@iflow-mcp/jkheadley-instar 0.26.2

package/upgrades/0.23.17.md

@@ -0,0 +1,23 @@
+# Upgrade Guide — NEXT
+
+<!-- bump: patch -->
+
+## What Changed
+
+- **Topic purpose awareness**: TopicSummarizer now extracts a `purpose` field alongside each summary — a one-line description of the topic's current focus. Sessions receive soft awareness ("Current focus: ...") so they know what the topic has been about without being constrained. Job sessions get stronger posting guidance via a `[TOPIC AWARENESS]` block to keep autonomous output contextually relevant.
+- **Prompt Gate**: Interactive prompt detection in sessions with auto-approval for safe operations and Telegram relay for risky ones. Includes InputClassifier, AutoApprover, and CallbackRegistry.
+- **Schema migration v4**: `topic_summaries` table gains a `purpose` column. Migration is automatic and non-breaking.
+
+## What to Tell Your User
+
+Sessions now understand what each topic is about. When a conversation drifts to a new subject, the purpose updates naturally on the next summary cycle. Job sessions are aware of where they're posting and will keep output relevant to the topic context. Interactive prompts in sessions can now be auto-approved or relayed to Telegram for review.
+
+## Summary of New Capabilities
+
+- Topic purpose extraction in TopicSummarizer (via LLM PURPOSE line parsing)
+- Soft "Current focus" awareness in session bootstrap
+- Job session topic awareness block in JobScheduler
+- TopicMemory schema v4 with purpose column + migration
+- Prompt Gate: InputDetector, InputClassifier, AutoApprover, CallbackRegistry
+- POST /topic/summary API accepts optional `purpose` field
+- GET /topic/context returns purpose in TopicContext

package/upgrades/0.23.18.md

@@ -0,0 +1,41 @@
+# Upgrade Guide — vNEXT
+
+## What Changed
+
+### Default Job Configuration Overhaul
+
+The default jobs template that ships with every new agent has been comprehensively reviewed and updated:
+
+- **Removed 4 legacy jobs**: `update-check`, `dispatch-check` (replaced by built-in AutoUpdater/AutoDispatcher), `self-diagnosis` (replaced by overseer pattern), `evolution-review` (split into two paired jobs)
+- **Added 9 new jobs**: `evolution-proposal-evaluate` (Sonnet), `evolution-proposal-implement` (Opus), `commitment-detection`, `dashboard-link-refresh`, and 5 category overseers (guardian, learning, maintenance, infrastructure, development)
+- **7 jobs migrated to skills**: coherence-audit, degradation-digest, state-integrity-check, memory-hygiene, guardian-pulse, session-continuity-check, git-sync — long inline prompts replaced with `.claude/skills/` files installed during init
+- **Priority bumps**: 7 jobs moved from `low` to `medium` or `high` to prevent quota-blocking
+- **Model right-sizing**: relationship-maintenance downgraded from opus to haiku (mechanical check), evolution-evaluate uses sonnet instead of opus
+- **Tag taxonomy**: All jobs now use standardized `cat:`, `role:`, `exec:` tags
+
+### Quota Threshold Improvements
+
+- **Raised baseline threshold** from 50% to 75% — jobs no longer get blocked at moderate usage
+- **Script bypass**: `exec:script` jobs (zero LLM tokens) skip quota gating entirely
+- **First-time notifications**: Users notified the first time each quota tier is crossed, with actionable messaging
+- **Migration blocking**: Only blocks during migration when there's actual quota pressure (prevents stale migration state from permanently blocking jobs)
+
+### New Built-in Skills
+
+7 new skills are installed automatically during `instar init`:
+- coherence-audit, degradation-digest, state-integrity-check, memory-hygiene, guardian-pulse, session-continuity-check, git-sync (with tiered model escalation)
+
+## What to Tell Your User
+
+- Jobs that were previously stuck due to low priority will now run reliably
+- Quota thresholds are more permissive — you'll see fewer "quota blocked" skips
+- The first time you cross a quota tier, you'll get a notification explaining what's happening
+- Five new overseer jobs provide meta-monitoring across job categories
+
+## Summary of New Capabilities
+
+- Paired audit-action job architecture (evaluate then implement)
+- Category overseers that spot cross-job patterns and contradictions
+- Commitment detection from Telegram messages
+- Tiered git-sync: haiku for clean merges, opus subagent only for complex conflicts
+- Script jobs bypass quota entirely (zero token cost)

