@caddis/cli 0.3.0 → 0.4.0

package/bundles/antigravity-plugin/agents/knowledge-transfer.md

@@ -56,36 +56,22 @@ Pre-empt it before writing the relay (use Grep/Read — you don't run the checke
 - Fix gaps directly — keeping these docs current is knowledge-doc work, squarely in your remit. Record the
   edits under `live_writes` in your return block.
 
-## Feed Dream Memory (the short-term fact store)
-Dream Memory (`.caddis/memory.jsonl`, one JSON fact per line) is caddis's automatic, **decaying
-short-term** memory — reinforced when a fact recurs, pruned when it doesn't. The Stop hook already
-captures the *mechanical* kinds (a command that failed, a build that went red→green). You are the only
-place that can capture the **reasoned** kinds, so add them as a byproduct — cheap insurance that a
-hard-won "don't do that" resurfaces next session even before it graduates into a curated doc. This does
-**not** replace the durable write above — do both; this is short-term, that is long-term.
+## Dream Memory — RETIRED 2026-08-26. Do not write to `.caddis/memory.jsonl`.
+Both halves of the mechanism this section used to describe are gone: the Stop hook's automatic
+capture (`dream_capture.py`) and the SessionStart surfacing in `inject_relay.py` were both removed
+the same day. Reason, verbatim from the retirement comment (`inject_relay.py`): it ranked facts by
+hit count, so the most-repeated shell typo always outranked a real insight — 128 of 131 records were
+`failure-mode`, the top six were a month stale with counts of 69-77, and the two genuinely useful
+records sat at `hitCount: 1` and never surfaced. A count of 77 meant that command failed 77 times
+**while its own warning held the top slot** — the mechanism never changed behaviour.
 
-**Append** — for each `rejected-approach` (an approach tried and abandoned, so no one retries it) or
-`repo-fact` (a non-obvious, durable truth about how this repo works) from this session, append ONE line
-to `.caddis/memory.jsonl` (create it if absent; never rewrite or reorder existing lines — the next
-Stop consolidates and dedups). Every field is required:
-```
-{"kind":"rejected-approach","key":"poll-api-for-status","summary":"don't poll the API for status — use the SSE stream","hitCount":1,"firstSeen":"2026-07-01T00:00:00Z","lastSeen":"2026-07-01T00:00:00Z","source":"knowledge-transfer"}
-```
-- `kind` is ONLY `rejected-approach` or `repo-fact` — the mechanical kinds are the hook's job, not yours.
-- `key` is a short, stable, lowercase dedup phrase (it's the identity; the same lesson must reuse the
-  same key so it reinforces instead of duplicating). `summary` is ≤1 imperative line.
-- **Redact secrets** — never put a token, password, env value, or credentialed URL in `key`/`summary`.
-  Store the lesson, not the secret; if you can't state it without a secret, don't store it.
-- Use today's date for both timestamps (a date-precision ISO like `2026-07-01T00:00:00Z` is fine).
-- Malformed lines are silently dropped on load, so a fumbled line is safe — but aim to get it right.
-
-**Promote (the boundary — do this when you see it):** if the store already holds a fact with
-`hitCount >= 3` (it has recurred across sessions), it has earned a permanent home. Write it into the
-right curated doc — a `.caddis/kb/*.md` note or the most specific live `AGENTS.md` (same routing as
-above) — then **delete that one fact's line** from `.caddis/memory.jsonl` so it doesn't double-surface.
-Promotion is one-way and one-at-a-time: the curated docs hold what survived; Dream Memory holds what's
-still proving itself. (Dream Memory is gitignored/local; promotion is how a repeatedly-useful fact
-becomes durable and shareable.)
+The file (`.caddis/memory.jsonl`) is left on disk but nothing reads it anymore — appending a fact
+there now has zero effect, good or bad. **Do not append to it.** Claude Code's own per-repo memory
+(`~/.claude/projects/<slug>/memory/*.md` + `MEMORY.md`) does the short-term/reasoned-fact job instead;
+it is maintained by the harness natively, not by this agent writing a file. If you land a
+`rejected-approach` or `repo-fact` this session, it goes straight into the durable write above (a
+`.caddis/kb/*.md` note or the most specific `AGENTS.md`) — there is no more intermediate decaying
+tier to stage it in first.
