subconscious-cli 4.1.2 → 4.1.4

package/bin/agents.js

@@ -32,6 +32,15 @@ import { compareVersions } from './update-check.js';
 
 export const MIN_CLAUDE_CODE_VERSION = '2.1.242';
 
+// First Codex that honours `multi_agent_version: "v2"` in the model catalog,
+// found by installing each release and reading what it advertises: 0.142.5 and
+// below offer no subagent namespace at all, 0.143.0 offers `collaboration`.
+//
+// Below this the catalog field is not rejected, it is ignored - the run simply
+// has no subagents and nothing says why. That silence is the reason for a
+// version gate rather than a note in the docs.
+export const MIN_CODEX_VERSION = '0.143.0';
+
 // --- Registry (single source of truth, generated copy shipped in the package).
 const registry = JSON.parse(
   readFileSync(new URL('./registry.generated.json', import.meta.url), 'utf-8'),
@@ -193,10 +202,10 @@ const AGENT_HELP = {
   codex: {
     usage: 'subc [-p NAME] codex [help|install|status|uninstall] [Codex arguments...]',
     behavior:
-      'Launches Codex with a temporary Subconscious provider catalog. Compaction hooks are merged into ~/.codex/hooks.json and removed with subc codex uninstall.',
+      'Launches Codex with a temporary Subconscious provider catalog. Compaction hooks are merged into ~/.codex/hooks.json on first launch and removed with subc codex uninstall; Codex will not run them until you trust them with /hooks once.',
     options: [
       ['help', 'Show this help'],
-      ['install', 'Install only the Subconscious compaction hooks'],
+      ['install', 'Install the Subconscious compaction hooks (then trust with /hooks)'],
       ['status', 'Inspect the installed compaction hooks'],
       ['uninstall', 'Remove only the Subconscious Codex hooks'],
       ['--model MODEL', 'Override the profile model for this launch'],
@@ -205,7 +214,9 @@ const AGENT_HELP = {
       ['--auto-compact-token-limit N', 'Override the automatic compaction threshold'],
       ['--reasoning-effort LEVEL', 'Use none, low, medium, high, or max'],
       ['--external-tools', 'Enable Codex apps/plugins for this launch'],
-      ['--subagents', 'Use the pinned legacy Codex subagent mode'],
+      ['--stream-idle-timeout MS', 'Allow a longer silent think before timing out'],
+      ['--max-subagents N', 'Subagents allowed to run at once (default 4)'],
+      ['--subagent-effort LEVEL', 'Effort for subagents, separate from the parent'],
       ['-- ARGS...', 'Pass remaining arguments to Codex'],
     ],
   },
@@ -625,6 +636,10 @@ export function parseClaudeVersion(text) {
 }
 
 export function claudeVersionNeedsUpgrade(installed, minimum = MIN_CLAUDE_CODE_VERSION) {
+  return versionNeedsUpgrade(installed, minimum);
+}
+
+export function versionNeedsUpgrade(installed, minimum) {
   return Boolean(installed && minimum && compareVersions(installed, minimum) < 0);
 }
 
@@ -649,13 +664,20 @@ export function readClaudeVersion(bin, binDir, options = {}) {
   }
 }
 
+const MINIMUM_VERSIONS = {
+  'claude-code': { minimum: MIN_CLAUDE_CODE_VERSION, label: 'Claude Code' },
+  // Below this, subagents are silently absent rather than broken.
+  codex: { minimum: MIN_CODEX_VERSION, label: 'Codex' },
+};
+
 async function ensureClaudeCompatible(agent, binDir) {
-  if (agent.id !== 'claude-code') return;
+  const requirement = MINIMUM_VERSIONS[agent.id];
+  if (!requirement) return;
   const version = readClaudeVersion(agent.bin, binDir);
-  if (!claudeVersionNeedsUpgrade(version)) return;
+  if (!versionNeedsUpgrade(version, requirement.minimum)) return;
 
   console.error(
-    `\n  Minimum supported Claude Code version is ${MIN_CLAUDE_CODE_VERSION}. Your version is ${version}. Upgrade to get the best experience.\n`,
+    `\n  Minimum supported ${requirement.label} version is ${requirement.minimum}. Your version is ${version}. Upgrade to get the best experience.\n`,
   );
   console.error(`  Upgrade it with:`);
   console.error(`    ${c.cyan}${agent.install}${c.reset}`);

package/bin/model-capabilities.js

@@ -0,0 +1,11 @@
+import { readFileSync } from 'node:fs';
+
+// The models endpoint currently supplies IDs, not modality metadata. Keep
+// explicit capabilities in the same registry used to generate the Unix lookup.
+const { modelCapabilities = {} } = JSON.parse(
+  readFileSync(new URL('./registry.generated.json', import.meta.url), 'utf8'),
+);
+
+export function modelSupportsVision(modelId) {
+  return Object.hasOwn(modelCapabilities, modelId) && modelCapabilities[modelId].vision === true;
+}