package/upgrades/0.23.2.md

@@ -0,0 +1,32 @@
+# Upgrade Guide — v0.23.2
+
+<!-- bump: patch -->
+
+## What Changed
+
+**Session respawn loop fixed — spawn-in-progress guard + user-visible failure messages.**
+
+Previously, when a session died and kept failing to restart (due to session limit, quota exhaustion, or other errors), the system entered a silent respawn loop:
+- Each incoming Telegram message triggered a new respawn attempt
+- Respawn failures were only logged to console — never reported to the user
+- The user saw repeated "🔄 Session restarting — message queued" messages with no explanation
+
+Two fixes:
+
+1. **Spawn-in-progress guard**: A per-topic lock prevents multiple concurrent spawn attempts for the same topic. If a spawn is already in progress, subsequent messages receive "Session is still starting up — please wait a moment" instead of triggering another spawn.
+
+2. **Failure reporting**: When a respawn fails, the user now receives a Telegram message explaining what happened:
+   - Session limit errors: "❌ Session restart failed — session limit reached. Close an existing session or increase maxSessions in your config, then try again."
+   - Other errors: "❌ Session restart failed. Try sending your message again in a moment."
+
+## What to Tell Your User
+
+- **Session restart reliability**: "When my session dies and I can't restart it, I'll now tell you exactly what went wrong — including if you've hit the session limit — instead of just silently failing. Expect clearer error messages."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Respawn failure messages | Automatic — user receives error via Telegram when restart fails |
+| Session limit error guidance | Automatic — specific message when session limit prevents restart |
+| Spawn deduplication | Automatic — rapid messages during startup no longer create duplicate sessions |

package/upgrades/0.23.4.md

@@ -0,0 +1,21 @@
+# Upgrade Guide — v0.23.4
+
+<!-- bump: patch -->
+
+## What Changed
+
+**Bug fix: MessageSentinel LLM classifier no longer misclassifies conversational "hold on" as pause.**
+
+- Fixed MessageSentinel LLM classifier prompt: moved "hold on let me think" from the pause examples to the normal (conversational) examples. Messages like "hold on let me think" (5+ words that bypass the fast-path word gate) were being sent to the LLM which classified them as `pause` because the prompt listed this exact phrase as a pause example. The fix adds a KEY DISTINCTION instruction and clearer directive-only pause examples.
+- Fixed the fresh-install integration test to expect 21 default jobs (was 20) after the identity-review job was added to the default job set in the previous release. No runtime behavior changed — the test expectation was simply out of date.
+- Updated blog page with images, corrected publish date, synced with latest draft, and added canonical URL metadata.
+
+## What to Tell Your User
+
+- **Reliability**: "Fixed a false positive where saying 'hold on let me think' (or similar conversational phrases) could accidentally pause your agent session. The MessageSentinel now correctly distinguishes between user thought-narration and actual pause directives."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| No behavior changes | This is a bug fix release |

package/upgrades/0.23.6.md

@@ -0,0 +1,19 @@
+# Upgrade Guide — v0.23.6
+
+<!-- bump: patch -->
+
+## What Changed
+
+The AgentRegistry heartbeat now includes automatic recovery from stale lock files. After a process crash (such as the mutex abort during auto-update restart), the `proper-lockfile` lock on `registry.json` could get stuck indefinitely, causing every heartbeat to fail with "Lock file is already being held." The heartbeat now tracks consecutive failures and after 3 failures, force-removes the stale lock and retries. This makes the registry self-healing after crash scenarios.
+
+A new `forceRemoveRegistryLock()` function is exported for programmatic use.
+
+## What to Tell Your User
+
+- **Self-healing registry**: "If you were seeing repeated 'Lock file is already being held' errors in the logs, that's now fixed. The registry automatically recovers from stale locks after a crash — no manual intervention needed."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Registry lock auto-recovery | Automatic — kicks in after 3 consecutive heartbeat failures |

