Skip to content

CHANGELOG.md

Latest commit

 

History

History
761 lines (513 loc) · 56.6 KB

CHANGELOG.md

File metadata and controls

761 lines (513 loc) · 56.6 KB
Raw

Changelog

All notable changes to iosm-cli will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[0.3.13] - 2026-04-11

Added

  • Cross-terminal keybinding fallbacks — added Alt+... fallback bindings for frequently unreliable terminal control combinations (expandTools, toggleThinking, model cycling, model selector).
  • Keybindings parser compatibility layer — keybindings.json now accepts both preferred action -> key(s) and legacy key -> action formats, with action aliases (nextModel, previousModel, openModelSelector, etc.).
  • Keybinding normalization — loader now normalizes common key spelling/casing variants (Control+, spacing, case differences) for consistent runtime matching.

Changed

  • Platform-aware suspend defaults — suspend hotkey is now disabled by default on Windows and remains enabled on Unix (Ctrl+Z + Alt+Z).
  • Hotkeys command output — /hotkeys now shows reverse model-cycle shortcut and renders active configured key for image paste instead of hardcoded value.
  • Keyboard documentation alignment — interactive/configuration/getting-started docs now document real keybinding format and cross-platform fallback shortcuts.

Fixed

  • Suspend runtime resilience — unsupported suspend environments now safely restore TUI state and show a warning instead of leaving UI in a broken state.
  • User keybinding migration regressions — legacy keybinding files from older docs/examples now resolve to expected actions instead of silently failing.

[0.3.12] - 2026-04-11

Added

  • Per-turn specialist system overlay — runtime-selected specialist instructions are now injected as an additive system layer for the active root run, including protocol auto-repair retries.
  • Routing provenance metadata — orchestration/task assignment records now persist selected specialist (agent) and routing source for traceable worker labeling.
  • Model-semantic specialist router hooks — SDK/task pipeline now exposes semantic specialist routing callback support with richer candidate hints (name/description/profile/instructions).

Changed

  • Runtime specialist routing strategy — removed embedding dependency from the interactive runtime path; routing now uses model-semantic selection with deterministic heuristic fallback and strict time/confidence budgets.
  • full profile delegation policy — root full profile prompt now explicitly prefers specialist delegation via task(agent=NAME) for actionable domain work and avoids task() for non-actionable chat/capability queries.
  • Task tool guidance quality — task schema/tool descriptions now prioritize agent= routing, include expanded built-in specialist catalog, and clarify profile semantics.
  • Orchestration panel UX — header now renders elapsed TIME stopwatch instead of ETA; worker rows include agent@profile and routing source label; empty QUEUED/COMPLETED/FAILED sections are omitted.

Fixed

  • Dynamic instruction application reliability — specialist instructions no longer rely on fragile custom-message layering; they now apply deterministically for the whole run and do not leak into subsequent turns.
  • Retry-path context loss — protocol auto-repair retries now preserve runtime specialist instruction context.
  • Small-talk misrouting regressions — short conversational turns are guarded from accidental specialist auto-selection.
  • Orchestration completion flow — completed orchestration panel now archives cleanly before continuing root-agent response rendering.

[0.3.11] - 2026-04-10

Added

  • Runtime specialist instruction overlay — full root session can now inject a selected custom/core agent instruction block at runtime (runtime_agent_context) so the same root conversation dynamically applies task-specific specialist guidance.
  • Assistant/runtime specialist labeling — interactive UI now renders specialist identity on assistant/tool frames when runtime specialist context is active (instead of always showing only the default root label).
  • Expanded core specialist catalog — added/extended practical specialist agents for workflow optimization, tooling evaluation, readiness validation, data engineering, brand consistency, and result analytics.

Changed

  • Root-agent specialist routing behavior — runtime specialist selection now prefers semantic routing using agent descriptions/instructions (embedding-backed when available) with deterministic fallback behavior.
  • Orchestration worker visibility — orchestration progress now surfaces explicit worker identity with agent@profile formatting where a specialized agent is chosen.
  • Task-agent auto-routing plumbing — task tool routing interfaces now carry richer agent hint metadata (name/description/instructions/profile) for higher-quality specialist selection.

Fixed

  • Runtime overlay activation gaps — specialist overlay path now covers non-imperative task requests more reliably instead of requiring narrow action-keyword patterns.
  • UI consistency for specialist context — runtime specialist metadata is now consistently queued/consumed with display alias handling, preventing mismatched labels during turn rendering.

[0.3.10] - 2026-04-06

Changed

  • Telegram live trace UX — running status card now presents richer execution telemetry (assistant/tool lifecycle events) in a compact streaming trace section instead of request→wait→final-only visibility.

Fixed

  • Streaming trace duplicate lines — assistant streaming trace now uses prefix-based upsert semantics, so the live card keeps one current assistant streaming line instead of appending duplicates on every update.
  • Trace payload matching internals — added dedicated trace payload extract/find helpers to update existing trace rows deterministically in both visible and full-history trace buffers.

[0.3.9] - 2026-04-06

Changed

  • Telegram completion contract hardening — bridge runtime prompt now explicitly requires a short final textual summary for each completed turn (prevents tool-only finishes in normal flow).

Fixed

  • Task completed with no assistant text output fallback quality — replaced generic empty-output message with contextual fallback summary including tool activity, elapsed time, and recovery hint for concise no-tool recap.
  • Turn-level telemetry for empty-output diagnosis — Telegram bridge now tracks per-turn tool call count and unique tool names and logs them when assistant final text is empty.
  • Assistant text capture resilience — bridge now also captures text from message_update events as an additional fallback path before agent_end.

[0.3.8] - 2026-04-06

Changed

  • Telegram runtime guidance injection — bridge now appends runtime execution guidance to child RPC system prompt to reduce unstable command patterns during remote execution (focused outputs, bounded scans, script-file fallback for complex shell flows).

