diff --git a/CHANGELOG.md b/CHANGELOG.md
index a2c99673..a922a616 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -13,6 +13,12 @@ upstream release: 1.0.8). Format follows [Keep a Changelog](https://keepachangel
 - **Quota failure throttling** — Back off repeated quota authentication failures
   for up to one hour, while retrying immediately after Claude credentials
   change.
+- **Usage dashboard recovery (#79, fixes #82)** — one oversized non-transcript
+  `.jsonl` can no longer abort the earliest-timestamp probe and blank the whole
+  dashboard. Thanks [@ptweezy](https://github.com/ptweezy).
+- **Opus 5 context window (#81, reported in #84)** — recognise the bare
+  `claude-opus-5` model id as a 1M-context model and remove its spurious
+  unknown-model pricing diagnostic. Thanks [@e7d](https://github.com/e7d).
 
 ## [2.2.1] — 2026-07-18
 
@@ -35,23 +41,6 @@ upstream release: 1.0.8). Format follows [Keep a Changelog](https://keepachangel
   are credited as development tools, separately from human contributors.
 
 ### Fixed
-- **Usage dashboard no longer blanks on one oversized non-transcript file** — a
-  single `.jsonl` with more than 1 MiB of leading lines that carry no
-  `timestamp` (such as a large workflow `journal.jsonl`, whose records hold
-  `type`/`key`/`result`/`agentId` and never a timestamp) made the
-  earliest-timestamp probe throw `TimestampProbeLimitError`. That rejection was
-  unhandled, so it aborted the whole load and left the dashboard empty on every
-  refresh (`failed=1`, `bytes=0` in diagnostics). The probe now fails soft for
-  that one file and the rest load normally.
-- **Opus 5 context window** — `claude-opus-5` reported a 200 000-token window
-  as if it were exact, so the status bar and dashboard showed a context
-  percentage five times too high. Claude Code writes the plain `claude-opus-5`
-  id to its JSONL even for the 1M-context variant, so the existing `[1m]`
-  marker never covered it. The major-version check now recognises Opus 5+
-  alongside Sonnet 5+, and `claude-opus-5` is an explicit entry in the pricing
-  table (same $5 / $25 current-Opus tier the family fallback already resolved,
-  minus the spurious `Unknown model` diagnostic).
-
 - **High-CPU refresh mitigation (#70)** — polling now always honors the
   configured 30–3600 second `refreshInterval`; file watching is quiet-debounce
   only and adds 60/120/300-second choices. First-timestamp reads stop after the