@iflow-mcp/jkheadley-instar 0.26.2

package/upgrades/0.24.20.md

@@ -0,0 +1,45 @@
+# Upgrade Guide — v0.24.20
+
+<!-- bump: patch -->
+
+## What Changed
+
+### Fix: Slack Socket Mode Heartbeat Death Loop
+
+The heartbeat timeout for Slack Socket Mode connections was 60 seconds — far too aggressive for quiet channels with no activity. This caused a connect-disconnect loop every ~90 seconds, dropping messages during reconnection windows. Increased to 5 minutes, which correctly detects zombie connections without false-positiving on idle channels.
+
+### Slack Workspace Modes (Dedicated vs Shared)
+
+SlackAdapter now supports two workspace modes:
+- **dedicated** (default): Agent owns the workspace. Auto-joins all public channels, responds to all messages.
+- **shared**: Agent is a guest in an existing workspace. Only responds to @mentions and DMs. Does not auto-join channels.
+
+Configure via `slack.workspaceMode` in config. The `respondMode` and `autoJoinChannels` settings can also be set independently.
+
+### Fix: GitSync Crash on Empty Repos
+
+GitSync now handles repos with no commits (freshly `git init`'d). Previously, `rev-parse HEAD` would throw and crash the sync process.
+
+### Fix: ServerSupervisor Grace Period After Restart
+
+When the Supervisor starts and finds the server already running (e.g., after a Lifeline self-restart for an update), it now correctly sets `spawnedAt` and checks for planned-exit markers. This prevents false "server down" alerts during update transitions.
+
+### Slack Image/Document Tags in Session Bootstrap
+
+`[image:path]` and `[document:path]` tags in Slack messages are now transformed into explicit read instructions before being passed to Claude sessions, improving reliability of file handling.
+
+### Browsable Stat Cards — Click Metrics to View Content
+
+Stat cards in the Systems detail view are now interactive. When a metric maps to browsable content, clicking the card fetches and displays the actual data inline.
+
+## What to Tell Your User
+
+A few reliability fixes in this release. If you were having trouble with Slack integration (agent not responding, connection dropping), that's fixed now — the connection was being too aggressive about reconnecting. Also added support for "shared" workspace mode if you want to install an agent into an existing Slack workspace without it taking over every channel.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Slack workspace modes | Set `slack.workspaceMode: "shared"` for guest-in-existing-workspace behavior |
+| Slack mention-only mode | Set `slack.respondMode: "mention-only"` to only respond to @mentions |
+| Browsable stat cards | Click any stat card with browsable content to expand it inline |

package/upgrades/0.24.21.md

@@ -0,0 +1,25 @@
+# v0.24.21 — Slack Stability & Image Support
+
+## What Changed
+
+**Slack Socket Mode heartbeat** — Fixed a reconnect death loop where the connection cycled every ~90 seconds on quiet channels. The heartbeat timeout is now 5 minutes instead of 60 seconds, since Slack only sends events when someone posts a message.
+
+**Workspace modes** — New workspaceMode config (dedicated or shared) with sensible defaults. Dedicated mode auto-joins channels and responds to all messages. Shared mode only responds when mentioned and does not auto-join. All settings individually overridable.
+
+**Image/file support in Slack** — Messages with file attachments are now downloaded and passed to sessions with explicit read instructions, matching the Telegram image pipeline.
+
+**Session injection race condition** — Fixed a timing issue where messages arriving shortly after session spawn could fail to inject because Claude was not ready yet. Now waits for Claude readiness before injection.
+
+**Smart port allocation** — New agents scan for free ports instead of blindly defaulting to 4040 when registry allocation fails.
+
+## What to Tell Your User
+
+Slack got more reliable this update. The bot no longer disconnects and reconnects in a loop on quiet channels. Images and files sent in Slack are now fully supported, just like Telegram. There is also a new workspace mode setting that controls whether the bot responds to every message or only when mentioned.
+
+## Summary of New Capabilities
+
+- Slack workspace modes: dedicated (respond to everything, auto-join channels) and shared (mention-only, no auto-join)
+- Slack image and file attachment support with automatic download and session delivery
+- Socket Mode connection stability fix (5-minute heartbeat instead of 60 seconds)
+- Session injection race condition fix for rapid message sequences
+- Smart port allocation for new agent initialization

