chore: release v8.0.0

Major, for one breaking change: `team(mode:"run")` no longer waits for the
models. It starts them and returns a slot map; the caller polls `mode:"status"`.

The deadline it replaces was actively harmful. In session team-20260827-0015 it
killed three of five slots that were all working — 45, 13 and 24 tool calls, up
to 30,557 output tokens at the moment they died — because its only progress
signal was token flow, which stops for the entire duration of a local tool call.
A model running `go test ./...` was indistinguishable from a hung one. Every kill
landed within 133ms of a 60s multiple: the watcher's poll cadence, not a
deadline at all.

`timeout` is removed with it, from the tool schema, `TeamRunOptions` and every
`runModels` call site. It controlled nothing once the reaper was gone.

Consumers must migrate. `/multimodel:team` in the magus marketplace repo reads
per-model results straight from the `run` response and passes `timeout`; both
break. Migration guide with before/after for every changed call:
ai-docs/reports/team-mcp-breaking-change-plugin-migration.md

Also ships two latent bugs found by comparing team's child supervision against
the channel's — non-ASCII answers measured wrong, and upstream-error capture
dead for every team child since the feature existed — plus `bun --cwd` script
ordering that had silently skipped packages/macos-bridge in test, lint and
format. Full detail in CHANGELOG.md.

Verified: typecheck clean; build clean; `bun run test:safe` green —
3017 pass / 17 skip / 0 fail across 193 files (packages/cli) plus 20 pass /
0 fail (packages/macos-bridge), exit 0.

Co-Authored-By: Magus <magus@madappgang.com>

Crafted with agentic harness Magus (https://github.com/MadAppGang/magus)

Author: erudenko

CHANGELOG.md

@@ -2,6 +2,70 @@
 
 All notable changes to [Claudish](https://github.com/MadAppGang/claudish).
 
+## [8.0.0] - 2026-08-27
+
+### BREAKING CHANGES
+
+- **`team(mode:"run")` no longer waits for the models.** It starts them and returns a slot
+  map immediately; poll `mode:"status"` for progress and results. A team slot is a full
+  Claude Code session and may work for a long time, and holding the tool call open made the
+  run's duration the client's problem — a real run was aborted at exactly 1800s of client
+  idle timeout. `mode:"run-and-judge"` is unchanged and still blocks.
+- **The `timeout` parameter is removed** from the `team` tool, `TeamRunOptions`, and the
+  `runModels` call sites. It controlled nothing once the deadline was deleted. `--timeout`
+  survives in `team-cli` for the magmux grid path, which has its own.
+- Consumers must migrate. See `ai-docs/reports/team-mcp-breaking-change-plugin-migration.md`.
+
+### Features
+
+- **Nothing terminates a team slot on a timer.** The stall reaper killed three of five
+  productive slots in session `team-20260827-0015`; its only progress signal was
+  `stats/<id>.json`, which advances on token flow and therefore freezes for the whole
+  duration of a local tool call. Every kill landed within 133ms of a 60s multiple — the
+  watcher's poll cadence, not a deadline. Claude Code already bounds every tool call itself
+  (Bash default 120s, max 600s), so the detector guarded a condition the harness prevents.
+- `team(mode:"cancel", slot?)` — the only thing that kills a slot now, and it kills the
+  process group. A cancelled slot records `reason: "cancelled"`, distinct from
+  `nonzero_exit`: it is a decision, not a defect.
+- `mode:"status"` reports `idle_seconds_by_slot` and `activity_by_slot`. Read together,
+  they separate a slot inside a build from one that stopped mid-answer. Reported, never
+  acted on. Once settled, `status` also carries `summary`, the result card `run` used to
+  return.
+- `input_file` on the `team` tool, preferred over `input`. A vote prompt is typically a
+  200-line brief, and inline text is echoed verbatim in the caller's terminal. Contained to
+  the working directory on the same terms as `path`.
+- Channel sessions report `idleSeconds` on `list_sessions`, `get_output` and
+  `get_diagnostics`.
+
+### Bug Fixes
+
+- **`team` mangled non-ASCII answers.** It read child stdout with `chunk.toString()`, which
+  replaces a codepoint split across a pipe read boundary with U+FFFD. Measured: a 4-byte
+  emoji became three replacement characters AND the byte count inflated 19 → 24 — and
+  `outputSize` feeds `minOutputBytes` and the empty check, so answers were MEASURED wrong,
+  not merely displayed wrong. Both subsystems now share `stdio-decode.ts`.
+- **`team`'s upstream-error capture never ran.** `captureUpstreamError` is gated on
+  `CLAUDISH_UPSTREAM_ERROR_LOG`, which `team` never set — a guaranteed no-op for every team
+  child since the feature existed. The provider body that separates a retryable rate limit
+  from a hard quota wall was discarded as soon as it was classified. Now set per slot, and
+  `ModelError.upstreamErrorLogPath` names the file only when one was written.
+- **`bun --cwd <dir> run <script>` runs nothing and exits 0.** Bun prints usage. All four
+  occurrences sat on the right of an `&&`, so `test`, `lint` and `format` silently skipped
+  `packages/macos-bridge` — 20 tests that had never run through the root scripts.
+
+### Refactor
+
+- One stream-json parser. `team` fed `createAssistantTextCapture()` directly and hand-rolled
+  the supervision; `StreamJsonReducer` is now its only production consumer. The parsers
+  disagreed on one case — valid JSON that is not stream-json vocabulary — which is now the
+  documented `keepUnrecognizedJson` option rather than one rule imposed on both.
+
+### Documentation
+
+- `ai-docs/architecture/team-lifecycle.md` — why nothing kills a slot, the measurement that
+  removed the deadline, and the `SessionManager` adoption route that was evaluated and
+  rejected.
+
 ## [7.67.1] - 2026-08-27
 