 
 ## Return format (always end with this)
 ```
@@ -94,10 +80,9 @@ knowledge_transfer:
     - file: <path>   section: <heading>   nugget: <one-line summary>
   secondary_writes:
     - <session-log entry, or "none">
-  dream_memory:
-    - <kind + key of each fact appended to .caddis/memory.jsonl, or "none">
   promotions:
-    - <fact key promoted to <curated doc> and removed from the store, or "none">
+    - <legacy field, kept for callers that still read it — Dream Memory retired 2026-08-26, so this is
+      always "none" now>
   adr_flag:
     - <decision worth formalizing as an ADR, or "none">
   skipped:

package/bundles/antigravity-plugin/caddis_statusline.py

@@ -631,9 +631,31 @@ def install(hosts, dry_run: bool = False) -> int:
     failures = 0
     for host in hosts:
         settings = HOST_SETTINGS[host]
-        # agy runs its status-line command through cmd.exe on Windows, where a quoted
-        # path containing spaces breaks. Home is space-free here; keep it native.
-        command = '%s "%s" --profile %s' % (py, str(INSTALLED_SCRIPT), host)
+        # QUOTING DIFFERS BY HOST, and getting it wrong breaks the line silently.
+        #
+        # Claude Code runs its statusLine command through a SHELL, so a quoted path is
+        # correct and a path containing spaces works.
+        #
+        # agy does NOT. It splits the string itself, so quotes survive as literal path
+        # characters and the result is resolved against the session cwd. Observed live
+        # 2026-08-27: agy reported it could not open a file whose path was the session
+        # cwd with the quoted absolute path appended inside it. The user own working agy
+        # line had always been unquoted; mine was not.
+        #
+        # Consequence to be honest about: an agy install whose home contains a space
+        # cannot be expressed this way. Detect it and say so, rather than writing a
+        # command that fails at render time with a confusing message.
+        target = str(INSTALLED_SCRIPT)
+        if host == "agy":
+            if " " in target:
+                print("  !! agy    home path contains a space (%s)." % target)
+                print("            agy does not parse statusLine through a shell, so the path")
+                print("            cannot be quoted. Wire agy statusLine by hand. Skipped.")
+                failures += 1
+                continue
+            command = "%s %s --profile %s" % (py, target, host)
+        else:
+            command = '%s "%s" --profile %s' % (py, target, host)
         try:
             failures += wire_host(host, settings, command, dry_run)
         except Exception as exc:  # never leave a half-written settings file unexplained

package/bundles/antigravity-plugin/guard_agy.py

@@ -303,8 +303,27 @@ def _hook_note(feature, exc, root=None):
                 sys.path.insert(0, _cand)
         import hook_log as _hl  # noqa: E402
         _hl.record(os.path.basename(__file__).replace(".py", ""), feature, exc, root)
-    except Exception as _exc:
-        _hook_note("agy guard config read", _exc)
+    except Exception:
+        pass  # the ledger must never recurse into itself
+
+
+def _workspace_roots(data):
+    """Every workspace root in the payload, as a list. Empty when the payload has none.
+
+    agy sends `workspacePaths` as a LIST. All three caddis agy hooks used to read
+    `workspacePaths[0]` and treat it as "the workspace" — silently, with no signal that a
+    choice had been made. In a single-root workspace that is correct and the behaviour here
+    is unchanged. In a MULTI-root workspace it is a coin toss: the hook acts on the first
+    root regardless of which one the session is actually working in.
+
+    Nothing on this machine is multi-root today, so this has never fired. It is a live trap
+    for exactly the multi-repo workspaces the leakage investigation was about, which is why
+    each caller below now handles the list rather than assuming its first element.