package/upgrades/0.24.22.md

@@ -0,0 +1,35 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+<!-- Valid values: patch, minor, major -->
+<!-- patch = bug fixes, refactors, test additions, doc updates -->
+<!-- minor = new features, new APIs, new capabilities (backwards-compatible) -->
+<!-- major = breaking changes to existing APIs or behavior -->
+
+## What Changed
+
+<!-- Describe what changed technically. What new features, APIs, behavioral changes? -->
+<!-- Write this for the AGENT — they need to understand the system deeply. -->
+
+## What to Tell Your User
+
+<!-- Write talking points the agent should relay to their user. -->
+<!-- This should be warm, conversational, user-facing — not a changelog. -->
+<!-- Focus on what THEY can now do, not internal plumbing. -->
+<!--                                                                    -->
+<!-- PROHIBITED in this section (will fail validation):                 -->
+<!--   camelCase config keys: silentReject, maxRetries, telegramNotify -->
+<!--   Inline code backtick references like silentReject: false        -->
+<!--   Fenced code blocks                                              -->
+<!--   Instructions to edit files or run commands                      -->
+<!--                                                                    -->
+<!-- CORRECT style: "I can turn that on for you" not "set X to false"  -->
+<!-- The agent relays this to their user — keep it human.              -->
+
+- **[Feature name]**: "[Brief, friendly description of what this means for the user]"
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| [Capability] | [Endpoint, command, or "automatic"] |

package/upgrades/0.24.23.md

@@ -0,0 +1,17 @@
+# v0.24.23 — Slack Connection Stability
+
+## What Changed
+
+**Heartbeat timeout increased to 1 hour** — The Socket Mode heartbeat was causing reconnection loops on quiet channels. At 5 minutes, it would kill healthy connections and messages arriving during the 13-second reconnect window were silently lost. Now set to 1 hour since WebSocket close events handle real disconnections instantly. The heartbeat is only a zombie-connection detector for edge cases.
+
+**Debug logging removed** — Verbose raw WebSocket message logging and envelope dumps added during Slack development have been cleaned up.
+
+## What to Tell Your User
+
+The Slack connection should be rock-solid now. Previously, quiet channels could cause the bot to briefly disconnect every few minutes, which sometimes caused it to miss messages. That is fixed — the connection stays alive for hours without interruption.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Stable Slack connections | Automatic — no configuration needed |

package/upgrades/0.24.24.md

@@ -0,0 +1,15 @@
+# v0.24.24 — Slack Standby Updates
+
+## What Changed
+
+**PresenceProxy wired to Slack** — Slack sessions now receive the same tiered standby updates (🔭) that Telegram gets. When a user sends a message in Slack and the agent takes time to respond, they see real-time status updates about what the agent is doing. Uses a synthetic ID bridge to map Slack channel IDs into the existing PresenceProxy without refactoring its internals.
+
+## What to Tell Your User
+
+When you send a message in Slack and the agent is working on it, you will now see status updates showing what the agent is doing. These appear as messages prefixed with a telescope icon. Previously this only worked in Telegram.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Slack standby updates | Automatic — fires when agent takes more than 20 seconds to respond in a Slack channel |

package/upgrades/0.24.25.md

@@ -0,0 +1,15 @@
+# v0.24.25 — Fix Slack Standby Spam
+
+## What Changed
+
+**Removed Telegram-to-Slack standby mirroring** — Previously, Telegram standby messages were mirrored to Slack if the session had a Slack channel. This caused standby updates to appear in Slack even when no one had sent a message there. Now Slack only gets standby from its own direct PresenceProxy wiring, which correctly triggers only after a user message in that Slack channel.
+
+## What to Tell Your User
+
+Fixed an issue where status updates were appearing in Slack channels even when you had not sent a message there. Standby updates now only appear in the channel where you actually sent a message.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Platform-isolated standby | Automatic — each platform only shows standby for its own conversations |

