@dsiloed/silo-link 1.8.3 → 1.8.4

package/dist/cli/claude-md-block.d.ts

@@ -4,5 +4,5 @@
  *
  * Keep this block SHORT. It loads into every SiloLink-launched agent's context.
  */
-export declare const SILOLINK_CLAUDE_MD_BLOCK = "## Remote Communication (SiloLink)\n\nApplies only when launched via SiloLink (prompt contains `conversation_id` or tells you to `remote_register`). Ignore otherwise.\n\n- Register with `remote_register({ conversation_id })`.\n- Call `remote_load_context()` ONLY on a fresh session. Skip it when you were resumed \u2014 you already have the history.\n- All communication via `remote_notify` / `remote_ask`. Never print to the terminal; the user is in Harmoniq, not watching.\n- **@-mention discipline.** Your launch prompt has a `**Dispatched by:** @<DispatcherHandle>` line. ANY message that needs your dispatcher to act \u2014 mid-task questions, blockers, errors, requests for direction, AND your final completion \u2014 MUST start with that `@<DispatcherHandle>`. Mention-dispatch is the ONLY way to wake your dispatcher's LlmAgent; without it your message lands silently in the conv and they won't know to respond. Routine progress notes (\"phase 2 done, starting tests\") can skip the mention.\n- Send a `remote_notify` after every major step (file edits, commits, merges, errors). Going silent = user blind.\n- Poll loop: `remote_poll()` blocks up to 30s server-side (longer blocks risk Claude killing the tool as hung \u2014 keep this value as-is). If `{pending:true}`, loop again. The poll loop is the **safety net**, not the primary exit path \u2014 see \"Self-kill when work is done\" below. After **~60 consecutive pending (~30 min idle)**, `remote_notify(\"Going idle\")`, `remote_unregister()`, exit.\n- Every polled message is new; the queue drops it on your next call. Never ignore one. `remote_load_context()` is history-only.\n- Prefer small/cheap models for mechanical work (running specs, PR creation, merges). Use one-shot sub-sessions when possible to keep the main polling session's context lean.\n\n### Two conversations are in play\n\nYou typically have TWO distinct conversation ids in your session:\n- **Dispatch conv** \u2014 the conv you `remote_register`'d on (e.g. `silolink-claude-convN`). SiloLink relays operator messages from this conv to your poll loop.\n- **Report conv** \u2014 the conv where the originating user is watching for your final findings. Passed to you as `report_to_conversation_id` in your launch prompt. Often different from the dispatch conv (e.g. dispatch is per-ticket, report is the user's group/DM channel).\n\nFor routine progress updates use `remote_notify` (lands in the dispatch conv automatically). For your **final findings**, post to the **report conv** via `mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })`.\n\n### Self-kill when work is done\n\nDon't sit in the poll loop after your work for this dispatch is finished. When you reach a true terminal state, call `remote_notify(\"Done \u2014 unregistering\")` then `remote_unregister()` and exit cleanly. Orphan idle CCs each burn ~$0.50/hr on the safety-net timeout, so explicit self-kill is the primary exit path \u2014 when it's actually time.\n\n**HARD PRECONDITIONS \u2014 verify ALL before calling `remote_unregister()`. If any fails, do NOT self-kill; stay in the poll loop and finish.**\n\n1. **No uncommitted changes.** `git status --porcelain` returns empty in your cwd (and in any worktree you opened). If you made edits, commit them first.\n2. **No unpushed commits.** `git log @{u}..` returns empty on every branch you touched. Push before you exit.\n3. **MR opened if your runbook requires one.** Verify via `git_provider_tool get_mr` that the MR exists with `state: \"opened\"`. \"I plan to open it\" is not enough \u2014 open it, then verify.\n4. **No subprocess you started is still running.** Tests, builds, long scripts: wait for the exit code and capture the result. **\"Now running tests\" is NOT a terminal state** \u2014 \"tests passed (or failed with documented next steps)\" is.\n5. **Final findings posted, starting with `@<DispatcherHandle>`** per the @-mention discipline above. Final goes to `report_to_conversation_id` if set, otherwise the dispatch conv via `remote_notify`. Same rule applies to *any* mid-flight message that needed dispatcher action \u2014 they all required the mention, not just this last one.