diff --git a/kernel/bin/agent b/kernel/bin/agent index fd6a3d8b..3052d975 100644 --- a/kernel/bin/agent +++ b/kernel/bin/agent @@ -3442,27 +3442,19 @@ async function bareAgentSkillsCompactPrompt(ctx, paths, maxChars) { const cap = Math.min(Math.max(Number(maxChars) || 4000, 500), 12000) const skills = await bareAgentDiscoverSkills(ctx, paths) let block = - '## Available skills (compact index)\n' + - 'Each skill is a directory with **SKILL.md** (optional YAML frontmatter: `name`, `description`, …).\n' + - '**Workspace skills** (`~/.agent/workspace/skills/`) override **global** (`~/.agent/skills/`) and walk-up `.grok/skills` / `.agents/skills` when names match.\n' + - 'To run one: call the **read_skill** tool with the skill id or frontmatter `name` before following its instructions.\n\n' + 'Skills (playbooks — NOT callable tools)\n' + + 'These ids are SKILL.md playbooks, not functions. Do not list them when asked about your tools.\n' + + 'Each skill is a directory with SKILL.md (optional YAML frontmatter: name, description).\n' + + 'Workspace skills (~/.agent/workspace/skills/) override global (~/.agent/skills/) and walk-up .grok/skills / .agents/skills when names match.\n' + + 'To use one, call the read_skill tool with the skill id or frontmatter name, then follow its instructions.\n\n' if (!skills.length) { - block += '(No skills discovered yet — add folders under `workspace/skills//SKILL.md`.)\n' + block += '(No skills discovered yet — add folders under workspace/skills//SKILL.md.)\n' return block.length > cap ? block.slice(0, cap) + '\n…\n' : block } - block += '| id | name | source | description |\n| --- | --- | --- | --- |\n' for (const s of skills) { - const desc = s.description.replace(/\|/g, '/').replace(/\r?\n/g, ' ').slice(0, 160) - block += - '| `' + - s.id.replace(/`/g, "'") + - '` | ' + - s.name.replace(/\|/g, '/').replace(/\r?\n/g, ' ') + - ' | ' + - s.source + - ' | ' + - desc + - ' |\n' + const desc = s.description.replace(/\r?\n/g, ' ').slice(0, 160) + const label = s.name && s.name !== s.id ? s.id + ' / ' + s.name : s.id + block += '- ' + label + ' (' + s.source + ')' + (desc ? ' — ' + desc : '') + '\n' } if (block.length > cap) block = block.slice(0, cap) + '\n… truncated\n' return block @@ -7987,6 +7979,23 @@ function bareAgentToolDefinitions() { } } }, + { + type: 'function', + function: { + name: 'list_agent_tools', + description: + 'Return the live callable tool registry (name, description, parameter keys). Use this when the user asks what tools you have. Skills and /bin names are not in this list.', + parameters: { + type: 'object', + properties: { + name: { + type: 'string', + description: 'Optional tool name or substring to describe one entry' + } + } + } + } + }, { type: 'function', function: { @@ -8303,6 +8312,78 @@ function bareAgentToolDefinitions() { ] } +/** + * Live tool registry derived from bareAgentToolDefinitions (never hand-maintained). + * @param {unknown[]} [tools] + * @returns {{ name: string, description: string, params: string[] }[]} + */ +function bareAgentToolCatalog(tools) { + const src = + Array.isArray(tools) && tools.length ? tools : bareAgentToolDefinitions() + /** @type {{ name: string, description: string, params: string[] }[]} */ + const out = [] + /** @type {Record} */ + const seen = {} + for (const t of src) { + if (!t || typeof t !== 'object') continue + const o = /** @type {Record} */ (t) + const fn = + o.function && typeof o.function === 'object' + ? /** @type {Record} */ (o.function) + : o + const name = typeof fn.name === 'string' ? fn.name.trim() : '' + if (!name || seen[name]) continue + seen[name] = 1 + const description = + typeof fn.description === 'string' + ? fn.description.replace(/\s+/g, ' ').trim() + : '' + const paramsRoot = + fn.parameters && typeof fn.parameters === 'object' + ? /** @type {Record} */ (fn.parameters) + : {} + const props = + paramsRoot.properties && + typeof paramsRoot.properties === 'object' && + !Array.isArray(paramsRoot.properties) + ? /** @type {Record} */ (paramsRoot.properties) + : {} + out.push({ name, description, params: Object.keys(props) }) + } + return out +} + +/** + * System-prompt block. Placed last so small local models see the real tool + * names instead of the skills index. + * @param {unknown[]} [tools] + */ +function bareAgentToolCatalogPrompt(tools) { + const cat = bareAgentToolCatalog(tools) + const lines = [ + 'AVAILABLE TOOLS (' + + cat.length + + ' callable functions — live registry)', + 'These names are the only functions you can call. Skills such as ctx-api-change, docs-contract-update, hdms, and holesail are playbooks: load them with read_skill. Guest /bin names such as ctx-baredoctor, hdms, and holepunch are commands for run_command, not tools.', + 'When the user asks what tools you have or what you can do, list every name below. Do not invent names. Do not list skills as tools.', + 'To call a tool, emit exactly:', + '{"name":"TOOL_NAME","arguments":{}}', + 'You may call list_agent_tools to reprint this catalog.', + '' + ] + for (const t of cat) { + const desc = + t.description.length > 110 + ? t.description.slice(0, 107) + '...' + : t.description + const params = t.params.length ? ' (' + t.params.join(', ') + ')' : '' + lines.push( + '- ' + t.name + params + (desc ? ' — ' + desc : '') + ) + } + return lines.join('\n') +} + /** Tool dispatch for /bin/agent (preamble; no import). */ /** @@ -10731,6 +10812,27 @@ async function bareAgentDispatchTool(o) { return bareAgentJsonResult({ ok: true, path: dest, kind, appended: line }) } + if (toolName === 'list_agent_tools') { + const cat = + typeof bareAgentToolCatalog === 'function' ? bareAgentToolCatalog() : [] + const want = typeof args.name === 'string' ? args.name.trim() : '' + const rows = want + ? cat.filter(function (t) { + return ( + t.name === want || + String(t.name).toLowerCase().indexOf(want.toLowerCase()) !== -1 + ) + }) + : cat + appendProgress('list_agent_tools n=' + String(rows.length)) + return bareAgentJsonResult({ + ok: true, + count: rows.length, + total: cat.length, + tools: rows + }) + } + if (toolName === 'list_skills') { const max = typeof args.max === 'number' && Number.isFinite(args.max) @@ -13217,7 +13319,7 @@ async function bareAgentPromptSetupLine(ctx, prompt, opts) { const BARE_AGENT_STATIC_SYSTEM = `You are the Bare OS coding agent — a senior implementer that lives inside the guest image (JavaScript POSIX on Hyperdrive + Hyperswarm, Pear/Bare runtime). You write, edit, debug, and verify code and OS state by calling tools. You are not a chatbot that narrates plans and waits. Lead with tools. Execute until the job is done, then call task_complete. -YOU RUN THE TOOLS. The user never runs your tools, commands, patches, or scripts. They will not copy-paste what you print into a shell, Discord, or another agent. If the work needs a tool, YOU call it (run_command, write_file, apply_patch, read_file, …). Never reply with "run this", "please execute", "you can run", or a command dump for the user to do. The user is not your tool runner. +YOU RUN THE TOOLS. The user never runs your tools, commands, patches, or scripts. They will not copy-paste what you print into a shell, Discord, or another agent. If the work needs a tool, YOU call