 ### Bug Fixes

ai-docs/reports/team-mcp-breaking-change-plugin-migration.md

@@ -0,0 +1,223 @@
+# Migration: `/multimodel:team` against the new `team` MCP contract
+
+**For:** whoever maintains `plugins/multimodel/commands/team.md` in the magus
+marketplace repo.
+**Why:** claudish's `team` tool changed in a way that breaks that command. The
+two must ship together.
+**File to change:** `plugins/multimodel/commands/team.md` (Step 2 and Step 3).
+
+---
+
+## What broke, in one sentence
+
+`team(mode:"run")` no longer waits for the models and no longer returns their
+results. It starts them and returns a slot map immediately.
+
+The command currently calls `run` and reads per-model results straight out of
+the response. Those results are not there any more, so vote parsing gets a JSON
+object with no votes in it and the panel reports INCONCLUSIVE every time.
+
+There is a second, quieter break: `timeout` was removed from the tool schema. The
+schema does not set `additionalProperties: false`, so passing it is not an error
+— it is silently ignored. The command will look like it still sets a deadline
+when nothing reads it.
+
+---
+
+## Why it changed
+
+A `team` slot is a full Claude Code session and can legitimately work for a long
+time. The old shape held the MCP tool call open for the whole run, which made the
+run's duration the client's problem — a real run was aborted at exactly 1800s of
+client idle timeout.
+
+The deadline that existed to bound it was worse: in session
+`team-20260827-0015` it killed three of five slots that were all actively
+working, because its only progress signal was token flow, which stops during a
+local tool call. A model running `go test ./...` looked identical to a hung one.
+
+Nothing terminates a slot on a timer now. The caller polls, looks at the
+evidence, and decides. Full rationale: `ai-docs/architecture/team-lifecycle.md`
+in the claudish repo.
+
+---
+
+## Step 2 — the call
+
+### Before
+
+```
+claudish team(mode="run", path=SESSION_DIR,
+  models=[...ALL resolved models, "internal" included...],
+  input=VOTE_PROMPT, timeout=180,
+  require_pattern="```vote", agent=RESOLVED_AGENT)
+```
+
+### After
+
+```
+# 1. Write VOTE_PROMPT to SESSION_DIR/input.md first.
+# 2. Then:
+claudish team(mode="run", path=SESSION_DIR,
+  models=[...ALL resolved models, "internal" included...],
+  input_file="SESSION_DIR/input.md",
+  require_pattern="```vote", agent=RESOLVED_AGENT)
+```
+
+Three changes:
+
+- **`timeout` is gone.** Remove it.
+- **`input_file` replaces `input`.** Both still work and passing BOTH is a hard
+  error, but prefer the file. A vote prompt is 100+ lines and passing it inline
+  echoes the whole thing verbatim in the user's terminal, burying every other
+  argument in the tool call. The path must be inside the working directory.
+- **`require_pattern` is unchanged** and still the point. Keep it.
+
+### What `run` returns now
+
+```json
+{
+  "started": true,
+  "team_session_id": "team-20260827-0015",
+  "session_path": "/abs/path/to/SESSION_DIR",
+  "slots": { "gpt-5.6-sol": "01", "grok-4.6": "02", "internal": "03" },
+  "next": { "status": "...", "cancel": "...", "judge": "..." },
+  "note": "..."
+}
+```
+
+`slots` maps the display model name to its anonymised slot id. That slot id
+addresses everything else on disk for that model:
+
+| Path | Contents |
+|---|---|
+| `<session_path>/response-<slot>.md` | the model's answer — parse votes from here |
+| `<session_path>/stats/<slot>.json` | tokens, cost, tool counts |
+| `<session_path>/errors/<slot>.log` | stderr and diagnostics, on failure |
+| `<session_path>/errors/<slot>-upstream.jsonl` | raw provider error bodies, when any |
+
+---
+
+## Step 2b — the new polling step (this is the part that did not exist)
+
+Between starting the run and parsing votes, poll until every slot has left
+`RUNNING`:
+
+```
+claudish team(mode="status", path=SESSION_PATH)
+```
+
+Returns the full `TeamStatus` plus three added fields:
+
+```json
+{
+  "startedAt": "...",
+  "models": {
+    "01": { "state": "COMPLETED", "exitCode": 0, "outputSize": 10988, ... },
+    "02": { "state": "RUNNING",   "exitCode": null, ... }
+  },
+  "idle_seconds_by_slot": { "02": 94 },
+  "activity_by_slot":     { "02": "tool_executing" },
+  "note": "...",
+  "summary": "<rendered result card — present ONLY once the run has settled>"
+}
+```
+
+**Settled means: no slot in `models` has `state === "RUNNING"`.** That is the
