- Added prompt_cache_hit_tokens / prompt_cache_miss_tokens parsing in
parseChunkUsage for DeepSeek's prompt cache format where
prompt_tokens = hit_tokens + miss_tokens.
- DeepSeek formula: input = prompt_tokens - hit_tokens (= miss, billed input),
total = input + output + hit_tokens (avoid double-counting miss in cacheWrite).
- Added cache_hit status line segment showing cache hit rate:
rate = cacheRead / (cacheRead + cacheWrite) x 100%.
- Added cache_hit to all status line presets.
Ghostty/cmux continued to show trails after hiding the hardware cursor because OMP switched to a blinking software cursor glyph. Keep editor cursor markers enabled when hardware cursor display is requested, but hide the actual hardware cursor on Ghostty; this gives the TUI a cursor position without emitting any cursor glyph. Also preserve the history-anchor cursor exit before public insertText, as flagged in PR review.
Constraint: Ghostty leaves visible trails for both hardware bar cursors and blinking software cursor cells during rapid input-row repaints.
Rejected: Only removing SGR blink | PTY writes still emitted the cursor glyph, so glyph afterimages could still accumulate.
Rejected: Ignoring the PR comment | public insertText callers bypass typed-character history-exit behavior.
Confidence: high
Scope-risk: moderate
Directive: Keep cursor-marker mode separate from actual hardware cursor visibility; callers that decide editor rendering must not use getShowHardwareCursor as a proxy.
Tested: bun test packages/tui/test/editor.test.ts packages/tui/test/render-regressions.test.ts; bun test packages/tui/test/*.test.ts; bun run check; Ghostty-env PTY capture of installed omp while typing abcdef emitted zero cursor-show writes, zero blink SGR, and zero input cursor glyph bytes before shell restore.
Four refinements to the sticky Todos panel on top of the live SessionObserverRegistry linkage:
- Cube animates whenever any visible open todo is "live" (in_progress, or a still-pending todo with a matching in-flight subagent). The previous subagent-only gate left lone in_progress rows on the static '⟳' fallback; ticking on an orphan in_progress row is the correct "still open" signal.
- 'normalizeForTodoMatch' now collapses any non-alphanumeric run to one space, so subagent descriptions with '#', '.', ':' etc. match todo content that omits the punctuation. Fixes the case where 3 subagents were spawned but only 2 of 3 matching todos lit up because the matcher's normalizer collapsed whitespace but left '#' intact.
- New '#reconcileTodosWithSubagents' runs on every observer-registry change and auto-checkmarks any pending/in_progress todo whose content matches a 'status === "completed"' subagent description. Failed/aborted subagents intentionally don't auto-flip - those stay open for the user (or next agent turn) to decide.
- All-done close animation: when every visible task is closed, fold the panel away over ~1.4s. A 900ms celebratory frame holds the bright bold "Todos ✓" header so the user can read the final checkmarks, then a fade through 'muted' / 'dim' with rows progressively dropped from the bottom. '#todoClosingState' state machine plays the animation exactly once per open->all-closed transition and aborts cleanly if a new open task arrives mid-animation.
Verification:
bun test test/tools/todo-write.test.ts → 24 pass / 0 fail (one new case for # punctuation tolerance)
bun run check → biome + tsgo clean
The always-on Todos panel above the editor pinned to the first 5 tasks of the active phase, so each todo_write flip mutated at most one row (color + strikethrough) and the +N more hint only shrank at end-of-phase. Marking task 1 done left tasks 6,7,... invisible until tasks 1-5 were all closed.
Introduce selectStickyTodoWindow(tasks, maxVisible=5) — returns up to 5 open (pending / in_progress) tasks in original phase order plus the count of remaining open tasks for +N more. When every task is closed, falls back to the trailing window (with +N more suppressed) so the panel keeps useful context until getActivePhase walks to the next phase. The collapsed branch of #renderTodoList now uses it; the expanded branch is untouched.
- Added a new ultrathink mode module with standalone, case-insensitive detection, rainbow editor highlighting, and a hidden notice payload.
- Updated `CustomEditor` and the shared TUI `Editor` to support optional zero-width text decoration and apply the `ultrathink` styling during input rendering.
- Extended `AgentSession` prompt handling to append the hidden ultrathink notice after user turns in both streaming and non-streaming message flows, excluding synthetic messages.