package/upgrades/0.23.7.md

@@ -0,0 +1,33 @@
+# Upgrade Guide — v0.23.7
+
+<!-- bump: patch -->
+
+## What Changed
+
+**Threadline agent-to-agent messaging fixes:**
+
+Local-first delivery for co-located agents. The `relay-send` endpoint now detects when the target agent is on the same machine and delivers directly via HTTP (`/messages/relay-agent`), bypassing the cloud relay entirely. This eliminates stale WebSocket connection issues that caused "delivered" messages to silently fail after server restarts.
+
+The cold-spawn prompt now correctly instructs spawned sessions to use the `threadline_send` MCP tool for replies (previously referenced a nonexistent `/msg reply` command). Template variable substitution also fixed — `{remote_agent}` and `{thread_id}` are now properly replaced in all occurrences.
+
+Relay auth rate-limit handling improved — when the relay rejects with "Too many auth attempts," the client now enforces longer backoff (~32s) before retrying, preventing retry storms during rapid server restarts.
+
+Also: CLI commands for inspecting job execution history and continuity data (`instar job history`, `instar job handoff`). Job execution now supports handoff notes for cross-execution continuity. New usage-based reflection metrics system. Test infrastructure improvements and a separate publish workflow for the threadline-mcp subpackage.
+
+## What to Tell Your User
+
+- **Threadline local delivery**: "Agent-to-agent messaging on the same machine now works reliably. Messages are delivered directly between agents without going through the cloud relay, which means no more silent delivery failures after server restarts."
+- **Threadline reply fix**: "Agents can now actually reply to threadline messages. The spawned sessions know to use the threadline_send MCP tool instead of a command that didn't exist."
+- **Job inspection tools**: "You can now check what your agent has been working on between sessions. The new job history and handoff commands show execution records and continuity notes."
+- **Reflection monitoring**: "Your agent now tracks reflection frequency, so you can see how often it pauses to learn from its work."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Threadline local delivery | Automatic — same-machine agents deliver directly via HTTP |
+| Threadline reply fix | Automatic — spawned sessions use threadline_send MCP tool |
+| Relay auth backoff | Automatic — rate-limited auth retries use longer backoff |
+| Job execution history | `instar job history [job-slug]` |
+| Job handoff inspection | `instar job handoff [job-slug]` |
+| Usage-based reflection metrics | Automatic — tracked during agent operation |

package/upgrades/0.23.8.md

@@ -0,0 +1,38 @@
+# Upgrade Guide — v0.23.8
+
+<!-- bump: patch -->
+
+## What Changed
+
+Three fixes for Threadline agent-to-agent messaging reliability:
+
+**Local-first delivery for co-located agents.** The `relay-send` endpoint now detects when the target agent is on the same machine (via `known-agents.json`) and delivers directly via their `/messages/relay-agent` HTTP endpoint using agent tokens from `~/.instar/agent-tokens/`. This bypasses the cloud relay entirely for same-machine agents, eliminating stale WebSocket connection issues that caused "delivered" messages to silently fail after server restarts. Falls back to relay if local delivery fails.
+
+**Cold-spawn prompt fix.** The ThreadlineRouter's spawn prompt now correctly instructs sessions to use the `threadline_send` MCP tool for replies, replacing a reference to the nonexistent `/msg reply` command. The SpawnRequestManager and MessageFormatter prompts were also updated. Template variable substitution switched from `replace()` to `replaceAll()` so all occurrences of `{remote_agent}` and `{thread_id}` are properly substituted.
+
+**Relay auth rate-limit backoff.** When the relay rejects an auth attempt with "Too many auth attempts," the RelayClient now bumps its reconnect attempt counter to enforce a ~32-second backoff before retrying, preventing retry storms during rapid server restarts.
+
+**Lifeline queue feedback loop prevention.** Three fixes to prevent the lifeline from getting stuck in a restart loop that floods Telegram with thousands of "Server is temporarily down" messages: (1) Queue acknowledgment messages are now rate-limited to one per topic per 2 minutes, and suppressed entirely when the queue exceeds 100 messages. (2) Messages that fail replay 3 times are dropped instead of re-queued indefinitely, preventing poison messages from crashing the server on every recovery. (3) Queue replay now stops immediately when the server goes unhealthy mid-replay, re-queuing remaining messages in bulk instead of processing them one-by-one.
+
+Also includes: CLI commands for inspecting job execution history and continuity data (`instar job history`, `instar job handoff`), handoff notes for cross-execution continuity, usage-based reflection metrics, test infrastructure improvements, and a separate publish workflow for the threadline-mcp subpackage.
+
+## What to Tell Your User
+
+- **Reliable agent-to-agent messaging**: "Agents on the same machine can now talk to each other reliably. Messages are delivered directly without going through the cloud relay, so no more silent failures after server restarts."
+- **Agents can reply**: "When one agent messages another, the receiving agent now knows how to reply properly. Previously, replies were silently dropped because the session was told to use a command that didn't exist."
+- **Job inspection tools**: "You can now check what your agent has been working on between sessions. The new job history and handoff commands show execution records and continuity notes."
+- **Reflection monitoring**: "Your agent now tracks reflection frequency, so you can see how often it pauses to learn from its work."
+- **No more restart spam**: "If the server gets stuck in a restart loop, the lifeline will no longer flood your Telegram with thousands of 'temporarily down' messages. Queue notifications are now rate-limited, and messages that keep failing to deliver are dropped after 3 attempts."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Threadline local delivery | Automatic for same-machine agents |
+| Threadline reply fix | Automatic in spawned sessions |
+| Relay auth backoff | Automatic on rate-limited connections |
+| Job execution history | `instar job history [job-slug]` |
+| Job handoff inspection | `instar job handoff [job-slug]` |
+| Usage-based reflection metrics | Automatic |
+| Lifeline queue ack rate-limiting | Automatic |
+| Lifeline poison message protection | Automatic (drops after 3 replay failures) |