package/bin/profiles.js

@@ -144,6 +144,28 @@ const SETTINGS = {
     type: 'integer',
     min: 1,
   },
+  CODEX_SUBAGENT_REASONING_EFFORT: {
+    key: 'CODEX_SUBAGENT_REASONING_EFFORT',
+    label: 'Subagent reasoning effort',
+    description:
+      'Effort for Codex subagents, set separately from the parent: a marathon model at max thinks for minutes without emitting a token, and a subagent turn is one uninterrupted think, so inheriting max makes every delegated subtask look like a hang.',
+    type: 'text',
+  },
+  CODEX_MULTI_AGENT_VERSION: {
+    key: 'CODEX_MULTI_AGENT_VERSION',
+    label: 'Codex multi-agent version',
+    description:
+      'Which subagent implementation Codex offers: v2 gives the collaboration tools. Leave empty to turn subagents off entirely.',
+    type: 'text',
+  },
+  CODEX_STREAM_IDLE_TIMEOUT_MS: {
+    key: 'CODEX_STREAM_IDLE_TIMEOUT_MS',
+    label: 'Stream idle timeout (ms)',
+    description:
+      'How long Codex waits on a silent stream before giving up. Needs to exceed the longest think the model will take.',
+    type: 'integer',
+    min: 1,
+  },
   MAX_SUBAGENT_SPAWN_DEPTH: {
     key: 'MAX_SUBAGENT_SPAWN_DEPTH',
     label: 'Subagent spawn depth',
@@ -305,6 +327,9 @@ const AGENT_SETTING_KEYS = {
     'CODEX_AUTO_COMPACT_TOKEN_LIMIT',
     'CODEX_REASONING_EFFORT',
     'CODEX_EXTERNAL_TOOLS',
+    'CODEX_STREAM_IDLE_TIMEOUT_MS',
+    'CODEX_MULTI_AGENT_VERSION',
+    'CODEX_SUBAGENT_REASONING_EFFORT',
     'MAX_CONCURRENT_SUBAGENTS',
   ],
   opencode: ['OPENCODE_API_KEY', 'OPENCODE_CONTEXT_LIMIT', 'OPENCODE_OUTPUT_LIMIT'],

package/bin/registry.generated.json

@@ -6,13 +6,19 @@
       "subconscious/glm-5.3-marathon",
       "subconscious/glm-5.2",
       "subconscious/tim-qwen3.6-27b",
-      "subconscious/deepseek-v4-flash-marathon"
+      "subconscious/deepseek-v4-flash-marathon",
+      "subconscious/deepseek-v4.1-flash-marathon"
     ],
     "baseUrl": "https://api.subconscious.dev",
     "baseUrlV1": "https://api.subconscious.dev/v1",
     "keyPlaceholder": "your_key",
     "platformUrl": "https://subconscious.dev/platform"
   },
+  "modelCapabilities": {
+    "subconscious/deepseek-v4.1-flash-marathon": {
+      "vision": true
+    }
+  },
   "agents": [
     {
       "id": "subconscious-code",
@@ -161,7 +167,7 @@
           "uninstall"
         ]
       },
-      "launch": "codex -c model={model} -c model_provider=subconscious -c web_search=disabled -c model_providers.subconscious.name=Subconscious -c model_providers.subconscious.base_url={baseUrlV1} -c model_providers.subconscious.wire_api=responses -c model_providers.subconscious.env_key=SUBCONSCIOUS_API_KEY -c model_providers.subconscious.stream_idle_timeout_ms=300000",
+      "launch": "codex -c model={model} -c model_provider=subconscious -c web_search=disabled -c model_providers.subconscious.name=Subconscious -c model_providers.subconscious.base_url={baseUrlV1} -c model_providers.subconscious.wire_api=responses -c model_providers.subconscious.env_key=SUBCONSCIOUS_API_KEY -c model_providers.subconscious.stream_idle_timeout_ms=900000",
       "env": {
         "SUBCONSCIOUS_API_KEY": "{apiKey}",
         "SUBCONSCIOUS_GATEWAY_URL": "{baseUrl}"
@@ -253,6 +259,24 @@
                       "context": 5000000,
                       "output": 65536
                     }
+                  },
+                  "subconscious/deepseek-v4.1-flash-marathon": {
+                    "name": "subconscious/deepseek-v4.1-flash-marathon",
+                    "tools": true,
+                    "attachment": true,
+                    "modalities": {
+                      "input": [
+                        "text",
+                        "image"
+                      ],
+                      "output": [
+                        "text"
+                      ]
+                    },
+                    "limit": {
+                      "context": 5000000,
+                      "output": 65536
+                    }
                   }
                 }
               }

package/bin/runbook/codex/hooks-lib.sh

@@ -61,6 +61,30 @@ EOF
   chmod 600 "$HOOKS_ENV_FILE"
 }
 
+# True when our hooks, hook script and env file are all already in place and
+# current, so there is nothing to write.
+#
+# This matters more than it looks. Codex asks the user to trust a hook before
+# it will run one, and it decides what is "new" from what is on disk. Rewriting
+# an identical hook on every single launch meant the trust prompt never went
+# away and the compaction hooks never actually ran - we were re-introducing
+# ourselves as a stranger several times a day.
+codex_hooks_current() {
+  [[ -f "$HOOKS_JSON" ]] || return 1
+  codex_hooks_json_valid || return 1
+  [[ -x "$HOOK_DST" ]] || return 1
+  [[ -f "${HOOK_SRC:-}" ]] || return 1
+  cmp -s "$HOOK_SRC" "$HOOK_DST" || return 1
+  [[ -f "$HOOKS_ENV_FILE" ]] || return 1
+  grep -q "SUBCONSCIOUS_GATEWAY_URL='${GATEWAY_URL%/}'" "$HOOKS_ENV_FILE" || return 1
+  grep -q "SUBCONSCIOUS_API_KEY='${API_KEY}'" "$HOOKS_ENV_FILE" || return 1
+  # Our entry present under both compaction events, and exactly once each.
+  jq -e --arg m "$MARKER" '
+    def ours: [ .[]?.hooks[]? | select((.command // "") | tostring | contains($m)) ] | length;
+    (.hooks.PreCompact | ours) == 1 and (.hooks.PostCompact | ours) == 1
+  ' "$HOOKS_JSON" >/dev/null 2>&1
+}
+
 # codex_ensure_hooks [strict|best-effort]
 codex_ensure_hooks() {
   local mode="${1:-strict}"
@@ -91,12 +115,18 @@ codex_ensure_hooks() {
   if [[ ! -f "$HOOKS_JSON" ]]; then
     printf '%s\n' '{"hooks":{}}' >"$HOOKS_JSON"
   fi
+  if codex_hooks_current; then
+    # Already installed and unchanged. Saying nothing is the point: a message
+    # on every launch trains the user to ignore it, and rewriting the file
+    # would reset the trust they have already granted.
+    return 0
+  fi
   codex_write_hook_script
   codex_write_hooks_env
   codex_strip_hook_entries
   codex_append_hook_groups
-  echo "Merged Subconscious Codex hooks into $HOOKS_JSON"
-  echo "Trust new hooks in Codex with /hooks (or --dangerously-bypass-hook-trust)."
+  echo "Installed Subconscious compaction hooks in $HOOKS_JSON"
+  echo "These do not run until you trust them: type /hooks inside Codex once."
 }
 
 codex_uninstall_hooks() {

package/bin/runbook/codex/install.sh

@@ -33,6 +33,16 @@ Usage:
 
