bb-cc-lite 0.1.0 → 0.1.2

package/README.md

@@ -16,6 +16,12 @@ npx bb-cc-lite install --scope local
 
 Restart Claude Code in the project. The line appears at the bottom.
 
+Install replaces any existing Claude Code `statusLine` by default and stores a backup so `uninstall` can restore it. It also builds a local personal baseline by default after it installs the statusline. It reads bounded local Claude Code JSONL once, newest eligible files first, extracts aggregate patterns from past sessions, and stores a small `baseline.json` under the `bb-cc-lite` app home. Use `--no-learn` to install the statusline without scanning old JSONL:
+
+```bash
+npx bb-cc-lite install --scope local --no-learn
+```
+
 For faster tool-loop detection:
 
 ```bash
@@ -28,8 +34,12 @@ Hooks are optional. They run in the background and skip `UserPromptSubmit`.
 
 ```text
 bb: Healthy | ctx 42% | $0.18 | cache warm | continue normally
-bb: Careful | ctx 82% | cache writes high | ask Claude for a 6-bullet handoff before more work
-bb: Stop | Bash failed 3x running tests | fix the test setup manually, then ask Claude to rerun only that test
+bb: Healthy | research phase: usually normal for you | continue
+bb: Careful | edits not checked yet | run focused check
+bb: Careful | tests failed twice; usually recovers after one fix | inspect first failure
+bb: Careful | edit lag unusual for you | run focused check
+bb: Careful | ctx 82% | Context is getting tight | ask Claude for a 6-bullet handoff before more work
+bb: Stop | why: test loop: usually unrecovered after 3x | do: inspect first failure
 ```
 
 `Healthy` means keep going. `Careful` means slow down. `Stop` means take over before Claude burns more turns.
@@ -43,20 +53,49 @@ bb-cc-lite why
 ```text
 Last decision: Stop.
 Reason: Bash failed 3x running tests. Claude is retrying a broken test loop.
-Next action: fix the test setup manually, then ask Claude to rerun only that test.
+Next action: inspect first failure.
+```
+
+By default, `why` explains the latest recorded decision. To inspect a specific Claude Code session, pass its session id:
+
+```bash
+bb-cc-lite why --session <session-id>
+```
+
+`why` may mention when the personal baseline influenced wording or priority, including safe recovery and edit-lag explanations recorded by the statusline. It reads only the derived local event store; it does not reopen old transcripts to expose raw content.
+
+## Personal Baseline
+
+`bb-cc-lite` learns weak Healthy-like, Careful-like, and Stop-like outcome patterns from your own past Claude Code sessions. It scans newest JSONL candidates first, applies the file and byte caps before summarizing, and reads transcript tails with bounded parallelism. Those weak labels can affect footer wording, priority, and confidence, but hard global `Stop` rules still win.
+
+Refresh, inspect, or clear the baseline:
+
+```bash
+bb-cc-lite doctor --build-baseline
+bb-cc-lite doctor --baseline
+bb-cc-lite doctor --clear-baseline
+bb-cc-lite unlearn
 ```
 
+`doctor --baseline` prints a safe aggregate summary, including recent window, validation category, and safe tool category names when present. `doctor --clear-baseline` and `unlearn` remove only `baseline.json`.
+
 ## Privacy
 
-By default, `bb-cc-lite` does not upload transcripts, store raw prompts, store raw tool output, store file contents, or print raw prompt/tool text.
+By default, `bb-cc-lite` does not upload transcripts, store raw prompts, store raw tool output, store raw shell commands, store file contents, store raw paths, or print raw prompt/tool text.
+
+For learning, it reads bounded local Claude Code JSONL history and stores only derived aggregate data: counts, rates, scenario counts, weak outcome-label counts, confidence buckets, recovery rates, validation categories, edit-to-validation lag buckets, safe tool category names, scan timestamps, and privacy flags. It uses file mtimes and raw shell commands only transiently for sorting and categorization. It never stores prompts, assistant text, commands, command arguments, tool output, file contents, transcript paths, scanned-file lists, per-file mtimes, workspace paths, raw session ids, API keys, or per-session transcript rows.
+
+For live rendering, statusline reads Claude Code status JSON from stdin, tails only the current transcript defensively, and may read the small `baseline.json`. It stores derived metadata only: counts, reason codes, token totals, costs, and hashed session ids.
 
-It reads Claude Code status JSON from stdin, tails the local transcript defensively, and stores derived metadata only: counts, reason codes, token totals, costs, and hashed session ids.
+LiteLLM is used only as public pricing data for cost estimates. `bb-cc-lite` does not run a LiteLLM proxy, gateway, or message router.
 
 ## Undo
 
 ```bash
 bb-cc-lite doctor
+bb-cc-lite doctor --clear-baseline
+bb-cc-lite unlearn
 bb-cc-lite uninstall --scope local
 ```
 
-Install preserves an existing Claude Code `statusLine` unless you pass `--replace`.
+Install backs up the previous Claude Code `statusLine` before replacing it.

package/dist/baseline-builder.d.ts

@@ -0,0 +1,13 @@
+import { type PersonalBaseline } from "./baseline.js";
+export interface BuildBaselineOptions {
+    homeDir?: string;
+    appHomePath?: string;
+    claudeProjectsDir?: string;
+    maxFiles?: number;
+    maxBytesPerTranscript?: number;
+    now?: Date;
+}
+export declare function buildBaseline(options?: BuildBaselineOptions): Promise<{
+    baseline: PersonalBaseline;
+    written: boolean;
+}>;