package/upgrades/0.23.9.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/0.24.1.md

@@ -0,0 +1,32 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+
+## What Changed
+
+### Session Resume: HTTP hooks replaced with command hooks
+
+Claude Code HTTP hooks (`type: "http"` in settings.json) silently fail to fire as of v2.1.78. This caused `claudeSessionId` to never be populated, making session resume fall back to the unreliable mtime heuristic — cross-contaminating UUIDs across topics.
+
+All 9 HTTP hooks replaced with command hooks running `hook-event-reporter.js`, which reads the hook payload from stdin, extracts the Claude Code session UUID, and POSTs it to `/hooks/events`. Command hooks reliably fire on every tool call.
+
+`PostUpdateMigrator.migrateHttpHooksToCommandHooks()` automatically replaces HTTP hooks in existing agents and installs the reporter script.
+
+### Session resume: authoritative claudeSessionId
+
+`beforeSessionKill` and the resume heartbeat now use `session.claudeSessionId` (populated by hook events) instead of the mtime heuristic. The mtime fallback is only used when exactly one session is active (no ambiguity).
+
+### Tests
+
+44 new tests across 4 files, 115 total resume-related tests passing. Covers: HTTP-to-command migration, claudeSessionId bridging, migration parity, hook installation, and e2e resume flows.
+
+## What to Tell Your User
+
+- **Session resume is more reliable now.** Previously, killing a session and sending a new message sometimes started a fresh conversation instead of resuming. The mechanism that tracks which conversation belongs to which session has been rebuilt and is now confirmed working.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Reliable session resume | Automatic — sessions correctly save and restore conversation context |
+| HTTP to command hook migration | Automatic — existing agents get patched on next update |

package/upgrades/0.24.10.md

@@ -0,0 +1,23 @@
+# Upgrade Guide — v0.24.11
+
+<!-- bump: patch -->
+
+## What Changed
+
+- **Health endpoint no longer blocks on stale sessions**: The health check previously called synchronous tmux operations for every tracked session, blocking the event loop. With many stale sessions (common after restart loops), this caused health check timeouts, which triggered MORE restarts — a death spiral. Now uses a cached session count updated asynchronously by the monitor tick.
+
+- **Startup session purge**: Before monitoring starts, the server now does a fast one-pass purge of all session records whose tmux sessions no longer exist. Uses a 1-second timeout per session to fail fast. This prevents the startup overload where the server spent all its boot time polling dead sessions.
+
+- **CoherenceMonitor suppresses known-pending updates**: When the AutoUpdater has already applied a version and a restart is pending (deferred for active sessions), the CoherenceMonitor no longer flags the version mismatch as a failure. This eliminates the race condition where both systems fought over restart decisions.
+
+## What to Tell Your User
+
+- **Stability improvement**: "The server should be much more resilient to restart loops now. Even if many sessions build up, the health check stays fast and the server won't get stuck in a crash-restart cycle. Stale sessions are cleaned up instantly on boot instead of dragging things down."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Non-blocking health endpoint | Automatic — no action needed |
+| Startup session purge | Automatic — runs on every server boot |
+| CoherenceMonitor update suppression | Automatic — no action needed |

