@magpie-community/opencode-workbuddy-auth 0.2.2 → 0.2.4

package/README.md

@@ -174,21 +174,33 @@ default one, as the app does.
 
 WorkBuddy refuses some agents' identity lines as an "unapproved channel"
 (11128), so those lines are reworded before a chat goes out. Since 0.1.27
-(#279) that includes DeepSeek Harness's opening line. A refusal of words
-not reworded yet is sent once more, since 0.1.29, with the system prompt
-moved into the first user message (inside `<instructions>` tags) and
-WorkBuddy's default system prompt in its place: WorkBuddy checks only the
-system and assistant messages, and the line it matched is then in neither.
-A system message later in the chat becomes a user message where it stands.
-The plugin remembers each refused system prompt while it is loaded, so that
-agent's next chats go moved at once. The move can't reach an assistant
-message, so words the model quotes back into one keep that chat refused
-until they are in the list reworded in every message (since 0.1.30, the "Web GUI whose URL the runtime
-context names" line, #291); otherwise only a new chat clears it. A refusal
-that stands after the move ends in a hint that names no one client
-(since 0.1.28): WorkBuddy refused words in this client's prompt; use it from
-Hermes, OpenCode or Pi, or add another provider to the group. Neither
-refusal rests the account.
+(#279) that includes DeepSeek Harness's opening line. Since 0.2.3 a refusal
+of words not reworded yet is answered first by saying the same line with one
+more comma — a vendor matches such a line whole and exactly, so the comma
+passes it while the line still names the harness (measured 2026-10-30; a word
+added after the product name did not pass, the comma did). The comma is set in
+the system and assistant messages only — the two WorkBuddy reads — so the
+user's own words and tools' results go out as they were sent. A system prompt
+whose comma passed is remembered while the plugin is loaded, as a moved one
+is, and its next chats go with the comma already in before the first send:
+otherwise every turn, and every step of a tool loop, would pay the refusal
+once. Only a chat refused still is sent once more, since 0.1.29, with the
+system prompt moved into the first user message (inside `<instructions>` tags)
+and WorkBuddy's default system prompt in its place: WorkBuddy checks only the
+system and assistant messages, and the line it matched is then in neither. A system message later
+in the chat becomes a user message where it stands. The plugin remembers each
+refused system prompt while it is loaded, so that agent's next chats go moved
+at once. The move can't reach an assistant message, so words the model quotes
+back into one keep that chat refused until they are in the list reworded in
+every message (since 0.1.30, the "Web GUI whose URL the runtime context names"
+line, #291); otherwise only a new chat clears it. A refusal that stands after
+the move ends in a hint that names no one client (since 0.1.28): WorkBuddy
+refused words in this client's prompt; use it from Hermes, OpenCode or Pi, or
+add another provider to the group. Since 0.2.3 that hint is preceded by the
+address of [REFUSED.md](REFUSED.md), which says what the check reads, how a
+line is reworded to pass it, and how to turn such a line off in an agent that
+has a switch for it — so the error carries its own fix, for the person who
+reads it or the agent they paste it to. Neither refusal rests the account.
 
 The loader also translates OpenAI-compatible request spellings WorkBuddy
 doesn't take: `max_completion_tokens` becomes `max_tokens`, `developer`
@@ -227,6 +239,17 @@ serves; a model that refused it once gets it that way from then on, until
 the plugin restarts. A chat ending with the assistant's message (a prefill)
 is refused as before.
 
+WorkBuddy's safety filter can refuse a reply it has begun:
+`deepseek-v4.1-flash` reasoned, said "Sorry, I can't respond to this
+question.", and ended with `finish_reason: "content_filter"` and no output
+tokens, all under a 200 (#540). Since 0.2.4 that line is taken out of a
+streamed chat ending `content_filter`, so the reply says nothing before the
+filter's stop: magpie hands such a refusal to the group's next member
+instead of passing it to the agent as an answer. Only text that is still
+the filter's line is held, so any other reply's words go as they come; the
+same words ending `stop`, or the filter cutting off a reply after other
+words, go through as they came.
+
 ## Models
 