\n\nWhat counts as \"done\" is agent-specific (your runbook may add more preconditions on top of the five above). When in doubt: stay in the poll loop, `remote_notify` what you're waiting on, and let the operator decide.\n\nIf the work fails or you hit a question the operator must answer, do NOT self-kill. Stay in the poll loop and wait \u2014 exiting with unanswered questions strands the operator.\n\n### If a report-conv post fails\n\nIf `remote_post` to your `report_to_conversation_id` returns an error (404 not found, 403 forbidden, etc.):\n1. Retry once. Transient errors do happen.\n2. If the second attempt fails the same way, do NOT silently drop your findings. Post them to your **dispatch conv** instead with a clear note: `\u26A0 remote_post to report conv <X> returned <error> \u2014 posting findings here for triage`. Then exit.\n\nFindings landing somewhere a human can see them matters more than landing in the \"right\" channel. Never let an unreachable report conv strand a completed task.\n\n### Permissions\n\nYou were launched with `--dangerously-skip-permissions` and silolink has pre-cleared the bypass-permissions consent gate, so per-command permission prompts should not appear. If you DO hit one mid-task (usually CC's hardcoded circuit breaker on patterns like `rm -rf /` or `rm -rf ~`):\n\n- Do NOT sit waiting \u2014 there's no human at the keyboard to click \"Yes\".\n- Call `remote_ask` with the prompt text so the operator can decide.\n- The operator can also drive the prompt directly from the web UI's peek dialog (the `send_keys` control command sends keystrokes into your tmux pane).\n\nIf a prompt appeared on a routine command (not a circuit-breaker pattern), report it via `remote_notify` \u2014 it likely means the bypass-consent didn't get pre-written and silolink needs to be restarted.\n";
+export declare const SILOLINK_CLAUDE_MD_BLOCK = "## Remote Communication (SiloLink)\n\nApplies only when launched via SiloLink (prompt contains `conversation_id` or tells you to `remote_register`). Ignore otherwise.\n\n- Register with `remote_register({ conversation_id })`.\n- Call `remote_load_context()` ONLY on a fresh session. Skip it when you were resumed \u2014 you already have the history.\n- All communication via `remote_notify` / `remote_ask`. Never print to the terminal; the user is in Harmoniq, not watching.\n- **@-mention discipline.** Your launch prompt has a `**Dispatched by:** @<DispatcherHandle>` line. ANY message that needs your dispatcher to act \u2014 mid-task questions, blockers, errors, requests for direction, AND your final completion \u2014 MUST start with that `@<DispatcherHandle>` **AND be posted to the report conv** (`mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })`), not via `remote_notify` to the dispatch conv. Mention-dispatch only wakes the dispatcher's LlmAgent when the @-mention lands in a conversation they're a member of \u2014 the dispatch conv is per-session and they're typically not subscribed; the report conv is the channel they read. Routine progress notes (\"phase 2 done, starting tests\") still go via `remote_notify` (no @-mention needed).\n- Send a `remote_notify` after every major step (file edits, commits, merges, errors). Going silent = user blind.\n- Poll loop: `remote_poll()` blocks up to 30s server-side (longer blocks risk Claude killing the tool as hung \u2014 keep this value as-is). If `{pending:true}`, loop again. The poll loop is the **safety net**, not the primary exit path \u2014 see \"Self-kill when work is done\" below. After **~60 consecutive pending (~30 min idle)**, `remote_notify(\"Going idle\")`, `remote_unregister()`, exit.\n- Every polled message is new; the queue drops it on your next call. Never ignore one. `remote_load_context()` is history-only.\n- Prefer small/cheap models for mechanical work (running specs, PR creation, merges). Use one-shot sub-sessions when possible to keep the main polling session's context lean.\n\n### Two conversations are in play\n\nYou typically have TWO distinct conversation ids in your session:\n- **Dispatch conv** \u2014 the conv you `remote_register`'d on (e.g. `silolink-claude-convN`). SiloLink relays operator messages from this conv to your poll loop.\n- **Report conv** \u2014 the conv where the originating user is watching for your final findings. Passed to you as `report_to_conversation_id` in your launch prompt. Often different from the dispatch conv (e.g. dispatch is per-ticket, report is the user's group/DM channel).