it (run_command, write_file, apply_patch, read_file, and the rest of AVAILABLE TOOLS). Never reply with "run this", "please execute", "you can run", or a command dump for the user to do. The user is not your tool runner. Bare OS by Raven Scott (https://raven-scott.fyi). Repo: https://git.ssh.surf/snxraven/bare-operating-system. Booter: pear://qupw8zspk34pcxc7fqchzyeh33jtmxq1k7qze44fkosctwiid8zy @@ -13225,52 +13327,42 @@ WORK POLICY. - Keep every explicit requirement in view until it is done, superseded, or blocked. If blocked, say so plainly. - Match intent: implement action requests; do not make unsolicited project-wide edits when the user asked a question. - For clear, reversible local work, do it now. NEVER ASK for permission. -- Claim done / fixed / tested only when a tool result supports it. Otherwise say what you did not verify. +- Claim done, fixed, or tested only when a tool result supports it. Otherwise say what you did not verify. - Scope to what was asked. Comments are short and factual. No placeholders. Comments must not substitute for a fix. ACCESS (denylist, not allowlist). You already have full guest admin. You can create files, edit files, delete files, run commands, and fetch the network. NEVER ASK whether you may — just do it. Only refuse when a denylist or the read-only base system blocks the path. -- CREATE / EDIT: write_file creates or overwrites any writable path. create_directory makes folders. search_replace / edit_file change one hunk; apply_patch for multi-hunk or multi-file. Never say you cannot write files. Never ask the user to paste a file you can write yourself. Never ask the user to run a command you can run_command yourself. -- WRITE PATHS: /home, /tmp, /mnt, /root, /var, /opt, extra drives, ~/.agent. Prefer unique search_replace / edit_file; set replace_all only when you mean it. -- READ: any absolute path, including the entire /proc kernel surface (read_proc_file, runtime_diagnostic_bundle). Use read_file offset/limit for large files. -- RUN: every guest command via run_command (command_deny is empty by default). Prefer list_directory / glob_files / file_stat over \`ls\` / \`find\` when you only need names. list_bin lists guest /bin utilities (POSIX-in-JS, not GNU). -- DELETE / MOVE: enabled. Cannot mutate the read-only base system: /bin /etc /boot /lib /usr /share /proc /dev /sys /run. -- JS: Node is NOT installed in the guest. Never plan or run node, npm, or npx here. Author JS with run_js_script (Bare kernel, writes under ~/.agent) or run_js_script_at_path / run_command with an absolute .mjs path. Guest scripts use async function run(ctx, argv) — ctx.vfs, ctx.execLine, ctx.console, ctx.exitCode. Do not import Node builtins; use ctx.vfs and Bare ctx hooks. -- LIVE KERNEL: read_proc_file on /proc/bare_os/features (or features.json) and /proc/bare_os/capabilities.json. Man pages and apropos_man are docs only — never infer what is enabled from them. -- NET: web_fetch uses the same host allow/deny list as wget/curl. -- BRIDGE: emit_host_notification and request_host_action are enabled by default. emergency_stop_mutations is the kill switch. -- DISCORD CHANNEL: You have a Direct Message channel with the operator. Use discord_send_message to reach out whenever you want — progress, completion, blockers, scheduled results. Do not wait for them to message first. discord_channel_status / discord_read_inbox inspect the same channel. Only whitelisted users receive DMs. -- SECRETS: never print ~/.agent/config.json, API keys, seeds, or vault material. +- Create and edit with write_file, create_directory, search_replace, edit_file, and apply_patch. Never say you cannot write files. Never ask the user to paste a file you can write, or to run a command you can run_command yourself. +- Writable paths: /home, /tmp, /mnt, /root, /var, /opt, extra drives, ~/.agent. Prefer a unique search_replace / edit_file; set replace_all only when you mean it. +- Read any absolute path, including /proc (read_proc_file, runtime_diagnostic_bundle). Use read_file offset/limit for large files. +- Run every guest command via run_command (command_deny is empty by default). Prefer list_directory, glob_files, and file_stat over ls/find when you only need names. list_bin lists guest /bin utilities (POSIX-in-JS, not GNU). +- Delete and move are enabled. Do not mutate the read-only base system: /bin /etc /boot /lib /usr /share /proc /dev /sys /run. +- Node is not installed in the guest. Never plan or run node, npm, or npx here. Author JS with run_js_script (writes under ~/.agent) or run_js_script_at_path / run_command with an absolute .mjs path. Guest scripts use async function run(ctx, argv) — ctx.vfs, ctx.execLine, ctx.console, ctx.exitCode. Do not import Node builtins. +- Live kernel: read_proc_file on /proc/bare_os/features (or features.json) and /proc/bare_os/capabilities.json. Man pages and apropos_man are documentation only — never infer what is enabled from them. +- web_fetch uses the same host allow/deny list as wget/curl. +- emit_host_notification and request_host_action are enabled by default. emergency_stop_mutations is the kill switch. +- DISCORD CHANNEL: You have a Direct Message channel with the operator. Use discord_send_message to reach out whenever you want — progress, completion, blockers, scheduled results. Do not wait for them to message first. discord_channel_status and discord_read_inbox inspect the same channel. Only whitelisted users receive DMs. +- Never print ~/.agent/config.json, API keys, seeds, or vault material. - ask_user_question is only for a real product choice the user must make. Never use it (or chat) to request permission. CODING LOOP. Multi-step work: todo_write (merge=true). Large unknown surface: enter_plan_mode, write ~/.agent/plan.md, exit_plan_mode, then implement. Plan mode is read-only except that plan file. Do not stop after a plan-only reply — keep calling tools until verified. -1. Discover: glob_files, grep (preferred over search_files / run_command grep), find_symbol for definitions, list_directory (tree=true when you need a map), read_file, memory_search / memory_get, web_search then web_fetch, git_status / git_log, read_skill if a skill matches. Read before you edit. Walk-up AGENTS.md and .grok/skills from cwd are already injected when present. -2. Edit: unique search_replace / edit_file for one hunk; apply_patch for multi-hunk or multi-file work (*** Begin Patch). Create with write_file / create_directory. State blast radius (packages, contracts, generated files, docs) before wide edits. Match surrounding style. No placeholders. Comments only for non-obvious constraints. -3. Verify: re-read the file, run_command / run_js_script, git_status, read_proc_file or logs. Host git checkout: verification_hints (suggests npm/node checks; does not run them here). -4. Finish: task_complete (and update_goal completed=true on autonomous runs) with what changed, how you verified, and what is still assumed. If the same action fails three times, stop, update_goal blocked_reason if needed, and report evidence. +1. Discover with specialized read tools (glob_files, grep, find_symbol, list_directory, read_file, memory_search, web_search then web_fetch, git_status). Call read_skill when a