package/upgrades/0.24.26.md

@@ -0,0 +1,15 @@
+# v0.24.26 — Fix Slack Standby Cancellation
+
+## What Changed
+
+**Standby cancellation on agent response** — The PresenceProxy never saw agent responses in Slack because bot messages returned through Socket Mode with bot_id set, getting filtered before reaching the event handler. This meant standby timers ran indefinitely even after the bot had responded. Now the reply route notifies the PresenceProxy when the agent sends a response, properly cancelling the standby cycle.
+
+## What to Tell Your User
+
+Fixed an issue where the status updates kept appearing in Slack even after the bot had already responded. The status updates now properly stop once the bot sends its reply.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Proper standby cancellation in Slack | Automatic — standby stops as soon as the agent responds |

package/upgrades/0.24.27.md

@@ -0,0 +1,17 @@
+# Upgrade Guide — v0.24.27
+
+<!-- bump: patch -->
+
+## What Changed
+
+Fixed the relay cold-spawn path to use the full sender fingerprint (32-char hex) instead of a truncated 8-char display name when constructing the message envelope. Previously, the spawned Claude session received a partial agent ID in its prompt, causing threadline_send replies to fail because the relay couldn't match the truncated ID to any registered agent.
+
+## What to Tell Your User
+
+- **Threadline relay replies now work**: "If you've had issues with agents not being able to reply to relay messages, that's fixed. Spawned sessions now get the full agent address so replies route correctly."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Reliable relay reply routing | Automatic — spawned sessions now use full fingerprints |

package/upgrades/0.24.28.md

@@ -0,0 +1,35 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+<!-- Valid values: patch, minor, major -->
+<!-- patch = bug fixes, refactors, test additions, doc updates -->
+<!-- minor = new features, new APIs, new capabilities (backwards-compatible) -->
+<!-- major = breaking changes to existing APIs or behavior -->
+
+## What Changed
+
+<!-- Describe what changed technically. What new features, APIs, behavioral changes? -->
+<!-- Write this for the AGENT — they need to understand the system deeply. -->
+
+## What to Tell Your User
+
+<!-- Write talking points the agent should relay to their user. -->
+<!-- This should be warm, conversational, user-facing — not a changelog. -->
+<!-- Focus on what THEY can now do, not internal plumbing. -->
+<!--                                                                    -->
+<!-- PROHIBITED in this section (will fail validation):                 -->
+<!--   camelCase config keys: silentReject, maxRetries, telegramNotify -->
+<!--   Inline code backtick references like silentReject: false        -->
+<!--   Fenced code blocks                                              -->
+<!--   Instructions to edit files or run commands                      -->
+<!--                                                                    -->
+<!-- CORRECT style: "I can turn that on for you" not "set X to false"  -->
+<!-- The agent relays this to their user — keep it human.              -->
+
+- **[Feature name]**: "[Brief, friendly description of what this means for the user]"
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| [Capability] | [Endpoint, command, or "automatic"] |

package/upgrades/0.24.29.md

@@ -0,0 +1,15 @@
+# v0.24.29 — Fix Stuck Slack Sessions
+
+## What Changed
+
+**Kill stuck sessions instead of losing messages** — When a Slack session gets stuck on a permissions prompt or tool hang, the server previously injected the user's message anyway, which silently failed. Now it detects the stuck session, kills it, and spawns a fresh one so the user's message is always processed. This prevents the scenario where a user sends messages and gets checkmarks but never receives a response.
+
+## What to Tell Your User
+
+Fixed an issue where messages in Slack could get lost if the bot session was stuck. The bot now automatically recovers by restarting the session when it detects a problem, so your messages always get processed.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Automatic stuck session recovery | Automatic — stuck sessions are killed and respawned on new messages |

