@cabane/companion 0.6.6 → 0.6.7

package/dist/cli.js

@@ -6290,7 +6290,7 @@ function createSummonMcpServer(summonState, skipState, askState, subAgentCreate,
       ...askState ? [
         tool(
           ASK_TOOL,
-          "Ask a HUMAN a structured question (or a short LIST of them) you need answered to continue, then END your turn \u2014 don't wait for the reply. Use it when you genuinely can't proceed without a person's input (a decision only they can make, a missing fact, a go/no-go). Pass `targetUserId` (a workspace member's user id \u2014 get it from `mcp__cabane__list_members`). Two forms: a SINGLE question \u2014 a `headline` (the actual question as one clear, capitalized sentence ending in `?`, \"Do we go to prod?\") plus a short `question` body for the framing the headline can't hold \u2014 OR, when a plan ends with SEVERAL bounded decisions at once (\"three calls before I build: A? B? C?\"), a `questions` array of 1\u20135 items, each `{ headline, body?, options? }`. **Prefer the list over cramming the extra decisions into prose or dropping them** \u2014 end the turn with one ask carrying every question, never pick one and bury the rest. Each question keeps the same form rules: a one-sentence `headline`, a short `body` frame (NOT a report \u2014 your status, links, and detail go in your REPLY, and the body renders inline markdown only: links/emphasis/inline code, no bulleted lists or headings), and 2\u20134 `options` when the answer is a bounded choice \u2014 for a yes/no go-ahead always pass them, so it's one click, not a typed reply. An option can be a short button label OR a sentence that carries its own context; short/binary sets render as inline buttons, long ones stack full-width. Provide EITHER `question` (single) or `questions` (array), never both. The ask is recorded as a first-class attention item aimed at that person; your final reply carries the surrounding CONTEXT (what you found, why you're stuck), the ask carries the QUESTION(S). An open ask marks you as blocked until EVERY question is answered, so raise one only when you truly can't proceed \u2014 never ceremonially. One ask per turn (last call wins). After asking, stop \u2014 when the person replies addressed to you, the ask resolves and you resume; other people's or agents' messages may wake you but leave it open. Targets a human only; to hand work to another AGENT use summon/dispatch instead.",
+          "Ask a HUMAN a structured question (or a short LIST of them) you need answered to continue, then END your turn \u2014 don't wait for the reply. Use it when you genuinely can't proceed without a person's input (a decision only they can make, a missing fact). Pass `targetUserId` (a workspace member's user id \u2014 get it from `mcp__cabane__list_members`). Two forms: a SINGLE question \u2014 a `headline` (the actual question as one clear, capitalized sentence ending in `?`, \"Do we go to prod?\") plus a short `question` body for the framing the headline can't hold \u2014 OR, when a plan ends with SEVERAL bounded decisions at once, a `questions` array of 1\u20135 items, each `{ headline, body?, options? }`. **Prefer the list over cramming the extra decisions into prose or dropping them** \u2014 end the turn with one ask carrying every question, never pick one and bury the rest. Each question keeps the same form rules: a one-sentence `headline`, a short `body` frame (NOT a report \u2014 your status, links, and detail go in your REPLY, and the body renders inline markdown only: links/emphasis/inline code, no bulleted lists or headings), and 2\u20134 `options` when the answer is a bounded choice \u2014 for a yes/no go-ahead always pass them, so it's one click, not a typed reply. An option can be a short button label or a whole sentence. Provide EITHER `question` (single) or `questions` (array), never both. The ask is a first-class attention item aimed at that person; your final reply carries the surrounding CONTEXT (what you found, why you're stuck), the ask carries the QUESTION(S). An open ask marks you as blocked until EVERY question is answered, so raise one only when you truly can't proceed \u2014 never ceremonially. One ask per turn (last call wins). After asking, stop \u2014 when the person replies addressed to you, the ask resolves and you resume; other people's or agents' messages may wake you but leave it open. Targets a human only; to hand work to another AGENT use summon/dispatch instead.",
           {
             targetUserId: z12.string().uuid().describe("The workspace member (human) to ask \u2014 a user id from `list_members`."),
             question: z12.string().min(1).max(400).optional().describe(
@@ -6397,7 +6397,7 @@ function createSummonMcpServer(summonState, skipState, askState, subAgentCreate,