\n\nRouting rule, by message kind:\n\n| Message kind | Tool | Lands in | @-mention? |\n|---|---|---|---|\n| Routine progress (\"phase 2 done\", logs, heartbeats) | `remote_notify` | dispatch conv | no |\n| **Mid-task questions / blockers / requests for direction** | `mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })` | **report conv** | **YES \u2014 start with `@<DispatcherHandle>`** |\n| Final findings on completion | `mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })` | report conv | YES \u2014 start with `@<DispatcherHandle>` |\n\nWhy mid-task questions go to the **report conv** and not the dispatch conv: the dispatch conv (`silolink-claude-convN`) is per-session and the dispatcher's LlmAgent isn't typically subscribed to it. `remote_notify` with an @-mention there lands silently and the dispatcher never wakes. The report conv is the channel they read \u2014 @-mention dispatch fires there.\n\n### Self-kill when work is done\n\nDon't sit in the poll loop after your work for this dispatch is finished. When you reach a true terminal state, call `remote_notify(\"Done \u2014 unregistering\")` then `remote_unregister()` and exit cleanly. Orphan idle CCs each burn ~$0.50/hr on the safety-net timeout, so explicit self-kill is the primary exit path \u2014 when it's actually time.\n\n**HARD PRECONDITIONS \u2014 verify ALL before calling `remote_unregister()`. If any fails, do NOT self-kill; stay in the poll loop and finish.**\n\n1. **No uncommitted changes.** `git status --porcelain` returns empty in your cwd (and in any worktree you opened). If you made edits, commit them first.\n2. **No unpushed commits.** `git log @{u}..` returns empty on every branch you touched. Push before you exit.\n3. **MR opened if your runbook requires one.** Verify via `git_provider_tool get_mr` that the MR exists with `state: \"opened\"`. \"I plan to open it\" is not enough \u2014 open it, then verify.\n4. **No subprocess you started is still running.** Tests, builds, long scripts: wait for the exit code and capture the result. **\"Now running tests\" is NOT a terminal state** \u2014 \"tests passed (or failed with documented next steps)\" is.\n5. **Final findings posted, starting with `@<DispatcherHandle>`** per the @-mention discipline above. Final goes to `report_to_conversation_id` if set, otherwise the dispatch conv via `remote_notify`. Same rule applies to *any* mid-flight message that needed dispatcher action \u2014 they all required the mention, not just this last one.\n\nWhat counts as \"done\" is agent-specific (your runbook may add more preconditions on top of the five above). When in doubt: stay in the poll loop, `remote_notify` what you're waiting on, and let the operator decide.\n\nIf the work fails or you hit a question the operator must answer, do NOT self-kill. Stay in the poll loop and wait \u2014 exiting with unanswered questions strands the operator.\n\n### If a report-conv post fails\n\nIf `remote_post` to your `report_to_conversation_id` returns an error (404 not found, 403 forbidden, etc.):\n1. Retry once. Transient errors do happen.\n2. If the second attempt fails the same way, do NOT silently drop your findings. Post them to your **dispatch conv** instead with a clear note: `\u26A0 remote_post to report conv <X> returned <error> \u2014 posting findings here for triage`. Then exit.\n\nFindings landing somewhere a human can see them matters more than landing in the \"right\" channel. Never let an unreachable report conv strand a completed task.\n\n### Permissions\n\nYou were launched with `--dangerously-skip-permissions` and silolink has pre-cleared the bypass-permissions consent gate, so per-command permission prompts should not appear. If you DO hit one mid-task (usually CC's hardcoded circuit breaker on patterns like `rm -rf /` or `rm -rf ~`):\n\n- Do NOT sit waiting \u2014 there's no human at the keyboard to click \"Yes\".\n- Call `remote_ask` with the prompt text so the operator can decide.\n- The operator can also drive the prompt directly from the web UI's peek dialog (the `send_keys` control command sends keystrokes into your tmux pane).\n\nIf a prompt appeared on a routine command (not a circuit-breaker pattern), report it via `remote_notify` \u2014 it likely means the bypass-consent didn't get pre-written and silolink needs to be restarted.\n";
 //# sourceMappingURL=claude-md-block.d.ts.map