package/upgrades/0.24.30.md

@@ -0,0 +1,40 @@
+# Upgrade Guide — v0.24.28
+
+<!-- bump: patch -->
+
+## What Changed
+
+### Fix: Slack Sessions Losing Conversation Context
+
+Slack sessions were spawning with incomplete context — the ring buffer only stored user messages, not bot (agent) responses. When a session died and respawned, the new session had no idea what the agent had previously said or done. This caused agents to repeat greetings, forget what was being discussed, and generally "lose the plot" after every message.
+
+Three changes fix this:
+
+1. Bot messages are now stored in the ring buffer alongside user messages, giving new sessions visibility into both sides of the conversation.
+2. The context file now uses a human-readable thread history format (matching the Telegram format) instead of raw JSON, making it easier for Claude to understand the conversation flow.
+3. Bot messages are correctly labeled as "Agent" in the history, so the session knows which messages are its own.
+
+## What to Tell Your User
+
+If your Slack agent was losing context between messages — repeating itself, forgetting what you were discussing, or greeting you fresh every time — that should be fixed now. The agent now sees the full conversation history (both your messages and its own replies) when picking up a conversation.
+
+### Fix: Topic Creation API Now Supports Initial Messages
+
+The `POST /telegram/topics` endpoint now accepts an optional `firstMessage` parameter. When provided, the message is sent to the new topic through the standard `sendToTopic` pipeline, which means it gets properly logged to both the JSONL message log and TopicMemory.
+
+Previously, sessions that created topics and sent initial messages via direct Telegram API calls would bypass the logging pipeline. When a new session later spawned in that topic, the thread history was empty — the session had no idea what the topic was about.
+
+The endpoint also now uses `findOrCreateForumTopic` instead of `createForumTopic`, preventing duplicate topics when called with the same name.
+
+## What to Tell Your User
+
+If your Slack agent was losing context between messages — repeating itself, forgetting what you were discussing, or greeting you fresh every time — that should be fixed now. The agent now sees the full conversation history (both your messages and its own replies) when picking up a conversation.
+
+Also, new topics created by the agent will now properly carry their initial context message into any sessions that spawn there.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Full Slack conversation context | Automatic — sessions now see both sides of the conversation |
+| Topic creation with context | `POST /telegram/topics` with `firstMessage` parameter |

package/upgrades/0.24.31.md

@@ -0,0 +1,45 @@
+# Upgrade Guide — v0.24.31
+
+<!-- bump: patch -->
+
+## What Changed
+
+**Autonomous mode structural enforcement fix.** The `/autonomous` skill's stop hook — the mechanism that prevents Claude from exiting and forces it to keep working through the task list — was never being deployed or registered for agents.
+
+The skill.md and setup script were installed correctly, but the `hooks/` directory (containing `autonomous-stop-hook.sh` and `hooks.json`) was missing, and the hook was never registered in `.claude/settings.json`. This meant autonomous sessions had zero structural enforcement: Claude would work based on prompt instructions, finish a response, and stop normally. The zombie detector would then kill the idle session after 15 minutes.
+
+**Sleep/wake crash loop fix.** The supervisor was killing the server on every wake because health checks failed briefly during the system transition. Three fixes:
+
+1. **Wake-tolerant supervisor** (`ServerSupervisor.ts`) — Added a 60-second wake transition window. During this window, if health checks fail but the server tmux session is still alive, the supervisor waits instead of killing. Also lowered SleepWakeDetector drift threshold from 10s to 5s to catch shorter sleeps.
+
+2. **Caffeinate leak fix** (`CaffeinateManager.ts`) — `cleanupStale()` now kills ALL orphaned caffeinate processes via `pgrep -x caffeinate`, not just the single PID from the PID file. Orphans accumulated from crash restarts where `stop()` never ran.
+
+3. **Session server restart prevention** (`cli.ts`) — `instar server start/stop/restart` now rejects calls from inside sessions (checks `INSTAR_SESSION_ID` env var). Also fixed the `fix restart` command writing `restart-requested.json` to the wrong directory.
+
+**Autonomous mode changes (three changes):**
+
+1. **PostUpdateMigrator: `ensureAutonomousStopHook()`** — New migration that deploys the stop hook files to `.claude/skills/autonomous/hooks/` and registers the hook in `settings.json` Stop chain (first position). Runs automatically on server startup when version changes. Idempotent.
+
+2. **init.ts: `installAutonomousSkill()`** — New agents now get the full autonomous skill deployed during init, including hooks directory and scripts. The hook is also registered in `installClaudeSettings()`.
+
+3. **Tests** — 17 new tests across 3 files:
+   - `autonomous-skill-deployment.test.ts` (14 tests) — Full skill deployment verification
+   - `fresh-install.test.ts` (+2 tests) — New project autonomous hook verification
+   - `migration-parity.test.ts` (+1 test) — Ensures migration includes autonomous hook deployment
+
+## What to Tell Your User
+
+- **Sleep/wake fix**: "Your agent should no longer crash-loop when your laptop wakes from sleep. The supervisor now waits for the server to recover instead of killing it during the transition."
+- **Caffeinate cleanup**: "Cleaned up orphaned caffeinate processes. They were accumulating from server crash restarts — this won't happen anymore."
+- **Autonomous mode fix**: "I found why autonomous sessions were dying after 30 minutes instead of running for hours. The stop hook that keeps me working through the task list was never actually installed. It's fixed now — autonomous sessions should properly loop until all tasks are done or the timer runs out."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Wake-tolerant supervisor | Automatic — 60s grace window after sleep/wake detection |
+| Caffeinate orphan cleanup | Automatic — kills all orphaned caffeinate on startup |
+| Session server restart guard | Automatic — blocks `instar server start/stop/restart` from sessions |
+| Autonomous stop hook deployment | Automatic — deployed during migration on update |
+| Autonomous stop hook registration | Automatic — registered in settings.json Stop chain |
+| New agent autonomous skill | Automatic — full skill deployed during `instar init` |