 Merges Subconscious compaction hooks into ~/.codex/hooks.json. Does not write
 config.toml. Launch Codex with subc codex.
+
+The hooks report compaction start and end to the gateway, which is what lets a
+conversation's turns stay correlated across a compaction boundary.
+
+Codex will not run a hook until you trust it. After installing, type /hooks
+inside Codex once and approve them; until then they are present but inert.
+Non-interactive runs have nobody to ask, so `codex exec` can only run them with
+--dangerously-bypass-hook-trust, which is yours to opt into and is never passed
+for you. Fleet installs should ship the hooks through Codex's managed
+(MDM) configuration, which is trusted without prompting.
 EOF
 }
 

package/bin/runbook/codex/run.sh

@@ -6,20 +6,24 @@
 # The model catalog (needed to suppress the "model metadata not found" warning)
 # is written to a temp file that is cleaned up on exit.
 #
-# Subagents are OFF by default. Current Codex (>=0.144) wraps subagent tools in a
-# `type: "namespace"` wire format that non-OpenAI providers can't resolve, causing
-# "unsupported call: spawn_agent" (upstream #32318, #26977). The fix (PR #29602) is
-# not yet merged.
+# Subagents are ON by default. The per-model `multi_agent_version` catalog
+# field selects WHICH implementation Codex offers, not whether it offers one:
+# "v2" gives the `collaboration` namespace, and an empty value falls back to
+# the older `multi_agent_v1` namespace rather than turning subagents off.
+# Only `agents.enabled=false` actually disables them, so clearing
+# CODEX_MULTI_AGENT_VERSION does both.
 #
-# To run with subagents, pass --subagents — this runs the pinned legacy
-# codex@0.132.0 with multi-agent v1 config (plain tool names) that the model can
-# resolve. Requires npx.
+# This needs a gateway that round-trips Codex's namespaced tools. Codex sends
+# them as `type: "namespace"` and resolves an incoming call by namespace plus
+# name, so a gateway that flattens the namespace for the model must restore it
+# on the way back, or every call fails with "unsupported call: spawn_agent".
 #
 # Usage:
 #   ./run.sh                         # uses GATEWAY_URL/API_KEY from ../.env
 #   ./run.sh --context-window 5000000 -- --resume
-#   ./run.sh --subagents              # run codex@0.132.0 with subagents enabled
-#   ./run.sh --subagents -- --resume  # subagents + passthrough args
+#   ./run.sh --stream-idle-timeout 1800000  # allow a 30min silent think
+#   ./run.sh --max-subagents 12        # raise the concurrent-subagent ceiling
+#   ./run.sh --subagent-effort low     # cheaper subagents than the parent
 #   ./run.sh --external-tools          # include Codex apps/plugins (may exceed gateway tool limits)
 #
 # Config: copy ../env.example to ../.env and edit. .env is gitignored.
@@ -31,6 +35,7 @@
 set -euo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../model-capabilities.generated.sh"
 
 # Load shared env from SUBC_ENV_FILE, or a sibling .env / env.example.
 SHARED_ENV="${SUBC_ENV_FILE:-${SCRIPT_DIR}/../.env}"
@@ -40,16 +45,31 @@ if [[ -f "$SHARED_ENV" ]]; then set -a; source "$SHARED_ENV"; set +a; fi
 GATEWAY_URL="${GATEWAY_URL:-}"
 API_KEY="${CODEX_API_KEY:-${API_KEY:-}}"
 MODEL="${MODEL:-subconscious/glm-5.3-marathon}"
-MAX_CONCURRENT_SUBAGENTS="${MAX_CONCURRENT_SUBAGENTS:-4}"
-SUBAGENTS=false
+MAX_CONCURRENT_SUBAGENTS="${MAX_CONCURRENT_SUBAGENTS:-}"
+# Effort for spawned agents, independent of the parent's. Empty inherits the
+# Codex default. See the note above SUBAGENT_ARGS for why max is a poor choice.
+#
+# Low rather than medium: the gateway rounds medium up to high for GLM, and at
+# high a subagent handed an open-ended task keeps deliberating without ever
+# writing the closing `</think>`. The turn then comes back as one long block of
+# prose with no tool call, so the agent reports success having produced no
+# files. Measured against the worker on one broad task: low closed and called a
+# tool in ~1.5k characters, high ran past 76k and was still going.
+CODEX_SUBAGENT_REASONING_EFFORT="${CODEX_SUBAGENT_REASONING_EFFORT-low}"
 EXTERNAL_TOOLS="${CODEX_EXTERNAL_TOOLS:-false}"
 CODEX_CONTEXT_WINDOW="${CODEX_CONTEXT_WINDOW:-5000000}"
 CODEX_MAX_CONTEXT_WINDOW="${CODEX_MAX_CONTEXT_WINDOW:-}"
 CODEX_AUTO_COMPACT_TOKEN_LIMIT="${CODEX_AUTO_COMPACT_TOKEN_LIMIT:-4500000}"
 CODEX_REASONING_EFFORT="${CODEX_REASONING_EFFORT:-max}"
-
-# Codex version that supports the legacy multi-agent v1 config (plain tool names).
-SUBAGENT_CODEX_VERSION="0.132.0"
+# How long Codex waits on a silent stream before giving up. The marathon models
+# at max reasoning effort routinely think for longer than five minutes without
+# emitting a token, and a subagent's turn is one uninterrupted think with no
+# visible progress, so the old 300s ceiling cut off work that was still running.
+CODEX_STREAM_IDLE_TIMEOUT_MS="${CODEX_STREAM_IDLE_TIMEOUT_MS:-900000}"
+# Codex gates its subagent tools on this per-model catalog field. Only "v2" is
+# honored by current releases; "v1" registers nothing. Set to empty to opt out.
+# Unset defaults to v2; an explicitly empty value opts out, so use `-` not `:-`.
+CODEX_MULTI_AGENT_VERSION="${CODEX_MULTI_AGENT_VERSION-v2}"
 