Fixed

  • Windows shell adaptation robustness — PowerShell-adapted commands now execute via -EncodedCommand, avoiding fragile quote/escape transport through bash wrappers.
  • Windows adapter false-positive routing — commands like ls C:\projects are no longer incorrectly forced into cmd.exe just because they include a Windows path argument.
  • Telegram long-message delivery stability — rich/fallback message sending now chunks long output payloads to stay within Telegram safe limits and reduce delivery failures.

[0.3.7] - 2026-04-06

Added

  • Telegram bridge module split — extracted dedicated Telegram runtime modules:
    • telegram-api client with serialized outbound requests and retry controls
    • prompt-queue with fair round-robin draining across chats
    • polling-state store for persisted long-polling offsets
  • Configurable Telegram retry policy — introduced telegram.retry settings block with API retry/backoff, polling backoff, and status-edit network retry controls

Changed

  • Bridge retry/backoff wiring — Telegram bridge now reads retry/backoff parameters from settings instead of hardcoded values for:
    • API request retries
    • polling-loop transient failure backoff
    • status edit network retry delay
  • Configuration docs — extended Telegram configuration examples and generated settings documentation with new retry policy fields

Tests

  • Added dedicated test suites for:
    • Telegram API retry and outbound serialization behavior
    • prompt queue fairness
    • polling offset persistence
  • Expanded settings-manager telegram settings tests to cover retry normalization and safe bounds

[0.3.6] - 2026-04-06

Added

  • Telegram polling trace mode — added verbose polling/update tracing for diagnostics via telegram.debug.pollingTrace=true and IOSM_TELEGRAM_POLLING_TRACE=1

Changed

  • Telegram live status update cadence — raised default status edit throttle to 3000ms, applied runtime bounds 1000..10000ms, and removed implicit animation-time override to reduce API pressure

Fixed

  • Stale callback query handling — answerCallbackQuery now safely ignores expired callback IDs (query is too old / invalid query) instead of surfacing noisy polling failures
  • Polling network backoff — transient fetch failed and related network errors now use exponential backoff (2s to 30s) in long polling loop
  • Status edit resilience on network loss — transient status edit network failures now schedule coalesced retries instead of immediate repeated edits
  • Telegram API transport diagnostics — fetch-level Telegram API failures now include method context (Telegram API <method> request failed) for faster root-cause analysis

[0.3.5] - 2026-04-06

Fixed

  • Telegram API 429 resilience — added automatic retry after N backoff/retry handling for all Telegram API operations (including sendMessage, editMessageText, sendDocument, and control actions)
  • Telegram status message recovery — when live status edit target becomes invalid (MESSAGE_ID_INVALID, deleted or non-editable message), bridge now recreates status message and continues updates on a fresh message_id
  • Bridge runtime stability under delivery failures — wrapped RPC event callbacks and final output delivery with fail-safe guards to prevent bridge process crashes from transient Telegram delivery errors

[0.3.4] - 2026-04-06

Fixed

  • Windows shell command compatibility — added runtime auto-adaptation for Windows command syntax, routing cmd-style commands (%VAR%, dir, C:\...) to cmd.exe and PowerShell-style commands ($env:..., Get-*) to powershell.exe
  • Telegram bridge command execution on Windows — fixed failures when users sent native Windows shell commands through Telegram-driven runs while IOSM executed under bash
  • Execution path consistency — applied shell adaptation uniformly across foreground bash execution, bash tool execution, and detached background processes

Tests

  • Added regression suite for Windows command adaptation and explicit-shell passthrough behavior

[0.3.3] - 2026-04-05

Fixed

  • META orchestration default fan-out — actionable META prompts without explicit agents=N now require at least two top-level task calls, preventing silent single-thread execution for orchestration-first requests
  • Locale-agnostic META fallback — replaced language-specific complexity keyword detection with runtime orchestration compliance signals (required task calls, nested delegation status) for stable behavior across languages
  • MCP/external tool output rendering — unknown external tools now use generic fallback rendering instead of empty tool boxes, restoring visible arguments/results in interactive UI

Tests

  • Updated META prompt fallback tests to cover actionable vs question-like prompts
  • Added regression test for unknown external tool result rendering in ToolExecutionComponent

[0.3.2] - 2026-04-05

Fixed

  • Playbook filename canonicalization — /init and iosm init now normalize lowercase variants (agents.md, iosm.md) to canonical AGENTS.md and IOSM.md
  • Cross-platform case handling — implemented robust case-only rename flow (with temporary hop) to reliably normalize filenames on case-insensitive filesystems
  • Profile context continuity — IOSM.md context is now kept loaded in the resource layer to avoid context loss after runtime profile switching

Changed

  • Standard init path resolution — standard /init now resolves and writes through canonical playbook path before reading/updating repository guidance
  • IOSM guide writers — IOSM/AGENTS guide writers now enforce canonical path selection prior to write operations

Tests

  • Added regression tests for lowercase-to-canonical playbook rename behavior in:
    • iosm guide writer flow
    • iosm init flow
    • interactive standard /init flow
  • Updated resource-loader profile test expectations for runtime profile switching behavior

[0.3.1] - 2026-04-05

Added

  • Profile-aware /init flow — /init now follows active profile semantics: full/plan/meta updates AGENTS.md; iosm initializes and synchronizes both AGENTS.md and IOSM.md
  • AGENTS managed sync block for IOSM — introduced deterministic managed section in AGENTS.md during IOSM init, preserving user-authored content outside the managed block

Changed

  • IOSM init file sync contract — iosm init now performs guaranteed create/update for both AGENTS.md and IOSM.md to avoid missing physical playbook files
  • Interactive init control flow — added dedicated standard-init session tracking with explicit “already running” guard and interrupt-safe cancellation path
  • CLI init status output — iosm init now reports per-file state for AGENTS.md and IOSM.md (created/updated/skipped/unchanged)

Documentation

  • Updated README release marker and highlights to 0.3.1
  • Updated docs/getting-started.md and docs/interactive-mode.md to document profile-scoped /init behavior