skill matches. Read before you edit. Walk-up AGENTS.md and .grok/skills from cwd are already injected when present. +2. Edit with unique search_replace / edit_file for one hunk; apply_patch for multi-hunk or multi-file work (*** Begin Patch). Create with write_file / create_directory. State blast radius before wide edits. Match surrounding style. +3. Verify by re-reading, run_command / run_js_script, git_status, and read_proc_file or logs. On a host git checkout, verification_hints suggests npm/node checks; it does not run them here. +4. Finish with task_complete (and update_goal completed=true on autonomous runs): what changed, how you verified, what is still assumed. If the same action fails three times, stop, set update_goal blocked_reason if needed, and report evidence. TOOL DISCIPLINE. - YOU call every tool. The user will never run them. Do not print commands, patches, or "run this" for the user. +- The live registry is AVAILABLE TOOLS at the end of this prompt. Those names are the only callable functions. Skills and /bin utilities are not tools. +- Prefer specialized tools over bash: grep not run_command grep; read_file not cat; apply_patch / search_replace not sed. Never use run_command to print thoughts. - Independent reads may be issued together; the harness may serialize them (tool_parallelism defaults to 1). - Do not paste huge files into the user reply — cite paths and show only the slice that matters. - Persist durable facts in MEMORY.md; older turns may be compacted. - Progress UI is automatic (tools write ~/.agent/progress.txt). Do not narrate tool chatter in the final answer. - Autonomous mode is on by default. Keep the ReAct loop going; autonomous_deny_ops is empty unless the operator set one. -TOOL CALLING. Prefer specialized tools over bash: grep not run_command grep; read_file not cat; apply_patch / search_replace not sed. Never use run_command to print thoughts. - COMMUNICATION. Write for a reader who has not seen tool calls. Lead with the answer. Define project terms on first use. State facts literally. The final message must stand alone. Do not invent acronyms. -TOOL MAP (schemas are already attached — use them): -- Files: read_file, read_many, write_file, edit_file, search_replace, apply_patch, undo_last_edit, create_directory, list_directory (tree=true for BFS), file_stat, glob_files, fuzzy_find, grep (output_mode content|files_with_matches|count), search_files, move_path, copy_path, diff_files, delete_path, find_symbol, list_bin -- Code / harness: run_command (optional cwd), run_js_script, run_js_script_at_path, todo_write, enter_plan_mode, exit_plan_mode, memory_search, memory_get, memory_append, remember, list_skills, read_skill, create_skill, edit_agent_config, git_status, git_diff, git_log, git_show, git_blame, history_search, rewind_session, export_session, schedule_task, unschedule_task, list_scheduled, wait_for -- Kernel / ops: read_proc_file, runtime_diagnostic_bundle, get_system_info, get_resource_limits, get_swarm_peers, list_services, service_status, list_timers, read_cron_log, read_audit_log, read_boot_policy, read_kernel_extension_resolution, get_initd_graph, read_unit_journal, inspect_ipc_backpressure, get_network_summary, tail_telemetry_streams, pkg_index_lookup -- Checks: list_verification_scripts, run_maintenance_gate, run_contract_checks, summarize_build_drift, verification_hints -- Docs: read_man_page, apropos_man (documentation search only) -- Bridge / web: web_search, web_fetch, get_hrpc_bridge_health, get_hrpc_allowlist_status, emit_host_notification, request_host_action, discord_send_message, discord_channel_status, discord_read_inbox -- Autonomy: autonomous_run, autonomous_run_status, autonomous_run_stop, update_goal -- Other: ask_user_question (product choice only, never permission), task_complete - -Skills live under ~/.agent/workspace/skills/ (and ~/.agent/skills/). The prompt includes a compact index — call read_skill and follow SKILL.md when a task matches (especially bare-os-super-developer, bareos-code-change, coreutils-command-change). +Skills live under ~/.agent/workspace/skills/ and ~/.agent/skills/. They are playbooks, not tools. When a task matches (especially bare-os-super-developer, bareos-code-change, coreutils-command-change), call read_skill and follow SKILL.md. Reply format (TTY, mandatory unless Discord override below): plain text only — no Markdown markup. Structure with blank lines, short paragraphs, ALL CAPS or dashed section breaks, bare URLs. Paths and commands as normal text.` @@ -13293,23 +13385,38 @@ const BARE_AGENT_OPERATING_CONTRACT = ` --- Operating contract (Bare OS repo alignment) --- -TWO RUNTIMES: Host tooling may use Node/npm at the git checkout only. Inside this guest image there is NO node/npm/npx — use run_js_script or run_command with /bin paths only. +TWO RUNTIMES. Host tooling may use Node/npm at the git checkout only. Inside this guest image there is no node, npm, or npx — use run_js_script or run_command with /bin paths only. -DOC READING ORDER (complex tasks): Prefer developer-guide README, then handbook chapter for the area, then docs/reference for numbers and env vars. Do not copy version tables into chat; link or cite paths. +DOC READING ORDER (complex tasks). Prefer the developer-guide README, then the handbook chapter for the area, then docs/reference for numbers and env vars. Do not copy version tables into chat; link or cite paths. -CONTRACT SPINE: For any behavior change, identify: source files, canonical reference doc under docs/reference or developer-guide, generated artifact if any (run pretest generators), verifier script name from scripts/README.md, and whether compatibility-matrix or CHANGELOG moves. +CONTRACT SPINE. For any behavior change, identify the source files, the canonical reference under docs/reference or developer-guide, any generated artifact (run pretest generators), the verifier script name from scripts/README.md, and whether compatibility-matrix or CHANGELOG moves. -LIVE KERNEL STATE: Use read_proc_file on /proc/bare_os/features or features.json and /proc/bare_os/capabilities.json. Never use apropos_man or read_man_page alone to decide what is enabled at runtime — those are documentation. +LIVE KERNEL STATE. Use read_proc_file on /proc/bare_os/features or features.json and /proc/bare_os/capabilities.json. Never use apropos_man or read_man_page alone to decide what is enabled at runtime — those are documentation. -GENERATED FILES: Do not hand-edit posix-dashboard, ctx-client-helper.generated.ts, kernel-extensions-generated-toc, bundle-health outputs — run npm scripts from repo root at checkout when working on the host tree. +GENERATED FILES. Do not hand-edit posix-dashboard, ctx-client-helper.generated.ts, kernel-extensions-generated-toc, or bundle-health outputs. Run npm scripts from repo root at checkout when working on the host tree. -TASK ROUTING: /bin or coreutils -> developer-guide 06 and man JSON; shell grammar -> developer-guide 18 and shell docs; seed RPC -> developer-guide 14 and protocol package; /proc node -> developer-guide 15; ctx API -> bare-os-ctx-api.js and compatibility-matrix; docs-only -> CONTRIBUTING-DOCS; Pear issues -> PEAR-RUN and ensure-pear-node-modules story. +TASK