 When you are signed in, the models are the ones WorkBuddy's product config

package/index.mjs

@@ -1008,6 +1008,79 @@ function systemAsUserOn(b) {
   return true
 }
 
+// OPENERS are the shapes in which an agent's opening says which harness it
+// runs in, as a vendor may have fingerprinted one. A vendor matches a refused
+// line whole and exactly — one extra space passes it, as FLAGGED's note above
+// has it — so the same line with one more comma passes while still saying
+// what it said. Measured 2026-10-30 against copilot.tencent.com, one account,
+// streamed, deepseek-v4.1-flash at max: DeepSeek Harness's opener answered
+// 400 code 11128 as written, and 200 with the comma, the harness's name still
+// in it; a suffix in its place (…Harness Desktop., …Harness CLI.) changed
+// nothing, so it is the line's shape the vendor knows, not the product alone.
+const OPENERS = [
+  [/(\bYou are (?:an?|the)\s[^\n.,]{0,60}?\b(?:agent|assistant))\s+(powered by|running on|running in|built on|based on)\b/gi, "$1, $2"],
+]
+
+function opened(text) {
+  for (const [re, to] of OPENERS) text = text.replace(re, to)
+  return text
+}
+
+// commaPaused is body with OPENERS' comma set in its system and assistant
+// messages — the two WorkBuddy reads (REFUSED.md). The user's messages and
+// tools' results are not among them: a comma in a message the vendor never
+// matches changes what someone said without changing what was refused, and
+// the user's own words come back altered. It is tried before systemAsUser,
+// which takes the system prompt out of the slot an agent's instructions are
+// read in. null when there was nothing to set a comma in.
+//
+// The two roles are read off the message's own role: a message with no role,
+// or one of another kind, is left as it is. roles, when given, collects the
+// roles the comma went into — the caller uses it to tell a comma that made a
+// chat pass in its system prompt from one that only set an assistant
+// message's (commaSystems).
+const PAUSED_ROLES = new Set(["system", "assistant"])
+
+function commaPaused(body, roles) {
+  const b = chatOf(body)
+  return commaPausedOn(b, roles) ? JSON.stringify(b) : null
+}
+
+// commaPausedOn makes commaPaused's change on a chat's object, in place; false
+// when there was nothing to set a comma in. commaPaused is this on a body, and
+// the first send of a system prompt whose comma passed is this too
+// (commaSystems), so the two can never drift apart.
+function commaPausedOn(b, roles) {
+  const msgs = b?.messages
+  if (!Array.isArray(msgs)) return false
+  let changed = false
+  for (const m of msgs) {
+    if (!PAUSED_ROLES.has(m?.role)) continue
+    let mine = false
+    if (typeof m.content === "string") {
+      const t = opened(m.content)
+      if (t !== m.content) {
+        m.content = t
+        mine = true
+      }
+    } else if (Array.isArray(m.content)) {
+      for (const p of m.content) {
+        if (typeof p?.text !== "string") continue
+        const t = opened(p.text)
+        if (t === p.text) continue
+        p.text = t
+        mine = true
+      }
+    }
+    // a message whose parts held no opening of their own is not changed, and
+    // is not reported as one that was
+    if (!mine) continue
+    changed = true
+    roles?.add(m.role)
+  }
+  return changed
+}
+
 // normalizeToolChoice converts the object form emitted by OpenAI clients to
 // WorkBuddy's string form. It runs after fitted(), so a named choice follows a
 // tool name that was made safe for WorkBuddy.
@@ -1526,6 +1599,106 @@ function renamed(res, names) {
   return new Response(body, { status: res.status, statusText: res.statusText, headers })
 }
 