 # Parse args (only when executed, not sourced)
 PASSTHRU=()
@@ -72,9 +92,17 @@ if [[ "${BASH_SOURCE[0]:-$0}" == "${0}" ]]; then
         CODEX_REASONING_EFFORT="${2:-}"
         shift 2
         ;;
-      --subagents)
-        SUBAGENTS=true
-        shift
+      --stream-idle-timeout)
+        CODEX_STREAM_IDLE_TIMEOUT_MS="${2:-}"
+        shift 2
+        ;;
+      --max-subagents)
+        MAX_CONCURRENT_SUBAGENTS="${2:-}"
+        shift 2
+        ;;
+      --subagent-effort)
+        CODEX_SUBAGENT_REASONING_EFFORT="${2:-}"
+        shift 2
         ;;
       --external-tools)
         EXTERNAL_TOOLS=true
@@ -105,7 +133,8 @@ esac
 DEFAULT_SUBCONSCIOUS_MODELS="subconscious/glm-5.3-marathon
 subconscious/glm-5.2
 subconscious/tim-qwen3.6-27b
-subconscious/deepseek-v4-flash-marathon"
+subconscious/deepseek-v4-flash-marathon
+subconscious/deepseek-v4.1-flash-marathon"
 SUPPORTED_MODELS=()
 
 add_supported_model() {
@@ -132,10 +161,17 @@ while IFS= read -r model_id; do
 done <<< "${SUBCONSCIOUS_MODELS:-$DEFAULT_SUBCONSCIOUS_MODELS}"
 
 write_model_catalog() {
-  local catalog_file="$1" model_id index=0
+  local catalog_file="$1" model_id vision_fields index=0 multi_agent_json="null"
+  if [[ -n "$CODEX_MULTI_AGENT_VERSION" ]]; then
+    multi_agent_json="\"${CODEX_MULTI_AGENT_VERSION}\""
+  fi
   {
     printf '{\n  "models": [\n'
     for model_id in "${SUPPORTED_MODELS[@]}"; do
+      vision_fields=""
+      if subc_model_supports_vision "$model_id"; then
+        vision_fields=', "input_modalities": ["text", "image"]'
+      fi
       if [[ "$index" -gt 0 ]]; then
         printf ',\n'
       fi
@@ -148,7 +184,13 @@ write_model_catalog() {
       "max_context_window": ${CODEX_MAX_CONTEXT_WINDOW},
       "auto_compact_token_limit": ${CODEX_AUTO_COMPACT_TOKEN_LIMIT},
       "effective_context_window_percent": 95,
-      "supported_reasoning_levels": [],
+      "supported_reasoning_levels": [
+        { "effort": "none", "description": "No additional reasoning" },
+        { "effort": "low", "description": "Fast responses with lighter reasoning" },
+        { "effort": "medium", "description": "Balances speed and reasoning depth" },
+        { "effort": "high", "description": "Greater reasoning depth for complex problems" },
+        { "effort": "max", "description": "Maximum reasoning depth for the hardest problems" }
+      ],
       "shell_type": "shell_command",
       "visibility": "list",
       "supported_in_api": true,
@@ -169,7 +211,9 @@ write_model_catalog() {
       "apply_patch_tool_type": "freeform",
       "truncation_policy": { "mode": "tokens", "limit": 10000 },
       "supports_parallel_tool_calls": true,
-      "experimental_supported_tools": []
+      "multi_agent_version": ${multi_agent_json},
+      "use_responses_lite": false,
+      "experimental_supported_tools": []${vision_fields}
     }
 EOF
       index=$((index + 1))
@@ -191,6 +235,41 @@ if [[ "$EXTERNAL_TOOLS" != "true" ]]; then
   )
 fi
 
+# Bound the subagent tree when the catalog turns it on. Codex still only spawns
+# a subagent when the user asks for one; these are ceilings, not defaults.
+#
+# Codex documents no nesting-depth limit, so a subagent can spawn its own
+# subagent and the tree can grow deeper than you asked for. Capping concurrency
+# is the only lever available.
+#
+# Subagent reasoning effort is set separately from the parent's on purpose: the
+# marathon models at max effort think for many minutes without emitting a
+# token, and a subagent turn is one uninterrupted think, so inheriting max
+# makes every delegated subtask look like a hang.
+# Applied after argument parsing so --max-subagents can win. Four is a floor
+# rather than a considered limit: Codex's own default is three, and exceeding
+# either one fails the spawn outright rather than queueing it, so a run that
+# wants more should say so instead of collecting retries.
+MAX_CONCURRENT_SUBAGENTS="${MAX_CONCURRENT_SUBAGENTS:-4}"
+
+SUBAGENT_ARGS=()
+if [[ -z "$CODEX_MULTI_AGENT_VERSION" ]]; then
+  # Clearing the catalog field alone only downgrades Codex to multi-agent v1,
+  # whose tools this gateway has never been exercised against. Turn the
+  # feature off outright instead.
+  SUBAGENT_ARGS=(-c agents.enabled=false)
+else
+  SUBAGENT_ARGS=(
+    -c agents.max_concurrent_threads_per_session="${MAX_CONCURRENT_SUBAGENTS}"
+    -c agents.interrupt_message=true
+  )
+  if [[ -n "$CODEX_SUBAGENT_REASONING_EFFORT" ]]; then
+    SUBAGENT_ARGS+=(
+      -c agents.default_subagent_reasoning_effort="${CODEX_SUBAGENT_REASONING_EFFORT}"
+    )
+  fi
+fi
+
 if [[ -z "$GATEWAY_URL" || -z "$API_KEY" ]]; then
   echo "error: GATEWAY_URL and API_KEY must be set in ../.env" >&2
   exit 1
@@ -200,6 +279,10 @@ export SUBCONSCIOUS_API_KEY="$API_KEY"
 export SUBCONSCIOUS_GATEWAY_URL="${GATEWAY_URL%/}"
 