+    """
+    roots = data.get("workspacePaths") if isinstance(data, dict) else None
+    if not isinstance(roots, list):
+        return []
+    return [str(r) for r in roots if isinstance(r, (str, bytes)) and str(r).strip()]
 
 def main() -> None:
     tier, reason = "allow", ""
@@ -314,7 +333,7 @@ def main() -> None:
             try:
                 _reconfig(encoding="utf-8")
             except Exception as _exc:
-                _hook_note("agy guard config read", _exc)
+                _hook_note("agy stdout utf-8 reconfigure", _exc)
         try:
             data = json.load(sys.stdin)
         except Exception:
@@ -324,12 +343,25 @@ def main() -> None:
         call = data.get("toolCall")
         if not isinstance(call, dict):
             return
-        workspaces = data.get("workspacePaths") or []
-        root = workspaces[0] if isinstance(workspaces, list) and workspaces else os.getcwd()
-        if guard_disabled(str(root)):
+        roots = _workspace_roots(data) or [os.getcwd()]
+
+        # Every per-repo guard setting RELAXES the guard: `enabled = false` bypasses all
+        # tiers, `mode = "deny-only"` drops the ask tier, and `[guard] allow` downgrades an
+        # ask. Reading them from workspacePaths[0] meant one repo's relaxation silently
+        # applied to every other repo in the workspace — repo A's kill switch unguarding
+        # repo B. Require UNANIMITY instead: a relaxation applies only when every root asks
+        # for it. One root is the common case and behaves exactly as before.
+        if all(guard_disabled(r) for r in roots):
             return  # kill switch: bypass every tier
-        tier, reason = decide(str(call.get("name") or ""), call.get("args"), _load_allow(str(root)))
-        if tier == "ask" and deny_only(str(root)):
+
+        # The allow-list downgrades an ask, so with several roots take only the entries
+        # present in ALL of them — the strictest reading, not the most permissive.
+        allow = _load_allow(roots[0])
+        for other in roots[1:]:
+            allow = [a for a in allow if a in _load_allow(other)]
+
+        tier, reason = decide(str(call.get("name") or ""), call.get("args"), allow)
+        if tier == "ask" and all(deny_only(r) for r in roots):
             tier = "allow"  # deny-only: keep catastrophe blocks, never force_ask (no plan interruption)
     except Exception:
         return  # ANY unexpected shape/error → no guard, never a crash and never a bad decision
@@ -337,7 +369,7 @@ def main() -> None:
         try:
             _emit(tier, reason)
         except Exception as _exc:
-            _hook_note("agy guard config read", _exc)
+            _hook_note("agy guard decision emit", _exc)
         # Always exit 0: a non-zero code from a PreToolUse hook is an agy hook *error*.
         sys.exit(0)
 

package/bundles/antigravity-plugin/plugin.json

@@ -1,5 +1,5 @@
 {
   "name": "caddis",
-  "version": "1.3.77",
+  "version": "1.3.83",
   "description": "caddis — agent-agnostic dev harness as an Antigravity (agy) plugin: TDD skills + commands, single-sourced with the Claude Code plugin."
 }

package/bundles/antigravity-plugin/scripts/caddis_inventory.py

@@ -290,12 +290,82 @@ def collect(root: Path, with_tests: bool = False) -> Inventory:
     if with_tests:
         # Run it. A handover once claimed 20 tests when there were 19, and the incoming Hub found
         # the discrepancy before anyone else did.
-        out = _run([sys.executable, "-m", "pytest", "-q", "--tb=no"], root)
-        tail = [l for l in out.splitlines() if "passed" in l or "failed" in l]
-        inv.tests = tail[-1] if tail else "suite did not report a summary — investigate"
+        summary, shown = _run_tests(_test_command(root), root)
+        # Name the command. "932 passed" means nothing on its own — a reader who cannot see
+        # the command cannot tell a scoped run from a whole-tree one, which is exactly the
+        # ambiguity that produced this defect.
+        inv.tests = f"{summary}  _(`{shown}`)_"
     return inv
 
 
+
+# Generated/vendored trees hold COPIES of the harness. They have no node_modules and no
+# import path, so collecting them is never right — it only produces collection errors that
+# mask the real result. Excluded by default; `[handover] test_cmd` overrides everything.
