Tools tab (gh) + prompt caching + conversation compaction

Dockerfile

@@ -13,6 +13,15 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 # Install Himalaya CLI (pre-built Rust binary for email management)
 RUN curl -sSL https://raw.githubusercontent.com/pimalaya/himalaya/master/install.sh | sh
 
+# Install GitHub CLI (gh) from the official apt repository (Tools tab: gh)
+RUN curl -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg \
+      -o /usr/share/keyrings/githubcli-archive-keyring.gpg \
+    && chmod go+r /usr/share/keyrings/githubcli-archive-keyring.gpg \
+    && echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" \
+      > /etc/apt/sources.list.d/github-cli.list \
+    && apt-get update && apt-get install -y --no-install-recommends gh \
+    && rm -rf /var/lib/apt/lists/*
+
 # Install uv
 COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
 

api/admin.py

@@ -8,6 +8,7 @@
 
 from __future__ import annotations
 
+import asyncio
 import collections
 import hashlib
 import json
@@ -36,6 +37,8 @@
     DEFAULT_TOOL_USAGE_BLOCK,
     build_prompt_sections,
 )
+from core.tools import registry as tool_registry
+from core.tools import tool_env
 from core.wacli import WacliManager
 
 if TYPE_CHECKING:
@@ -545,6 +548,8 @@ async def _lifespan(app: FastAPI):  # noqa: ANN001
     _HISTORY_PREFIX = "history."
     _EMAIL_PREFIX = "email."
     _PROMPT_PREFIX = "prompt."
+    _TOOLS_PREFIX = "tools."
+    _COMPACTION_PREFIX = "compaction."
 
     def _is_managed_key(key: str) -> bool:
         """Return True if this key is managed by a dedicated tab (not Config)."""
@@ -562,6 +567,8 @@ def _is_managed_key(key: str) -> bool:
             _HISTORY_PREFIX,
             _EMAIL_PREFIX,
             _PROMPT_PREFIX,
+            _TOOLS_PREFIX,
+            _COMPACTION_PREFIX,
         ):
             if key.startswith(prefix):
                 return True
@@ -801,6 +808,8 @@ async def partial_llm() -> HTMLResponse:
         tr_enabled = tr_enabled if tr_enabled is not None else "true"
         tr_provider = await config_store.get("task_reflection.provider") or "anthropic"
         tr_model = await config_store.get("task_reflection.model") or "claude-haiku-4-5"
+        compaction_provider = await config_store.get("compaction.provider") or "anthropic"
+        compaction_model = await config_store.get("compaction.model") or "claude-haiku-4-5"
         prompt_tool_usage_override = await config_store.get("prompt.tool_usage_override") or ""
         prompt_history_override = await config_store.get("prompt.history_handling_override") or ""
         prompt_capture_enabled = await config_store.get("admin.capture_prompts")
@@ -830,13 +839,58 @@ async def partial_llm() -> HTMLResponse:
             tr_enabled=tr_enabled,
             tr_provider=tr_provider,
             tr_model=tr_model,
+            compaction_provider=compaction_provider,
+            compaction_model=compaction_model,
             prompt_tool_usage_override=prompt_tool_usage_override,
             prompt_history_override=prompt_history_override,
             default_tool_usage=DEFAULT_TOOL_USAGE_BLOCK,
             default_history_handling=DEFAULT_HISTORY_HANDLING_BLOCK,
             prompt_capture_enabled=prompt_capture_enabled,
         )
 