       ...wakeState ? [
         tool(
           WAKE_ME_TOOL,
-          'Wake yourself later \u2014 end this turn now and be re-dispatched at a time you pick, with a note you write to yourself. Use it for "wait until X": when the thing you need hasn\'t happened yet (a conversation isn\'t done, a PR isn\'t merged, a human hasn\'t answered), arm a wake, end your turn, and you\'re woken later to CHECK \u2014 read the workspace, and either act or re-arm. This is the loop behind "check back in five minutes", "keep checking until {condition}", and a scheduled re-try ("re-send once that agent\'s limit resets"). Ground the delay before you arm it. Almost every wake is short \u2014 seconds to a couple of hours \u2014 waiting on a condition you can name: a session limit resetting, a PR merging, a deploy landing, a throttled dispatch to re-send. Reach past a few hours only when (a) a human asked for that timing, or (b) the wait is pinned to a real external event you can name \u2014 a report that only runs Mondays, a scheduled appointment, a known reset time. A speculative far-future check-in you invented yourself \u2014 "in two weeks I\'ll see whether this feature is used" \u2014 is the one thing not to arm: if no one asked and you can\'t name both what clears the wait and why it takes that long, don\'t arm it \u2014 finish now, or raise an `ask`. Pass EXACTLY ONE of `afterSeconds` (a relative delay \u2014 `300` for five minutes) or `at` (an absolute ISO-8601 timestamp WITH a zone, e.g. `2026-07-16T09:00:00-07:00` \u2014 YOU compute it from a phrase like "tomorrow morning"; the system never parses natural-language time). `note` is a message to your future self \u2014 it becomes the body of the wake message that re-dispatches you, so write the condition to re-check ("check whether CT441 merged yet"). The wake is armed when your turn SETTLES, not now, so the delay counts from the turn ending; one wake per turn (last call wins). This is the sanctioned way to schedule your own continuation \u2014 the ONLY one; never reach for a host cron/scheduler. Guardrails: at least 60s out, at most 14 days; widen the interval as a loop ages (5m \u2192 15m \u2192 1h\u2026) rather than hammering; after many consecutive re-arms with no other activity you\'ll be steered to raise an `ask` to the human instead. If a wake can\'t be armed you\'re re-dispatched with a note explaining why \u2014 never a silent drop.',
+          "Wake yourself later \u2014 end this turn now and be re-dispatched at a time you pick, with a note you write to yourself. Use it for \"wait until X\": when the thing you need hasn't happened yet (a PR isn't merged, a human hasn't answered), arm a wake, end your turn, and you're woken later to CHECK \u2014 read the workspace, and either act or re-arm. Ground the delay before you arm it. Almost every wake is short \u2014 seconds to a couple of hours \u2014 waiting on a condition you can name: a session limit resetting, a PR merging. Reach past a few hours only when (a) a human asked for that timing, or (b) the wait is pinned to a real external event you can name \u2014 a report that only runs Mondays, a known reset time. A speculative far-future check-in you invented yourself is the one thing not to arm: if no one asked and you can't name both what clears the wait and why it takes that long, don't arm it \u2014 finish now, or raise an `ask`. Pass EXACTLY ONE of `afterSeconds` (a relative delay \u2014 `300` for five minutes) or `at` (an absolute ISO-8601 timestamp WITH a zone, e.g. `2026-07-16T09:00:00-07:00` \u2014 YOU compute it from a phrase like \"tomorrow morning\"; the system never parses natural-language time). `note` is a message to your future self \u2014 it becomes the body of the wake message that re-dispatches you, so write the condition to re-check (\"check whether CT441 merged yet\"). The wake is armed when your turn SETTLES, not now, so the delay counts from the turn ending; one wake per turn (last call wins). This is the sanctioned way to schedule your own continuation \u2014 the ONLY one; never reach for a host cron/scheduler. Guardrails: at least 60s out, at most 14 days; widen the interval as a loop ages (5m \u2192 15m \u2192 1h\u2026) rather than hammering; after many consecutive re-arms with no other activity you'll be steered to raise an `ask` to the human instead. If a wake can't be armed you're re-dispatched with a note explaining why \u2014 never a silent drop.",
           {
             afterSeconds: z12.number().int().positive().optional().describe(
               "Relative delay in seconds from when this turn ends (e.g. 300 = five minutes). Provide EITHER this or `at`, not both. Floor 60s, horizon 14 days \u2014 enforced server-side."