 # Merge compaction hooks without replacing ~/.codex/hooks.json or config.toml.
+# `codex_ensure_hooks` returns early when they are already current, so a launch
+# that changes nothing writes nothing: Codex decides what counts as a new hook
+# from what is on disk, and rewriting an identical one reset the trust the user
+# had already granted.
 HOOK_SRC="${SCRIPT_DIR}/hook.sh"
 # shellcheck source=hooks-lib.sh
 source "${SCRIPT_DIR}/hooks-lib.sh"
@@ -214,43 +297,22 @@ write_model_catalog "$CATALOG_FILE"
 
 # If sourced, just export env and return.
 if [[ "${BASH_SOURCE[0]:-$0}" != "${0}" ]]; then
-  export GATEWAY_URL CATALOG_FILE MAX_CONCURRENT_SUBAGENTS SUBAGENTS
+  export GATEWAY_URL CATALOG_FILE MAX_CONCURRENT_SUBAGENTS CODEX_SUBAGENT_REASONING_EFFORT
   return 0 2>/dev/null || true
 fi
 
 # Ephemeral config via -c flags — nothing is written to ~/.codex/config.toml.
-if [[ "$SUBAGENTS" == "true" ]]; then
-  # Legacy multi-agent v1 config — plain tool names the model can resolve.
-  echo "Starting codex@${SUBAGENT_CODEX_VERSION} with subagents enabled (max ${MAX_CONCURRENT_SUBAGENTS} threads)..." >&2
-  exec npx -y "@openai/codex@${SUBAGENT_CODEX_VERSION}" \
-    -c model="${MODEL}" \
-    -c model_provider=subconscious \
-    -c model_catalog_json="${CATALOG_FILE}" \
-    -c model_reasoning_effort="${CODEX_REASONING_EFFORT}" \
-    -c web_search=disabled \
-    ${EXTERNAL_TOOL_ARGS[@]+"${EXTERNAL_TOOL_ARGS[@]}"} \
-    -c features.multi_agent=true \
-    -c agents.max_threads="${MAX_CONCURRENT_SUBAGENTS}" \
-    -c agents.max_depth=1 \
-    -c agents.interrupt_message=true \
-    -c model_providers.subconscious.name=Subconscious \
-    -c model_providers.subconscious.base_url="${GATEWAY_URL}/v1" \
-    -c model_providers.subconscious.wire_api=responses \
-    -c model_providers.subconscious.env_key=SUBCONSCIOUS_API_KEY \
-    -c model_providers.subconscious.stream_idle_timeout_ms=300000 \
-    ${PASSTHRU[@]+"${PASSTHRU[@]}"}
-else
-  exec codex \
-    -c model="${MODEL}" \
-    -c model_provider=subconscious \
-    -c model_catalog_json="${CATALOG_FILE}" \
-    -c model_reasoning_effort="${CODEX_REASONING_EFFORT}" \
-    -c web_search=disabled \
-    ${EXTERNAL_TOOL_ARGS[@]+"${EXTERNAL_TOOL_ARGS[@]}"} \
-    -c model_providers.subconscious.name=Subconscious \
-    -c model_providers.subconscious.base_url="${GATEWAY_URL}/v1" \
-    -c model_providers.subconscious.wire_api=responses \
-    -c model_providers.subconscious.env_key=SUBCONSCIOUS_API_KEY \
-    -c model_providers.subconscious.stream_idle_timeout_ms=300000 \
-    ${PASSTHRU[@]+"${PASSTHRU[@]}"}
-fi
+exec codex \
+  -c model="${MODEL}" \
+  -c model_provider=subconscious \
+  -c model_catalog_json="${CATALOG_FILE}" \
+  -c model_reasoning_effort="${CODEX_REASONING_EFFORT}" \
+  -c web_search=disabled \
+  ${EXTERNAL_TOOL_ARGS[@]+"${EXTERNAL_TOOL_ARGS[@]}"} \
+  ${SUBAGENT_ARGS[@]+"${SUBAGENT_ARGS[@]}"} \
+  -c model_providers.subconscious.name=Subconscious \
+  -c model_providers.subconscious.base_url="${GATEWAY_URL}/v1" \
+  -c model_providers.subconscious.wire_api=responses \
+  -c model_providers.subconscious.env_key=SUBCONSCIOUS_API_KEY \
+  -c model_providers.subconscious.stream_idle_timeout_ms="${CODEX_STREAM_IDLE_TIMEOUT_MS}" \
+  ${PASSTHRU[@]+"${PASSTHRU[@]}"}

package/bin/runbook/copilot/install.sh

@@ -64,6 +64,7 @@
 set -euo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../model-capabilities.generated.sh"
 HOOK_SRC="${SCRIPT_DIR}/hook.sh"
 HOOKS_TEMPLATE="${SCRIPT_DIR}/hooks.json"
 
@@ -177,7 +178,8 @@ done
 DEFAULT_SUBCONSCIOUS_MODELS="subconscious/glm-5.3-marathon
 subconscious/glm-5.2
 subconscious/tim-qwen3.6-27b
-subconscious/deepseek-v4-flash-marathon"
+subconscious/deepseek-v4-flash-marathon
+subconscious/deepseek-v4.1-flash-marathon"
 SUPPORTED_MODELS=()
 
 add_supported_model() {
@@ -294,13 +296,16 @@ write_config() {
   local existing
   existing="$(strip_subconscious "$models_json")"
 