+// FILTER_SAYS is the text WorkBuddy's safety filter puts in a reply it ends
+// with finish_reason "content_filter": deepseek-v4.1-flash on
+// copilot.tencent.com reasoned, said this, and ended content_filter with
+// no output tokens counted (itspig, plugins #540).
+const FILTER_SAYS = ["Sorry, I can't respond to this question."]
+
+// filterSaid tells whether text said so far may still be FILTER_SAYS.
+const filterSaid = (text) => {
+  const t = text.replace(/’/g, "'").trimStart()
+  return t !== "" && FILTER_SAYS.some((s) => s.startsWith(t))
+}
+
+// unsaid is a streamed chat whose reply WorkBuddy's filter refused with its
+// text (FILTER_SAYS) and finish_reason content_filter, with the text taken
+// out: a reply that says nothing before content_filter is the refusal magpie
+// asks the group's next member for, where a 200 that said something is an
+// answer the agent fails on ("The model refused to complete the request",
+// plugins #540). Only text that is still the filter's is held, so any other
+// reply's first words go as they come, and one that only begins like it is
+// let through at its first word that isn't, or at its end.
+function unsaid(res) {
+  if (!res.body || !/event-stream/i.test(res.headers.get("content-type") ?? "")) return res
+  const headers = new Headers(res.headers)
+  headers.delete("content-length")
+  headers.delete("content-encoding")
+  const dec = new TextDecoder()
+  const enc = new TextEncoder()
+  let rest = ""
+  let open = true // the reply may still be the filter's
+  let said = ""
+  let held = [] // events held while it may be
+  const chunkOf = (ev) => {
+    const m = /^data:\s?(.*)$/m.exec(ev)
+    if (!m || m[1] === "[DONE]") return null
+    try {
+      const v = JSON.parse(m[1])
+      return v && typeof v === "object" ? v : null
+    } catch {
+      return null
+    }
+  }
+  // without is ev with the filter's text taken out, or "" when nothing
+  // else is in it
+  const without = (ev) => {
+    const v = chunkOf(ev)
+    const c = v?.choices?.[0]
+    if (!c?.delta || typeof c.delta.content !== "string") return ev
+    delete c.delta.content
+    const left = Object.keys(c.delta).filter((k) => k !== "role")
+    if (!left.length && c.finish_reason == null && v.usage == null) return ""
+    return `data: ${JSON.stringify(v)}`
+  }
+  const event = (ev) => {
+    if (!open) return [ev]
+    const c = chunkOf(ev)?.choices?.[0]
+    const text = typeof c?.delta?.content === "string" ? c.delta.content : ""
+    const finish = c?.finish_reason
+    if (c?.delta?.tool_calls || (text && !filterSaid(said + text))) {
+      open = false
+      const out = [...held, ev]
+      held = []
+      return out
+    }
+    said += text
+    if (finish === "content_filter" && said) {
+      open = false
+      const out = [...held, ev].map(without).filter(Boolean)
+      held = []
+      return out
+    }
+    if (finish) {
+      open = false
+      const out = [...held, ev]
+      held = []
+      return out
+    }
+    if (!said) return [ev] // reasoning, the role, before any text
+    held.push(ev)
+    return []
+  }
+  const body = res.body.pipeThrough(
+    new TransformStream({
+      transform(chunk, ctl) {
+        rest = (rest + dec.decode(chunk, { stream: true })).replace(/\r\n/g, "\n")
+        const evs = rest.split("\n\n")
+        rest = evs.pop()
+        const out = evs.flatMap(event)
+        if (out.length) ctl.enqueue(enc.encode(out.map((e) => e + "\n\n").join("")))
+      },
+      flush(ctl) {
+        rest += dec.decode()
+        const out = held.map((e) => e + "\n\n").join("") + rest
+        held = []
+        if (out) ctl.enqueue(enc.encode(out))
+      },
+    }),
+  )
+  return new Response(body, { status: res.status, statusText: res.statusText, headers })
+}
+
 // ---- signing in -------------------------------------------------------------
 