- Handled "incomplete" stop reasons in session recovery and auto-compaction workflows.
- Dropped the prior assistant turn before attempting recovery on incomplete-length stops.
- Expanded auto-compaction reason types and triggers to include "incomplete".
- Updated internal URLs parsing internals, export order, tests, and Obsidian URI prompt docs.
- Replaced the `keepaliveWhile` Promise wrapper with a new `EventLoopKeepalive` class that registers and disposes an interval timer through `Symbol.dispose`.
- Updated `Agent` to instantiate `EventLoopKeepalive` via `using` during prompt execution instead of manually managing an interval.
- Wrapped interactive mode's await path with the new helper and removed redundant `keepaliveWhile` usage from the CLI entrypoint.
- Redesigned hashline patch syntax from anchor-based (`A-B:`) to hunk-header format (`@@ A..B @@`) with unified-diff compatibility.
- Removed `autoDropPureInsertDuplicates` option and simplified apply behavior to preserve duplicated boundary and context lines.
- Changed repeat operator from `^A-B` to `&A..B` and range separator from `-` to `..` for consistency with hunk-header syntax.
- Added image resizing and dimension notes to eval tool output; improved write tool hashline header sanitation for legacy formats.
- Removed 521 lines of boundary-duplicate absorption code and simplified parser to auto-convert bare body rows and unified-diff contamination.
- Replaced 4-hex content-derived file hashes with 3-hex opaque tags minted by InMemorySnapshotStore, making tags session-bound pointers rather than content fingerprints.
- Removed lru-cache dependency; replaced LRU-bounded per-path rings with a flat 4096-slot global ring using a scrambled permutation to prevent LLM tag extrapolation.
- Made SnapshotStore required in Patcher (was optional); tag resolution now drives stale-anchor detection instead of recomputing hashes at apply time.
- Changed literal payload sigil from `|` to `+` and accepted `^A` shorthand for `^A-A`; added lenient recovery for bare bodies, lone `-` rows, and overlapping bare/concrete block pairs.
Patch axis: extend
Displacement: net-zero; reuses existing plan reference state instead of adding persistence or overwriting approved artifacts
Rule violations averted: no approved-plan overwrite, no transcript format migration, no public CLI/API expansion
PASS/FAIL: PASS after plan-mode focused tests and package check. Note: system-prompt-templates has an unrelated HOME=/tmp path-shortening expectation failure.
- Added `extraBody` merging into OpenAI Responses request params.
- Fixed xAI OAuth redirect URI to fail fast on port conflicts.
- Exposed `antigravity` and `xai` as explicit `providers.image` options.
- Added `isImageProviderPreference` guard, replacing inline string checks.
- Fixed TTS tool to resolve output path relative to cwd and require write approval.
- Imported isPromise from node:util/types in four modules.
- Replaced four instanceof Promise checks with isPromise calls for more reliable promise detection.
- Added optional skipRoleRank controls to model and canonical-model sort helpers to make role ranking conditional.
- Re-sorted fuzzy-filtered model and canonical model results with role rank skipped so stronger query matches are not outranked by weak default-role matches.
- Removed post-filter re-sorting for canonical model matches so fuzzyFilter's relevance order is preserved.
- Stopped reapplying role, MRU, and version sorting after fuzzy filtering canonical and non-canonical models.
- Documented the stable-sort rationale so fuzzy text relevance remains the primary ordering.
- Clamped `formatDuration` to return `0ms` for non-positive, NaN, or infinite inputs.
- Updated usage report rendering to suppress reset countdowns when `resetsAt` is absent or no longer in the future.
- Added unit tests for `formatDuration` covering clamped values and standard duration formatting.
- Extract resource_metadata URL from WWW-Authenticate and follow RFC 9728 chain
- Add buildWellKnownUrls with path-prefixed well-known fallback for gateways
- Fix resolveRegistrationEndpoint to try path-prefixed well-known (was missing await)
- Support relative Mcp-Auth-Server URL resolution against server URL
- Pass resourceMetadataUrl through all discoverOAuthEndpoints call sites
- Add comprehensive tests for path-prefixed, resource_metadata, and relative URL flows
- Replaced external watchdog timers with per-request SDK timeouts for first-event budget across OpenAI, Anthropic, and Azure providers.
- Keyed Python shared kernels by (sessionId, cwd) to prevent cross-directory state bleed.
