@ai-agent-forge/agent-forge 0.86.1 → 0.86.3

package/CHANGELOG.md

@@ -2,6 +2,40 @@
 
 ## [Unreleased]
 
+## [0.86.3] - 2026-10-01
+### Added
+- The session picker loads progressively: partial session lists publish to the UI as they are discovered (sorted by activity), loading is cancellable through an `AbortSignal` threaded from the selector through `SessionManager.list`/`listAll`, and file stats are reused between discovery and header reads (upstream pi `dfbf793b7`).
+- HTML exports gained a hidden-message toggle (H key): messages hidden in the terminal (`display: false`, e.g. turn-interrupted markers) render with a marker by default and can be revealed; deep links to hidden messages expand them automatically (upstream pi `b2bd111f2`).
+- Successful RPC `prompt`, `steer`, and `follow_up` responses now include `data.disposition` (`"handled"` / `"queued"` / `"started"`), so clients can tell whether a run started and whether to wait for `agent_settled`; `AgentSession.steer()`/`followUp()` and `RpcClient.prompt()`/`steer()`/`followUp()` return the disposition, and `RpcClient.promptAndWait()` no longer waits for `agent_settled` when the disposition is `"handled"`. The `preflightResult` prompt hook now receives the disposition instead of a boolean and is not called for rejected prompts (upstream pi `e473b5cd8`).
+### Fixed
+- New session files are now created at the first user message instead of the first assistant response, so interrupting the first turn (double Ctrl+C, process kill) no longer loses the prompt and the session stays visible to resume/fork; setup-only sessions still leave no file behind (upstream pi `ff72faba2`).
+- Pinned temporary git plugin sources (`git:...@ref`) key their cache folder by host, path, and ref, so changing the pinned ref no longer keeps loading the previously downloaded commit; unpinned sources keep the existing folder (upstream pi `f444ea5ea`).
+- Signal-terminated shell commands now report the conventional `128 + signal` exit code instead of `null`, and custom shell operations returning a null exit code fail closed with partial output preserved, so killed commands are no longer reported to the model as successful (upstream pi `a8b3dd199`).
+- `RpcClient` event dispatch iterates a listener snapshot, so a listener unsubscribing during dispatch no longer makes later listeners miss that event (upstream pi `92e8d4f02`).
+- Compaction cancellation races are closed: summarization auth waits honor `AbortSignal` (and abort is never swallowed into an unauthenticated fallback), branch-summary auth takes the branch controller's signal, and a cancelled auto-compaction settles as `compaction_end aborted:true` instead of reporting "Auto-compaction failed" and dispatching the failed lifecycle (upstream pi `de2de549b`).
+- Split-turn summarization prompts no longer trigger reasoning-model refusals: the conversation payload and instructions are separated into `# Conversation` / `# Instructions` sections with continuation-oriented wording, replacing the `<conversation>` wrapper (upstream pi `d192bd6dc`).
+- Recent-session discovery sorts candidates by mtime and reads headers until the first cwd match instead of reading every header (upstream pi `dd01f5b24`).
+- Resuming by an exact session id (or validating a fork target) reads a single header through the new `SessionManager.findById` instead of scanning every transcript (upstream pi `9b791a4cc`).
+- Anthropic thinking-drop notices are shortened to `Anthropic dropped N thinking block(s) (details in session)` and shown only when the count grows over the previous assistant message, instead of repeating a full reason/path list after every turn and during session replay (upstream pi `13784598d`, `465853498`).
+- Stale tool-image conversions no longer render after a result update: the Kitty conversion cache validates the source image data and drops conversions whose source was replaced mid-flight (upstream pi `21b8cc1a4`).
+- Clipboard writes report real failures: a local session with a display fails closed instead of counting an unverifiable OSC 52 write as success, backend failures surface platform-specific setup hints, and the fullscreen copy flash shows the concrete error for 5 seconds (upstream pi `3349e1db1`, `60e7e76bd`).
+- Headless and WSL clipboard writes work again: display-less sessions keep the OSC 52 fallback, and WSL without WSLg routes through `powershell.exe Set-Clipboard` with a UTF-8 file handoff so non-ASCII text no longer corrupts through `clip.exe`'s code page (upstream pi `6dff740fa`).
+- Bash tool durations format as hours/minutes/seconds once a command runs past a minute (upstream pi `fe219d7f8`).
+- Image sniffing requires complete GIF signatures (`GIF87a`/`GIF89a`), so text files starting with `GIF` are no longer treated as images (upstream pi `47a18e37b`).
+- Invalid prompt-template frontmatter is reported as a warning diagnostic with the parse location instead of the template silently disappearing (upstream pi `b6419322e`).
+- Invalid `--mode` values are rejected with the list of valid modes instead of being silently ignored (upstream pi `e40126f57`).
+- The `opencode-go` and `together` default models now resolve to `kimi-k3` / `moonshotai/Kimi-K3` after Kimi K2.6 retired from the catalog, instead of pointing at a nonexistent model (upstream pi `e1787702d`).
+
+- Small context windows no longer spin the compaction loop: `keepRecentTokens` is now clamped per model window (`effectiveKeepRecentTokens`: half the summarizable window `contextWindow - reserveTokens`, floor 1024, result floor 1; unknown and large windows keep the configured value — 200K stays 20000, a 30K window resolves 20000 → 6808), applied uniformly to threshold decisions, manual and automatic preparations, and the overflow trim stage-0 target. Seat-measured on a 30K gateway window: the default 20000 budget protected the entire history, so threshold hits answered `compact: true` with a zero-length summarizable prefix until the real ceiling overflowed. Deterministic coverage in `test/suite/agent-session-compaction.test.ts`.
+- A recoverable length stop (truncated response) whose compact-and-retry recovery failed now enters the overflow trim ladder instead of dead-ending: turn-boundary projection trims free input room for the truncated output, and the failure prefix is split by cause (the existing `Context overflow persists...` wording is unchanged; truncated responses report `Truncated response persists after compaction and N trim attempts. `). Deterministic coverage in `test/suite/agent-session-compaction.test.ts`.
+### Changed
+- The pre-response threshold check (`compactBeforeNextAssistantResponse`, the default `prepareNextTurn` link between agent turns) now carries `branchEntries`/`keepRecentTokens`/`microcompactEnabled` in its decision request and lands `decision.relief` exactly like the post-run path (entry append + context reprojection + 3-response cooldown), so microcompact relief is no longer unreachable on that path. Deterministic coverage in `test/agent-session-microcompact.test.ts`.
+
+## [0.86.2] - 2026-10-01
+### Changed
+
+- Microcompact relief hardening (cooldown guard / setting / media protection): threshold decisions now carry `microcompactEnabled` (settings `compaction.microcompactEnabled`, default `true`, AND a per-model-request cooldown), so a landed relief starts a 3-assistant-response cooldown (one decrement per `message_end`, floor 0, not reset by user `message_start`) during which threshold hits go straight to full compaction — every microcompact is a mid-history rewrite and this gateway's prefix cache keeps only 40-50% on those, so the cooldown caps the worst-case repeat-trigger churn on the threshold band. Tool results whose content carries non-text parts (images) are never elided (originals stay until full compaction). Deterministic coverage in `test/agent-session-microcompact.test.ts`.
+
 ## [0.86.1] - 2026-09-30
 ### Added