ROUTING. /bin or coreutils: developer-guide 06 and man JSON. Shell grammar: developer-guide 18 and shell docs. Seed RPC: developer-guide 14 and the protocol package. /proc node: developer-guide 15. ctx API: bare-os-ctx-api.js and compatibility-matrix. Docs-only: CONTRIBUTING-DOCS. Pear: PEAR-RUN and the ensure-pear-node-modules story. -P2P / TEARDOWN: Hyperswarm peer wait vs offline LKG boot are different — see environment appendix. When diagnosing replication, prefer runtime_diagnostic_bundle and any /proc/bare_os file; closing order is swarm before drives when changing booter lifecycle code. +P2P / TEARDOWN. Hyperswarm peer wait and offline LKG boot are different. When diagnosing replication, prefer runtime_diagnostic_bundle and any /proc/bare_os file. Closing order is swarm before drives when changing booter lifecycle code. -HOST CHECKS: Use verification_hints for suggested repo-root npm checks when the user describes changed paths (host checkout). It does not execute npm inside the guest. Use runtime_diagnostic_bundle for one-shot live /proc and resource snapshot. Use read_skill for workflow skills under workspace/skills/. +HOST CHECKS. Use verification_hints for suggested repo-root npm checks when the user describes changed paths (host checkout). It does not execute npm inside the guest. Use runtime_diagnostic_bundle for a one-shot live /proc and resource snapshot. Use read_skill for workflow skills under workspace/skills/. -SELF-REVIEW before task_complete: docs updated? generated regen? wrong runtime assumption (guest vs host)? same failing action tried fewer than three times?` +SELF-REVIEW before task_complete. Docs updated? Generated files regenerated? Wrong runtime assumption (guest vs host)? Same failing action tried fewer than three times?` + +/** + * @param {string} text + */ +function bareAgentLooksLikeToolInventoryQuestion(text) { + const s = String(text || '').trim().toLowerCase() + if (!s) return false + return ( + /\b(your tools|what tools|which tools|list tools|available tools|tool list|what can you do)\b/.test( + s + ) || + /^tools\??$/.test(s) || + /tell me about your tools/.test(s) + ) +} /** * Session-specific HOME / tilde context (injected every run so the model uses real paths). @@ -13324,21 +13431,21 @@ function bareAgentSessionHomeBlock(ctx, home, paths) { : {} const homeEnv = String(env.HOME || '').trim() return ( - '## This session: home directory and paths\n' + - '- **Resolved user home (this session):** `' + + 'This session — home directory and paths\n' + + '- Resolved user home: ' + home + - '`\n' + - '- **HOME in the environment:** `' + + '\n' + + '- HOME in the environment: ' + (homeEnv || home) + - '`\n' + - '- **Tilde \`~\`:** In shell and in user docs, \`~\` means this home directory. Examples: \`~/.agent\` == `' + + '\n' + + '- Tilde ~ means this home directory. ~/.agent is ' + paths.dir + - '`, agent config `' + + '; agent config is ' + paths.config + - '`. Always expand \`~\` to `' + + '. Always expand ~ to ' + home + - '\` when constructing absolute paths for tools.\n' + - '- **Reminder:** \`node\` is unavailable; use **run_js_script** for JS you author in this agent session.\n' + ' when constructing absolute paths for tools.\n' + + '- node is unavailable; use run_js_script for JavaScript you author in this session.\n' ) } @@ -14522,13 +14629,13 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { manDigest.slice(0, manBudget) if (instructions) systemContent += - '\n\n## Session notes\n' + instructions.slice(0, instructionsBudget) + '\n\nSession notes\n' + instructions.slice(0, instructionsBudget) if (typeof bareAgentRunHooks === 'function' && paths.hooks) { try { const startHook = await bareAgentRunHooks(ctx, paths.hooks, 'SessionStart', '', {}, '') if (startHook && startHook.inject) { systemContent += - '\n\n## SessionStart hook\n' + String(startHook.inject).slice(0, 1500) + '\n\nSessionStart hook\n' + String(startHook.inject).slice(0, 1500) } } catch { /* optional */ @@ -14541,7 +14648,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { const extra = await bareAgentReadTextFile(ctx, extraPath) if (extra && extra.trim()) { systemContent += - '\n\n## Extra instructions (--system)\n' + + '\n\nExtra instructions (--system)\n' + extra.trim().slice(0, instructionsBudget) } } @@ -14579,7 +14686,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { if (config.plan_mode_active) { systemContent += - '\n\n## PLAN MODE is ON\n' + + '\n\nPLAN MODE is ON\n' + 'Read-only tools only. The only allowed write is ' + String(paths.plan || home + '/.agent/plan.md') + '. Draft the plan there, then call exit_plan_mode before implementing.\n' @@ -14591,7 +14698,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { const sum = bareAgentTodoSummarize(todos) if (sum.total) { systemContent += - '\n\n## Session todos\n' + + '\n\nSession todos\n' + sum.text + '\n(open=' + String(sum.open) + @@ -14604,7 +14711,22 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { } } + // Catalog last so small local models see callable tools after skills/notes. + if (typeof bareAgentToolCatalogPrompt === 'function') { + systemContent += '\n\n' + bareAgentToolCatalogPrompt() + } + let userTask = String(task || '') + if ( + typeof bareAgentLooksLikeToolInventoryQuestion === 'function' + ? bareAgentLooksLikeToolInventoryQuestion(userTask) + : /\b(your tools|what tools|which tools|list tools|available tools|tool list|what can you do)\b/i.test( + userTask + ) + ) { + userTask += + '\n\n[harness] Answer from AVAILABLE TOOLS / list_agent_tools. List every callable tool name. Do not list skills or /bin names as tools.' + } if (config.autonomous_active) { userTask = 'AUTONOMOUS RUN. Execute until the goal is done. Do not stop after a plan — use tools, then call task_complete.\n' + @@ -14683,6 +14805,13 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { bareAgentNormalizeBaseUrl(String(configRef.current.rest_base_url || '')) + '/chat/completions' const tools = bareAgentToolDefinitions() + appendProgress( + 'tools_attached n=' + + String(tools.length) + + (typeof bareAgentToolCatalog === 'function' + ? ' catalog=' + String(bareAgentToolCatalog(tools).length) + : '') + ) let reasoningSettings = bareAgentReasoningSettings(configRef.current) let autonomousSettings = bareAgentAutonomousSettings(configRef.current) let reasoningCharCount = 0 diff --git a/packages/bare-os-coreutils/lib/agent/agent-skills.js b/packages/bare-os-coreutils/lib/agent/agent-skills.js index d2916be8..89483215 100644 --- a/packages/bare-os-coreutils/lib/agent/agent-skills.js +++ b/packages/bare-os-coreutils/lib/agent/agent-skills.js @@ -131,27 +131,19 @@ async function bareAgentSkillsCompactPrompt(ctx, paths, maxChars) { const cap = Math.min(Math.max(Number(maxChars) || 4000, 500), 12000) const skills = await bareAgentDiscoverSkills(ctx, paths) let block = - '## Available skills (compact index)\n' + - 'Each skill is a directory with **SKILL.md** (optional YAML frontmatter: `name`, `description`, …).\n' + - '**Workspace skills** (`~/.agent/workspace/skills/`) override **global** (`~/.agent/skills/`) and walk-up `.grok/skills` / `.agents/skills` when names match.\n' + - 'To run one: call the **read_skill** tool with the skill id or frontmatter `name` before following its instructions.