMetrics.ts`, `src/cli/runEvaluation.ts`, `scripts/validate-eval-paths.js` (new), `scripts/render-icon.js` (new), `TODO-0.1.6.md` (new), `src/pipeline/taskUnderstanding.ts`, `src/pipeline/ranker.ts`, `src/pipeline/runPipeline.ts`, `src/cli/mcpServer.ts`, `src/mcp/mcpTools.ts`, `notes/Req5.txt`, `notes/mcp-config-paths.md`, `.gitignore`, `.vscodeignore`, `.npmignore` (new), `README.md`, `ARCHITECTURE.md`, `COMPATIBILITY.md`, `MARKETPLACE_DEMO.md`, `UPDATE_GUIDE.md`, `plan.md`
+
+## 0.1.5 - Agent UX Stability, Cross-Host Compatibility, and MCP Server
+
+- **MCP Server:** Added standalone MCP server (`flowseeker-mcp`) exposing FlowSeeker retrieval as MCP tools. Works with Cursor, Antigravity, Windsurf, Claude Desktop, and any MCP-compliant AI client.
+- **MCP tools:** `flowseeker_retrieve` (full Solve Packet), `flowseeker_guide` (agent prompt), `flowseeker_files` (lightweight file list).
+- **Cross-IDE notification:** FlowSeeker now detects Cursor, Antigravity, and Windsurf hosts at startup and shows a one-time guidance notification pointing users to the sidebar chat or MCP setup.
+- **Headless agent prompt:** Added `agentPromptHeadless.ts` — agent solve prompt generator without VS Code dependency, used by MCP server and CLI.
+- **Standalone CLI binary:** `flowseeker-mcp` bin for running the FlowSeeker MCP server outside of any IDE.
+- **Host detection:** FlowSeeker now identifies the editor host at startup (VS Code, Cursor, Antigravity, Windsurf, VSCodium, Theia) and logs the identity + capabilities to the Output channel.
+- **Runtime probing:** Beyond `typeof` checks, FlowSeeker now calls `vscode.lm.selectChatModels()` to count actual available models and vendors — a reliable signal of whether the host exposes usable AI models.
+- **Tier table:** Single source of truth for host capabilities — `native-full` (VS Code/Insiders), `native-unverified` (Cursor, Antigravity, Windsurf — API exists but not confirmed to render), `sidebar-primary` (VSCodium, Theia — no API).
+- **Shared command runner:** Core retrieve→ask flow extracted into `flowCommandRunner.ts` so both native chat and sidebar use the same logic.
+- **Enhanced compatibility check:** `FlowSeeker: Check Host Compatibility` now shows host identity, tier, live model probe results, and offers human-in-the-loop verification for unverified hosts.
+- **Status bar indicator:** Shows current tier mode (`FlowSeeker` for native, `FlowSeeker: Sidebar` for sidebar-only) with tooltip explaining why. Click opens compatibility report.
+- **JSON prompt hardening:** Edit generation instruction moved to END of prompt (after context pack), strengthened wording, added `response_format: { type: "json_object" }` for OpenAI-compatible providers, improved JSON extraction from prose responses.
+- **Per-file apply status:** Edits applied file-by-file with individual success/failure reporting including file name and reason.
+- **Graded provider diagnosis:** `testProvider` classifies errors (401/403 credentials, 429 quota vs rate-limit, 5xx server, network, missing CLI) with actionable fix steps.
+- **MCP server visibility:** Agent policy tooltip shows each MCP server name, type, and status (configured/disabled/misconfigured).
+- **Feedback hardening:** Save confirmation includes file path and replay hint. CLI replay shows original top files for drift detection.
+- Added `COMPATIBILITY.md` with detailed host compatibility matrix and model sharing explanation.
+
+### 0.1.5 Files changed
+`src/runtime/hostDetect.ts` (new), `src/runtime/hostTier.ts` (new), `src/runtime/statusBar.ts` (new), `src/agent/flowCommandRunner.ts` (new), `src/mcp/mcpTools.ts` (new), `src/pipeline/agentPromptHeadless.ts` (new), `src/cli/mcpServer.ts` (new), `src/commands/checkHostCompatibility.ts`, `src/extension.ts`, `src/ui/chatViewProvider.ts`, `src/gateway/aiGateway.ts`, `src/gateway/aiProviders.ts`, `src/gateway/agentPrompts.ts`, `src/gateway/editProposalStore.ts`, `src/cli/runReplay.ts`, `src/pipeline/tokenSavings.ts`, `package.json`, `COMPATIBILITY.md` (new)
+
+## 0.1.4 - Token Proof Panel and Feedback Hardening
+
+- Added sidebar proof panel showing token proof status (complete/partial/unavailable) with differentiated visual states per retrieval card.