+    @app.get("/partials/tools", dependencies=[Depends(auth)])
+    async def partial_tools() -> HTMLResponse:
+        """Tools tab partial — manage optional external CLI tools (e.g. gh)."""
+        gh_enabled = await config_store.get("tools.gh.enabled")
+        gh_enabled = gh_enabled if gh_enabled is not None else "false"
+        gh_token = await config_store.get("tools.gh.token") or ""
+        return _render_partial(
+            "partials/tools.html",
+            tools=tool_registry(),
+            gh_enabled=gh_enabled,
+            gh_token=gh_token,
+        )
+
+    @app.post("/tools/gh/test", dependencies=[Depends(auth)])
+    async def test_gh_tool(request: Request) -> dict:
+        """Verify a GitHub token by calling the GitHub API as that token."""
+        body = await request.json()
+        token = str(body.get("token", "")).strip()
+        if not token:
+            return {"ok": False, "error": "Token is required."}
+        try:
+            resp = await asyncio.to_thread(
+                http_requests.get,
+                "https://api.github.com/user",
+                headers={
+                    "Authorization": f"Bearer {token}",
+                    "Accept": "application/vnd.github+json",
+                    "X-GitHub-Api-Version": "2022-11-28",
+                },
+                timeout=10,
+            )
+        except Exception as exc:  # noqa: BLE001 — surface any network error to the UI
+            return {"ok": False, "error": str(exc)}
+        if resp.status_code == 200:
+            login = resp.json().get("login", "")
+            return {"ok": True, "login": login}
+        if resp.status_code in (401, 403):
+            return {
+                "ok": False,
+                "error": "Token rejected by GitHub (invalid or insufficient scope).",
+            }
+        return {"ok": False, "error": f"GitHub returned HTTP {resp.status_code}."}
+
     @app.get("/partials/search", dependencies=[Depends(auth)])
     async def partial_search() -> HTMLResponse:
         """Search tab partial."""
@@ -890,10 +944,23 @@ async def partial_history() -> HTMLResponse:
         """History tab partial."""
         mode = await config_store.get("history.mode") or "injection"
         max_turns = await config_store.get("history.max_turns") or "10"
+        c_enabled = await config_store.get("compaction.enabled")
+        c_enabled = c_enabled if c_enabled is not None else "true"
+        c_threshold_type = await config_store.get("compaction.threshold_type") or "percent"
+        c_threshold_percent = await config_store.get("compaction.threshold_percent") or "80"
+        c_threshold_tokens = await config_store.get("compaction.threshold_tokens") or "150000"
+        c_context_window = await config_store.get("compaction.context_window") or "200000"
+        c_keep_recent_turns = await config_store.get("compaction.keep_recent_turns") or "4"
         return _render_partial(
             "partials/history.html",
             mode=mode,
             max_turns=max_turns,
+            compaction_enabled=c_enabled,
+            compaction_threshold_type=c_threshold_type,
+            compaction_threshold_percent=c_threshold_percent,
+            compaction_threshold_tokens=c_threshold_tokens,
+            compaction_context_window=c_context_window,
+            compaction_keep_recent_turns=c_keep_recent_turns,
         )
 
     @app.get("/partials/logs", dependencies=[Depends(auth)])
@@ -1174,6 +1241,7 @@ async def patch_config(body: ConfigPatchIn) -> dict:
                 new_config = await config_store.export_to_config()
                 agent.config = new_config
                 agent.llm = LLMClient.from_agent_config(new_config.agent)
+                agent.executor.tool_env = tool_env(new_config)
                 agent.history_mode = new_config.history.mode
                 agent.memory.long_term_limit = new_config.memory.long_term_limit
                 agent.reflections.max_reflections = new_config.task_reflection.max_reflections

api/templates/base.html

@@ -42,7 +42,7 @@
     }
     window.showToast = showToast;
 
-    function llmTab(provider, apiKey, model, openaiKey, openaiBaseUrl, googleKey, googleBaseUrl, grokKey, grokBaseUrl, deepseekKey, deepseekBaseUrl, extractionProvider, extractionModel, consolidationProvider, consolidationModel, gdEnabled, gdProvider, gdModel, trEnabled, trProvider, trModel, promptToolUsageOverride, promptHistoryOverride, defaultToolUsage, defaultHistoryHandling, promptCaptureEnabled) {
+    function llmTab(provider, apiKey, model, openaiKey, openaiBaseUrl, googleKey, googleBaseUrl, grokKey, grokBaseUrl, deepseekKey, deepseekBaseUrl, extractionProvider, extractionModel, consolidationProvider, consolidationModel, gdEnabled, gdProvider, gdModel, trEnabled, trProvider, trModel, promptToolUsageOverride, promptHistoryOverride, defaultToolUsage, defaultHistoryHandling, promptCaptureEnabled, compactionProvider, compactionModel) {
       const currentProvider = provider || 'anthropic';
       return {
         providerOptions: [
@@ -97,6 +97,10 @@
         gdResultOk: false,
         trResult: '',
         trResultOk: false,
+        compactionProvider: compactionProvider || 'anthropic',
+        compactionModel: compactionModel || 'claude-haiku-4-5',
+        compactionResult: '',
+        compactionResultOk: false,
         result: '',
         resultOk: false,
         tests: {

api/templates/dashboard.html

@@ -80,6 +80,10 @@ <h2 class="text-base text-accent">Agent Control</h2>
             @click="select('search')">
       Search
     </button>
+    <button class="tab-link" :class="tab === 'tools' && 'tab-link-active'"
+            @click="select('tools')">
+      Tools
+    </button>
     <button class="tab-link" :class="tab === 'permissions' && 'tab-link-active'"
             @click="select('permissions')">
       Permissions
@@ -675,6 +679,7 @@ <h2 class="text-base text-accent">Agent Control</h2>
       history: '/partials/history',
       channels: '/partials/channels',
       search: '/partials/search',
+      tools: '/partials/tools',
       permissions: '/partials/permissions',
       skills: '/partials/skills',
       jobs: '/partials/jobs',