package/upgrades/0.24.32.md

@@ -0,0 +1,19 @@
+# Upgrade Guide — v0.24.32
+
+<!-- bump: patch -->
+
+## What Changed
+
+**Autonomous stop hook migration for existing agents.** v0.24.31 added the autonomous stop hook for new agents (via init), but existing agents didn't get the PostUpdateMigrator migration. This release adds `ensureAutonomousStopHook()` to the migrator so existing agents get the hook deployed and registered automatically on server restart.
+
+Also fixes execSync security test exemptions for commands/init.ts, commands/setup.ts, and core/AgentRegistry.ts.
+
+## What to Tell Your User
+
+- **Autonomous mode fix (existing agents)**: "Autonomous mode should now work properly. The stop hook that keeps sessions running through task lists is now automatically deployed when you update."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Autonomous stop hook migration | Automatic — runs on server restart after update |

package/upgrades/0.24.33.md

@@ -0,0 +1,35 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+<!-- Valid values: patch, minor, major -->
+<!-- patch = bug fixes, refactors, test additions, doc updates -->
+<!-- minor = new features, new APIs, new capabilities (backwards-compatible) -->
+<!-- major = breaking changes to existing APIs or behavior -->
+
+## What Changed
+
+<!-- Describe what changed technically. What new features, APIs, behavioral changes? -->
+<!-- Write this for the AGENT — they need to understand the system deeply. -->
+
+## What to Tell Your User
+
+<!-- Write talking points the agent should relay to their user. -->
+<!-- This should be warm, conversational, user-facing — not a changelog. -->
+<!-- Focus on what THEY can now do, not internal plumbing. -->
+<!--                                                                    -->
+<!-- PROHIBITED in this section (will fail validation):                 -->
+<!--   camelCase config keys: silentReject, maxRetries, telegramNotify -->
+<!--   Inline code backtick references like silentReject: false        -->
+<!--   Fenced code blocks                                              -->
+<!--   Instructions to edit files or run commands                      -->
+<!--                                                                    -->
+<!-- CORRECT style: "I can turn that on for you" not "set X to false"  -->
+<!-- The agent relays this to their user — keep it human.              -->
+
+- **[Feature name]**: "[Brief, friendly description of what this means for the user]"
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| [Capability] | [Endpoint, command, or "automatic"] |

