thinkpool-pair 0.7.179 → 0.7.180

package/README.md

@@ -120,10 +120,17 @@ ThinkPool Code runs **any model you choose** — not just Anthropic. The agent
 talks the Anthropic Messages API, so you point it at any endpoint that speaks
 that format:
 
-- **Anthropic-compatible endpoints directly** — Z.ai GLM, OpenRouter, your own proxy.
-- **Any other model via a gateway** — put a translator like **LiteLLM** or
-  OpenRouter in front and run **GPT, Gemini, Llama, DeepSeek, or a local model**;
-  the gateway presents the Anthropic Messages API while calling whatever you pick.
+- **Anthropic-compatible endpoints directly** — Z.ai GLM, Moonshot Kimi, OpenRouter, your own proxy.
+- **Any other model via a translating gateway** — put **LiteLLM** (or a proxy of
+  your own) in front and run **GPT, Gemini, Llama, DeepSeek, or a local model**;
+  it presents the Anthropic Messages API while calling whatever you pick.
+
+> OpenRouter is the exception worth knowing: its `/v1/messages` surface is a
+> pass-through for **Anthropic models only**. Per OpenRouter's own docs, "Claude
+> Code expects Anthropic request semantics, so non-Anthropic models aren't
+> supported through the native endpoint" — and this bridge spawns the `claude`
+> CLI. To reach a non-Anthropic model, use a translating gateway (LiteLLM),
+> not OpenRouter's Anthropic endpoint.
 
 ```bash
 npx thinkpool-pair provider custom --base <url> --token <key> [--model <name>]
@@ -134,7 +141,8 @@ Common base urls (the agent appends `/v1/messages`):
 | Provider / gateway | Base url                          | Models                       |
 |--------------------|-----------------------------------|------------------------------|
 | Z.ai GLM           | `https://api.z.ai/api/anthropic`  | GLM                          |
-| OpenRouter         | `https://openrouter.ai/api`       | many, incl. non-Anthropic    |
+| Moonshot Kimi      | `https://api.moonshot.ai/anthropic` | Kimi                       |
+| OpenRouter         | `https://openrouter.ai/api`       | Anthropic models only        |
 | LiteLLM (your own) | `http://localhost:4000`           | GPT, Gemini, Llama, local, … |
 
 > The endpoint must serve the Anthropic Messages API (`/v1/messages`). A raw

package/bridge.mjs

@@ -1566,7 +1566,7 @@ function worktreeSnapshot(cwd) {
 // relay STRUCTURED events. onEvent → broadcast `code-event` + print locally +
 // persist to the host file; tool calls round-trip through the perm card; the
 // rolling log replays to joiners and survives bridge restarts (session-store).
-function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rolePrompt, flowSessionId, flowTaskKey, cwd, reviewSliceRoots, openedAt, defer, provider, carryRecap }) {
+function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rolePrompt, flowSessionId, flowTaskKey, cwd, reviewSliceRoots, openedAt, defer, provider, carryRecap, lastUsage }) {
   if (sessions.has(id)) return
   // The lane's EFFECTIVE model — computed ONCE and used for BOTH the truthful display
   // label (entry.model) and the SDK's `model` option below. These used to be computed
@@ -1608,6 +1608,12 @@ function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rol
   // event-id.mjs + src/pages/code/seqDedup.js). pushLog() is the single stamp point.
   entry.seq = makeSeqCounter(maxSeq(entry.log))
   entry.rolePrompt = rolePrompt || null   // FL-M6 — persisted so a bridge restart restores the flow role prompt
+  // lastUsage — the terminal's most recent usage/ctx meter. Persisted (sessionData) +
+  // restored so a bridge restart can re-emit it on replay: usage is chrome (kept out of
+  // the replayed transcript log), so without this a (re)joiner sees "—" for context until
+  // the terminal's next turn. correctContext nulls ctx for models whose true window we
+  // can't determine; those restore a null ctx and replay nothing (no misleading %).
+  entry.lastUsage = lastUsage || null
   // S5 (slice 1b) — a REVIEW lane's write-block scope: the worktree root(s) of the slice(s)
   // it reviews (its deps). Persisted so a bridge restart re-arms the structural gate. Empty/
   // absent on builder lanes → reviewGate stays null → builder toolset UNCHANGED.
@@ -1775,7 +1781,7 @@ function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rol
   // restart. Without this, sessionData omitted it → on restart the resumed session
   // re-launched on the host default (Opus) regardless of the last switch, and the
   // switch looked like it "never changed the model" (Max 2026-07-02). Restored below.
-  const sessionData = () => ({ sessionId: entry.session?.sessionId || resume || null, log: entry.log, commands: entry.commands, mode: entry.mode, model: entry.model || null, provider: entry.provider || null, spawnedBy: entry.spawnedBy, flowSessionId: entry.flowSessionId, flowTaskKey: entry.flowTaskKey, cwd: entry.cwd, rolePrompt: entry.rolePrompt, flowReviewTarget: entry.flowReviewTarget || null, revertTarget: entry.revertTarget || null, reviewSliceRoots: entry.reviewSliceRoots || [], openedAt: entry.openedAt || null })
+  const sessionData = () => ({ sessionId: entry.session?.sessionId || resume || null, log: entry.log, commands: entry.commands, mode: entry.mode, model: entry.model || null, provider: entry.provider || null, spawnedBy: entry.spawnedBy, flowSessionId: entry.flowSessionId, flowTaskKey: entry.flowTaskKey, cwd: entry.cwd, rolePrompt: entry.rolePrompt, flowReviewTarget: entry.flowReviewTarget || null, revertTarget: entry.revertTarget || null, reviewSliceRoots: entry.reviewSliceRoots || [], openedAt: entry.openedAt || null, lastUsage: entry.lastUsage || null })