 // poll asks path until WorkBuddy answers with data, retrying the codes
@@ -3040,6 +3213,13 @@ function makePlugin(site) {
   // the system prompts this run WorkBuddy refused as an unapproved channel:
   // their next chats go with systemAsUser at once, not refused first
   const refusedSystems = new Set()
+  // the system prompts this run WorkBuddy refused as written and took with
+  // OPENERS' comma: their next chats go with the comma set before the first
+  // send, as a moved one goes moved. Without it every turn of that agent —
+  // and every step of its tool loop — pays a refused round trip first, which
+  // is latency on each request and a pattern of channel refusals on the
+  // account. Bounded as refusedSystems is.
+  const commaSystems = new Set()
   return async ({ client, magpie }) => {
   // each account's list as last read, for FREE_ONLY: {at, list}
   const lists = new Map()
@@ -3290,6 +3470,10 @@ function makePlugin(site) {
             // words at once (see systemAsUser)
             const sysKey = native ? "" : systemKeyOf(chat)
             if (sysKey && refusedSystems.has(sysKey) && systemAsUserOn(chat)) changed = true
+            // a system prompt whose comma passed once goes with the comma
+            // set before the first send (see commaSystems), the way one
+            // refused before goes moved above
+            if (sysKey && commaSystems.has(sysKey) && commaPausedOn(chat)) changed = true
             // shaped is a body as such a model takes it. The first try's
             // is made on the chat already parsed, after the move above,
             // as shaped would make it of the body written; a resend with
@@ -3317,7 +3501,27 @@ function makePlugin(site) {
               else res = new Response(text, { status: res.status, statusText: res.statusText, headers: res.headers })
             }
             if (res.status === 400 && !native) {              let text = await res.text()
-              const moved = REFUSED.test(text) && sysKey && !refusedSystems.has(sysKey) ? systemAsUser(body) : null
+              // the vendor may know a shape of the agent's opening that FLAGGED
+              // doesn't: the same words said with one more comma pass it, and
+              // that is tried before the system prompt is moved out of its slot.
+              // The roles the comma went into are kept: one set in the system
+              // prompt is what its next chats go with, one set in an assistant
+              // message alone says nothing about the next chat's.
+              const roles = new Set()
+              const paused = REFUSED.test(text) ? commaPaused(body, roles) : null
+              if (paused) {
+                body = paused
+                res = await send(shaped(body))
+                if (res.status === 400) text = await res.text()
+                // not refused this time: the comma is what passed the
+                // vendor's check, so this system prompt's next chats go
+                // with it set before the first send
+                else if (sysKey && roles.has("system")) {
+                  commaSystems.add(sysKey)
+                  if (commaSystems.size > 256) commaSystems.delete(commaSystems.values().next().value)
+                }
+              }
+              const moved = res.status === 400 && REFUSED.test(text) && sysKey && !refusedSystems.has(sysKey) ? systemAsUser(body) : null
               if (moved) {
                 refusedSystems.add(sysKey)
                 if (refusedSystems.size > 256) refusedSystems.delete(refusedSystems.values().next().value)
@@ -3335,7 +3539,7 @@ function makePlugin(site) {
             }
             if (ours.has(res)) return res
             // renewed here: a mark the account had comes off
-            return res.status >= 400 ? explained(res) : kept(native ? res : await renamed(res, names), renewed ? "renewed" : "kept")
+            return res.status >= 400 ? explained(res) : kept(native ? res : unsaid(await renamed(res, names)), renewed ? "renewed" : "kept")
           },
         }
       },
@@ -3446,6 +3650,13 @@ function makePlugin(site) {
 // apart from WorkBuddy's own words.
 const REFUSED_HINT =
   "WorkBuddy refused words in this client's prompt (most often its system prompt); use it from Hermes, OpenCode or Pi, or add another provider to this group"
+
+// REFUSED_DOC is where a refusal's fix is written down, for whoever holds the
+// error next: the person reading it, or the agent they paste it to. It goes
+// before the hint, never after, because magpie's Routing page tells the hint
+// from WorkBuddy's own words by the error ending in it.
+const REFUSED_DOC = "https://github.com/magpie-community/plugins/blob/main/packages/workbuddy/REFUSED.md"
+
 const REFUSED = /unapproved channel|illegal api invocation/i
 
 // BAD_SIGNATURE is WorkBuddy's refusal of a thinking block whose signature it
@@ -3524,7 +3735,11 @@ async function explained(res) {
   const hint = isRefused ? REFUSED_HINT : isRestricted ? RESTRICTED_HINT : EXHAUSTED_HINT
   const type = isRefused || isRestricted ? "permission_error" : "insufficient_quota"
   const errCode = isRefused ? null : isRestricted ? RESTRICTED_CODE : 14018
-  msg = `${msg} — ${hint}`
+  // a refusal says where its fix is written down, so whoever holds the error
+  // — the person, or the agent they paste it to — can go and read it. It goes
+  // before the hint, never after: the hint is the message's end.
+  const where = isRefused ? `${REFUSED_DOC} — ` : ""
+  msg = `${msg} — ${where}${hint}`
   const headers = new Headers(res.headers)
   headers.set("content-type", "application/json")
   headers.delete("content-length")
@@ -3537,4 +3752,4 @@ export const WorkBuddyAuthPlugin = makePlugin(SITES.workbuddy)
 export const WorkBuddyAIAuthPlugin = makePlugin(SITES["workbuddy-ai"])
 