- Deduplicated concurrent cold-start session acquisition for JS and Python executors.
- Moved `isOpenAIResponsesProgressEvent` to shared module and scoped display output routing per run for interleaved async cells.
- Removed deprecated MCP-specific type aliases and functions from tool-discovery module, consolidating to unified generic tool discovery API.
- Migrated session and SDK code to use generic filterBySource() and collectDiscoverableTools() instead of MCP-specific variants.
- Removed deprecated interface members including hasQueuedMessages(), FocusPane, AcpBuiltinCommandRuntime, and legacy settings methods.
- Updated test suites to use renamed generic discovery methods and removed back-compat test coverage for legacy MCP shapes.
Avoid rendering the pre-reset transcript during /new and force the final new-session frame through the clear-scrollback path so the cursor and prompt land with the replacement transcript.
Fixes#1295
Add an explicit clear-scrollback render option for intentional transcript replacement and use it when switching, branching, reloading, or creating sessions. Regular active redraws continue to preserve terminal scrollback.
Fixes#1295
- Added Symbol-keyed sidecar on each AgentMessage to memoize estimateTokens, with a cheap content fingerprint to detect in-place mutations.
- Fixed stale cache on same-length replaceMessages, post-hoc error attachment, and branch rebuild edge cases.
- Fixed usage fetch error backoff: stamped fetchedAt on failure so the 5-min TTL also gates retries during outages.
- Extracted computeNonMessageBreakdown as shared helper to prevent drift between status-line and context panel token counts.
Separated model selector provider tab labels from provider ids so human-readable labels like Ollama Cloud refresh and filter the underlying ollama-cloud models.
Fixes#1153
- Removed the exported formatBashFixupNotice helper from bash command fixup utilities.
- Removed BashTool's one-time bash-fixup notice tracking and stopped emitting those notices when fixups were applied.
Root cause (verified on user's environment):
- User commit `296641213` swapped status-line's context% computation from cheap `calculatePromptTokens(lastAssistantMessage.usage)` to `computeContextBreakdown(session)`, which walks EVERY message and runs native `countTokens` (~0.5 ms per message).
- The 2-second TTL cache helps for steady-state idle but every cache MISS is a full sweep.
- `updateEditorTopBorder()` is invoked on EVERY agent event (event-controller.ts:163 — `agent_start`, `delta`, `agent_end`, `tool_*`). Each delta during streaming can trigger a cache miss.
- User session has 2,312 messages → each full sweep is ~1,120 ms blocking.
- During streaming the UI freezes for ~1.1 s every ~2 s, producing the user-visible 'jittery rendering' ("버벅거림") and 'status bar disappearing' symptoms.
Fix:
`StatusLineComponent.getCachedContextBreakdown()` (renamed from `#getCachedContextBreakdown` so unit tests can exercise it directly) now uses an incremental per-message token cache that exploits the append-only nature of `session.messages`:
1. Message tokens (the dominant cost): cached per-index. New messages are tokenized as they arrive; previously-cached messages are reused. The LAST message is always recomputed because its content may still be growing during streaming. Compaction (messages.length shrinks) resets the cache.
2. Non-message tokens (system prompt + tools + skills): cached separately, invalidated only when a cheap inputs-identity fingerprint changes (model swap, skill toggle, tool registration). These rarely change during a session.
Required exposing three helpers from `modes/utils/context-usage.ts` (`estimateSkillsTokens`, `estimateToolSchemaTokens`, `computeNonMessageTokens`) so the status-line cache can call them directly.
Performance (2,300-message synthetic session, measured on user's M-series Mac):
- COLD warm-up call: ~75 ms (one-time, runs at OMP startup before any streaming)
- WARM refresh, no new message: ~0.04 ms (20 calls = 0.7 ms total)
- WARM refresh, 1 new message: ~0.02 ms
vs. prior implementation:
- Per cache-miss call: ~1,120 ms blocking
- 28,000× speedup on warm-state refresh
`computeContextBreakdown` itself is untouched — `/context` slash command continues to use it, and its output matches the status-line context% for the same session state (parity preserved).
Tests: 6 new cases in `packages/coding-agent/test/status-line-context-cache.test.ts` covering cold/warm/append/compaction/non-message-invalidation/zero-messages and a perf smoke test asserting 20 warm refreshes on a 200-message session complete in <100 ms.