+_GENERATED_TREES = ("vscode-extensions", "dist", "node_modules", ".venv")
+
+# The suite is the one thing here that is not a directory listing, so it gets its own
+# runner rather than `_run`. Three reasons `_run` is wrong for it:
+#   * it returns "" on a non-zero exit — and pytest exits non-zero when tests FAIL, so a
+#     genuinely red suite became "did not report a summary", indistinguishable from a
+#     broken invocation. A red suite MUST report "3 failed".
+#   * its 30s timeout is shorter than a real suite. This repo's takes 112s, so
+#     `--with-tests` could never have produced a number here whatever else was fixed.
+#   * it discards stderr, where pytest writes collection errors.
+def _run_tests(cmd: list[str], cwd: Path, timeout: int = 900) -> tuple[str, str]:
+    """Return ``(summary_line, command_string)``. Never raises."""
+    shown = " ".join(cmd[1:]) if cmd and cmd[0] == sys.executable else " ".join(cmd)
+    shown = shown.replace("-m pytest", "pytest", 1)
+    try:
+        r = subprocess.run(cmd, cwd=str(cwd), capture_output=True, text=True,
+                           encoding="utf-8", errors="replace", timeout=timeout)
+    except subprocess.TimeoutExpired:
+        return (f"suite exceeded {timeout}s — not run to completion", shown)
+    except Exception as exc:
+        return (f"could not run the suite ({type(exc).__name__})", shown)
+
+    blob = (r.stdout or "") + chr(10) + (r.stderr or "")
+    # pytest's summary line is the last one naming passed/failed/error.
+    tail = [l.strip() for l in blob.splitlines()
+            if ("passed" in l or "failed" in l or "error" in l.lower()) and "==" not in l[:2]]
+    if tail:
+        return (tail[-1], shown)
+    return (f"suite did not report a summary (exit {r.returncode}) — investigate", shown)
+
+
+def _test_command(root: Path) -> list[str]:
+    """The repo's test command.
+
+    A repo's test command is a PROJECT FACT, and this used to guess it. Bare `pytest` is a
+    guess that is right in a single-package repo and wrong in any repo carrying a vendored
+    or generated tree — which describes caddis itself and every consumer that vendors
+    anything. Set `[handover] test_cmd` in `.caddis/config.toml` to state it.
+    """
+    # In the SHIPPED bundle this file and claudster_config.py are co-located under
+    # scripts/. In the source repo they are not: this lives in scripts/, that one in
+    # claude-harness/scripts/. Try both, or the config is silently never read — which is
+    # how the first version of this fix looked like it worked and did not.
+    here = Path(__file__).resolve().parent
+    configured = ""
+    for cand in (here, here.parent / "claude-harness" / "scripts"):
+        try:
+            if str(cand) not in sys.path:
+                sys.path.insert(0, str(cand))
+            import claudster_config as _cc  # noqa: E402
+            configured = _cc.get_str(_cc.load_config(root, "handover"), "test_cmd", "")
+            break
+        except Exception:
+            continue
+    if configured:
+        import shlex
+        return shlex.split(configured)
+    cmd = [sys.executable, "-m", "pytest", "-q", "--tb=no"]
+    for name in _GENERATED_TREES:
+        if (root / name).exists():
+            cmd += ["--ignore", name]
+    return cmd
+
+
 def _table(rows: list[dict], cols: list[tuple[str, str]]) -> list[str]:
     head = "| " + " | ".join(c[0] for c in cols) + " |"
     sep = "|" + "|".join("---" for _ in cols) + "|"

package/bundles/antigravity-plugin/scripts/caddis_statusline.py

@@ -631,9 +631,31 @@ def install(hosts, dry_run: bool = False) -> int:
     failures = 0
     for host in hosts:
         settings = HOST_SETTINGS[host]
-        # agy runs its status-line command through cmd.exe on Windows, where a quoted
-        # path containing spaces breaks. Home is space-free here; keep it native.
-        command = '%s "%s" --profile %s' % (py, str(INSTALLED_SCRIPT), host)
+        # QUOTING DIFFERS BY HOST, and getting it wrong breaks the line silently.
+        #
+        # Claude Code runs its statusLine command through a SHELL, so a quoted path is
+        # correct and a path containing spaces works.