api/templates/partials/history.html

@@ -2,8 +2,16 @@
 <div class="space-y-6" x-data="{
   mode: {{ mode|default('injection', true)|tojson|forceescape }},
   maxTurns: {{ max_turns|default('10', true)|tojson|forceescape }},
+  compactionEnabled: {{ compaction_enabled|default('true', true)|tojson|forceescape }},
+  thresholdType: {{ compaction_threshold_type|default('percent', true)|tojson|forceescape }},
+  thresholdPercent: {{ compaction_threshold_percent|default('80', true)|tojson|forceescape }},
+  thresholdTokens: {{ compaction_threshold_tokens|default('150000', true)|tojson|forceescape }},
+  contextWindow: {{ compaction_context_window|default('200000', true)|tojson|forceescape }},
+  keepRecentTurns: {{ compaction_keep_recent_turns|default('4', true)|tojson|forceescape }},
   result: '',
-  resultOk: false
+  resultOk: false,
+  cResult: '',
+  cResultOk: false
 }">
   <div class="card">
     <h2 class="text-base mb-1">Conversation History</h2>
@@ -66,4 +74,104 @@ <h2 class="text-base mb-1">Conversation History</h2>
       <div :class="resultOk ? 'alert-success' : 'alert-error'" class="mt-2" x-text="result"></div>
     </template>
   </div>
+
+  {# Compaction — only relevant in session mode #}
+  <div class="card">
+    <div class="flex items-center justify-between gap-3 mb-1">
+      <h2 class="text-base">Context compaction</h2>
+      <div class="flex items-center gap-2">
+        <label class="label mb-0">Enabled</label>
+        <input type="checkbox" x-model="compactionEnabled"
+               :checked="compactionEnabled === 'true' || compactionEnabled === true">
+      </div>
+    </div>
+    <p class="text-muted text-xs mb-4">
+      In <strong>session</strong> mode, when the conversation approaches the model's
+      context window, the oldest turns are summarized by a small model (configured in
+      the <strong>LLM</strong> tab) while recent turns are kept verbatim. The user is
+      notified with a system message when this happens. Has no effect in injection mode.
+    </p>
+
+    <div class="space-y-3" :class="(compactionEnabled === 'true' || compactionEnabled === true) ? '' : 'opacity-50'">
+      <div>
+        <label class="label">Trigger threshold</label>
+        <div class="flex gap-4 mt-1">
+          <label class="flex items-center gap-2 cursor-pointer">
+            <input type="radio" name="threshold_type" value="percent"
+                   x-model="thresholdType" class="accent-primary">
+            <span class="text-sm">% of context window</span>
+          </label>
+          <label class="flex items-center gap-2 cursor-pointer">
+            <input type="radio" name="threshold_type" value="tokens"
+                   x-model="thresholdType" class="accent-primary">
+            <span class="text-sm">Absolute tokens</span>
+          </label>
+        </div>
+      </div>
+
+      <div x-show="thresholdType === 'percent'" x-transition class="flex gap-4 flex-wrap">
+        <div>
+          <label class="label">Percent</label>
+          <input type="number" class="input-sm" style="max-width:120px"
+                 x-model="thresholdPercent" min="10" max="99">
+          <p class="text-muted text-xs mt-1">Compact at this % of the window.</p>
+        </div>
+        <div>
+          <label class="label">Context window (fallback)</label>
+          <input type="number" class="input-sm" style="max-width:160px"
+                 x-model="contextWindow" min="8000" step="1000">
+          <p class="text-muted text-xs mt-1">Used when the model's window isn't known (e.g. 200000, 1000000).</p>
+        </div>
+      </div>
+
+      <div x-show="thresholdType === 'tokens'" x-transition>
+        <label class="label">Tokens</label>
+        <input type="number" class="input-sm" style="max-width:160px"
+               x-model="thresholdTokens" min="8000" step="1000">
+        <p class="text-muted text-xs mt-1">Compact once the context reaches this many tokens (e.g. 150000).</p>
+      </div>
+
+      <div>
+        <label class="label">Keep recent turns</label>
+        <input type="number" class="input-sm" style="max-width:120px"
+               x-model="keepRecentTurns" min="1" max="20">
+        <p class="text-muted text-xs mt-1">Most-recent user turns kept verbatim; everything older is summarized.</p>
+      </div>
+    </div>
+
+    <div class="flex items-center gap-2 mt-4">
+      <button class="btn-primary btn-sm"
+              @click="
+                const vals = {
+                  'compaction.enabled': String(compactionEnabled === true || compactionEnabled === 'true'),
+                  'compaction.threshold_type': thresholdType,
+                  'compaction.threshold_percent': String(thresholdPercent),
+                  'compaction.threshold_tokens': String(thresholdTokens),
+                  'compaction.context_window': String(contextWindow),
+                  'compaction.keep_recent_turns': String(keepRecentTurns)
+                };
+                fetch('/config', {
+                  method: 'PATCH',
+                  headers: {
+                    'Content-Type': 'application/json',
+                    'Authorization': 'Bearer ' + (localStorage.getItem('admin_api_key') || '')
+                  },
+                  body: JSON.stringify({values: vals})
+                })
+                .then(r => { cResultOk = r.ok; return r.json(); })
+                .then(d => {
+                  cResult = cResultOk ? 'Compaction settings saved' : (d.detail || 'Error');
+                  if (cResultOk && window.showToast) { window.showToast('Compaction settings saved'); }
+                })
+                .catch(e => { cResultOk = false; cResult = e.message; })
+              ">
+        Save
+      </button>
+      <span class="text-muted text-xs">Compaction model is set in the LLM tab.</span>
+    </div>
+
+    <template x-if="cResult">
+      <div :class="cResultOk ? 'alert-success' : 'alert-error'" class="mt-2" x-text="cResult"></div>
+    </template>
+  </div>
 </div>