+loop condition.
+
+**`summary` is the string the old `run` used to return.** Once the run settles,
+`status` carries the same rendered result card — `N/M succeeded`,
+`reason=shape_mismatch`, and the rest. If the command's Step 3 was matching on
+that text, it can keep doing so; it just reads it from a settled `status`
+instead of from `run`.
+
+Bound the loop, and fail loudly rather than looping forever. There is no
+server-side deadline any more, so an unbounded poll is an unbounded wait.
+
+---
+
+## Step 2c — deciding whether a quiet slot is stuck
+
+This is the capability the deadline used to take away from you.
+
+- `idle_seconds_by_slot` — seconds since that slot's child last wrote anything.
+- `activity_by_slot` — what it is doing: `running`, `tool_executing`,
+  `waiting_for_input`, or a terminal state.
+
+**Read them together.** Ninety seconds of silence in `tool_executing` is a build
+or a test suite running, and is completely normal. The same ninety seconds in
+`running` is a model that stopped mid-answer. The old reaper had no state at all,
+which is exactly why it killed the first kind.
+
+Nothing cancels on your behalf. If you decide a slot is wedged:
+
+```
+claudish team(mode="cancel", path=SESSION_PATH, slot="02")   # one slot
+claudish team(mode="cancel", path=SESSION_PATH)              # the whole run
+```
+
+A cancelled slot is recorded with `error.reason === "cancelled"`, which is
+distinct from `nonzero_exit`. It is not a defect — it is a decision, and the
+report should say so rather than showing the model as crashed.
+
+Suggested default for a vote panel: do not cancel automatically. Report the slot
+as still running and let the user decide. Losing a vote to an impatient
+auto-cancel is the same failure the deadline used to cause.
+
+---
+
+## Step 3 — parsing votes
+
+Unchanged in substance, different in where the text comes from.
+
+**Before:** per-model results were read from the `run` response.
+
+**After:** read `<session_path>/response-<slot>.md` for each slot in the map
+returned by `run`, after `status` reports settled. The vote regex is unchanged:
+
+```
+/```vote\s*\n([\s\S]*?)\n\s*```/
+```
+
+Failure handling is unchanged too: a slot whose `models[<slot>].state` is
+`FAILED` or `EMPTY` did not vote. `error.reason` tells you why —
+`shape_mismatch` means it answered but never produced the required block, which
+still must NOT be counted as a vote.
+
+---
+
+## What did NOT change
+
+- `require_pattern` and `min_output_bytes` — same semantics, still recommended.
+- `agent` — same, still applies to every model in the run.
+- `mode:"judge"` — unchanged.
+- `mode:"run-and-judge"` — still BLOCKING, and still returns the verdict. If the
+  command would rather not implement polling at all, this mode is the drop-in
+  path. The trade is that it holds the tool call open for the whole run, which is
+  the shape that hit the client's 1800s idle abort.
+- Native Claude names (`internal`, `default`, `opus`, …) are still ordinary
+  slots and still belong in `models`.
+
+---
+
+## Minimum viable change
+
+If you want the smallest possible diff and are willing to keep the blocking
+behaviour:
+
+1. Remove `timeout` from the call.
+2. Change `mode:"run"` to `mode:"run-and-judge"`, or keep `run` and add the poll.
+
+If you want the shape the tool is now designed around, implement the poll in
+Step 2b — it is what lets the panel survive a slow model instead of losing its
+vote.
+
+---
+
+## Version
+
+Ships in claudish **v8.0.0**. The major bump is for this change specifically.
+Pin or require `>=8.0.0` once the command is updated, and note that a plugin
+updated for v8 will not work against v7.67.x — `input_file` does not exist there
+and `run` still blocks.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "claudish-monorepo",
-  "version": "7.67.1",
+  "version": "8.0.0",
   "private": true,
   "description": "Monorepo for Claudish - Run Claude Code with any model",
   "type": "module",

packages/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claudish",
-  "version": "7.67.1",
+  "version": "8.0.0",
   "description": "Run Claude Code with any model - OpenRouter, Ollama, LM Studio & local models",
   "type": "module",
   "main": "./dist/index.js",

packages/cli/src/version.ts

@@ -1,2 +1,2 @@
 // Auto-generated by scripts/generate-version.ts — do not edit
-export const VERSION = "7.67.1";
+export const VERSION = "8.0.0";