 // for tests
-export const _internal = { idFor, replyAfterToolsOn, withDefaultEffort, defaultEffortOf, gzipped, FREE_ONLY, PRICE_TTL, keyAccount, remoteConfig, cycleOf, typeUnions, systemAsUser, systemKey, DEFAULT_SYSTEM, EDGE_READS, fitShape, replyAfterTools, AFTER_TOOLS_REFUSED, enterpriseUsageOf, accountName, notDesktop, textOnly, runsOutOf, packagesOf, endOf, enterpriseHeaders, enterpriseOf, liveModels, browserSignIn, withSystem, asReasoning, normalizeToolChoice, fitTools, fitted, renamed, unflagged, usageOf, desktopHeld, explained, REFUSED_HINT, RESTRICTED_HINT, EXHAUSTED_HINT, freeCredits, creditsOf, priceOf, fresh, SITES, renewing, renewed, SESSION, attend, scopeOf, turnKey, conversationKey, signatureOf, isMessages, withoutUnsigned, withoutSignedThinking, prepared, answeredBefore, nativeOf, modelOf, configModels, freeWindowOf, tokensShort, freeWindowRows, freeWindowLedger, readLedgerTail, pinned, rowTime, rowAccount, ofAccount, windowOffset, windowReset, FREE_WINDOW_MODELS, FREE_WINDOW_SPAN, FREE_WINDOW_GAP, FREE_WINDOW_LEAD, checkinOf, API_MESSAGES: NPM_MESSAGES, API_CHAT: NPM }
+export const _internal = { idFor, replyAfterToolsOn, withDefaultEffort, defaultEffortOf, gzipped, FREE_ONLY, PRICE_TTL, keyAccount, remoteConfig, cycleOf, typeUnions, systemAsUser, systemKey, DEFAULT_SYSTEM, EDGE_READS, fitShape, replyAfterTools, AFTER_TOOLS_REFUSED, enterpriseUsageOf, accountName, notDesktop, textOnly, runsOutOf, packagesOf, endOf, enterpriseHeaders, enterpriseOf, liveModels, browserSignIn, withSystem, asReasoning, normalizeToolChoice, fitTools, fitted, renamed, unsaid, FILTER_SAYS, unflagged, usageOf, desktopHeld, explained, REFUSED_HINT, RESTRICTED_HINT, EXHAUSTED_HINT, REFUSED_DOC, commaPaused, commaPausedOn, opened, freeCredits, creditsOf, priceOf, fresh, SITES, renewing, renewed, SESSION, attend, scopeOf, turnKey, conversationKey, signatureOf, isMessages, withoutUnsigned, withoutSignedThinking, prepared, answeredBefore, nativeOf, modelOf, configModels, freeWindowOf, tokensShort, freeWindowRows, freeWindowLedger, readLedgerTail, pinned, rowTime, rowAccount, ofAccount, windowOffset, windowReset, FREE_WINDOW_MODELS, FREE_WINDOW_SPAN, FREE_WINDOW_GAP, FREE_WINDOW_LEAD, checkinOf, API_MESSAGES: NPM_MESSAGES, API_CHAT: NPM }

package/package.json

@@ -1,12 +1,24 @@
 {
   "name": "@magpie-community/opencode-workbuddy-auth",
-  "version": "0.2.2",
+  "version": "0.2.4",
   "description": "OpenCode provider plugins for WorkBuddy's plans: WorkBuddy (CodeBuddy, copilot.tencent.com) and WorkBuddy AI (www.workbuddy.ai)",
   "type": "module",
   "main": "./index.mjs",
   "exports": "./index.mjs",
-  "files": ["index.mjs", "README.md"],
-  "keywords": ["opencode", "opencode-plugin", "workbuddy", "codebuddy"],
+  "files": [
+    "index.mjs",
+    "README.md"
+  ],
+  "keywords": [
+    "opencode",
+    "opencode-plugin",
+    "workbuddy",
+    "codebuddy"
+  ],
   "license": "MIT",
-  "repository": { "type": "git", "url": "git+https://github.com/magpie-community/plugins.git", "directory": "packages/workbuddy" }
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/magpie-community/plugins.git",
+    "directory": "packages/workbuddy"
+  }
 }