package/dist/cli/claude-md-block.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"claude-md-block.d.ts","sourceRoot":"","sources":["../../src/cli/claude-md-block.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,eAAO,MAAM,wBAAwB,24LAsDpC,CAAC"}
+{"version":3,"file":"claude-md-block.d.ts","sourceRoot":"","sources":["../../src/cli/claude-md-block.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,eAAO,MAAM,wBAAwB,o3NA8DpC,CAAC"}

package/dist/cli/claude-md-block.js

@@ -11,7 +11,7 @@ Applies only when launched via SiloLink (prompt contains \`conversation_id\` or
 - Register with \`remote_register({ conversation_id })\`.
 - Call \`remote_load_context()\` ONLY on a fresh session. Skip it when you were resumed — you already have the history.
 - All communication via \`remote_notify\` / \`remote_ask\`. Never print to the terminal; the user is in Harmoniq, not watching.
-- **@-mention discipline.** Your launch prompt has a \`**Dispatched by:** @<DispatcherHandle>\` line. ANY message that needs your dispatcher to act — mid-task questions, blockers, errors, requests for direction, AND your final completion — MUST start with that \`@<DispatcherHandle>\`. Mention-dispatch is the ONLY way to wake your dispatcher's LlmAgent; without it your message lands silently in the conv and they won't know to respond. Routine progress notes ("phase 2 done, starting tests") can skip the mention.
+- **@-mention discipline.** Your launch prompt has a \`**Dispatched by:** @<DispatcherHandle>\` line. ANY message that needs your dispatcher to act — mid-task questions, blockers, errors, requests for direction, AND your final completion — MUST start with that \`@<DispatcherHandle>\` **AND be posted to the report conv** (\`mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })\`), not via \`remote_notify\` to the dispatch conv. Mention-dispatch only wakes the dispatcher's LlmAgent when the @-mention lands in a conversation they're a member of — the dispatch conv is per-session and they're typically not subscribed; the report conv is the channel they read. Routine progress notes ("phase 2 done, starting tests") still go via \`remote_notify\` (no @-mention needed).
 - Send a \`remote_notify\` after every major step (file edits, commits, merges, errors). Going silent = user blind.
 - Poll loop: \`remote_poll()\` blocks up to 30s server-side (longer blocks risk Claude killing the tool as hung — keep this value as-is). If \`{pending:true}\`, loop again. The poll loop is the **safety net**, not the primary exit path — see "Self-kill when work is done" below. After **~60 consecutive pending (~30 min idle)**, \`remote_notify("Going idle")\`, \`remote_unregister()\`, exit.
 - Every polled message is new; the queue drops it on your next call. Never ignore one. \`remote_load_context()\` is history-only.
@@ -23,7 +23,15 @@ You typically have TWO distinct conversation ids in your session:
 - **Dispatch conv** — the conv you \`remote_register\`'d on (e.g. \`silolink-claude-convN\`). SiloLink relays operator messages from this conv to your poll loop.
 - **Report conv** — the conv where the originating user is watching for your final findings. Passed to you as \`report_to_conversation_id\` in your launch prompt. Often different from the dispatch conv (e.g. dispatch is per-ticket, report is the user's group/DM channel).
 
-For routine progress updates use \`remote_notify\` (lands in the dispatch conv automatically). For your **final findings**, post to the **report conv** via \`mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })\`.
+Routing rule, by message kind:
+
+| Message kind | Tool | Lands in | @-mention? |
+|---|---|---|---|
+| Routine progress ("phase 2 done", logs, heartbeats) | \`remote_notify\` | dispatch conv | no |
+| **Mid-task questions / blockers / requests for direction** | \`mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })\` | **report conv** | **YES — start with \`@<DispatcherHandle>\`** |
+| Final findings on completion | \`mcp__silolink__remote_post({ conversation_id: <report_to_conversation_id> })\` | report conv | YES — start with \`@<DispatcherHandle>\` |
+
+Why mid-task questions go to the **report conv** and not the dispatch conv: the dispatch conv (\`silolink-claude-convN\`) is per-session and the dispatcher's LlmAgent isn't typically subscribed to it. \`remote_notify\` with an @-mention there lands silently and the dispatcher never wakes. The report conv is the channel they read — @-mention dispatch fires there.
 
 ### Self-kill when work is done
 

package/dist/cli/claude-md-block.js.map

@@ -1 +1 @@
-{"version":3,"file":"claude-md-block.js","sourceRoot":"","sources":["../../src/cli/claude-md-block.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,MAAM,CAAC,MAAM,wBAAwB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAsDvC,CAAC"}
+{"version":3,"file":"claude-md-block.js","sourceRoot":"","sources":["../../src/cli/claude-md-block.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,MAAM,CAAC,MAAM,wBAAwB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8DvC,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dsiloed/silo-link",
-  "version": "1.8.3",
+  "version": "1.8.4",
   "description": "Claude Code Remote Bridge — connects Claude Code sessions to DSiloed via MCP + ActionCable",
   "type": "module",
   "license": "MIT",