api/templates/partials/llm.html

@@ -25,7 +25,9 @@
   {{ prompt_history_override|default('', true)|tojson|forceescape }},
   {{ default_tool_usage|default('', true)|tojson|forceescape }},
   {{ default_history_handling|default('', true)|tojson|forceescape }},
-  {{ prompt_capture_enabled|default(false, true)|tojson|forceescape }}
+  {{ prompt_capture_enabled|default(false, true)|tojson|forceescape }},
+  {{ compaction_provider|default('anthropic', true)|tojson|forceescape }},
+  {{ compaction_model|default('claude-haiku-4-5', true)|tojson|forceescape }}
 )">
   <div class="card">
     <h2 class="text-base mb-1">System Prompt Controls</h2>
@@ -415,6 +417,61 @@ <h2 class="text-base mb-1">Task Reflection</h2>
     </template>
   </div>
 
+  <div class="card">
+    <h2 class="text-base mb-1">History Compaction</h2>
+    <p class="text-muted text-xs mb-4">Model used to summarize older conversation turns when a session's context grows large. Enable compaction and set the trigger threshold in the <strong>History</strong> tab.</p>
+
+    <form class="space-y-4" @submit.prevent>
+      <div>
+        <label class="label">Provider</label>
+        <select class="input-sm" style="max-width:500px" x-model="compactionProvider" x-effect="$el.value = compactionProvider">
+          <template x-for="item in providerOptions" :key="'cp-' + item.value">
+            <option :value="item.value" :selected="item.value === compactionProvider" x-text="item.label"></option>
+          </template>
+        </select>
+      </div>
+      <div>
+        <label class="label">Model</label>
+        <input type="text" class="input-sm" style="max-width:500px" x-model="compactionModel"
+               list="dl-model-compaction" placeholder="Type or pick a model">
+        <datalist id="dl-model-compaction">
+          <template x-for="item in modelOptions(compactionProvider)" :key="'cp-m-' + item.value">
+            <option :value="item.value"></option>
+          </template>
+        </datalist>
+        <p class="text-muted text-xs mt-1">A small, fast model is recommended (summarization is cheap).</p>
+      </div>
+    </form>
+
+    <div class="flex items-center gap-2 mt-4">
+      <button class="btn-primary btn-sm"
+              @click="
+                fetch('/config', {
+                  method: 'PATCH',
+                  headers: {
+                    'Content-Type': 'application/json',
+                    'Authorization': 'Bearer ' + (localStorage.getItem('admin_api_key') || '')
+                  },
+                  body: JSON.stringify({values: {
+                    'compaction.provider': compactionProvider,
+                    'compaction.model': compactionModel
+                  }})
+                })
+                .then(r => { compactionResultOk = r.ok; return r.json(); })
+                .then(d => {
+                  compactionResult = compactionResultOk ? 'Compaction model saved' : (d.detail || 'Error');
+                  if (compactionResultOk && window.showToast) { window.showToast('Compaction model saved'); }
+                })
+                .catch(e => { compactionResultOk = false; compactionResult = e.message; })
+              ">
+        Save compaction model
+      </button>
+    </div>
+    <template x-if="compactionResult">
+      <div :class="compactionResultOk ? 'alert-success' : 'alert-error'" class="mt-2" x-text="compactionResult"></div>
+    </template>
+  </div>
+
   <div class="card">
     <h2 class="text-base mb-1">Anthropic</h2>
     <p class="text-muted text-xs mb-4">Configure the Anthropic API key and test the connection.</p>