Tests

  • Expanded coverage for IOSM/AGENTS guide writers and profile-aware /init behavior in interactive and init test suites

[0.3.0] - 2026-04-05

Added

  • Unified framed message windows — introduced consistent full-contour windows (╭╮│╰╯) with labels and padding for interactive conversation output
  • Input area redesign — replaced the plain editor strip with a framed, mode-aware input box that reflects active mode (input, command, bash)
  • Reasoning preview stream UX — added collapsed reasoning summary with live spinner animation during assistant streaming
  • Compact footer setting — added configurable compact footer mode in settings and interactive selector

Changed

  • Assistant response branding — assistant window label now renders as IOSM Agent
  • Tool execution visuals — migrated tool call/result rendering from filled background blocks to green-framed windows with consistent spacing
  • Orchestration panel visuals — migrated orchestration status panel to orange-framed IOSM Orchestration window and aligned inner section header colors
  • Message surface consistency — removed legacy filled-background user/orchestration/tool blocks in favor of border-first presentation

Documentation

  • Updated README release marker and release highlights to 0.3.0

Tests

  • Added and expanded coverage for framed message/input rendering, assistant label updates, tool window framing, and orchestration panel presentation

[0.2.17] - 2026-04-05

Fixed

  • Startup crash after npm install — fixed packaging regression where dist/core/settings.schema.json was not included in published artifacts, causing ENOENT on startup in some global installers (including Volta)
  • Settings schema load resilience — added safe fallback validator path when settings schema file is missing or unreadable, so runtime no longer hard-crashes at import time

[0.2.16] - 2026-04-05

Added

  • Policy Engine v2 — added TOML-backed layered policy resolver with deterministic precedence and legacy compatibility bridge for existing permission rules
  • ACP mode — added --mode acp adapter over RPC/event bus with capability negotiation and compatibility degradation handling
  • Tooling expansion — added built-in apply_patch, tool_search, and tool_suggest tools
  • Unified execution runtime — added PTY-capable unified execution support for interactive command sessions (exec + stdin streaming workflow)
  • Session resume indexing — added indexed session lookup to avoid full filesystem scans on large session sets
  • Schema-driven settings docs pipeline — added generated configuration docs and CI validation scripts for settings/schema drift detection

Changed

  • Permission flow — unified permission evaluation across interactive and RPC runtimes through policy-backed evaluator
  • Permission UX — added turn-scoped and session-scoped approval behavior to reduce repeated prompts in ask mode while preserving safety controls
  • MCP governance — enforced source trust checks and logged per-tool policy decision traces for MCP tool calls
  • Task UI in interactive mode — task-state operations now render as structured checklists (done / in progress / pending) instead of raw JSON argument dumps
  • Package/tool security baseline — install/update/extensions/tool-manager paths now use trust-ledger and source security checks (host allowlist, fingerprint/integrity, consent gates)
  • Linux sandbox execution — added explicit opt-in sandbox path using bwrap with hard failure when required runtime is unavailable
  • Command rollback safety — integrated filesystem checkpoint snapshot/restore with deterministic rollback ordering in command dispatch flow

Documentation

  • Updated README release marker and release highlights for 0.2.16
  • Added docs navigation entries for ACP mapping and generated configuration reference
  • Refreshed configuration documentation from schema generation pipeline

Tests

  • Expanded coverage for policy engine behavior, ACP mode, apply_patch grammar/runtime, tool_search/tool_suggest flow, session index resume path, and unified execution runtime
  • Revalidated interactive task-plan/task-checklist rendering and orchestration-related interactive regressions

[0.2.15] - 2026-04-04

Changed

  • Telegram stop/resume contour — when bridge is stopped, Hub now switches to a dedicated stop-state action set (Start, Refresh, Help) instead of showing full active-run controls
  • Telegram quick keyboard state — reply keyboard now also switches to stop-state layout (Hub, Start, Help) and restores full actions after resume
  • Telegram Hub navigation after restart — /start now posts a fresh Hub card at the bottom of chat for immediate visibility

Fixed

  • Telegram quick-action reliability — removed over-aggressive quick-action debounce that could drop first button press and make actions appear to require a second tap
  • Slash command parsing in Telegram — /start@botname and similar bot-suffixed slash commands now route correctly
  • Stopped-bridge menu routing — command/model callbacks in stopped mode now return cleanly to Hub instead of leaving stale menu state

Tests

  • Revalidated telegram-related runtime changes with TypeScript check and interactive/command regression suites

[0.2.14] - 2026-04-04

Added

  • Telegram bridge mode — added mobile remote-control runtime (iosm telegram / --mode telegram) with long-polling transport, single-operator allowlist guard, Control Hub navigation, live status updates, quick-action keyboard, model picker, and confirmation flow for dangerous tool calls
  • Headless built-in command dispatcher — extracted built-in slash command execution into a shared dispatcher for non-TUI runtimes
  • RPC built-in command contract — added get_builtin_commands and run_builtin_command RPC commands so external clients can invoke core slash commands without TUI routing
  • RPC permission events — added requires_confirmation event and confirm_permission extension UI request contract for ask-mode tool confirmations in external clients
  • Subagent background run registry — added persistent metadata/log tracking for background subagent execution with stop/stop-all/prune helpers and interactive command surface

Changed

  • CLI mode surface — expanded --mode to include telegram and added top-level iosm telegram alias command
  • RPC startup behavior — RPC mode now starts without a preselected model so remote clients can choose model later (for example via Telegram /model)
  • Interactive subagent commands — extended /subagent-runs with bg management commands (list/running/queued/done/error/cancelled/status/logs/stop/stop-all/prune)
  • Provider policy filtering — disabled google-antigravity provider from auth/model/login catalogs and default model resolution in this build
  • Example set refresh — removed deprecated examples/extensions/antigravity-image-gen.ts and updated extension example count

Documentation

  • Updated README integration/mode coverage with Telegram bridge examples and profile-based launch commands
  • Updated docs/cli-reference.md with --mode telegram and iosm telegram --profile ... usage
  • Updated docs/configuration.md with telegram settings schema (enabled, botToken, allowedUserIds, transport, chatDefaults)
  • Updated docs/rpc-json-sdk.md with built-in command RPC calls and confirmation event/response protocol