package/upgrades/0.24.12.md

@@ -0,0 +1,17 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+
+## What Changed
+
+- **Fix: User messages lost when Lifeline forwards** — When Lifeline won the Telegram polling race (during startup, sleep/wake, or server restarts), it consumed user messages and forwarded them via `/internal/telegram-forward`. That route delivered messages to sessions but never logged them to JSONL or TopicMemory. Agent responses were still logged, creating one-sided conversation histories. The forward route now records inbound messages before routing them.
+
+## What to Tell Your User
+
+- **Bug fix**: "Your messages in topic history should now be complete. There was a bug where messages forwarded by Lifeline weren't being saved to conversation memory — so the agent could respond to you, but the history only showed the agent's side. This is fixed now."
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Lifeline message logging | Automatic — no action needed |

package/upgrades/0.24.13.md

@@ -0,0 +1,16 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+
+## What Changed
+
+<!-- Describe changes in the next release here -->
+
+## What to Tell Your User
+
+<!-- Brief, user-friendly summary -->
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|

package/upgrades/0.24.14.md

@@ -0,0 +1,26 @@
+# Upgrade Guide — v0.24.14
+
+## What Changed
+
+### Server Supervisor Self-Healing
+
+The ServerSupervisor now runs preflight diagnostics before every server spawn attempt. Previously, if the shadow install was missing or the node symlink was broken, the server would crash immediately on startup — and restarting via `/lifeline restart` would just repeat the same failure.
+
+Now, before each spawn, the supervisor checks and auto-repairs:
+
+- **Missing shadow install** — reinstalls via npm automatically
+- **Broken node symlink** — recreates it from available node binaries
+- **Stale lifeline locks** — removes locks older than 10 minutes
+
+This means `/lifeline restart` now actually fixes the problem instead of blindly retrying a broken state. Agents on remote machines can self-recover from common post-update failures without manual SSH intervention.
+
+## What to Tell Your User
+
+Your agent's server is now more resilient. If the server goes down due to a missing or corrupted install, the system will automatically detect and fix the issue before restarting. The /lifeline restart command in Telegram now does real recovery, not just a blind retry.
+
+## Summary of New Capabilities
+
+- **Preflight self-healing**: Server supervisor diagnoses and fixes common startup failures before each spawn
+- **Shadow install auto-repair**: Missing shadow installs are reinstalled automatically
+- **Node symlink recovery**: Broken node symlinks are recreated from available binaries
+- **Stale lock cleanup**: Old lifeline locks are removed to prevent restart deadlocks

package/upgrades/0.24.15.md

@@ -0,0 +1,49 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: minor -->
+
+## What Changed
+
+### Slack Messaging Adapter
+
+Instar now includes a full Slack messaging adapter (`SlackAdapter`), enabling agents to send and receive messages via Slack workspaces. The adapter supports:
+
+- Sending messages to channels and threads
+- Token validation and secure credential handling
+- Atomic write operations for message persistence
+- CLI tooling for Slack interaction (`slack-cli`)
+- Server-side routes for Slack webhook integration
+
+The adapter follows the same pattern as existing messaging adapters (Telegram, WhatsApp) and plugs into the unified channel routing system.
+
+### Server-Side Process Review Dashboard
+
+A new Systems tab on the dashboard provides real-time visibility into server processes. This includes:
+
+- Process telemetry collection and display
+- Health status monitoring for running services
+- Integration with the existing dashboard infrastructure
+
+### Architecture Documentation
+
+A new "Under the Hood" page on the documentation site explains Instar's internal architecture, giving developers and agents a clearer picture of how the system is structured.
+
+### CI Stability Improvements
+
+Several CI fixes improve reliability: resolved flaky test exclusions, fixed WhatsApp string test, added `skipStallClear` option to prevent stall-clear interference during topic sends, and fixed test failures blocking the push gate.
+
+## What to Tell Your User
+
+- **Slack support is here**: Your agent can now send and receive messages through Slack. If your team uses Slack, you can connect it as a messaging channel alongside Telegram and other platforms. Ask your agent to set it up for you.
+
+- **Better visibility into what is running**: There is a new Systems tab on your dashboard that shows what server processes are active and their health status. Useful for troubleshooting if something feels off.
+
+- **More reliable updates**: Several under-the-hood fixes make the CI pipeline more stable, which means smoother and more frequent releases going forward.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Slack messaging | Configure Slack token via CLI, then messages route automatically |
+| Process review dashboard | Visit the Systems tab on your Instar dashboard |
+| Architecture docs | Check the "Under the Hood" page on the docs site |