+- Added context coverage display in the proof panel showing found/missing required slots.
+- Added proof warnings in the metric card for partial scans, capped evidence, and missing slots.
+- Added one-click "Copy token proof" action that copies a formatted token savings summary including proof status, coverage, and warnings.
+- Feedback save confirmation now includes the saved file path and a replay hint.
+- CLI replay reports now show original top files from feedback snapshots alongside current top files for drift visibility.
+
+## 0.1.3 - Token Savings Proof
+
+- Added measured token savings to FlowSeeker results, comparing Solve Packet tokens against indexed workspace text tokens.
+- Uses local BPE token counting with `cl100k_base` for headless, sidebar, and report metrics, with heuristic fallback only when tokenization fails.
+- Counts the actual prompt through VS Code `LanguageModelChat.countTokens()` when the selected native chat model exposes token counting.
+- Added generic Context Blueprint coverage so FlowSeeker checks required flow slots for each task shape instead of ranking only by keyword score.
+- Updated Solve Packet generation to preserve required context slots, report missing links, and mark likely noise.
+- Improved plan approval safety: responses that only ask to read more files are no longer treated as approvable plans.
+- Added bounded context expansion and retry when the AI explains that the first Solve Packet is insufficient.
+- Added structured-edit handling with diff preview, scope lock, apply actions, and verification command proposals.
+- Added sidebar chat history, provider profiles, model switching, provider model fetch, and compact quota/usage status.
+- Added agent controls for workspace edit approval, safe command proposals, MCP/tool proposals, and user-defined prompt rules.
+- Added host compatibility reporting for Native Chat, Language Model, and MCP support across VS Code-compatible editors.
+- Added sidebar feedback capture and real-world replay support so retrieval misses can become regression cases.
+- Normalized feedback kinds from the sidebar and legacy records so replay metrics classify missing context consistently.
+- Shows compact token reduction in native VS Code Chat after retrieval.
+- Added a sidebar chat fallback for VS Code-compatible editors that do not expose the native VS Code Chat participant API.
+- Added token savings and context coverage metrics to headless and VS Code evaluation reports alongside hit@K, MRR, and scan duration.
+- Keeps token reduction wording technical, with retrieval quality measured separately.
+
+## 0.1.2 - Scan UX and Logo Polish
+
+- Refined the FlowSeeker logo assets for extension icon, activity bar, and native chat usage.
+- Kept chat scan progress compact by updating status lines instead of flooding the conversation.
+- Improved AI plan language behavior so generated plans follow the user's request language when set to auto.
+- Tightened package excludes so Marketplace builds do not ship local caches, source maps, or reference raster assets.
+
+## 0.1.1 - Large Repo Stabilization
+
+- Added persistent workspace indexing and dependency expansion to reduce full-codebase scans.
+- Added scan budgets for max files, bytes, duration, raw evidence, and indexed candidates.
+- Improved Solve Packet output with clearer trace, context buckets, uncertainty notes, and agent guidance.
+- Added headless evaluation support for repeatable retrieval tests across sample workspaces.
+
+## 0.1.0 - Initial Release
+
+- **Core retrieval pipeline**: task understanding, workspace indexing, universal text scan, evidence ranking, evidence graph with framework edges
+- **Solve Packet v1**: structured reasoning context with flow buckets (entry/handler/service/schema/tests), problem hypothesis, noise risk detection, missing link identification, agent instructions
+- **Native VS Code Chat participant**: `@flowseeker /guide`, `/auto`, `/apply` commands with plan-first editing guardrail
+- **Framework adapters**: Laravel (PHP), NestJS (TypeScript), Express (JavaScript/TypeScript) - route->controller, event->listener, controller->service, view rendering, DI injection edges
+- **Structured extraction**: regex-based code structure extraction for 11 languages (TS, JS, Python, PHP, Java, Go, C#, Ruby, Rust, Kotlin) - declarations, calls, string literals, comments
+- **Scope lock**: edits are validated against approved plan scope before applying
+- **Diff preview**: file-level change preview before applying structured edits
+- **Verification loop**: auto-detection of test/lint/typecheck commands after edits, with pass/fail reporting
+- **Headless evaluation**: repeatable retrieval benchmarking with hit@K, MRR, expected-path coverage
+- **VS Code Language Model API integration**: works with the selected VS Code Chat model; AI Gateway fallback
+- **Result sidebar**: ranked evidence tree with checkbox selection, context pack copy, agent prompt copy
+- **Configuration**: `.flowseeker/config.json` workspace config + VS Code settings overrides
+- **Scan budget safety**: max files, max bytes, max time limits prevent runaway scans on large repos