Tests

  • Added command dispatcher coverage in test/command-dispatcher.test.ts
  • Added subagent background run storage/control coverage in test/subagent-background-runs.test.ts
  • Expanded RPC/interactive/settings/model/auth regressions for telegram mode wiring, built-in command dispatch, permission-mode/state handling, and provider policy filtering

[0.2.13] - 2026-04-02

Added

  • Interactive extension lifecycle commands — added /extensions with /ext alias for list, install, update, remove, enable, disable, and help flows in interactive mode
  • Background prune command — added safe cleanup support for old completed background process records via /bg prune [hours]
  • Git snapshot runtime context capture — added bounded git snapshot composition (status, diff --stat, diff --cached --stat) wired into prompt context when enabled
  • Prompt-context setting for git snapshot size — added promptContext.gitSnapshotMaxChars (default 2000)

Changed

  • Prompt context telemetry — system_prompt_context_compose trace events now include git snapshot diagnostics (git_snapshot_chars_before, git_snapshot_chars_after, git_snapshot_truncated, git_snapshot_max_chars)
  • Background command UX — /bg usage/help/menu now includes prune flow and clearer cleanup guidance
  • System prompt guidance — root-agent guidance now includes explicit /extensions lifecycle routing for extension management requests

Documentation

  • Updated README version marker to 0.2.13
  • Updated docs/interactive-mode.md and docs/cli-reference.md with /extensions (/ext) and /bg prune command coverage
  • Updated docs/configuration.md for promptContext.gitSnapshotMaxChars and bounded git snapshot behavior
  • Updated docs/sessions-traces-export.md with expanded prompt-context trace event fields
  • Refreshed task.md and improvement-checklist.md with v2 execution status and risky-feature rollout policy

Tests

  • Added git snapshot context tests in test/system-prompt.test.ts
  • Added background prune coverage in test/background-processes.test.ts
  • Added interactive command coverage for /bg prune and /extensions lifecycle paths in test/interactive-mode-status.test.ts

[0.2.12] - 2026-04-01

Changed

  • Startup command execution policy — system prompt guidance now explicitly routes "start/run project", dev-server, and watcher requests to detached bash execution (run_in_background=true) by default, with foreground mode reserved for immediate-output needs
  • Subagent runtime capability hint — delegated task prompts now include the same detached-startup guidance so child agents avoid blocking turns when launching persistent processes

Documentation

  • Updated README usage guidance to document default detached handling for project/server startup requests
  • Updated docs/cli-reference.md and docs/interactive-mode.md with explicit startup-to-background behavior and /bg follow-up flow

Tests

  • Added system-prompt regression asserting explicit startup background guidance (run_in_background=true, /bg status|logs|stop)
  • Expanded subagent orchestration tests to assert delegated runtime hints include startup/background routing rules

[0.2.11] - 2026-04-01

Added

  • Prompt context preprocessing controls — added deterministic promptContext pipeline (normalize -> dedupe -> per-file cap -> total cap) with optional git snapshot plumbing and context preprocessing trace metrics
  • Background shell process runtime — added persistent detached process manager with metadata/log tracking under .iosm/background/processes
  • Interactive background command surface — added /bg command family (list, status, logs, stop) and detached shell submission syntax via ! <command> &
  • Bash detached execution parameter — added run_in_background support in built-in bash tool and session-level bash execution path with returned backgroundTaskId/status/log metadata
  • Extension tool permission tiers — added optional requiredPermission metadata (read-only, workspace-write, danger-full-access) for extension tools and runtime propagation through permission flow
  • Subagent tool-name normalization — added normalization/filtering for custom agent tools and disallowed_tools (lowercase, - -> _, unknown removal with diagnostics)

Changed

  • Permission flow unification — consolidated tool permission evaluation for built-ins and extensions with tier/source-aware request metadata and stricter extension enforcement branch behind permissions.extensionToolEnforcement (default off)
  • Compaction continuation framing — compaction summary prefix now explicitly instructs continuation from current state without full recap unless user asks
  • System prompt guidance updates — bash guidance now includes detached execution usage (run_in_background) for long-running non-blocking commands

Documentation

  • Updated README header/version marker to 0.2.11
  • Updated interactive/CLI/configuration/session-trace/orchestration docs with:
    • background process workflow (! ... &, /bg)
    • promptContext settings and defaults
    • extension permission tier semantics and permissions.extensionToolEnforcement
    • subagent frontmatter tool normalization behavior
  • Updated extension authoring docs with requiredPermission examples for registerTool

Tests

  • Added test/background-processes.test.ts coverage for start/list/log-tail/stop lifecycle
  • Added interactive /bg coverage in test/interactive-mode-status.test.ts
  • Added run_in_background coverage for bash in test/tools.test.ts
  • Expanded regressions for context preprocessing, extension permission-tier behavior, compaction continuation framing, and subagent tool normalization

[0.2.10] - 2026-03-18

Added

  • /ultrathink built-in command — added deep multi-iteration analysis mode with robust slash parsing (-q/--iterations, -- separator), context-aware no-query fallback, and shared behavior through AgentSession.prompt across interactive/print/json/rpc flows
  • Ultrathink checkpoint engine — added structured checkpoint state (Goal, Verified Facts, Rejected Hypotheses, Open Questions, Next Checks) with carry-forward summaries and checkpoint compression support for long runs
  • Ultrathink read-only execution policy — added strict temporary read-only tool filtering during ultrathink runs with guaranteed restoration of the original active tool set
  • Ultrathink evidence/runtime tests — added dedicated parser and session-flow test suites covering iteration loops, no-query objective resolution, streaming guards, budget/stagnation paths, evidence-policy fallback, and tool-set restoration