-  local provider_models='[]' model_id
+  local provider_models='[]' model_id vision
   for model_id in "${SUPPORTED_MODELS[@]}"; do
+    vision=false
+    if subc_model_supports_vision "$model_id"; then vision=true; fi
     provider_models=$(jq -cn \
       --argjson models "$provider_models" \
       --arg modelId "$model_id" \
       --arg modelName "Subconscious ${model_id}" \
       --arg url "$messages_url" \
+      --argjson vision "$vision" \
       --argjson maxIn "$MAX_INPUT_TOKENS" \
       --argjson maxOut "$MAX_OUTPUT_TOKENS" \
       '$models + [{
@@ -308,7 +313,7 @@ write_config() {
         name: $modelName,
         url: $url,
         toolCalling: true,
-        vision: false,
+        vision: $vision,
         maxInputTokens: $maxIn,
         maxOutputTokens: $maxOut,
         thinking: true,

package/bin/runbook/cursor/install.sh

@@ -59,7 +59,8 @@ COMMAND="install"
 DEFAULT_SUBCONSCIOUS_MODELS="subconscious/glm-5.3-marathon
 subconscious/glm-5.2
 subconscious/tim-qwen3.6-27b
-subconscious/deepseek-v4-flash-marathon"
+subconscious/deepseek-v4-flash-marathon
+subconscious/deepseek-v4.1-flash-marathon"
 
 SUPPORTED_MODELS=()
 add_supported_model() {

package/bin/runbook/deepseek-harness/run.sh

@@ -4,6 +4,9 @@
 
 set -euo pipefail
 
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../model-capabilities.generated.sh"
+
 GATEWAY_URL="${GATEWAY_URL:-}"
 API_KEY="${DEEPSEEK_HARNESS_API_KEY:-${API_KEY:-}}"
 MODEL="${MODEL:-subconscious/glm-5.3-marathon}"
@@ -13,7 +16,8 @@ MAX_TOKENS="${DEEPSEEK_HARNESS_MAX_TOKENS:-65536}"
 DEFAULT_SUBCONSCIOUS_MODELS="subconscious/glm-5.3-marathon
 subconscious/glm-5.2
 subconscious/tim-qwen3.6-27b
-subconscious/deepseek-v4-flash-marathon"
+subconscious/deepseek-v4-flash-marathon
+subconscious/deepseek-v4.1-flash-marathon"
 
 if [[ -z "$GATEWAY_URL" || -z "$API_KEY" ]]; then
   echo "error: GATEWAY_URL and API_KEY are required to launch DeepSeek Harness" >&2
@@ -79,6 +83,9 @@ trap cleanup EXIT HUP INT TERM
   for model_id in "${SUPPORTED_MODELS[@]}"; do
     printf "          - id: '%s'\n" "$model_id"
     printf "            name: '%s'\n" "$model_id"
+    if subc_model_supports_vision "$model_id"; then
+      printf '%s\n' '            input: [text, image]'
+    fi
     printf "            contextWindow: %s\n" "$CONTEXT_WINDOW"
     printf "            maxTokens: %s\n" "$MAX_TOKENS"
   done

package/bin/runbook/model-capabilities.generated.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+# Generated from agents/registry.json. Do not edit by hand.
+# Exact IDs only: similarly named models do not inherit capabilities.
+subc_model_supports_vision() {
+  case "${1:-}" in
+    'subconscious/deepseek-v4.1-flash-marathon') return 0 ;;
+    *) return 1 ;;
+  esac
+}

package/bin/runbook/opencode/install.sh

@@ -54,6 +54,7 @@
 set -euo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../model-capabilities.generated.sh"
 
 # Load shared env from SUBC_ENV_FILE, or a sibling .env / env.example.
 SHARED_ENV="${SUBC_ENV_FILE:-${SCRIPT_DIR}/../.env}"
@@ -122,7 +123,8 @@ done
 DEFAULT_SUBCONSCIOUS_MODELS="subconscious/glm-5.3-marathon
 subconscious/glm-5.2
 subconscious/tim-qwen3.6-27b
-subconscious/deepseek-v4-flash-marathon"
+subconscious/deepseek-v4-flash-marathon
+subconscious/deepseek-v4.1-flash-marathon"
 SUPPORTED_MODELS=()
 
 add_supported_model() {
@@ -147,7 +149,11 @@ done <<< "${SUBCONSCIOUS_MODELS:-$DEFAULT_SUBCONSCIOUS_MODELS}"
 
 MODELS_JSON=""
 for model_id in "${SUPPORTED_MODELS[@]}"; do
-  model_json="\"${model_id}\":{\"name\":\"${model_id}\",\"tools\":true,\"limit\":{\"context\":${CONTEXT_LIMIT},\"output\":${OUTPUT_LIMIT}}}"
+  vision_fields=""
+  if subc_model_supports_vision "$model_id"; then
+    vision_fields=',"attachment":true,"modalities":{"input":["text","image"],"output":["text"]}'
+  fi
+  model_json="\"${model_id}\":{\"name\":\"${model_id}\",\"tools\":true,\"limit\":{\"context\":${CONTEXT_LIMIT},\"output\":${OUTPUT_LIMIT}}${vision_fields}}"
   if [[ -n "$MODELS_JSON" ]]; then
     MODELS_JSON="${MODELS_JSON},${model_json}"
   else

package/bin/runbook/opencode/run.sh

@@ -18,6 +18,7 @@
 set -euo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../model-capabilities.generated.sh"
 
 # Load shared env from SUBC_ENV_FILE, or a sibling .env / env.example.
 SHARED_ENV="${SUBC_ENV_FILE:-${SCRIPT_DIR}/../.env}"
@@ -33,7 +34,8 @@ OUTPUT_LIMIT="${OPENCODE_OUTPUT_LIMIT:-65536}"
 DEFAULT_SUBCONSCIOUS_MODELS="subconscious/glm-5.3-marathon
 subconscious/glm-5.2
 subconscious/tim-qwen3.6-27b
-subconscious/deepseek-v4-flash-marathon"
+subconscious/deepseek-v4-flash-marathon
+subconscious/deepseek-v4.1-flash-marathon"
 SUPPORTED_MODELS=()
 
 add_supported_model() {
@@ -58,7 +60,11 @@ done <<< "${SUBCONSCIOUS_MODELS:-$DEFAULT_SUBCONSCIOUS_MODELS}"
 