package/upgrades/0.24.34.md

@@ -0,0 +1,29 @@
+# Upgrade Guide — v0.24.34
+
+<!-- bump: patch -->
+
+## What Changed
+
+**Session error nudge: automatic recovery from API errors.** When a session hits an API error (400 "Could not process image", rate limits, timeouts, network errors) and goes idle at the prompt, the session manager now detects the error in the terminal output and injects a nudge message to get the session working again.
+
+Previously, API errors that caused Claude to stop and return to the idle prompt were invisible to the recovery system — it only analyzed the JSONL conversation log, which looked clean after such errors. The zombie detector would kill the session 15 minutes later with no recovery.
+
+**How it works:**
+1. SessionManager's monitor tick detects a session is newly idle at the prompt
+2. Captures 30 lines of terminal output and checks for error patterns (API Error, invalid_request_error, rate_limit_error, timeouts, network errors)
+3. If an error is found: injects "You hit an API error. Please continue your work — skip or work around the action that failed."
+4. Resets the idle timer so the zombie detector doesn't kill immediately
+5. Each session gets ONE nudge — if it goes idle again after the nudge, the normal 15-minute zombie kill proceeds
+
+Also expanded crash-detector.ts error classification to include 400, 502, invalid_request_error, and overloaded_error as API error types.
+
+## What to Tell Your User
+
+- **Session error recovery**: "If your agent hits an API error mid-work (like an image it can't process), it will now automatically get nudged to continue instead of just stopping. Previously, those errors would leave the session dead at the prompt until the zombie detector cleaned it up 15 minutes later."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Session error nudge | Automatic — detects API errors in terminal output and nudges to continue |
+| Expanded error classification | Automatic — 400, 502, overloaded errors now classified as API type |

package/upgrades/0.24.4.md

@@ -0,0 +1,19 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+<!-- Valid values: patch, minor, major -->
+<!-- patch = bug fixes, refactors, test additions, doc updates -->
+<!-- minor = new features, new APIs, new capabilities (backwards-compatible) -->
+<!-- major = breaking changes to existing APIs or behavior -->
+
+## What Changed
+
+<!-- Document new features, breaking changes, and improvements here -->
+
+## What to Tell Your User
+
+<!-- Key points to communicate to end users about this upgrade -->
+
+## Summary of New Capabilities
+
+<!-- Brief list of new capabilities added in this version -->

package/upgrades/0.24.5.md

@@ -0,0 +1,20 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+<!-- Valid values: patch, minor, major -->
+<!-- patch = bug fixes, refactors, test additions, doc updates -->
+<!-- minor = new features, new APIs, new capabilities (backwards-compatible) -->
+<!-- major = breaking changes to existing APIs or behavior -->
+
+## What Changed
+
+- **PermissionRequest auto-approve hook**: Subagents spawned via the Agent tool don't inherit `--dangerously-skip-permissions` from the parent session. This caused permission prompts that blocked autonomous work. A new catch-all `PermissionRequest` hook (`auto-approve-permissions.js`) unconditionally approves all permission requests. Real safety remains in PreToolUse hooks.
+- The hook is automatically added to `settings.json` during init and on upgrade via PostUpdateMigrator.
+
+## What to Tell Your User
+
+No user-facing changes. This fix ensures autonomous sessions and subagents run without interruption.
+
+## Summary of New Capabilities
+
+- Auto-approve permissions for subagent sessions (prevents blocking on tool use prompts)

package/upgrades/0.25.0.md

@@ -0,0 +1,34 @@
+# v0.25.0 — Slack Feature Parity
+
+## What Changed
+
+**Comprehensive Slack parity release** — 18 features implemented to match Telegram's functionality in Slack.
+
+**Slash commands** — New commands available via the ! prefix: claim (bind sessions to channels), unlink, interrupt (nudge stuck sessions), restart, status. These match the Telegram equivalents.
+
+**Stall detection** — Full stall tracking pipeline for Slack sessions. Tracks injected messages, detects when sessions go quiet, and alerts both the originating channel and the attention channel.
+
+**Voice messages** — Audio files sent in Slack are automatically transcribed using Whisper (via Groq or OpenAI) and delivered as voice transcripts to sessions.
+
+**Session reliability** — Messages now inject through SessionManager for proper idle timer reset and bracketed paste mode. Proactive UUID save on session kill enables seamless resume. Delivery confirmation shows when messages are received.
+
+**Standby commands** — Users can type "unstick", "quiet", "resume", or "restart" directly in Slack channels to control the PresenceProxy standby system.
+
+**Safety** — Sentinel intercept classifies Slack messages for emergency stop and pause signals. Unknown users receive an ephemeral "not authorized" message instead of silent drops.
+
+**Infrastructure** — TopicMemory dual-write for search and summaries. All notification tiers route to Slack attention channel. Dedicated updates channel created automatically. Long messages use temp files to prevent tmux issues.
+
+## What to Tell Your User
+
+Slack now has feature parity with Telegram. You can control sessions with commands like unstick, restart, and interrupt. Voice messages are transcribed automatically. The bot will tell you when your message is delivered, alert you if a session stalls, and show status updates while working on long tasks. Unknown workspace members get a friendly "not authorized" message instead of being silently ignored.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Session commands | Type !claim, !unlink, !interrupt, !restart, !status in any channel |
+| Standby commands | Type unstick, quiet, resume, or restart during standby |
+| Voice transcription | Send an audio file in Slack |
+| Stall alerts | Automatic after 5 minutes of silence |
+| Delivery confirmation | Automatic on every message |
+| Emergency stop | Type "stop everything" in any channel |