Changed

  • Ultrathink runtime hardening — added budget guardrails (per-iteration input, run input/total tokens, run cost), stagnation early-stop behavior, and evidence-catalog carry-forward between passes
  • Ultrathink anti-hallucination policy — added quantitative-claim evidence tagging rules, verify/synthesis no-new-evidence marker handling, and compliance-repair pass support
  • Ultrathink grounding behavior — when early passes produce no tool evidence, runtime now injects an internal grounding retry that explicitly forces live read-only workspace probes before continuing
  • Interactive slash UX — added /ultrathink to built-in slash registry and interactive autocomplete argument hints (-q, --iterations, common iteration counts)

Fixed

  • Ultrathink hard-stop on evidence mismatch — repeated evidence-policy mismatch no longer aborts the entire command; runtime now degrades gracefully and returns a best-effort final response instead of throwing
  • Internal prompt visibility leakage — ultrathink internal retries (iteration/grounding/policy-repair prompts) are now consistently routed through hidden orchestration aliases, so users see clean progress text instead of raw directives
  • Budget accounting with internal retries — per-iteration budget checks now account for cumulative input tokens across the main pass plus internal retry prompts

Documentation

  • Updated README header/version marker to 0.2.10
  • Added /ultrathink command coverage to interactive and CLI references, including read-only behavior, context fallback, and grounding-retry semantics

Tests

  • Added test/ultrathink.test.ts parser/validation coverage for /ultrathink
  • Added test/agent-session-ultrathink.test.ts runtime coverage for q-iteration flow, early-stop, budget-cutoff, no-query fallback, evidence-policy graceful fallback, and restoration guarantees
  • Expanded semantic regressions to assert /ultrathink built-in slash discoverability

[0.2.9] - 2026-03-15

Added

  • Structured verification/data tools — added built-in test_run, lint_run, typecheck_run, and db_run with runner/adapter auto-detection, normalized statuses, bounded output capture, and tool-registry/SDK exports
  • DB runtime settings layer — added dbTools settings (defaultConnection, named connections, adapter-specific fields, migrate script options) wired into session/runtime resolution for db_run
  • Universal terminal theme — added built-in universal theme and made it the default/fallback theme for interactive mode
  • Protocol and stall auto-repair flow — added bounded automatic recovery for raw pseudo tool markup and silent stop responses, including interactive recovery actions (retry, repeat prompt, switch model + retry, keep session)

Changed

  • Profile tool policy expansion — enabled test_run, lint_run, typecheck_run, and db_run across write-capable engineering profiles (full, meta, iosm); enabled typecheck_run for iosm_verifier
  • Interactive UX readability pass — updated dark/light palettes, introduced universal box colors, and normalized box paddings/section spacing for user/custom/tool/plan/subagent/summary messages
  • System prompt hardening — added explicit guidance for structured verification/data tools, instruction-priority handling, untrusted tool-output/web-content handling, completion checks before final success claims, and stricter pseudo-markup prohibition
  • Doctor diagnostics coverage — expanded interactive /doctor CLI-toolchain checks to include verification and DB client commands used by new structured tools

Fixed

  • Abort continuation wording — when user interrupts execution, recovery selector now shows a user-action title (You stopped the current run) instead of model-failure wording
  • Protocol false positives — inline explanatory mentions like raw <tool_call>/<function=...> markup no longer trigger protocol auto-repair; only executable-looking pseudo-blocks are repaired
  • Dark theme contrast in boxes — fixed low-contrast text-on-dark-box cases for user/custom/tool blocks

Documentation

  • Updated README header/version marker to 0.2.9
  • Updated CLI/config/interactive/development docs with test_run/lint_run/typecheck_run/db_run, profile policy updates, dbTools configuration, and /doctor toolchain scope

Tests

  • Added dedicated tool coverage for test_run, lint_run, typecheck_run, and db_run
  • Added protocol auto-repair and recovery-selector coverage in interactive/session tests (raw markup, silent stop, false-positive guard, model-switch recovery)
  • Expanded regressions for profiles, SDK exports, settings manager, shadow guard, system prompt guidance, and theme defaults/colors

[0.2.8] - 2026-03-14

Added

  • web_search built-in tool — added structured web discovery with provider chaining (Tavily -> SearXNG -> DuckDuckGo), include/exclude domain filters, recency/topic/depth hints, runtime configuration hooks, and permission-guard integration
  • git_write built-in tool — added structured git mutation actions (add, restore, reset_index, commit, switch, branch_create, fetch, pull, push, stash_push, stash_pop, stash_apply, stash_drop, stash_list) with action-specific validation, permission-guard integration, and safe argv execution (no raw shell passthrough)
  • GitHub tools settings — added persistent githubTools settings block (networkEnabled, token) and interactive settings submenu for enabling git network actions and managing GitHub token
  • Expanded git_read actions — added read-only actions show, branch_list, remote_list, and rev_parse while preserving backwards compatibility for existing status, diff, log, and blame contracts

Changed

  • Shared git tool runtime — consolidated git process execution/capture/truncation/error handling into a common internal helper used by git_read and git_write
  • Profile and mutation policy updates — enabled git_write by default in write-capable profiles (full, meta, iosm), included web_search in read-oriented profiles, and updated mutation classification in task/shadow-guard checks
  • System prompt and tool-routing guidance — added explicit guidance to prefer git_write over ad-hoc git bash mutations and to use web_search for discovery plus fetch for source validation
  • CLI/SDK tool surface expansion — updated tool registry, factories, and exports to include web_search and git_write in all relevant creation paths and public SDK/index entrypoints
  • Settings/runtime wiring — threaded web search runtime config (provider/fallback/safe-search/max-results/timeout/credentials) and GitHub network/token policy from settings into tool execution
  • Interactive menu hint UX polish — normalized selector/menu control hints across settings, model/oauth/MCP selectors, tree/session views, and config/model-scoping panels for consistent navigation/action/search/exit guidance

