update QoL and ai integration

README.md

@@ -273,15 +273,24 @@ max_calls = 8
 max_request_bytes = 180000
 max_run_bytes = 900000
 max_file_bytes = 150000
+max_context_rounds = 2          # automatic file-request rounds for a thread
+max_context_files = 8           # additional files per thread discussion
 store_directory = ""
 exclude = [
   "*.lock", "go.sum", "package-lock.json", "vendor/", "node_modules/",
   "dist/", "build/", "generated/", "coverage/", "*.generated.*",
-  "*_generated.*", "*.min.js", "*.map", ".env", ".env.*", "*.pem",
-  "*.key", "*.p12", "*.pfx", "*credentials*",
+  "*_generated.*", "*.min.js", "*.map",
+]
+sensitive_paths = [
+  ".env", ".env.*", "*.pem", "*.key", "*.p12", "*.pfx",
+  "*credentials*",
 ]
 ```
 
+As with other TOML arrays, setting `exclude` or `sensitive_paths` replaces its
+default list. Copy the defaults you still want before adding project-specific
+patterns.
+
 `dashboard_mode = "hotkey"` opens threads directly from the picker and leaves
 the dashboard on `d`. `"intermediate"` places the dashboard between the picker
 and thread viewer.
@@ -298,6 +307,13 @@ Thread categories are:
 - `outdated`: unresolved threads attached to outdated code; and
 - `resolved`: all resolved threads, including resolved-and-outdated threads.
 
+New threads retain a `NEW THREAD` marker. When an existing thread receives new
+comments, diple places a `NEW MESSAGES` divider before the first unread comment
+and emphasizes the unread comment rail. Moving the thread-list cursor does not
+clear this state. It is cleared after the last unread comment becomes visible
+while scrolling the focused detail pane, or manually with
+`keybindings.threads.mark_read` (`m` by default).
+
 Within a category, `"file"` keeps paths together and `"timestamp"` sorts by
 the time the thread was opened.
 
@@ -394,6 +410,7 @@ search = ["/"]
 clear_filter = ["F"]
 next_unread = ["n"]
 previous_unread = ["N"]
+mark_read = ["m"]
 reply = ["c"]
 resolve = ["R"]
 toggle = ["enter"]
@@ -490,19 +507,31 @@ The `A` menu can:
 - review the current PR and create local-only review threads;
 - discuss an existing local AI thread with the same selected model;
 - add local-only context to existing unresolved GitHub threads;
+- let a focused thread discussion request bounded, exact-head repository files;
 - produce small GitHub-style suggestion blocks for contained changes;
 - refresh provider status without making an inference call; and
 - run one explicitly confirmed, minimal provider test that consumes quota but
   sends no PR contents.
 
-Before a review, diple shows the exact head commit, selected model, included and
-excluded files, byte count, maximum model-call count, and redaction count.
-Every run requires confirmation.
+Before a review, diple shows the exact head commit, selected model, initial
+included and excluded files, byte count, maximum model-call count, and
+redaction count. Every run requires confirmation. A focused thread confirmation
+also shows its repository-tree summary and the configured automatic
+file-request limits.
 
-AI input comes from the authenticated GitHub PR diff and PR metadata, not from
-the local checkout. diple excludes configured sensitive, generated, vendored,
-lock, binary, and oversized files; redacts secret-like values; chunks bounded
-requests; and validates findings against lines changed in the prepared head.
+Full reviews use the authenticated GitHub PR diff. Focused discussions instead
+send only the selected thread, its hunk, the complete target file when allowed,
+minimal PR identifiers, and a bounded tree for the exact PR head. The model can
+request additional paths from that tree, but diple validates and retrieves
+their committed blobs through GitHub; the provider never receives local
+checkout access.
+
+`sensitive_paths` are absent from the model-visible tree and can never be
+requested. `exclude` paths may appear as unavailable tree entries but their
+contents are not sent. Binary, submodule, oversized, generated, vendored, and
+lock-file content remains unavailable. All supplied content is bounded,
+control-sanitized, and checked for secret-like values. Full-review findings
+remain restricted to visibly changed lines in the prepared head.
 
 The Codex process runs ephemerally in an empty temporary directory with:
 
@@ -516,7 +545,9 @@ The Codex process runs ephemerally in an empty temporary directory with:
 
 Attempted tool or file-change events fail the run. Provider output is bounded
 and sanitized. Progress may display a provider-exposed reasoning summary, but
-diple neither requests nor displays hidden chain-of-thought.
+diple neither requests nor displays hidden chain-of-thought. Progress and
+Health report GitHub, filtering, and provider timing without storing prompts or
+repository contents.
 
 AI findings are stored locally, deduplicated deterministically, and marked
 outdated when the PR head changes. Resolving a local AI thread remains local.