Full suite: 3,199 tests, 26 pre-existing failures (status-line accent / log_experiment timing-flaky / skills / github tool / workspace-tree / tool path — all unrelated and baseline-confirmed). Lint: 1 pre-existing import-order issue in `event-controller-plan-ready.test.ts` unchanged.
Status-line's context_pct segment was computing tokens via
calculatePromptTokens(lastAssistantMessage.usage), which sums input +
cacheRead + cacheWrite from the Anthropic API usage object. The /context
slash command is computed by computeContextBreakdown, an offline estimate
over the live session state (systemPrompt + tools + skills + messages).
Both numbers are correct under their own definition, but they can
diverge by 2x+ on the same session when a turn rotates cache tiers
(e.g. 5m → 1h ephemeral re-cache) and cache_creation_input_tokens spikes.
Users read the two surfaces as one consistent dashboard and treat the
mismatch as a bug.
Repro: same session at the same moment reports 212K (21.2%) in /context
and 44.2%/1M in the status line — ~230K gap driven by per-turn
cache_creation on a system-prompt boundary.
This change makes status-line use the same computeContextBreakdown
source as /context so both surfaces stay consistent. The breakdown
result is cached with a 2s TTL inside the component so the per-frame
status-line render does not re-walk every message via
estimateMessagesTokens on long sessions. The Anthropic API per-turn
prompt size remains observable via existing token_in / cache_read /
cache_write / token_total segments.
- New 'usage' status-line segment showing Anthropic 5h/7d quota
- Background refresh (5min TTL) via fetchUsageReports
- thinking.max symbol added to UNICODE_SYMBOLS
Personal patch consolidated into branch.
Bug: Ctrl+C on the ask tool selector threw ToolAbortError, the turn
ended with stopReason === "aborted", and handleBackgroundEvent fired
sendCompletionNotification() unconditionally — producing a misleading
"Task complete" desktop toast for a turn that never actually completed.
Fix mirrors the stopReason filter already used by
#currentContextTokens, #handleMessageEnd, and the retry / TTSR /
compaction skip paths across agent-session.ts: check the most recent
assistant message via session.getLastAssistantMessage() and return
early when stopReason is "aborted" or "error".
Test coverage (event-controller-abort-guard.test.ts, 6 cases):
- aborted -> 0 sendNotification calls
- error -> 0 calls
- stop -> 1 call (normal completion)
- no last assistant message -> proceeds (defensive)
- isBackgrounded=false (foreground) -> still 0
- completion.notify=off -> still 0
Matching guard applied to the standalone desktop-notify extension
(~/.omp/agent/extensions/desktop-notify/index.ts) which is currently
the live producer of completion toasts after Phase 1 of
seed_0ca7e1143ac1.
ImmutablePrefix caches system prompt + tool specs after first build()
so subsequent turns reuse identical byte sequences. AppendOnlyLog
converts messages once via syncMessages() and only appends deltas
on further turns — prior-turn bytes stay stable.
- New module: packages/agent/src/append-only-context.ts
StablePrefix, AppendOnlyLog, AppendOnlyContextManager
- AppendOnlyContextManager added to AgentLoopConfig
- Wired into streamAssistantResponse in agent-loop.ts
- Toggleable via provider.appendOnlyContext setting (auto/on/off)
- Default auto enables for deepseek provider
- 38 tests covering prefix, log, sync, compaction handling
- /session info surfaces current active state
Keep chat notifications emitted during session_start visible after the initial transcript render rebuilds the chat container. Add regression coverage for preserving startup notifications during initial render.
Fixes#1316
When session.prompt() returns, idle-flush tasks for async-job result
deliveries are scheduled via #schedulePostPromptTask (1ms delay) and
added to #postPromptTasks immediately. The 800ms loop timer could fire
in that window before isStreaming became true, causing the loop to
submit the next prompt while the delivery turn was still pending. The
delivery then hit AgentBusyError and the job result was silently dropped.
Add AgentSession.hasPostPromptWork (= #postPromptTasks.size > 0) and
include it in #isLoopAutoSubmitBlocked() alongside isStreaming and
isCompacting. Add a regression test that verifies the loop defers when
hasPostPromptWork is true and fires once it becomes false.
Fixes#1294
Allowed /goal set to replace the current active goal instead of rejecting and discarding the command input.
Added goal runtime and interactive-mode regression coverage for active replacements.
Fixes#1293