Documentation

  • Updated README header/version marker to 0.2.8
  • Updated CLI/config/interactive/development/RPC docs for web_search behavior, git_write network actions, GitHub tools settings (networkEnabled, token), and expanded git_read action coverage
  • Updated help text and tool listings to reflect the new structured git/web workflow (web_search + fetch, git_read + git_write)

Tests

  • Added dedicated web_search and git_write tool coverage, plus expanded git_read tests for new actions (show, branch_list, remote_list, rev_parse)
  • Added/updated regression checks for profile tool membership, SDK defaults, settings manager behavior, system prompt guidance, shadow-guard/task classification, and settings/menu hint behavior

[0.2.7] - 2026-03-14

Added

  • fetch built-in tool — added structured HTTP tool with fields url, method, headers, body, timeout, max_bytes, response_format, max_redirects; includes manual redirect loop control, bounded body capture, and auto JSON/text formatting by content-type
  • git_read built-in tool — added read-only structured git introspection actions (status, diff, log, blame) with action-specific validation and safe argv execution (no raw shell passthrough)
  • fs_ops built-in tool — added structured filesystem mutation tool (mkdir, move, copy, delete) with explicit recursive/force safety gates and EXDEV move fallback (copy + delete)
  • SDK/public exports for new tools — exported factories, tool singletons, and typed input/options/details surfaces for fetch, git_read, and fs_ops

Changed

  • Profile tool policy expansion — fetch and git_read are now included in read-only exploration profiles; fs_ops is included in write-capable engineering profiles (full, meta, iosm)
  • Dynamic fetch method policy by active profile — read-only profiles (explore, plan, iosm_analyst) allow only GET|HEAD|OPTIONS; write-capable profiles allow full method set (GET|POST|PUT|PATCH|DELETE|HEAD|OPTIONS)
  • Unified tool permission/pre-hook integration — fetch and fs_ops now pass through the same session permission and pre-tool hook pipeline used by write-capable built-ins
  • Mutation classification updates — fs_ops is now classified as mutating in shadow guard and task write-capable tool checks

Documentation

  • Updated CLI help, system prompt tool guidance, and docs pages to include fetch, git_read, and fs_ops with usage/policy notes
  • Updated profile/tool tables and architecture snippets to reflect the expanded built-in tool layer

Tests

  • Added dedicated coverage for fetch, git_read, and fs_ops
  • Added regression assertions for profile membership, read-only/write-capable classification, shadow guard behavior, and CLI help/tool listings

[0.2.6] - 2026-03-14

Added

  • Shared memory scope policy — new IOSM_SHARED_MEMORY_SCOPE_POLICY environment variable (legacy / warn / enforce) controls how missing scope arguments are handled in shared_memory_read / shared_memory_write; meta profile automatically activates warn mode so omitted scopes surface a warning in tool output and details
  • Shared memory usage analytics — new summarizeSharedMemoryUsage() API aggregates write counts by scope, unique writers, unique keys, and per-task/delegate breakdown for observability in orchestrated runs
  • Nested delegation detection — promptMetaWithParallelismGuard now tracks nestedDelegationMissing: when top-level fan-out is satisfied but no nested delegates were observed for multi-stream tasks, the parallelism correction prompt and TUI warning fire explicitly
  • Workstream semantic deduplication — semanticallyDeduplicateWorkstreamTitles() uses Jaccard token similarity (threshold 0.82) to eliminate near-duplicate delegate workstream titles before dispatch
  • Duplicate delegated section detection — detectDuplicateDelegatedSections() compares normalized section bodies to catch copy-pasted or near-identical delegate blocks with ≥92% coverage overlap
  • Workstream title uniquification — uniquifyWorkstreamTitles() appends ordinal suffixes to disambiguate repeated titles in fan-out plans
  • Coordination details in task tool output — TaskToolDetails now surfaces a coordination object with sharedMemoryWrites, currentTaskWrites, currentTaskDelegateWrites, runScopeWrites, taskScopeWrites, duplicatesDetected, claimKeysMatched, and claimCollisions fields for post-run auditing
  • Swarm progress shared memory integration — TUI swarm progress reporter now reads results/ prefix keys from shared memory to enrich per-task summary display with delegated totals

Changed

  • META profile evidence policy — meta profile system prompt and subagent task prompt now require that metrics (speedup, compliance scores, conflict counts) are backed only by observed runtime evidence; unknown values must be marked as unknown rather than inferred
  • META profile artifact claims — meta and meta-subagent prompts now prohibit claiming report files or artifacts exist unless they were produced in the current run or verified on disk
  • resolveScope replaces normalizeScope — shared memory tool's scope defaulting logic refactored into resolveScope() with policy-aware warning output and enforce mode that throws on missing explicit scope
  • completedTaskToolCalls tracking — parallelism guard now separately tracks completed (resolved) task calls so nested delegation assessment waits for actual task completion rather than firing prematurely on partial state

Fixed

  • False nested-delegation compliance — guard no longer silently passes when top-level fan-out count is met but zero nested delegates exist inside multi-stream tasks; correction prompt now fires
  • Scope warning surface — shared_memory_write and shared_memory_read tool results now include scopePolicy and scopeWarning in their details payload for agent-side introspection

Documentation

  • README redesign — complete rewrite with professional positioning, IOSM methodology section with 4-phase table and 6 metrics, architecture ASCII diagram, profile split into primary/specialist, integration modes with CI row, extensibility as runtime platform, accurate install/extension syntax from docs

[0.2.5] - 2026-03-13

Added

  • Orchestrate parallel fan-out defaults — /orchestrate --parallel now auto-sets --max-parallel to the selected agent count when omitted, reducing accidental single-lane execution
  • Parallel worker profile auto-selection — when no worker profile is provided in parallel orchestration, assignments default to meta in write-capable host contexts for stronger orchestration behavior
  • Delegate hint propagation for orchestrate assignments — assignment generation now injects delegate_parallel_hint guidance to drive nested delegate fan-out inside child tasks
  • Swarm dispatch timeout controls — scheduler now supports bounded dispatch timeouts (including IOSM_SWARM_DISPATCH_TIMEOUT_MS) to avoid silent long stalls
  • Interactive swarm progress surfaces — improved live subagent task/delegate progress rendering and swarm-aware footer busy state in TUI