 MODELS_JSON=""
 for model_id in "${SUPPORTED_MODELS[@]}"; do
-  model_json="\"${model_id}\":{\"name\":\"${model_id}\",\"tools\":true,\"limit\":{\"context\":${CONTEXT_LIMIT},\"output\":${OUTPUT_LIMIT}}}"
+  vision_fields=""
+  if subc_model_supports_vision "$model_id"; then
+    vision_fields=',"attachment":true,"modalities":{"input":["text","image"],"output":["text"]}'
+  fi
+  model_json="\"${model_id}\":{\"name\":\"${model_id}\",\"tools\":true,\"limit\":{\"context\":${CONTEXT_LIMIT},\"output\":${OUTPUT_LIMIT}}${vision_fields}}"
   if [[ -n "$MODELS_JSON" ]]; then
     MODELS_JSON="${MODELS_JSON},${model_json}"
   else

package/bin/runbook/pi/install.sh

@@ -47,6 +47,7 @@
 set -euo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../model-capabilities.generated.sh"
 EXTENSION_SRC="${SCRIPT_DIR}/subconscious-compaction.ts"
 
 # Load shared env from SUBC_ENV_FILE, or a sibling .env / env.example.
@@ -117,7 +118,8 @@ done
 DEFAULT_SUBCONSCIOUS_MODELS="subconscious/glm-5.3-marathon
 subconscious/glm-5.2
 subconscious/tim-qwen3.6-27b
-subconscious/deepseek-v4-flash-marathon"
+subconscious/deepseek-v4-flash-marathon
+subconscious/deepseek-v4.1-flash-marathon"
 SUPPORTED_MODELS=()
 
 add_supported_model() {
@@ -142,7 +144,11 @@ done <<< "${SUBCONSCIOUS_MODELS:-$DEFAULT_SUBCONSCIOUS_MODELS}"
 
 MODEL_ENTRIES_JSON=""
 for model_id in "${SUPPORTED_MODELS[@]}"; do
-  model_json="{\"id\":\"${model_id}\",\"contextWindow\":${CONTEXT_WINDOW},\"maxTokens\":${MAX_TOKENS},\"compat\":{\"sendSessionAffinityHeaders\":true,\"sessionAffinityFormat\":\"openai-nosession\"}}"
+  vision_fields=""
+  if subc_model_supports_vision "$model_id"; then
+    vision_fields=',"input":["text","image"]'
+  fi
+  model_json="{\"id\":\"${model_id}\",\"contextWindow\":${CONTEXT_WINDOW},\"maxTokens\":${MAX_TOKENS},\"compat\":{\"sendSessionAffinityHeaders\":true,\"sessionAffinityFormat\":\"openai-nosession\"}${vision_fields}}"
   if [[ -n "$MODEL_ENTRIES_JSON" ]]; then
     MODEL_ENTRIES_JSON="${MODEL_ENTRIES_JSON},${model_json}"
   else

package/bin/windows/launch.js

@@ -3,6 +3,7 @@ import fs from 'node:fs/promises';
 import os from 'node:os';
 import path from 'node:path';
 import { claudeNativeLaunch } from './claude.js';
+import { modelSupportsVision } from '../model-capabilities.js';
 import { defaults, value, origin, modelIds, positiveInteger, parseOptions, writeJson } from './common.js';
 
 function connection(env, keyName) {
@@ -25,7 +26,10 @@ export async function windowsLaunch(id, argv, env, { tempRoot = os.tmpdir() } =
   if (id === 'opencode') {
     const context = positiveInteger(value(env, 'OPENCODE_CONTEXT_LIMIT', '5000000'), 'OPENCODE_CONTEXT_LIMIT');
     const output = positiveInteger(value(env, 'OPENCODE_OUTPUT_LIMIT', '65536'), 'OPENCODE_OUTPUT_LIMIT');
-    const models = Object.fromEntries(modelIds(env).map(id => [id, { name: id, tools: true, limit: { context, output } }]));
+    const models = Object.fromEntries(modelIds(env).map(id => [id, {
+      name: id, tools: true, limit: { context, output },
+      ...(modelSupportsVision(id) ? { attachment: true, modalities: { input: ['text', 'image'], output: ['text'] } } : {}),
+    }]));
     childEnv.OPENCODE_CONFIG_CONTENT = JSON.stringify({
       $schema: 'https://opencode.ai/config.json', disabled_providers: ['subconscious'],
       provider: { 'subconscious-cli': { npm: '@ai-sdk/openai-compatible', name: 'Subconscious Gateway', options: {
@@ -51,6 +55,7 @@ export async function windowsLaunch(id, argv, env, { tempRoot = os.tmpdir() } =
       supports_reasoning_summaries: false, support_verbosity: false, default_verbosity: null,
       apply_patch_tool_type: 'freeform', truncation_policy: { mode: 'tokens', limit: 10000 },
       supports_parallel_tool_calls: true, experimental_supported_tools: [],
+      ...(modelSupportsVision(id) ? { input_modalities: ['text', 'image'] } : {}),
     })) };
     // Validate all flags before creating a temporary directory.
     const threads = positiveInteger(value(env, 'MAX_CONCURRENT_SUBAGENTS', '4'), 'MAX_CONCURRENT_SUBAGENTS');
@@ -81,7 +86,7 @@ export async function windowsLaunch(id, argv, env, { tempRoot = os.tmpdir() } =
     const cleanup = () => fs.rm(dir, { recursive: true, force: true });
     try {
       const file = path.join(dir, 'subconscious.cordis.yml');
-      const text = `- id: llm-pi-ai\n  config:\n    providers:\n      subconscious:\n        apiKeyEnv: SUBCONSCIOUS_API_KEY\n        displayName: Subconscious Gateway\n        api: openai-completions\n        baseURL: !!js process.env.SUBCONSCIOUS_DSH_BASE_URL\n        headers:\n          x-subconscious-client: deepseek-harness\n        compat:\n          supportsDeveloperRole: false\n          maxTokensField: max_tokens\n        defaultContextWindow: ${context}\n        defaultMaxTokens: ${maxTokens}\n        models:\n${models.map(id => `          - id: '${id}'\n            name: '${id}'\n            contextWindow: ${context}\n            maxTokens: ${maxTokens}`).join('\n')}\n- id: agent-default-model\n  config:\n    provider: subconscious\n    model: '${model}'\n`;
+      const text = `- id: llm-pi-ai\n  config:\n    providers:\n      subconscious:\n        apiKeyEnv: SUBCONSCIOUS_API_KEY\n        displayName: Subconscious Gateway\n        api: openai-completions\n        baseURL: !!js process.env.SUBCONSCIOUS_DSH_BASE_URL\n        headers:\n          x-subconscious-client: deepseek-harness\n        compat:\n          supportsDeveloperRole: false\n          maxTokensField: max_tokens\n        defaultContextWindow: ${context}\n        defaultMaxTokens: ${maxTokens}\n        models:\n${models.map(id => `          - id: '${id}'\n            name: '${id}'\n${modelSupportsVision(id) ? '            input: [text, image]\n' : ''}            contextWindow: ${context}\n            maxTokens: ${maxTokens}`).join('\n')}\n- id: agent-default-model\n  config:\n    provider: subconscious\n    model: '${model}'\n`;
       await fs.writeFile(file, text, { mode: 0o600, flag: 'wx' });
       const mode = ['web', 'headless'].includes(argv[0]) ? argv[0] : 'web';
       const rest = ['web', 'headless'].includes(argv[0]) ? argv.slice(1) : argv;