+        #
+        # agy does NOT. It splits the string itself, so quotes survive as literal path
+        # characters and the result is resolved against the session cwd. Observed live
+        # 2026-08-27: agy reported it could not open a file whose path was the session
+        # cwd with the quoted absolute path appended inside it. The user own working agy
+        # line had always been unquoted; mine was not.
+        #
+        # Consequence to be honest about: an agy install whose home contains a space
+        # cannot be expressed this way. Detect it and say so, rather than writing a
+        # command that fails at render time with a confusing message.
+        target = str(INSTALLED_SCRIPT)
+        if host == "agy":
+            if " " in target:
+                print("  !! agy    home path contains a space (%s)." % target)
+                print("            agy does not parse statusLine through a shell, so the path")
+                print("            cannot be quoted. Wire agy statusLine by hand. Skipped.")
+                failures += 1
+                continue
+            command = "%s %s --profile %s" % (py, target, host)
+        else:
+            command = '%s "%s" --profile %s' % (py, target, host)
         try:
             failures += wire_host(host, settings, command, dry_run)
         except Exception as exc:  # never leave a half-written settings file unexplained

package/bundles/antigravity-plugin/scripts/claudster_config.py

@@ -23,11 +23,6 @@ Config schema (see ``.caddis/config.toml.example``)::
     page_guide = "UI_PAGE_GUIDE.md"
     claude_md_budget = 200
     ignore_routes = ["/health"]
-
-    [dream_memory]
-    prune_age_days = 14
-    max_facts = 200
-    surface_limit = 5
 """
 
 from __future__ import annotations

package/bundles/antigravity-plugin/scripts/hook_log.py

@@ -8,7 +8,7 @@ they fail open is ``except Exception: pass``, which is wrong: a hook that is bro
 and a hook that had nothing to say produce the identical result — nothing.
 
 There were 20 such swallows across 6 hook files. Each one guards a named feature.
-When ``dream_memory``'s surface stops printing, nobody learns that it stopped; they
+When the tidy nudge stops printing, nobody learns that it stopped; they
 learn that they have not seen it lately, months later, if at all.
 
 The fix keeps fail-open and adds a record. ``record()`` never raises, never prints,

package/bundles/antigravity-plugin/scripts/new-machine.ps1

@@ -0,0 +1,170 @@
+<#
+.SYNOPSIS
+  Set up a new Windows machine with the caddis toolchain, in the right order.
+
+.DESCRIPTION
+  The ORDER matters, and it is the only thing this script really encodes:
+
+    1. Agents first.  caddis is a META-installer — `caddis init` detects which agents are
+       present and installs into each. An agent installed AFTER it is simply not seen,
+       and you get a silent no-op rather than an error.
+
+       Note the rule is about `caddis init`, NOT about installing the CLI. The npm package
+       has no postinstall step; it only puts binaries on PATH. Installing it early is safe.
+    2. caddis CLI second, which brings the model-lane binaries with it.
+    3. Keys and status line last. Both depend on the CLI being on PATH.
+
+  Fetch this script on a machine that has no caddis yet (the SOURCE repo is private; the
+  public mirror is what a new machine can reach):
+
+    irm https://raw.githubusercontent.com/saajunaid/caddis-plugin/main/plugin/scripts/new-machine.ps1 -OutFile $env:TEMP
+ew-machine.ps1
+
+  Everything is idempotent: re-running skips what is already correct.
+
+  This script NEVER installs an agent for you. Each vendor's installer is theirs, changes
+  without notice, and may need an interactive login. It checks, reports, and tells you the
+  one command to run — then verifies afterwards.
+
+.PARAMETER SkipKeys
+  Do not run `caddis keys` (which prompts). Use on a machine that only needs Claude.
+
+.PARAMETER WhatIf
+  Report what would happen; change nothing.