\n\n' + 'Skills (playbooks — NOT callable tools)\n' + + 'These ids are SKILL.md playbooks, not functions. Do not list them when asked about your tools.\n' + + 'Each skill is a directory with SKILL.md (optional YAML frontmatter: name, description).\n' + + 'Workspace skills (~/.agent/workspace/skills/) override global (~/.agent/skills/) and walk-up .grok/skills / .agents/skills when names match.\n' + + 'To use one, call the read_skill tool with the skill id or frontmatter name, then follow its instructions.\n\n' if (!skills.length) { - block += '(No skills discovered yet — add folders under `workspace/skills//SKILL.md`.)\n' + block += '(No skills discovered yet — add folders under workspace/skills//SKILL.md.)\n' return block.length > cap ? block.slice(0, cap) + '\n…\n' : block } - block += '| id | name | source | description |\n| --- | --- | --- | --- |\n' for (const s of skills) { - const desc = s.description.replace(/\|/g, '/').replace(/\r?\n/g, ' ').slice(0, 160) - block += - '| `' + - s.id.replace(/`/g, "'") + - '` | ' + - s.name.replace(/\|/g, '/').replace(/\r?\n/g, ' ') + - ' | ' + - s.source + - ' | ' + - desc + - ' |\n' + const desc = s.description.replace(/\r?\n/g, ' ').slice(0, 160) + const label = s.name && s.name !== s.id ? s.id + ' / ' + s.name : s.id + block += '- ' + label + ' (' + s.source + ')' + (desc ? ' — ' + desc : '') + '\n' } if (block.length > cap) block = block.slice(0, cap) + '\n… truncated\n' return block diff --git a/packages/bare-os-coreutils/lib/agent/agent-tool-definitions.js b/packages/bare-os-coreutils/lib/agent/agent-tool-definitions.js index d3a13392..17bdb0da 100644 --- a/packages/bare-os-coreutils/lib/agent/agent-tool-definitions.js +++ b/packages/bare-os-coreutils/lib/agent/agent-tool-definitions.js @@ -1221,6 +1221,23 @@ function bareAgentToolDefinitions() { } } }, + { + type: 'function', + function: { + name: 'list_agent_tools', + description: + 'Return the live callable tool registry (name, description, parameter keys). Use this when the user asks what tools you have. Skills and /bin names are not in this list.', + parameters: { + type: 'object', + properties: { + name: { + type: 'string', + description: 'Optional tool name or substring to describe one entry' + } + } + } + } + }, { type: 'function', function: { @@ -1536,3 +1553,75 @@ function bareAgentToolDefinitions() { } ] } + +/** + * Live tool registry derived from bareAgentToolDefinitions (never hand-maintained). + * @param {unknown[]} [tools] + * @returns {{ name: string, description: string, params: string[] }[]} + */ +function bareAgentToolCatalog(tools) { + const src = + Array.isArray(tools) && tools.length ? tools : bareAgentToolDefinitions() + /** @type {{ name: string, description: string, params: string[] }[]} */ + const out = [] + /** @type {Record} */ + const seen = {} + for (const t of src) { + if (!t || typeof t !== 'object') continue + const o = /** @type {Record} */ (t) + const fn = + o.function && typeof o.function === 'object' + ? /** @type {Record} */ (o.function) + : o + const name = typeof fn.name === 'string' ? fn.name.trim() : '' + if (!name || seen[name]) continue + seen[name] = 1 + const description = + typeof fn.description === 'string' + ? fn.description.replace(/\s+/g, ' ').trim() + : '' + const paramsRoot = + fn.parameters && typeof fn.parameters === 'object' + ? /** @type {Record} */ (fn.parameters) + : {} + const props = + paramsRoot.properties && + typeof paramsRoot.properties === 'object' && + !Array.isArray(paramsRoot.properties) + ? /** @type {Record} */ (paramsRoot.properties) + : {} + out.push({ name, description, params: Object.keys(props) }) + } + return out +} + +/** + * System-prompt block. Placed last so small local models see the real tool + * names instead of the skills index. + * @param {unknown[]} [tools] + */ +function bareAgentToolCatalogPrompt(tools) { + const cat = bareAgentToolCatalog(tools) + const lines = [ + 'AVAILABLE TOOLS (' + + cat.length + + ' callable functions — live registry)', + 'These names are the only functions you can call. Skills such as ctx-api-change, docs-contract-update, hdms, and holesail are playbooks: load them with read_skill. Guest /bin names such as ctx-baredoctor, hdms, and holepunch are commands for run_command, not tools.', + 'When the user asks what tools you have or what you can do, list every name below. Do not invent names. Do not list skills as tools.', + 'To call a tool, emit exactly:', + '{"name":"TOOL_NAME","arguments":{}}', + 'You may call list_agent_tools to reprint this catalog.', + '' + ] + for (const t of cat) { + const desc = + t.description.length > 110 + ? t.description.slice(0, 107) + '...' + : t.description + const params = t.params.length ? ' (' + t.params.join(', ') + ')' : '' + lines.push( + '- ' + t.name + params + (desc ? ' — ' + desc : '') + ) + } + return lines.join('\n') +} diff --git a/packages/bare-os-coreutils/lib/agent/agent-tool-dispatch.js b/packages/bare-os-coreutils/lib/agent/agent-tool-dispatch.js index 07273560..3149b7e2 100644 --- a/packages/bare-os-coreutils/lib/agent/agent-tool-dispatch.js +++ b/packages/bare-os-coreutils/lib/agent/agent-tool-dispatch.js @@ -2426,6 +2426,27 @@ async function bareAgentDispatchTool(o) { return bareAgentJsonResult({ ok: true, path: dest, kind, appended: line }) } + if (toolName === 'list_agent_tools') { + const cat = + typeof bareAgentToolCatalog === 'function' ? bareAgentToolCatalog() : [] + const want = typeof args.name === 'string' ? args.name.trim() : '' + const rows = want + ? cat.filter(function (t) { + return ( + t.name === want || + String(t.name).toLowerCase().indexOf(want.toLowerCase()) !== -1 + ) + }) + : cat + appendProgress('list_agent_tools n=' + String(rows.length)) + return bareAgentJsonResult({ + ok: true, + count: rows.length, + total: cat.length, + tools: rows + }) + } + if (toolName === 'list_skills') { const max = typeof args.max === 'number' && Number.isFinite(args.max) diff --git a/packages/bare-os-coreutils/lib/agent/agent-tui.js b/packages/bare-os-coreutils/lib/agent/agent-tui.js index 920709f5..5d2fa96d 100644 --- a/packages/bare-os-coreutils/lib/agent/agent-tui.js +++ b/packages/bare-os-coreutils/lib/agent/agent-tui.js @@ -245,7 +245,7 @@ async function bareAgentPromptSetupLine(ctx, prompt, opts) { const BARE_AGENT_STATIC_SYSTEM = `You are the Bare OS coding agent — a senior implementer that lives inside the guest image (JavaScript POSIX on Hyperdrive + Hyperswarm, Pear/Bare runtime). You write, edit, debug, and verify code and OS state by calling tools. You are not a chatbot that narrates plans and waits. Lead with tools. Execute until the job is done, then call task_complete. -YOU RUN THE TOOLS. The user never runs your tools, commands, patches, or scripts. They will not copy-paste what you print into a shell, Discord, or another agent. If the work needs a tool, YOU call it (run_command, write_file, apply_patch, read_file, …). Never reply with "run this", "please execute", "you can run", or a command dump for the user to do. The user is not your tool runner. +YOU RUN THE TOOLS. The user never runs your tools, commands, patches, or scripts. They will not copy-paste what you print into a shell, Discord, or