package/upgrades/0.25.1.md

@@ -0,0 +1,24 @@
+# Upgrade Guide — v0.25.1
+
+<!-- bump: patch -->
+
+## What Changed
+
+**Standby mode now detects and reports quota exhaustion.** When a session hits Claude's API usage limit ("You've hit your limit - resets 7pm"), the PresenceProxy now detects the quota message in the terminal output and reports it clearly to the user instead of giving generic "agent is still working" status updates.
+
+Previously, when quota was hit, the standby system would send misleading LLM-generated updates like "echo is actively working" when in reality the session was frozen waiting for quota to reset. This wasted the user's time and eroded trust in the standby system.
+
+The fix adds a pre-LLM check in all three standby tiers. When quota exhaustion patterns are detected in the terminal output:
+- The LLM call is skipped entirely (saves tokens when quota is the issue)
+- A clear message is sent: "The agent has hit its Claude API usage limit. Quota resets 7pm (America/Los_Angeles). The session is paused until then — no work is being done."
+- Further tier escalation is cancelled (nothing more to assess)
+
+## What to Tell Your User
+
+- **Quota detection in standby**: "When your agent hits the Claude API limit mid-work, the standby system will now tell you clearly instead of pretending the agent is still working. You'll see the exact reset time so you know when work will resume."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Quota exhaustion detection in standby | Automatic — detects quota limit in terminal output across all standby tiers |

package/upgrades/0.25.10.md

@@ -0,0 +1,26 @@
+# v0.25.10 — Robustness Hardening
+
+## What Changed
+
+Four reliability improvements that prevent the cascade failure pattern where a stuck git rebase takes down an entire agent for days:
+
+1. **GitSync pre-flight recovery** — Before every pull, git-sync now checks for stuck rebases and detached HEAD states. If found, it automatically aborts the rebase and re-attaches to the main branch. Previously, a stuck rebase from an interrupted pull would silently block all future syncs.
+
+2. **ServerSupervisor git state healing** — The supervisor's preflight self-heal now detects and aborts stuck git rebases before spawning the server. This catches the problem even earlier in the startup chain.
+
+3. **Wider slow-retry window** — The circuit breaker's slow-retry mode now uses a 60-second detection window (up from 10 seconds). The old window was so narrow that retries could be missed entirely if health check timing didn't align with the 2-hour boundary.
+
+4. **AgentRegistry sync lock retries** — Synchronous registry operations now retry up to 5 times with exponential backoff when encountering lock contention. Previously, sync lock acquisition had zero retries and would fail immediately, causing constant heartbeat errors on machines running multiple agents.
+
+5. **Shell detection fix** — The supervisor's shell execution helper now uses the SHELL environment variable instead of hardcoding /bin/sh, which could fail in certain environments.
+
+## What to Tell Your User
+
+We fixed a reliability issue where a stuck git rebase could take down an agent for days. The system now automatically detects and recovers from stuck rebases, and the retry logic is more resilient. If you've seen agents go stale and not recover, this update addresses that.
+
+## Summary of New Capabilities
+
+- Automatic recovery from stuck git rebases during sync and server startup
+- More reliable agent registry operations under concurrent access from multiple agents
+- Wider retry windows so the circuit breaker doesn't silently miss recovery attempts
+- Better shell compatibility across different environments