package/upgrades/0.24.16.md

@@ -0,0 +1,48 @@
+# Upgrade Guide — v0.24.16
+
+## What Changed
+
+### Presence Proxy (Standby)
+
+New monitoring subsystem that provides intelligent status updates when an agent session is too busy to respond to Telegram messages.
+
+**Three-tiered response system:**
+- **Tier 1 (20s):** Haiku summarizes what the agent is currently doing based on terminal output
+- **Tier 2 (2min):** Haiku compares progress since Tier 1, reports what changed
+- **Tier 3 (5min):** Sonnet performs deep stall assessment — determines if the agent is genuinely stuck or running a legitimate long process. Offers user recovery options if stalled.
+
+**Conversation mode:** When the proxy is active and the user sends follow-up messages, the proxy can answer questions about what the agent is doing based on terminal output.
+
+**Security hardening:**
+- Tmux output sanitized before LLM calls (ANSI codes, credentials, injection patterns stripped)
+- LLM output guarded (URLs, commands, credential requests blocked)
+- Telegram sender authentication for action commands (unstick/restart/quiet)
+- Proxy messages don't reset StallDetector timers (isProxy flag)
+- Race condition guard checks message log before each tier fires
+
+**Integration:**
+- Coordinates with StallTriageNurse via triage mutex (prevents double-intervention)
+- State persists to disk for restart recovery
+- LLM calls capped at 3 concurrent with queue, rate-limited per topic
+
+**Configuration:** Activates automatically when `sharedIntelligence` and Telegram are available. Configurable under `monitoring.presenceProxy` in config.json.
+
+### StallDetector Integration Fix
+
+`POST /telegram/reply/:topicId` now accepts optional `metadata` field with `isProxy: true`. When set, the endpoint skips `clearInjectionTracker()` and `clearStallForTopic()` — ensuring proxy messages don't interfere with existing stall detection.
+
+### TelegramAdapter
+
+Added `skipStallClear` option to `sendToTopic()` to prevent proxy messages from resetting stall tracking timers.
+
+## What to Tell Your User
+
+Your agent now has a "standby assistant" that keeps you informed when the agent is busy working on something. If you send a message and don't get a response within 20 seconds, you'll see a 🔭 status update describing what the agent is doing. You can even ask follow-up questions and the standby will answer based on what it can see in the agent's terminal. At 5 minutes, it checks whether the agent is genuinely stuck and offers recovery options if needed.
+
+## Summary of New Capabilities
+
+- **Presence Proxy:** Tiered status updates (20s/2min/5min) when agent is busy
+- **Conversation mode:** Proxy can answer user questions about agent activity
+- **Stall assessment:** Intelligent stuck-vs-working detection at 5 minutes
+- **User commands:** quiet, resume, unstick, restart — with authentication
+- **42 e2e tests** covering all tiers, sanitization, conversation mode, and edge cases

package/upgrades/0.24.17.md

@@ -0,0 +1,40 @@
+# Upgrade Guide — v0.24.18
+
+## What Changed
+
+### Systems Dashboard v3 — Rich Card Grid with Detail Views
+
+The Systems tab has been completely redesigned from a cramped pill grid to an interactive card-based layout with drill-down detail views.
+
+**Overview page:**
+- Responsive card grid (2 columns on desktop, 1 on mobile)
+- Each card shows the capability name, a 2-line description, key stat chips, and a live metric summary
+- Issues panel only appears when there are actual problems
+- Events section collapsed by default to reduce noise
+
+**Detail views:**
+- Click any capability card to see its full detail page
+- Stat cards grid showing all numeric metrics individually
+- Component process list with status indicators
+- Last activity timestamp
+- Back navigation to return to overview
+
+**Rich telemetry backend:**
+- Each capability now returns live metrics: recoveries, interventions, coherence checks, memory pressure, quota usage, topic counts, evolution stats
+- New `GET /systems/capability/:id` endpoint for individual capability details
+- Richer capability descriptions (2-3 sentences each)
+
+### SlackAdapter: Missing Methods Fixed
+
+The SlackAdapter was missing several methods (channel-session registry, resume map, stall detection callback) that server routes depended on. This caused TypeScript build failures in CI. The methods are now committed and the build is clean.
+
+## What to Tell Your User
+
+The Systems dashboard got a major upgrade. Instead of a flat grid of status pills, it now shows interactive cards with live metrics for each capability — things like how many session recoveries happened, current quota usage, and Telegram connection uptime. You can click any card to drill into the full details. It's a much clearer picture of what your agent is actually doing at a glance.
+
+## Summary of New Capabilities
+
+- **Systems dashboard v3:** Interactive card grid with drill-down detail views and live metric summaries
+- **Per-capability telemetry:** Each system reports real-time stats (recoveries, quota %, topic counts, etc.)
+- **Capability detail endpoint:** `GET /systems/capability/:id` for individual system deep-dives
+- **SlackAdapter completeness:** All adapter methods now committed, CI build clean