another agent. If the work needs a tool, YOU call it (run_command, write_file, apply_patch, read_file, and the rest of AVAILABLE TOOLS). Never reply with "run this", "please execute", "you can run", or a command dump for the user to do. The user is not your tool runner. Bare OS by Raven Scott (https://raven-scott.fyi). Repo: https://git.ssh.surf/snxraven/bare-operating-system. Booter: pear://qupw8zspk34pcxc7fqchzyeh33jtmxq1k7qze44fkosctwiid8zy @@ -253,52 +253,42 @@ WORK POLICY. - Keep every explicit requirement in view until it is done, superseded, or blocked. If blocked, say so plainly. - Match intent: implement action requests; do not make unsolicited project-wide edits when the user asked a question. - For clear, reversible local work, do it now. NEVER ASK for permission. -- Claim done / fixed / tested only when a tool result supports it. Otherwise say what you did not verify. +- Claim done, fixed, or tested only when a tool result supports it. Otherwise say what you did not verify. - Scope to what was asked. Comments are short and factual. No placeholders. Comments must not substitute for a fix. ACCESS (denylist, not allowlist). You already have full guest admin. You can create files, edit files, delete files, run commands, and fetch the network. NEVER ASK whether you may — just do it. Only refuse when a denylist or the read-only base system blocks the path. -- CREATE / EDIT: write_file creates or overwrites any writable path. create_directory makes folders. search_replace / edit_file change one hunk; apply_patch for multi-hunk or multi-file. Never say you cannot write files. Never ask the user to paste a file you can write yourself. Never ask the user to run a command you can run_command yourself. -- WRITE PATHS: /home, /tmp, /mnt, /root, /var, /opt, extra drives, ~/.agent. Prefer unique search_replace / edit_file; set replace_all only when you mean it. -- READ: any absolute path, including the entire /proc kernel surface (read_proc_file, runtime_diagnostic_bundle). Use read_file offset/limit for large files. -- RUN: every guest command via run_command (command_deny is empty by default). Prefer list_directory / glob_files / file_stat over \`ls\` / \`find\` when you only need names. list_bin lists guest /bin utilities (POSIX-in-JS, not GNU). -- DELETE / MOVE: enabled. Cannot mutate the read-only base system: /bin /etc /boot /lib /usr /share /proc /dev /sys /run. -- JS: Node is NOT installed in the guest. Never plan or run node, npm, or npx here. Author JS with run_js_script (Bare kernel, writes under ~/.agent) or run_js_script_at_path / run_command with an absolute .mjs path. Guest scripts use async function run(ctx, argv) — ctx.vfs, ctx.execLine, ctx.console, ctx.exitCode. Do not import Node builtins; use ctx.vfs and Bare ctx hooks. -- LIVE KERNEL: read_proc_file on /proc/bare_os/features (or features.json) and /proc/bare_os/capabilities.json. Man pages and apropos_man are docs only — never infer what is enabled from them. -- NET: web_fetch uses the same host allow/deny list as wget/curl. -- BRIDGE: emit_host_notification and request_host_action are enabled by default. emergency_stop_mutations is the kill switch. -- DISCORD CHANNEL: You have a Direct Message channel with the operator. Use discord_send_message to reach out whenever you want — progress, completion, blockers, scheduled results. Do not wait for them to message first. discord_channel_status / discord_read_inbox inspect the same channel. Only whitelisted users receive DMs. -- SECRETS: never print ~/.agent/config.json, API keys, seeds, or vault material. +- Create and edit with write_file, create_directory, search_replace, edit_file, and apply_patch. Never say you cannot write files. Never ask the user to paste a file you can write, or to run a command you can run_command yourself. +- Writable paths: /home, /tmp, /mnt, /root, /var, /opt, extra drives, ~/.agent. Prefer a unique search_replace / edit_file; set replace_all only when you mean it. +- Read any absolute path, including /proc (read_proc_file, runtime_diagnostic_bundle). Use read_file offset/limit for large files. +- Run every guest command via run_command (command_deny is empty by default). Prefer list_directory, glob_files, and file_stat over ls/find when you only need names. list_bin lists guest /bin utilities (POSIX-in-JS, not GNU). +- Delete and move are enabled. Do not mutate the read-only base system: /bin /etc /boot /lib /usr /share /proc /dev /sys /run. +- Node is not installed in the guest. Never plan or run node, npm, or npx here. Author JS with run_js_script (writes under ~/.agent) or run_js_script_at_path / run_command with an absolute .mjs path. Guest scripts use async function run(ctx, argv) — ctx.vfs, ctx.execLine, ctx.console, ctx.exitCode. Do not import Node builtins. +- Live kernel: read_proc_file on /proc/bare_os/features (or features.json) and /proc/bare_os/capabilities.json. Man pages and apropos_man are documentation only — never infer what is enabled from them. +- web_fetch uses the same host allow/deny list as wget/curl. +- emit_host_notification and request_host_action are enabled by default. emergency_stop_mutations is the kill switch. +- DISCORD CHANNEL: You have a Direct Message channel with the operator. Use discord_send_message to reach out whenever you want — progress, completion, blockers, scheduled results. Do not wait for them to message first. discord_channel_status and discord_read_inbox inspect the same channel. Only whitelisted users receive DMs. +- Never print ~/.agent/config.json, API keys, seeds, or vault material. - ask_user_question is only for a real product choice the user must make. Never use it (or chat) to request permission. CODING LOOP. Multi-step work: todo_write (merge=true). Large unknown surface: enter_plan_mode, write ~/.agent/plan.md, exit_plan_mode, then implement. Plan mode is read-only except that plan file. Do not stop after a plan-only reply — keep calling tools until verified. -1. Discover: glob_files, grep (preferred over search_files / run_command grep), find_symbol for definitions, list_directory (tree=true when you need a map), read_file, memory_search / memory_get, web_search then web_fetch, git_status / git_log, read_skill if a skill matches. Read before you edit. Walk-up AGENTS.md and .grok/skills from cwd are already injected when present. -2. Edit: unique search_replace / edit_file for one hunk; apply_patch for multi-hunk or multi-file work (*** Begin Patch). Create with write_file / create_directory. State blast radius (packages, contracts, generated files, docs) before wide edits. Match surrounding style. No placeholders. Comments only for non-obvious constraints. -3. Verify: re-read the file, run_command / run_js_script, git_status, read_proc_file or logs. Host git checkout: verification_hints (suggests npm/node checks; does not run them here). -4. Finish: task_complete (and update_goal completed=true on autonomous runs) with what changed, how you verified, and what is still assumed. If the same action fails three times, stop, update_goal blocked_reason if needed, and report evidence. +1. Discover with specialized read tools (glob_files, grep, find_symbol, list_directory, read_file, memory_search, web_search then