Changed

  • Task profile defaulting — task tool now defaults missing profile to current host profile (fallback full) instead of always forcing full
  • Delegation depth baseline — max delegation depth default increased to 2 for better nested decomposition capacity
  • Shared memory read behavior — shared_memory_read now returns metadata-only by default (include_values=false) with safe value preview details when requested
  • Swarm planning fan-out quality — planner now prioritizes code-relevant touches and partitions work into multiple workstreams more aggressively for parallel execution
  • Singular run id generation — /singular run ids now include milliseconds and random suffix for collision-resistant rapid runs

Fixed

  • Dependent-task dead-end behavior — scheduler now marks downstream tasks as blocked when dependencies fail, preventing ambiguous pending states
  • Status update loss under file lock contention — team task status writes now queue and retry asynchronously instead of being dropped during temporary lock conflicts
  • Steering skip false errors — parallel task agent no longer marks steering-driven tool skips as execution errors
  • Swarm-from-singular startup guard — execution now fails fast with a clear warning if no active model is configured
  • Strict delegation in orchestrated contexts — nested delegation contract now also applies in run/task orchestrated contexts when delegate hints indicate required fan-out

Documentation

  • Updated README to v0.2.5 and added a focused "What's New in v0.2.5" section
  • Expanded orchestration docs (interactive-mode, cli-reference, orchestration-and-subagents) with /orchestrate parallel defaults and delegation guidance

[0.2.4] - 2026-03-12

Added

  • META profile onboarding UX — switching to meta now shows an explicit runtime hint describing orchestration-first usage and when to switch back to full
  • META interruption fallback hint — when a run ends without any assistant message in meta, the UI now emits a recovery warning with concrete prompt guidance
  • Task host-profile runtime getter — task tool integration now supports dynamic host profile reads (getHostProfileName) so orchestration pressure follows live profile changes during a session

Changed

  • META directive policy (chat-safe classification) — meta orchestration directive now explicitly classifies non-repository prompts as direct chat responses and limits orchestration rules to actionable repository work
  • META profile system prompt alignment — profile-level prompt now mirrors the chat-safe classification behavior to avoid conflicting orchestration instructions
  • Internal orchestration metadata aliasing — hidden orchestration UI metadata now always persists display aliases (including streaming paths) for safer prompt display substitution

Fixed

  • Profile switch propagation into task orchestration — runtime profile changes now propagate through session config-change events and task-tool host-profile resolution
  • Invisible assistant responses in meta chat prompts — assistant prose suppression no longer triggers for META_ORCHESTRATION_DIRECTIVE metadata (suppression remains for legacy ORCHESTRATION_DIRECTIVE blocks only)
  • Meta interruption messaging consistency — differentiated interruption guidance now appears for both assistant-level abort/error and run-level early termination scenarios

Documentation

  • Updated README version markers and added a dedicated Modes At A Glance block
  • Added explicit META model requirements guidance (modern models, large context windows, high output limits) in README, CLI reference, interactive mode docs, and configuration docs

[0.2.3] - 2026-03-11

Fixed

  • Startup model restore after restart — createAgentSession() now hydrates missing saved provider/model definitions from models.dev before resolving default model, so previously selected providers such as coding-plan providers are restored automatically on relaunch
  • Stale startup warning suppression — interactive startup no longer shows stale No models available... warning when model restore succeeds during session initialization
  • Restart UX consistency — startup header and active session state now align with restored saved provider/model selection without requiring manual /model re-selection

[0.2.2] - 2026-03-11

Added

  • Models.dev provider+model catalog runtime — added full catalog parsing (providers + models metadata) with timeout/fallback behavior for interactive auth/model flows
  • Automatic provider model hydration after /login — when a provider has credentials but no built-in model definitions, iosm-cli now registers models from models.dev so /model is immediately usable (including coding-plan providers such as zai-coding-plan)
  • Startup/on-demand auth model hydration — /model now attempts to hydrate missing models for saved authenticated providers before rendering provider/model choices

Changed

  • Provider/model visibility in status line — footer and model-switch status now display provider/model to make cross-provider switches explicit even when model IDs are identical
  • Auth UX feedback — login flow now reports a clear warning when credentials are stored but no models can be loaded yet

Fixed

  • API-key login crash — fixed unbound registry method usage that caused TypeError: Cannot read properties of undefined (reading 'models') in interactive login flows
  • Empty model selector after provider login — fixed cases where /model stayed empty after successful API-key auth for providers not shipped in the built-in registry

Documentation

  • Updated README and docs (getting-started, interactive-mode, configuration) to reflect full models.dev-backed provider/model availability via /login and /model

[0.2.1] - 2026-03-11

Added

  • Run/task shared memory runtime — introduced .iosm/subagents/shared-memory/*.json state with versioned entries and history for cross-task coordination
  • Shared memory tools — added shared_memory_write and shared_memory_read tools for subagent orchestration (run and task scopes, CAS support, append/set modes)
  • Canonical /swarm command surface — added dedicated runtime commands: /swarm run, /swarm from-singular, /swarm watch, /swarm retry, /swarm resume (with bounded parallelism and budget controls)
  • Swarm scheduler reliability modules — added dedicated scheduler/locks/gates/state-store/spawn/retry components for stable multi-task dispatch under contention
  • Swarm lock + gate execution model — introduced hierarchical touch locks and contract-aware task/run gates for Scopes -> Touches -> Locks -> Gates -> Done
  • Swarm runtime artifacts — added persisted run state in .iosm/orchestrate/<run-id>/ (run.json, dag.json, state.json, events.jsonl, checkpoints, reports)
  • Swarm watch telemetry — added runtime visibility for ready/running/blocked/done distribution, budget usage, lock snapshot, ETA/throughput, critical path, and theoretical speedup
  • Swarm spawn policy controls — added high-risk spawn candidate classification with confirmation-gated fan-out behavior
  • Project index subsystem — introduced repository indexing (.iosm/project-index/index.json) for scale-aware planning and targeted file selection
  • Failure retrospective engine — added failure-cause classification and retry directive generation for smarter follow-up attempts

