Fix high prio recommendations, add error / health screen

README.md

@@ -97,6 +97,7 @@ within_status = "file"          # "file" or "timestamp" (oldest first)
 enabled = true                 # instant stale view plus offline fallback
 max_age = "168h"               # 7 days; 0 means no age limit
 directory = ""                 # defaults to the OS user cache directory
+max_entries = 200              # bounded oldest-first pruning; 10-10000
 
 [editing]
 mode = "vim"                   # "vim" or "standard"; description field only for now
@@ -123,6 +124,7 @@ page_up = ["ctrl+u", "pgup"]
 [keybindings.views]
 open = ["enter", "l"]
 dashboard = ["d"]
+health = ["H"]
 edit = ["e"]
 toggle_list = ["tab"]
 
@@ -196,14 +198,25 @@ can be active together reports the context and both conflicting actions.
 When cached data exists, the picker and PR details are rendered immediately
 from that snapshot while a live GitHub refresh runs in the background. Cached
 screens are labelled with their save time and are replaced automatically when
-fresh data arrives. Check annotations are fetched separately only for failed
-checks so they do not inflate the initial PR query.
+fresh data arrives. Core PR and review data is rendered before check
+annotations and conflict-file analysis finish. A failed subsection keeps its
+last complete value, is marked partial, and does not discard the rest of a
+successful refresh. Check annotations are fetched separately only for failed
+checks and are reused by immutable check ID.
 
 The cache uses separate JSON files for the picker and each visited PR. Cache
 content is hashed before writing: unchanged responses do not rewrite their
 files. Their modification time is touched at most once per day (or half the
 configured maximum age, when shorter) so recently validated snapshots remain
-usable without writing on every poll. Changed files are replaced atomically.
+usable without writing on every poll. Changed files are replaced atomically,
+and oldest cache entries are pruned at the configured bound. Read state and
+recoverable reply/metadata drafts use versioned, atomic files beside the
+configuration.
+
+Polling adapts to GitHub's reported rate-limit budget. It backs off as the
+remaining budget gets low, honors server retry windows, and adds jitter to
+avoid synchronized clients. Opening another PR or starting another refresh
+cancels the superseded request.
 
 GitHub's public APIs report whether a PR conflicts but do not expose its
 conflicting file paths. For conflicting PRs only, `gh-threads` performs a
@@ -240,6 +253,7 @@ history and metadata.
 | `h` / `l` | Focus the thread list / thread detail |
 | `j` / `k` | Move between items or scroll the dashboard/focused detail |
 | `?` | Show contextual keybinding help |
+| `H` | Open application health and diagnostics |
 | `d` | Open the current pull request dashboard |
 | `e` | Edit the current PR title, target branch, and description from its dashboard |
 | `/` | Fuzzy-search paths and filter with `status:`, `author:`, `updated:true` |
@@ -258,6 +272,15 @@ history and metadata.
 | `r` | Refresh now |
 | `q` | Quit |
 
+The Health modal reports the interactive loop, configuration, GitHub API,
+rate-limit budget and reset/retry time, disk cache, unread-state persistence,
+draft recovery, core PR data, and secondary enrichment. Session warnings and
+errors are retained there with their component and timestamp. Long diagnostics
+wrap to the modal width. Refresh activity occupies a stable informational row
+so polling does not reorder the report. Press `H` from the picker, dashboard,
+or thread view; `b` or `esc` closes it without changing the underlying scroll
+position.
+
 The reply composer appears inline beneath the selected thread so its code and
 comments remain visible while writing. It supports multiple lines: `enter`
 inserts a newline, `ctrl-s` opens the rendered confirmation preview, and `esc`