   const persist = () => saveSession(room, id, sessionData())
   // Synchronous flush of this session's record. Used on open (so a brand-new session
   // has a file under its id BEFORE its first event — surviving a restart inside the
@@ -2227,6 +2233,10 @@ function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rol
         const m = onBuiltin ? (evt.model || evt.ctx?.model) : null
         if (m && m !== entry.model) { entry.model = m; announce() }
         if (evt.kind === 'models' && Array.isArray(evt.models) && evt.models.length) { roomModels = evt.models; announce() } }
+      // Cache the last usage/ctx meter on the entry (persisted + replayed — see
+      // openStructured). Chrome events bypass pushLog/persist in emitTail below, so
+      // usage needs an explicit persist() here to survive a bridge restart.
+      if (evt.kind === 'usage') { entry.lastUsage = evt; persist() }
       // FL-B2 — fold this flow lane's completed-turn output tokens into its budget so the
       // autopilot cap can halt the next wave before overrun (output_tokens = the billed
       // reasoning+output spend the indicator already tracks; conservative enough for a guard).
@@ -2712,8 +2722,16 @@ channel
       // the terminal. chunkReplayEvents splits it so nothing in the recent window is lost;
       // classifyReplay applies the chunks in arrival order. Older history beyond the chunk
       // budget is covered by the client's DB transcript snapshot.
+      // Attach the terminal's last usage/ctx meter to the FIRST replay chunk. usage is
+      // chrome (filtered out of `events` by both bridge + client), so without this a
+      // (re)joiner's ctx meter is empty ("—") until the terminal's next turn. The client
+      // applies `payload.lastUsage` in its code-replay handler (which is `to`-targeted, so
+      // only the joiner gets it). Gated on `.ctx` — a suppressed meter (correctContext
+      // nulled it for an unknown non-Claude model) replays nothing, never a misleading %.
+      let firstChunk = true
       for (const events of chunkReplayEvents(tail)) {
-        bcast('code-replay', { to: payload?.to ?? null, term: id, events, ...(ahead ? { reset: true } : {}) })
+        bcast('code-replay', { to: payload?.to ?? null, term: id, events, ...(ahead ? { reset: true } : {}), ...(firstChunk && s.lastUsage?.ctx ? { lastUsage: s.lastUsage } : {}) })
+        firstChunk = false
       }
     }
     // Re-send any still-pending permission/question cards. They ride a one-shot
@@ -3060,7 +3078,7 @@ channel
             // FL-M6 — restore the flow context (id/role/cwd) so an in-flight flow survives a
             // bridge restart: the conductor keeps its subagent-block + plan interception, and
             // lanes keep their worktree cwd + the ability to mark done.
-            openStructured({ id: rec.id, model: rec.model || undefined, provider: rec.provider || undefined, resume: canResume(rec) ? rec.sessionId : undefined, log: rec.log, commands: rec.commands, mode: rec.mode, spawnedBy: rec.spawnedBy, flowSessionId: rec.flowSessionId, flowTaskKey: rec.flowTaskKey, cwd: rec.cwd, rolePrompt: rec.rolePrompt, reviewSliceRoots: rec.reviewSliceRoots, openedAt: rec.openedAt,
+            openStructured({ id: rec.id, model: rec.model || undefined, provider: rec.provider || undefined, resume: canResume(rec) ? rec.sessionId : undefined, log: rec.log, commands: rec.commands, mode: rec.mode, spawnedBy: rec.spawnedBy, flowSessionId: rec.flowSessionId, flowTaskKey: rec.flowTaskKey, cwd: rec.cwd, rolePrompt: rec.rolePrompt, reviewSliceRoots: rec.reviewSliceRoots, openedAt: rec.openedAt, lastUsage: rec.lastUsage,
               // Lazy-boot restored terminals that were IDLE + not part of a flow: their transcript
               // shows immediately; the query boots on first turn. Mid-turn + flow terminals boot now
               // (mid-turn needs auto-resume; flow needs its lane live).

package/context-windows.mjs

@@ -86,7 +86,13 @@ export function correctContext(raw, env = process.env) {
   const correctedMax = override ?? mapped
 
   if (correctedMax == null) {
-    // No correction available — but never ship a >100% pct.
+    // No correction available. For a NON-Claude model the SDK's max is computed for the
+    // wrong model id (it believes it's talking to Claude over ANTHROPIC_BASE_URL) → a %
+    // off that wrong max misleads (the GLM-ran-past-100 bug, 2026-07-04). Rather than ship
+    // a wrong number, SUPPRESS the meter: return null so the mode row + Session info show
+    // nothing / "—" instead. We only display ctx when we know the true window — a Claude
+    // id (SDK max trusted), a known-map model (corrected above), or TP_CONTEXT_MAX (override).
+    if (!isClaudeModel(raw.model)) return null
     const pct = Number(raw.pct)
     if (Number.isFinite(pct) && pct > 100) return { ...raw, pct: 100, over: true }
     return raw

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thinkpool-pair",
-  "version": "0.7.179",
+  "version": "0.7.180",
   "description": "Share a local coding-agent CLI (Claude Code, Codex, Gemini, Aider, …) into a ThinkPool Code room, live.",
   "type": "module",
   "bin": {