+
+.EXAMPLE
+  .\new-machine.ps1
+  .\new-machine.ps1 -SkipKeys
+#>
+[CmdletBinding(SupportsShouldProcess)]
+param(
+    [switch]$SkipKeys
+)
+
+$ErrorActionPreference = 'Stop'
+$script:Problems = @()
+
+function Step($n, $text) { Write-Host "`n[$n] $text" -ForegroundColor Cyan }
+function Ok($text)       { Write-Host "   OK    $text" -ForegroundColor Green }
+function Missing($text)  { Write-Host "   TODO  $text" -ForegroundColor Yellow }
+function Bad($text)      { Write-Host "   FAIL  $text" -ForegroundColor Red; $script:Problems += $text }
+function Note($text)     { Write-Host "         $text" -ForegroundColor DarkGray }
+
+function Test-Bin($name) { $null -ne (Get-Command $name -ErrorAction SilentlyContinue) }
+
+# ── 0. prerequisites ────────────────────────────────────────────────────────────────────
+Step 0 'Prerequisites'
+
+foreach ($p in @(
+    @{ bin = 'node';   why = 'runs the caddis CLI and every agent installer'; get = 'winget install OpenJS.NodeJS.LTS' },
+    @{ bin = 'python'; why = 'runs the caddis hooks and the status line';     get = 'winget install Python.Python.3.12' },
+    @{ bin = 'git';    why = 'the status line and most commands read git state'; get = 'winget install Git.Git' }
+)) {
+    if (Test-Bin $p.bin) { Ok "$($p.bin) — $($p.why)" }
+    else { Bad "$($p.bin) is missing — $($p.why).  Install: $($p.get)" }
+}
+
+if ($script:Problems.Count -gt 0) {
+    Write-Host "`nStopping: the prerequisites above are required before anything else works." -ForegroundColor Red
+    exit 1
+}
+
+# ── 1. the agents — BEFORE caddis, or caddis cannot see them ────────────────────────────
+Step 1 'Agents (install these BEFORE caddis)'
+Note 'caddis installs INTO an agent. One added later needs a re-run of `caddis init`.'
+
+$agents = @(
+    @{ bin = 'claude'; name = 'Claude Code'; get = 'npm i -g @anthropic-ai/claude-code'; login = 'claude  (log in on first run)' },
+    @{ bin = 'agy';    name = 'Antigravity (agy)'; get = 'download from antigravity.google'; login = 'agy  (log in on first run)' },
+    @{ bin = 'codex';  name = 'Codex';       get = 'npm i -g @openai/codex'; login = 'codex login   (INTERACTIVE — a script cannot do this)' }
+)
+
+$present = @()
+foreach ($a in $agents) {
+    if (Test-Bin $a.bin) {
+        $ver = (& $a.bin --version 2>&1 | Select-Object -First 1)
+        Ok "$($a.name) — $ver"
+        $present += $a.name
+    } else {
+        Missing "$($a.name) not found.  Install: $($a.get)"
+        Note "then: $($a.login)"
+    }
+}
+
+if ($present.Count -eq 0) {
+    Write-Host "`nStopping: no agents found, so caddis would install into nothing." -ForegroundColor Red
+    exit 1
+}
+
+# ── 2. caddis CLI ───────────────────────────────────────────────────────────────────────
+Step 2 'caddis CLI'
+Note 'Brings claude-oss / claude-glm / claude-deepseek onto PATH as well — no profile edit.'
+
+if (Test-Bin 'caddis') {
+    Ok "caddis CLI present — $(caddis --version 2>&1 | Select-Object -First 1)"
+} elseif ($PSCmdlet.ShouldProcess('npm', 'install -g @caddis/cli')) {
+    npm i -g '@caddis/cli'
+    if (Test-Bin 'caddis') { Ok 'caddis CLI installed' } else { Bad 'npm reported success but `caddis` is not on PATH — open a new shell and re-run' ; exit 1 }
+}
+
+# ── 3. drive each agent ─────────────────────────────────────────────────────────────────
+Step 3 'Install caddis into every detected agent'
+if ($PSCmdlet.ShouldProcess('agents', 'caddis init')) {
+    caddis init
+}
+
+# ── 4. provider keys ────────────────────────────────────────────────────────────────────
+Step 4 'Provider API keys (GLM / DeepSeek)'
+if ($SkipKeys) {
+    Note 'Skipped (-SkipKeys). Run `caddis keys` later to enable the OSS model lanes.'