package/bin/windows/setup.js

@@ -1,6 +1,7 @@
 import fs from 'node:fs/promises';
 import os from 'node:os';
 import path from 'node:path';
+import { modelSupportsVision } from '../model-capabilities.js';
 import { defaults, origin, value, modelIds, positiveInteger, parseOptions, object, readJson, writeJson } from './common.js';
 
 const owned = command => {
@@ -92,7 +93,7 @@ export async function windowsSetup(id, action, argv, environment, { home = os.ho
         const context = positiveInteger(options['--max-input-tokens'] || value(env, 'COPILOT_MAX_INPUT_TOKENS', '5000000'), 'max-input-tokens');
         const maxTokens = positiveInteger(options['--max-output-tokens'] || value(env, 'COPILOT_MAX_OUTPUT_TOKENS', '65536'), 'max-output-tokens');
         providers.push({ name: providerName, vendor: 'customendpoint', apiKey: '${input:chat.lm.secret.subconscious-gateway}', apiType: 'messages', models: modelIds(env).map(id => ({
-          id, name: `Subconscious ${id}`, url: `${gatewayUrl}/v1/messages`, toolCalling: true, vision: false,
+          id, name: `Subconscious ${id}`, url: `${gatewayUrl}/v1/messages`, toolCalling: true, vision: modelSupportsVision(id),
           maxInputTokens: context, maxOutputTokens: maxTokens, thinking: true, streaming: true,
           requestHeaders: { 'x-subconscious-client': 'copilot' },
         })) });
@@ -131,7 +132,10 @@ export async function windowsSetup(id, action, argv, environment, { home = os.ho
     if (installing) {
       const context = positiveInteger(options['--context-window'] || value(env, 'PI_CONTEXT_WINDOW', '5000000'), 'context-window');
       const maxTokens = positiveInteger(options['--max-tokens'] || value(env, 'PI_MAX_TOKENS', '65536'), 'max-tokens');
-      providers.subconscious = { baseUrl: `${gatewayUrl}/v1`, api: 'openai-completions', apiKey, headers: { 'x-subconscious-client': 'pi' }, models: modelIds(env).map(id => ({ id, contextWindow: context, maxTokens, compat: { sendSessionAffinityHeaders: true, sessionAffinityFormat: 'openai-nosession' } })) };
+      providers.subconscious = { baseUrl: `${gatewayUrl}/v1`, api: 'openai-completions', apiKey, headers: { 'x-subconscious-client': 'pi' }, models: modelIds(env).map(id => ({
+        id, contextWindow: context, maxTokens, compat: { sendSessionAffinityHeaders: true, sessionAffinityFormat: 'openai-nosession' },
+        ...(modelSupportsVision(id) ? { input: ['text', 'image'] } : {}),
+      })) };
       await fs.mkdir(path.dirname(extension), { recursive: true });
       await fs.copyFile(new URL('./pi-compaction.ts', import.meta.url), extension);
       await writeJson(config, { gatewayUrl, apiKey });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "subconscious-cli",
-  "version": "4.1.2",
+  "version": "4.1.4",
   "description": "CLI for Subconscious — run Subconscious Code, Claude Code, Codex, OpenCode, DeepSeek Harness, Cursor, Copilot, and Pi",
   "bin": {
     "subc": "bin/cli.js"
@@ -9,7 +9,7 @@
     "build:tui": "node scripts/build-tui.js",
     "build:tui:host": "node scripts/build-tui.js --host",
     "test": "node --test",
-    "test:windows": "node --test test/windows.test.js test/models.test.js test/auth.test.js test/tui.test.js test/update-check.test.js test/sessions.test.js test/branding.test.js test/claude-version.test.js test/usage.test.js test/release.test.js",
+    "test:windows": "node --test test/windows.test.js test/models.test.js test/model-capabilities.test.js test/auth.test.js test/tui.test.js test/update-check.test.js test/sessions.test.js test/branding.test.js test/claude-version.test.js test/usage.test.js test/release.test.js",
     "test:tui": "go -C tui test ./...",
     "prepack": "npm run build:tui",
     "prepublishOnly": "node scripts/require-publish-script.js"