docs(changelog): stamp #7654 Codex commentary MEDIA auth + snapshots

Co-authored-by: happy5318 <happy5318@users.noreply.github.com>

CHANGELOG.md

@@ -5,6 +5,8 @@
 
 ### Fixed
 
+- **Images and files produced during Codex commentary are viewable again.** The Agent's OpenAI Codex Responses adapter persists user-visible assistant progress in `codex_message_items` with `phase: "commentary"` while the outer `content` field stays empty. WebUI's session MEDIA authorization and its snapshot capture both read only that outer field, so a `MEDIA:` token emitted during commentary was invisible to both — the artifact was never authorized for the session token and never snapshotted, leaving the user unable to open a file the assistant had just produced. Both consumers now inspect the commentary sidecars as well. The widening is confined to genuinely assistant-authored, correctly-phased items: role, type, phase and content checks all fail closed, so a user-authored message, a tool result or a malformed item cannot grant a path token. Every existing guard is unchanged — exact-path matching, MIME allow-listing, session ownership, profile visibility and the state/secret hard-deny all still run on anything discovered through the new route, and snapshot capture keeps its confinement without double-capturing an artifact already taken from outer content. Verified against the Agent's real producer shape plus an adversarial probe covering user/wrong-role/wrong-phase, denied-state, wrong-path, duplicate, malformed and 300k-item inputs. Thanks @happy5318. (#7654, #7565)
+
 - **Custom model names containing colons are no longer truncated in the picker.** A model id like `ollamacloud/qwen3.5:397b` was cut at its internal colon, so the picker showed a mangled label and two models that differed only after the colon could render identically. The colon is overloaded here — it separates a `@provider:model` routing lane, and it also appears inside perfectly ordinary model names — so the label builder now takes the catalog as the authority for `@custom` entries and leaves a plain-lane model id whole. Labels are display-only: the original option value survives selection and is dispatched unchanged, so no routing behaviour moves. Catalogued, cold, empty, stale, host:port, leading- and trailing-colon and multi-colon names were each checked to stay non-blank and exception-free. Thanks @webtecnica. (#7401, #7240)
 
 - **Context-compaction cards stay visible in long transcripts.** When a transcript grew past the virtualization window, the compaction markers explaining where context was compressed scrolled out of the rendered range and vanished, leaving no indication that compaction had occurred. Pre-window markers are now preserved and placed deterministically, and card placement is stable across repeated renders and window changes — verified to neither lose nor duplicate a card as the virtual window moves. Hardened during review: the settled current-summary card was created only when *no* compaction marker was loaded at all, so a transcript carrying older markers plus a newer summary that none of them referenced silently dropped the summary the session was actually operating under — marker presence is not proof that any marker matches. The fallback now keys on whether a loaded marker actually references the current summary, and participates in the single preserved-task-owner selection so the repair cannot duplicate the task card. Mid-compression, empty-summary, unmatched-anchor-key, first-compaction and matched-marker cases were each checked, and reverting the guard reproduces the original suppression. Thanks @ruizanthony. (#7124)