+} elseif ($PSCmdlet.ShouldProcess('~/.caddis/keys.env', 'caddis keys')) {
+    Note 'Prompts per provider, validates against the live endpoint, never prints a key.'
+    caddis keys
+}
+
+# ── 5. status line ──────────────────────────────────────────────────────────────────────
+Step 5 'Status line (Claude Code + agy, user scope, once per machine)'
+$statusline = Join-Path $HOME '.caddis\statusline.py'
+if (Test-Path $statusline) {
+    if ($PSCmdlet.ShouldProcess($statusline, 'install')) { python $statusline --install }
+} else {
+    Missing "statusline.py not found at $statusline"
+    Note 'It ships with the plugin. Run `/caddis:statusline` in a Claude Code session, or'
+    Note 'python PLUGIN_ROOT/scripts/caddis_statusline.py --install'
+}
+
+# ── 6. verify ───────────────────────────────────────────────────────────────────────────
+Step 6 'Verify'
+
+caddis status
+
+if (Test-Path $statusline) {
+    python $statusline --check
+}
+
+if (-not $SkipKeys -and (Test-Bin 'caddis')) {
+    Note 'Validating keys against the live endpoints...'
+    caddis keys --check
+}
+
+Write-Host "`n-- Per repo, not per machine --" -ForegroundColor Cyan
+Note 'In each project, once:   /caddis:setup-project-ai'
+Note 'The status line is deliberately NOT per-repo — one renderer serves every project.'
+
+Write-Host "`n-- Known gotchas on this platform --" -ForegroundColor Cyan
+Note 'codex headless: pass -c model_reasoning_effort=low. The default `high` can exceed 300s'
+Note '                on a trivial call.'
+Note '`codex doctor` is NOT an auth check — it validates the auth FILE, not the token, and'
+Note '                reports healthy against a dead one. Real check:'
+Note '                codex exec --skip-git-repo-check -s read-only "Reply OK"'
+Note 'agy statusLine is NOT parsed through a shell, so its path must be UNQUOTED. --install'
+Note '                handles this; if you wire it by hand, do not add quotes.'
+Note 'Codex has NO status line and cannot have one — no status_line key in config.toml and'
+Note '                nothing in `codex --help` (measured on codex-cli 0.150.1). Upstream limit,'
+Note '                not a caddis gap. Two profiles exist because two hosts can call them.'
+
+if ($script:Problems.Count -gt 0) {
+    Write-Host "`n$($script:Problems.Count) problem(s) above." -ForegroundColor Red
+    exit 1
+}
+Write-Host "`nDone." -ForegroundColor Green

package/bundles/antigravity-plugin/scripts/oss_ask.py

@@ -28,7 +28,7 @@ import oss_model  # noqa: E402  (same-directory import: shared key resolution)
 # which holds the Anthropic-protocol endpoints for Claude Code.
 CHAT_PROVIDERS: dict[str, dict[str, str]] = {
     "deepseek":   {"base_url": "https://api.deepseek.com",            "model": "deepseek-v4-flash",          "key_env": "DEEPSEEK_API_KEY"},
-    "glm":        {"base_url": "https://api.z.ai/api/coding/paas/v4", "model": "glm-5.2",                    "key_env": "GLM_API_KEY"},
+    "glm":        {"base_url": "https://api.z.ai/api/coding/paas/v4", "model": "glm-5.3",                    "key_env": "GLM_API_KEY"},
     "openrouter": {"base_url": "https://openrouter.ai/api/v1",        "model": "deepseek/deepseek-v4-flash", "key_env": "OPENROUTER_API_KEY"},
 }
 

package/bundles/antigravity-plugin/scripts/oss_model.py

@@ -34,9 +34,17 @@ from pathlib import Path
 # "openrouter" has no Anthropic-protocol endpoint: it only works through an
 # Anthropic-compatible gateway (e.g. LiteLLM — see claude-harness/LOCAL-MODELS.md) whose
 # URL you supply via OSS_BASE_URL.
+# The `[1m]` suffix on a model id is a CLAUDE CODE convention, not part of the model name.