package/upgrades/0.24.18-beta.0.md

@@ -0,0 +1,35 @@
+# v0.24.18-beta.0 — Slack Messaging Adapter (Beta)
+
+## What Changed
+
+Native Slack support joins Telegram and WhatsApp as a first-class messaging adapter. Your agent can now live in Slack — receiving messages via Socket Mode, responding in channels, and managing conversations with the same features Telegram has.
+
+This is a beta release for early testing. The core adapter is complete and working, with setup wizard support, session management, and feature parity across most Telegram capabilities.
+
+## What to Tell Your User
+
+Your agent can now talk to you through Slack. If you already use Slack for work, you can add it alongside Telegram — no need to switch. Setup takes about 5 minutes: you log into Slack, and the agent handles creating the workspace, app, and channels automatically.
+
+Slack sessions show up in the dashboard with their own platform badge, and you can create new sessions linked to Slack channels right from the dashboard. Commands work with an exclamation mark prefix — type !sessions or !new in any channel the agent is in.
+
+This also includes a new autonomous mode skill that lets the agent work independently for hours with structural enforcement — it literally cannot stop until the work is done.
+
+## Summary of New Capabilities
+
+- Slack messaging adapter with Socket Mode (real-time WebSocket, no webhooks needed)
+- DIY app model — each user creates their own Slack app, no shared infrastructure
+- Browser-automated setup wizard (Playwright drives workspace creation, app config, token extraction)
+- Session channel registry with bidirectional channel-to-session mapping
+- Channel resume map for conversation continuity across session restarts
+- Attention channel for critical alerts (auto-created, users auto-invited)
+- Dashboard link broadcast to Slack dashboard channel (update-in-place, no spam)
+- Slash commands via ! prefix (!sessions, !new, !help)
+- Platform badges on dashboard session cards
+- Platform dropdown when creating new sessions
+- Job scheduler integration (per-job Slack channels)
+- Cross-platform alerts (Slack added alongside Telegram and WhatsApp)
+- Presence proxy routing to correct session channels
+- SlackLifeline process for persistent Socket Mode connection
+- Autonomous mode skill with stop hook enforcement and session isolation
+- Unanswered message detection in Slack context hook
+- Message logging, search, and retention for Slack

package/upgrades/0.24.18.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/0.24.19.md

@@ -0,0 +1,21 @@
+# Upgrade Guide — vNEXT
+
+<!-- bump: patch -->
+
+## What Changed
+
+### Systems Dashboard CSS Fix — v3 Styles Restored
+
+The Systems dashboard CSS was still referencing old v2 class names (`.capability-card`, `.cap-info`, `.cap-label`) while the JavaScript had been updated to generate v3 classes (`.cap-card`, `.cap-card-name`, `.cap-card-desc`, `.cap-grid`). This mismatch caused the dashboard cards to render without proper styling.
+
+The CSS has been fully replaced with v3 styles covering: card grid layout, detail views, stat cards, process list indicators, and responsive grid breakpoints.
+
+## What to Tell Your User
+
+The Systems dashboard styling has been fixed. If the cards and layout looked broken or unstyled, that's now resolved — everything renders cleanly with the new card grid design.
+
+## Summary of New Capabilities
+
+| Capability | How to Use |
+|-----------|-----------|
+| Systems dashboard CSS fix | Automatic — dashboard renders correctly now |

package/upgrades/0.24.2.md

@@ -0,0 +1,13 @@
+# Upgrade Guide — vNEXT
+
+## 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 -->