package/dist/runtime.js

@@ -5943,7 +5943,7 @@ function createSummonMcpServer(summonState, skipState, askState, subAgentCreate,
       ...askState ? [
         tool(
           ASK_TOOL,
-          "Ask a HUMAN a structured question (or a short LIST of them) you need answered to continue, then END your turn \u2014 don't wait for the reply. Use it when you genuinely can't proceed without a person's input (a decision only they can make, a missing fact, a go/no-go). Pass `targetUserId` (a workspace member's user id \u2014 get it from `mcp__cabane__list_members`). Two forms: a SINGLE question \u2014 a `headline` (the actual question as one clear, capitalized sentence ending in `?`, \"Do we go to prod?\") plus a short `question` body for the framing the headline can't hold \u2014 OR, when a plan ends with SEVERAL bounded decisions at once (\"three calls before I build: A? B? C?\"), a `questions` array of 1\u20135 items, each `{ headline, body?, options? }`. **Prefer the list over cramming the extra decisions into prose or dropping them** \u2014 end the turn with one ask carrying every question, never pick one and bury the rest. Each question keeps the same form rules: a one-sentence `headline`, a short `body` frame (NOT a report \u2014 your status, links, and detail go in your REPLY, and the body renders inline markdown only: links/emphasis/inline code, no bulleted lists or headings), and 2\u20134 `options` when the answer is a bounded choice \u2014 for a yes/no go-ahead always pass them, so it's one click, not a typed reply. An option can be a short button label OR a sentence that carries its own context; short/binary sets render as inline buttons, long ones stack full-width. Provide EITHER `question` (single) or `questions` (array), never both. The ask is recorded as a first-class attention item aimed at that person; your final reply carries the surrounding CONTEXT (what you found, why you're stuck), the ask carries the QUESTION(S). An open ask marks you as blocked until EVERY question is answered, so raise one only when you truly can't proceed \u2014 never ceremonially. One ask per turn (last call wins). After asking, stop \u2014 when the person replies addressed to you, the ask resolves and you resume; other people's or agents' messages may wake you but leave it open. Targets a human only; to hand work to another AGENT use summon/dispatch instead.",
+          "Ask a HUMAN a structured question (or a short LIST of them) you need answered to continue, then END your turn \u2014 don't wait for the reply. Use it when you genuinely can't proceed without a person's input (a decision only they can make, a missing fact). Pass `targetUserId` (a workspace member's user id \u2014 get it from `mcp__cabane__list_members`). Two forms: a SINGLE question \u2014 a `headline` (the actual question as one clear, capitalized sentence ending in `?`, \"Do we go to prod?\") plus a short `question` body for the framing the headline can't hold \u2014 OR, when a plan ends with SEVERAL bounded decisions at once, a `questions` array of 1\u20135 items, each `{ headline, body?, options? }`. **Prefer the list over cramming the extra decisions into prose or dropping them** \u2014 end the turn with one ask carrying every question, never pick one and bury the rest. Each question keeps the same form rules: a one-sentence `headline`, a short `body` frame (NOT a report \u2014 your status, links, and detail go in your REPLY, and the body renders inline markdown only: links/emphasis/inline code, no bulleted lists or headings), and 2\u20134 `options` when the answer is a bounded choice \u2014 for a yes/no go-ahead always pass them, so it's one click, not a typed reply. An option can be a short button label or a whole sentence. Provide EITHER `question` (single) or `questions` (array), never both. The ask is a first-class attention item aimed at that person; your final reply carries the surrounding CONTEXT (what you found, why you're stuck), the ask carries the QUESTION(S). An open ask marks you as blocked until EVERY question is answered, so raise one only when you truly can't proceed \u2014 never ceremonially. One ask per turn (last call wins). After asking, stop \u2014 when the person replies addressed to you, the ask resolves and you resume; other people's or agents' messages may wake you but leave it open. Targets a human only; to hand work to another AGENT use summon/dispatch instead.",
           {
             targetUserId: z12.string().uuid().describe("The workspace member (human) to ask \u2014 a user id from `list_members`."),
             question: z12.string().min(1).max(400).optional().describe(
@@ -6050,7 +6050,7 @@ function createSummonMcpServer(summonState, skipState, askState, subAgentCreate,
       ...wakeState ? [
         tool(
           WAKE_ME_TOOL,