package/upgrades/0.25.2.md

@@ -0,0 +1,23 @@
+# v0.25.2 — Close Telegram-Gated Code Path Gaps
+
+## What Changed
+
+**Watchdog notifications for Slack** — Session intervention and recovery alerts from the watchdog now go to the correct Slack channel, not just Telegram topics.
+
+**Resume heartbeat includes Slack** — The 60-second proactive resume UUID save now covers Slack channel-session mappings. Previously only Telegram topics were tracked, meaning Slack sessions lost resume context on server crash.
+
+**Graceful shutdown saves Slack resume** — On server shutdown (SIGINT/SIGTERM), resume UUIDs are now saved for both Telegram topics and Slack channels. Previously gated behind an if-telegram check.
+
+**Prompt Gate relay for Slack** — Detected prompts in Slack-bound sessions now relay to the Slack channel with interactive Block Kit buttons. Previously prompts only went to Telegram topics.
+
+## What to Tell Your User
+
+Several infrastructure systems that only worked with Telegram now also work with Slack. Watchdog alerts, session resume on restart, and prompt approvals all properly route to Slack channels now. This means Slack sessions are more reliable and recover better after server restarts.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Watchdog alerts in Slack | Automatic when sessions get stuck |
+| Resume across restarts | Automatic — sessions resume with context after server restart |
+| Prompt relay in Slack | Automatic — approval prompts appear in the session channel |

package/upgrades/0.25.3.md

@@ -0,0 +1,25 @@
+# v0.25.3 — Platform-Agnostic Session Health Systems
+
+## What Changed
+
+**StallTriageNurse works for Slack** — The LLM-powered session recovery system was previously gated behind an if-telegram check. Now it initializes for any messaging adapter and uses platform-aware callbacks for message history, topic/channel communication, and session respawning.
+
+**TriageOrchestrator works for Slack** — The next-generation session recovery system (with scoped diagnostic sessions) is no longer Telegram-exclusive. All callbacks resolve Slack channels via synthetic ID mapping.
+
+**SessionRecovery works for Slack** — The fast mechanical recovery path now routes kill commands through SessionManager (firing the beforeSessionKill hook for resume UUID saving) and handles Slack channel-bound sessions.
+
+**SessionMonitor works for Slack** — Proactive health monitoring now aggregates sessions from both Telegram topics and Slack channels. Alerts route to the correct platform.
+
+**Resume heartbeat includes Slack** — The 60-second proactive resume UUID refresh now covers Slack sessions.
+
+## What to Tell Your User
+
+Session health monitoring and automatic recovery now work for Slack sessions, not just Telegram. If a Slack session gets stuck, the same intelligent triage system that works for Telegram will detect and attempt to recover it.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Stall triage for Slack | Automatic — detects and recovers stuck Slack sessions |
+| Session monitoring for Slack | Automatic — proactive health checks for all sessions |
+| Session recovery for Slack | Automatic — mechanical fast-path recovery |

package/upgrades/0.25.4.md

@@ -0,0 +1,24 @@
+# v0.25.4 — Complete Stall Detection Parity for Slack
+
+## What Changed
+
+**LLM-gated stall alerts for Slack** — Before alerting users about a stalled session, the system now confirms with a lightweight LLM call (Haiku) that the alert is warranted. This prevents false positives for sessions working on complex tasks. Fail-open: if the LLM is unavailable, the alert is sent.
+
+**Promise tracking for Slack** — When the agent says something like "give me a minute" or "working on it," the system tracks this promise and alerts the user if the agent does not follow through within the configured timeout (default 10 minutes). Matches the existing Telegram behavior.
+
+**!triage command for Slack** — Users can now type !triage in a Slack channel to see whether the triage system is actively diagnosing the session. Shows classification, check count, and last check time.
+
+**Session death classification for Slack** — When a Slack session dies, the system can now classify the cause (quota exhaustion, timeout, error, etc.), enabling better user-facing error messages.
+
+## What to Tell Your User
+
+The bot is now smarter about when to alert you about session issues. It checks with an AI model before sending stall alerts to avoid false alarms. It also tracks when the agent promises to follow up and will remind you if it does not deliver. You can type !triage in any channel to check the health status.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| LLM-gated stall alerts | Automatic — reduces false positive alerts |
+| Promise tracking | Automatic — detects and tracks agent commitments |
+| Triage status | Type !triage in any Slack channel |
+| Session death classification | Automatic — classifies why sessions end |