+# Claude Code strips it before calling the API; it only tells the client the real context
+# window. Without it, an id Claude Code does not recognise is assumed to be 200k and
+# auto-compact throws away context the model could still hold. Both providers are 1M
+# (GLM-5.3: docs.z.ai; DeepSeek V4 Flash: 1,048,576 positions) — verified 2026-08-23 by
+# running with and without the suffix and diffing the cap warning.
+# NEVER put this suffix in the OpenAI-dialect tables (oss_review.py / oss_ask.py): those
+# send the id raw to /chat/completions, which would reject it.
 PROVIDERS: dict[str, dict[str, str]] = {
-    "deepseek":   {"base_url": "https://api.deepseek.com/anthropic",  "model": "deepseek-v4-flash",          "key_env": "DEEPSEEK_API_KEY"},
-    "glm":        {"base_url": "https://api.z.ai/api/anthropic",      "model": "glm-5.2",                    "key_env": "GLM_API_KEY"},
+    "deepseek":   {"base_url": "https://api.deepseek.com/anthropic",  "model": "deepseek-v4-flash[1m]",          "key_env": "DEEPSEEK_API_KEY"},
+    "glm":        {"base_url": "https://api.z.ai/api/anthropic",      "model": "glm-5.3[1m]",                    "key_env": "GLM_API_KEY"},
     "openrouter": {"base_url": "",                                    "model": "deepseek/deepseek-v4-flash", "key_env": "OPENROUTER_API_KEY"},
 }
 DEFAULT_PROVIDER = "deepseek"

package/bundles/antigravity-plugin/scripts/oss_review.py

@@ -62,7 +62,7 @@ EXIT_CONFIG = 3
 # provider (Qwen, a local vLLM, …) is one new row. Callers can always bypass this via env/flags.
 PROVIDERS: dict[str, dict[str, str]] = {
     "deepseek":   {"base_url": "https://api.deepseek.com",            "model": "deepseek-v4-flash"},
-    "glm":        {"base_url": "https://api.z.ai/api/coding/paas/v4", "model": "glm-5.2"},
+    "glm":        {"base_url": "https://api.z.ai/api/coding/paas/v4", "model": "glm-5.3"},
     "openrouter": {"base_url": "https://openrouter.ai/api/v1",        "model": "deepseek/deepseek-v4-flash"},
 }
 DEFAULT_PROVIDER = "deepseek"

package/bundles/antigravity-plugin/scripts/setup_project_ai.py

@@ -941,11 +941,14 @@ allow = [
 #                                                # accepted back-compat alias for this key.
 # ignore_routes = ["/health", "/__internal"]     # routes intentionally left out of the page guide
 
-# ── [dream_memory] — short-term fact store (scripts/dream_memory.py) ── LIVE ──
-# [dream_memory]
-# prune_age_days = 14    # single-hit facts older than this decay out
-# max_facts = 200        # hard cap on the store
-# surface_limit = 5      # how many facts SessionStart prints
+# ── [handover] — what /caddis:handoff measures (scripts/caddis_inventory.py) ── LIVE ──
+# [handover]
+# test_cmd = "pytest tests/ -q"   # THIS REPO'S test command. Without it the inventory
+#                                 # guesses `pytest` at the root, which is right in a
+#                                 # single-package repo and wrong in any repo carrying a
+#                                 # vendored or generated tree. Generated trees
+#                                 # (vscode-extensions/, dist/, node_modules/, .venv/) are
+#                                 # excluded automatically; set this when that is not enough.
 """
 
 
@@ -1132,7 +1135,7 @@ def scaffold_artifact_dir(target: Path, dry: bool) -> list[str]:
     is gitignored. .caddis/ is the default home for every working-artifact kind (Track A
     Phase A3) — kb/ and prompts/ round out plans/prd/agent-docs/reviews so nothing has to
     scatter to the repo root or .github/. Also drops a documented `config.toml.example`
-    (guard/doc_coverage/dream_memory). Idempotent; never clobbers an existing .gitignore or
+    (guard/doc_coverage). Idempotent; never clobbers an existing .gitignore or
     config example.
 
     Writes into the repo's `.caddis/`.