-          'Wake yourself later \u2014 end this turn now and be re-dispatched at a time you pick, with a note you write to yourself. Use it for "wait until X": when the thing you need hasn\'t happened yet (a conversation isn\'t done, a PR isn\'t merged, a human hasn\'t answered), arm a wake, end your turn, and you\'re woken later to CHECK \u2014 read the workspace, and either act or re-arm. This is the loop behind "check back in five minutes", "keep checking until {condition}", and a scheduled re-try ("re-send once that agent\'s limit resets"). Ground the delay before you arm it. Almost every wake is short \u2014 seconds to a couple of hours \u2014 waiting on a condition you can name: a session limit resetting, a PR merging, a deploy landing, a throttled dispatch to re-send. Reach past a few hours only when (a) a human asked for that timing, or (b) the wait is pinned to a real external event you can name \u2014 a report that only runs Mondays, a scheduled appointment, a known reset time. A speculative far-future check-in you invented yourself \u2014 "in two weeks I\'ll see whether this feature is used" \u2014 is the one thing not to arm: if no one asked and you can\'t name both what clears the wait and why it takes that long, don\'t arm it \u2014 finish now, or raise an `ask`. Pass EXACTLY ONE of `afterSeconds` (a relative delay \u2014 `300` for five minutes) or `at` (an absolute ISO-8601 timestamp WITH a zone, e.g. `2026-07-16T09:00:00-07:00` \u2014 YOU compute it from a phrase like "tomorrow morning"; the system never parses natural-language time). `note` is a message to your future self \u2014 it becomes the body of the wake message that re-dispatches you, so write the condition to re-check ("check whether CT441 merged yet"). The wake is armed when your turn SETTLES, not now, so the delay counts from the turn ending; one wake per turn (last call wins). This is the sanctioned way to schedule your own continuation \u2014 the ONLY one; never reach for a host cron/scheduler. Guardrails: at least 60s out, at most 14 days; widen the interval as a loop ages (5m \u2192 15m \u2192 1h\u2026) rather than hammering; after many consecutive re-arms with no other activity you\'ll be steered to raise an `ask` to the human instead. If a wake can\'t be armed you\'re re-dispatched with a note explaining why \u2014 never a silent drop.',
+          "Wake yourself later \u2014 end this turn now and be re-dispatched at a time you pick, with a note you write to yourself. Use it for \"wait until X\": when the thing you need hasn't happened yet (a PR isn't merged, a human hasn't answered), arm a wake, end your turn, and you're woken later to CHECK \u2014 read the workspace, and either act or re-arm. Ground the delay before you arm it. Almost every wake is short \u2014 seconds to a couple of hours \u2014 waiting on a condition you can name: a session limit resetting, a PR merging. Reach past a few hours only when (a) a human asked for that timing, or (b) the wait is pinned to a real external event you can name \u2014 a report that only runs Mondays, a known reset time. A speculative far-future check-in you invented yourself is the one thing not to arm: if no one asked and you can't name both what clears the wait and why it takes that long, don't arm it \u2014 finish now, or raise an `ask`. Pass EXACTLY ONE of `afterSeconds` (a relative delay \u2014 `300` for five minutes) or `at` (an absolute ISO-8601 timestamp WITH a zone, e.g. `2026-07-16T09:00:00-07:00` \u2014 YOU compute it from a phrase like \"tomorrow morning\"; the system never parses natural-language time). `note` is a message to your future self \u2014 it becomes the body of the wake message that re-dispatches you, so write the condition to re-check (\"check whether CT441 merged yet\"). The wake is armed when your turn SETTLES, not now, so the delay counts from the turn ending; one wake per turn (last call wins). This is the sanctioned way to schedule your own continuation \u2014 the ONLY one; never reach for a host cron/scheduler. Guardrails: at least 60s out, at most 14 days; widen the interval as a loop ages (5m \u2192 15m \u2192 1h\u2026) rather than hammering; after many consecutive re-arms with no other activity you'll be steered to raise an `ask` to the human instead. If a wake can't be armed you're re-dispatched with a note explaining why \u2014 never a silent drop.",
           {
             afterSeconds: z12.number().int().positive().optional().describe(
               "Relative delay in seconds from when this turn ends (e.g. 300 = five minutes). Provide EITHER this or `at`, not both. Floor 60s, horizon 14 days \u2014 enforced server-side."

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cabane/companion",
-  "version": "0.6.6",
+  "version": "0.6.7",
   "type": "module",
   "description": "The Cabane Companion (headless): connect a coding agent on your machine to your Cabane workspace as a responder — drive work against your own codebase, files, and MCP servers without putting any of it in Cabane.",
   "license": "UNLICENSED",