web_fetch, git_status). Call read_skill when a skill matches. Read before you edit. Walk-up AGENTS.md and .grok/skills from cwd are already injected when present. +2. Edit with unique search_replace / edit_file for one hunk; apply_patch for multi-hunk or multi-file work (*** Begin Patch). Create with write_file / create_directory. State blast radius before wide edits. Match surrounding style. +3. Verify by re-reading, run_command / run_js_script, git_status, and read_proc_file or logs. On a host git checkout, verification_hints suggests npm/node checks; it does not run them here. +4. Finish with task_complete (and update_goal completed=true on autonomous runs): what changed, how you verified, what is still assumed. If the same action fails three times, stop, set update_goal blocked_reason if needed, and report evidence. TOOL DISCIPLINE. - YOU call every tool. The user will never run them. Do not print commands, patches, or "run this" for the user. +- The live registry is AVAILABLE TOOLS at the end of this prompt. Those names are the only callable functions. Skills and /bin utilities are not tools. +- Prefer specialized tools over bash: grep not run_command grep; read_file not cat; apply_patch / search_replace not sed. Never use run_command to print thoughts. - Independent reads may be issued together; the harness may serialize them (tool_parallelism defaults to 1). - Do not paste huge files into the user reply — cite paths and show only the slice that matters. - Persist durable facts in MEMORY.md; older turns may be compacted. - Progress UI is automatic (tools write ~/.agent/progress.txt). Do not narrate tool chatter in the final answer. - Autonomous mode is on by default. Keep the ReAct loop going; autonomous_deny_ops is empty unless the operator set one. -TOOL CALLING. Prefer specialized tools over bash: grep not run_command grep; read_file not cat; apply_patch / search_replace not sed. Never use run_command to print thoughts. - COMMUNICATION. Write for a reader who has not seen tool calls. Lead with the answer. Define project terms on first use. State facts literally. The final message must stand alone. Do not invent acronyms. -TOOL MAP (schemas are already attached — use them): -- Files: read_file, read_many, write_file, edit_file, search_replace, apply_patch, undo_last_edit, create_directory, list_directory (tree=true for BFS), file_stat, glob_files, fuzzy_find, grep (output_mode content|files_with_matches|count), search_files, move_path, copy_path, diff_files, delete_path, find_symbol, list_bin -- Code / harness: run_command (optional cwd), run_js_script, run_js_script_at_path, todo_write, enter_plan_mode, exit_plan_mode, memory_search, memory_get, memory_append, remember, list_skills, read_skill, create_skill, edit_agent_config, git_status, git_diff, git_log, git_show, git_blame, history_search, rewind_session, export_session, schedule_task, unschedule_task, list_scheduled, wait_for -- Kernel / ops: read_proc_file, runtime_diagnostic_bundle, get_system_info, get_resource_limits, get_swarm_peers, list_services, service_status, list_timers, read_cron_log, read_audit_log, read_boot_policy, read_kernel_extension_resolution, get_initd_graph, read_unit_journal, inspect_ipc_backpressure, get_network_summary, tail_telemetry_streams, pkg_index_lookup -- Checks: list_verification_scripts, run_maintenance_gate, run_contract_checks, summarize_build_drift, verification_hints -- Docs: read_man_page, apropos_man (documentation search only) -- Bridge / web: web_search, web_fetch, get_hrpc_bridge_health, get_hrpc_allowlist_status, emit_host_notification, request_host_action, discord_send_message, discord_channel_status, discord_read_inbox -- Autonomy: autonomous_run, autonomous_run_status, autonomous_run_stop, update_goal -- Other: ask_user_question (product choice only, never permission), task_complete - -Skills live under ~/.agent/workspace/skills/ (and ~/.agent/skills/). The prompt includes a compact index — call read_skill and follow SKILL.md when a task matches (especially bare-os-super-developer, bareos-code-change, coreutils-command-change). +Skills live under ~/.agent/workspace/skills/ and ~/.agent/skills/. They are playbooks, not tools. When a task matches (especially bare-os-super-developer, bareos-code-change, coreutils-command-change), call read_skill and follow SKILL.md. Reply format (TTY, mandatory unless Discord override below): plain text only — no Markdown markup. Structure with blank lines, short paragraphs, ALL CAPS or dashed section breaks, bare URLs. Paths and commands as normal text.` @@ -321,23 +311,38 @@ const BARE_AGENT_OPERATING_CONTRACT = ` --- Operating contract (Bare OS repo alignment) --- -TWO RUNTIMES: Host tooling may use Node/npm at the git checkout only. Inside this guest image there is NO node/npm/npx — use run_js_script or run_command with /bin paths only. +TWO RUNTIMES. Host tooling may use Node/npm at the git checkout only. Inside this guest image there is no node, npm, or npx — use run_js_script or run_command with /bin paths only. -DOC READING ORDER (complex tasks): Prefer developer-guide README, then handbook chapter for the area, then docs/reference for numbers and env vars. Do not copy version tables into chat; link or cite paths. +DOC READING ORDER (complex tasks). Prefer the developer-guide README, then the handbook chapter for the area, then docs/reference for numbers and env vars. Do not copy version tables into chat; link or cite paths. -CONTRACT SPINE: For any behavior change, identify: source files, canonical reference doc under docs/reference or developer-guide, generated artifact if any (run pretest generators), verifier script name from scripts/README.md, and whether compatibility-matrix or CHANGELOG moves. +CONTRACT SPINE. For any behavior change, identify the source files, the canonical reference under docs/reference or developer-guide, any generated artifact (run pretest generators), the verifier script name from scripts/README.md, and whether compatibility-matrix or CHANGELOG moves. -LIVE KERNEL STATE: Use read_proc_file on /proc/bare_os/features or features.json and /proc/bare_os/capabilities.json. Never use apropos_man or read_man_page alone to decide what is enabled at runtime — those are documentation. +LIVE KERNEL STATE. Use read_proc_file on /proc/bare_os/features or features.json and /proc/bare_os/capabilities.json. Never use apropos_man or read_man_page alone to decide what is enabled at runtime — those are documentation. -GENERATED FILES: Do not hand-edit posix-dashboard, ctx-client-helper.generated.ts, kernel-extensions-generated-toc, bundle-health outputs — run npm scripts from repo root at checkout when working on the host tree. +GENERATED FILES. Do not hand-edit posix-dashboard, ctx-client-helper.generated.ts, kernel-extensions-generated-toc, or bundle-health outputs. Run npm scripts from repo root at checkout when working on the host tree. -TASK ROUTING: /bin or coreutils -> developer-guide 06 and man JSON; shell grammar -> developer-guide 18 and shell docs; seed RPC -> developer-guide 14 and protocol package; /proc node -> developer-guide 15; ctx API -> bare-os-ctx-api.js and compatibility-matrix; docs-only -> CONTRIBUTING-DOCS; Pear issues -> PEAR-RUN