Changed

  • Swarm-first orchestration flow — /singular execution handoff now supports Start with Swarm (Recommended) and routes selected options to /swarm from-singular ...
  • Command separation — /orchestrate --swarm removed; /swarm is now the canonical gated runtime while /orchestrate remains manual legacy team splitting
  • Task orchestration contract — task tool/runtime now carries richer run/task metadata and improved scheduling context for delegated execution
  • Interactive swarm observability — expanded interactive mode status/watch output with deeper swarm runtime diagnostics and task progress details

Fixed

  • Swarm retry stability — improved retry bucket handling (permission, dependency/import, test, timeout, unknown) to reduce noisy re-runs
  • Lock/contention handling — improved execution behavior for conflicting touches and blocked tasks in DAG scheduling scenarios

Documentation

  • Expanded README, CLI reference, interactive mode, and orchestration docs for swarm runtime semantics, shared-memory collaboration, and reliability controls

[0.2.0] - 2026-03-11

Added

  • Interactive engineering contract manager (/contract) — field-by-field contract editing with immediate save-on-enter and automatic JSON generation for project scope
  • Layered contract model — explicit project, session, and effective contract layers with copy/delete flows and merged runtime enforcement
  • Singular feasibility mode (/singular) — command-first feasibility analysis that combines repository baseline scan with a standard agent pass and returns exactly three implementation options
  • Option-driven execution handoff — /singular now produces concrete file targets, step plans, trade-offs, and decision guidance before implementation starts
  • Regression coverage for large paste UX — multiline unbracketed paste now covered by dedicated tests to ensure one submission flow and compact marker rendering

Changed

  • Feasibility workflow naming — /blast replaced by /singular for feature feasibility decisions
  • Profile cleanup — /shadow workflow removed to avoid duplication with plan-oriented analysis
  • Contract interaction model — removed extra save step in field editor; entering value immediately persists to selected scope

Fixed

  • TUI width safety — startup resources block now truncates long lines to terminal width, preventing render crashes on narrow terminals
  • Paste queue behavior — large pasted multiline input is treated as a single paste event instead of fragmented queued submissions

Documentation

  • Expanded README with dedicated decision workflow section (/contract vs /singular), command migration notes, and clearer contract layer distinctions
  • Extended interactive mode docs with explicit effective/session/project explanations and migration guidance from removed commands
  • Updated CLI reference with interactive feasibility/contract command behavior and migration notes

[0.1.3] - 2026-03-10

Added

  • Semantic search runtime — Added built-in semantic_search tool (index, query, status, rebuild), interactive /semantic manager, and top-level iosm semantic command
  • Semantic setup UX upgrades — Added provider model discovery (OpenRouter/Ollama) and setup-flow guidance for optional fields
  • Search/analysis toolchain expansion — Added separate built-in tools: rg, fd, ast_grep, comby, jq, yq, semgrep, sed
  • Doctor CLI diagnostics for tools — /doctor now reports external CLI toolchain availability (rg, fd, ast_grep, comby, jq, yq, semgrep, sed) in text and JSON output
  • CLI entry point — Standalone iosm binary with full argument parsing
  • Interactive mode — Multi-turn terminal agent with keyboard-driven controls
  • Print mode — One-shot iosm -p for scripting and CI
  • JSON stream mode — Machine-readable --mode json event output
  • RPC mode — stdio JSON-RPC server for IDE integrations
  • Built-in tools — read, bash, edit, write, grep, find, ls, rg, fd, ast_grep, comby, jq, yq, semgrep, sed
  • Multi-provider support — Anthropic, OpenAI, Gemini, Groq, Cerebras, xAI, OpenRouter, Mistral, AWS Bedrock, Azure OpenAI, and more
  • Model cycling — --models flag and Ctrl+P for model rotation
  • Thinking levels — --thinking off|minimal|low|medium|high|xhigh
  • Agent profiles — full, plan, iosm + advanced profiles (explore, iosm_analyst, iosm_verifier, cycle_planner)
  • IOSM workspace — iosm init bootstraps iosm.yaml, IOSM.md, .iosm/ artifacts
  • IOSM cycles — iosm cycle plan|report|status|list for systematic improvement
  • IOSM metrics — Six canonical metrics (semantic, logic, performance, simplicity, modularity, flow) with IOSM-Index
  • Subagent orchestration — Parallel/sequential delegation with /orchestrate
  • Custom agents — Markdown agent definitions in .iosm/agents/
  • Extension system — TypeScript extensions with tools, commands, hooks, and UI
  • Skills — Markdown workflow modules
  • Prompt templates — Reusable prompt snippets as slash commands
  • Themes — JSON TUI theme customization
  • Package manager — iosm install|remove|update|list from npm/git/local
  • Session persistence — Automatic session saving and recovery
  • Session branching — /tree, /fork for conversation tree navigation
  • HTML export — /export with themed, self-contained HTML output
  • Session sharing — /share via GitHub Gists
  • Session trace — --session-trace for full JSONL audit logging
  • Context compaction — Automatic context summarization when approaching limits
  • File attachments — @file syntax for including files in prompts
  • Programmatic SDK — createAgentSession() for embedding in custom apps
  • Keybinding customization — ~/.iosm/agent/keybindings.json
  • OAuth support — /login for provider authentication
  • Permission controls — /permissions and /yolo for tool approval management

Changed

  • Switched from monorepo tsgo scripts to standalone tsc-based build system
  • Promoted CLI source, docs, examples, and tests to repository root

Documentation

  • Comprehensive README with feature showcase, provider reference, and examples
  • 10 detailed documentation files covering all features
  • CONTRIBUTING.md with development setup and PR guidelines
  • 66 extension examples with categorized README
  • 12 SDK examples with quick reference guide