and ensure-pear-node-modules story. +TASK ROUTING. /bin or coreutils: developer-guide 06 and man JSON. Shell grammar: developer-guide 18 and shell docs. Seed RPC: developer-guide 14 and the protocol package. /proc node: developer-guide 15. ctx API: bare-os-ctx-api.js and compatibility-matrix. Docs-only: CONTRIBUTING-DOCS. Pear: PEAR-RUN and the ensure-pear-node-modules story. -P2P / TEARDOWN: Hyperswarm peer wait vs offline LKG boot are different — see environment appendix. When diagnosing replication, prefer runtime_diagnostic_bundle and any /proc/bare_os file; closing order is swarm before drives when changing booter lifecycle code. +P2P / TEARDOWN. Hyperswarm peer wait and offline LKG boot are different. When diagnosing replication, prefer runtime_diagnostic_bundle and any /proc/bare_os file. Closing order is swarm before drives when changing booter lifecycle code. -HOST CHECKS: Use verification_hints for suggested repo-root npm checks when the user describes changed paths (host checkout). It does not execute npm inside the guest. Use runtime_diagnostic_bundle for one-shot live /proc and resource snapshot. Use read_skill for workflow skills under workspace/skills/. +HOST CHECKS. Use verification_hints for suggested repo-root npm checks when the user describes changed paths (host checkout). It does not execute npm inside the guest. Use runtime_diagnostic_bundle for a one-shot live /proc and resource snapshot. Use read_skill for workflow skills under workspace/skills/. -SELF-REVIEW before task_complete: docs updated? generated regen? wrong runtime assumption (guest vs host)? same failing action tried fewer than three times?` +SELF-REVIEW before task_complete. Docs updated? Generated files regenerated? Wrong runtime assumption (guest vs host)? Same failing action tried fewer than three times?` + +/** + * @param {string} text + */ +function bareAgentLooksLikeToolInventoryQuestion(text) { + const s = String(text || '').trim().toLowerCase() + if (!s) return false + return ( + /\b(your tools|what tools|which tools|list tools|available tools|tool list|what can you do)\b/.test( + s + ) || + /^tools\??$/.test(s) || + /tell me about your tools/.test(s) + ) +} /** * Session-specific HOME / tilde context (injected every run so the model uses real paths). @@ -352,21 +357,21 @@ function bareAgentSessionHomeBlock(ctx, home, paths) { : {} const homeEnv = String(env.HOME || '').trim() return ( - '## This session: home directory and paths\n' + - '- **Resolved user home (this session):** `' + + 'This session — home directory and paths\n' + + '- Resolved user home: ' + home + - '`\n' + - '- **HOME in the environment:** `' + + '\n' + + '- HOME in the environment: ' + (homeEnv || home) + - '`\n' + - '- **Tilde \`~\`:** In shell and in user docs, \`~\` means this home directory. Examples: \`~/.agent\` == `' + + '\n' + + '- Tilde ~ means this home directory. ~/.agent is ' + paths.dir + - '`, agent config `' + + '; agent config is ' + paths.config + - '`. Always expand \`~\` to `' + + '. Always expand ~ to ' + home + - '\` when constructing absolute paths for tools.\n' + - '- **Reminder:** \`node\` is unavailable; use **run_js_script** for JS you author in this agent session.\n' + ' when constructing absolute paths for tools.\n' + + '- node is unavailable; use run_js_script for JavaScript you author in this session.\n' ) } @@ -1550,13 +1555,13 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { manDigest.slice(0, manBudget) if (instructions) systemContent += - '\n\n## Session notes\n' + instructions.slice(0, instructionsBudget) + '\n\nSession notes\n' + instructions.slice(0, instructionsBudget) if (typeof bareAgentRunHooks === 'function' && paths.hooks) { try { const startHook = await bareAgentRunHooks(ctx, paths.hooks, 'SessionStart', '', {}, '') if (startHook && startHook.inject) { systemContent += - '\n\n## SessionStart hook\n' + String(startHook.inject).slice(0, 1500) + '\n\nSessionStart hook\n' + String(startHook.inject).slice(0, 1500) } } catch { /* optional */ @@ -1569,7 +1574,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { const extra = await bareAgentReadTextFile(ctx, extraPath) if (extra && extra.trim()) { systemContent += - '\n\n## Extra instructions (--system)\n' + + '\n\nExtra instructions (--system)\n' + extra.trim().slice(0, instructionsBudget) } } @@ -1607,7 +1612,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { if (config.plan_mode_active) { systemContent += - '\n\n## PLAN MODE is ON\n' + + '\n\nPLAN MODE is ON\n' + 'Read-only tools only. The only allowed write is ' + String(paths.plan || home + '/.agent/plan.md') + '. Draft the plan there, then call exit_plan_mode before implementing.\n' @@ -1619,7 +1624,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { const sum = bareAgentTodoSummarize(todos) if (sum.total) { systemContent += - '\n\n## Session todos\n' + + '\n\nSession todos\n' + sum.text + '\n(open=' + String(sum.open) + @@ -1632,7 +1637,22 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { } } + // Catalog last so small local models see callable tools after skills/notes. + if (typeof bareAgentToolCatalogPrompt === 'function') { + systemContent += '\n\n' + bareAgentToolCatalogPrompt() + } + let userTask = String(task || '') + if ( + typeof bareAgentLooksLikeToolInventoryQuestion === 'function' + ? bareAgentLooksLikeToolInventoryQuestion(userTask) + : /\b(your tools|what tools|which tools|list tools|available tools|tool list|what can you do)\b/i.test( + userTask + ) + ) { + userTask += + '\n\n[harness] Answer from AVAILABLE TOOLS / list_agent_tools. List every callable tool name. Do not list skills or /bin names as tools.' + } if (config.autonomous_active) { userTask = 'AUTONOMOUS RUN. Execute until the goal is done. Do not stop after a plan — use tools, then call task_complete.\n' + @@ -1711,6 +1731,13 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { bareAgentNormalizeBaseUrl(String(configRef.current.rest_base_url || '')) + '/chat/completions' const tools = bareAgentToolDefinitions() + appendProgress( + 'tools_attached n=' + + String(tools.length) + + (typeof bareAgentToolCatalog === 'function' + ? ' catalog=' + String(bareAgentToolCatalog(tools).length) + : '') + ) let reasoningSettings = bareAgentReasoningSettings(configRef.current) let autonomousSettings = bareAgentAutonomousSettings(configRef.current) let reasoningCharCount = 0 diff --git a/packages/bare-os-coreutils/test/agent-config-surface.test.mjs b/packages/bare-os-coreutils/test/agent-config-surface.test.mjs index bbe03846..06c99443 100644 --- a/packages/bare-os-coreutils/test/agent-config-surface.test.mjs +++ b/packages/bare-os-coreutils/test/agent-config-surface.test.mjs @@ -157,6 +157,7 @@ test('agent-tools exposes agent ops tools', async (t) => { 'git_status', 'memory_append', 'list_skills', + 'list_agent_tools', 'schedule_task', 'unschedule_task', 'list_scheduled', diff --git a/packages/bare-os-coreutils/test/agent-qvac.test.mjs b/packages/bare-os-coreutils/test/agent-qvac.test.mjs index 46fdacb8..613202f5 100644 --- a/packages/bare-os-coreutils/test/agent-qvac.test.mjs +++ b/packages/bare-os-coreutils/test/agent-qvac.test.mjs @@ -33,6 +33,7 @@ test('full agent tool list flattens to QVAC-safe property schemas', async (t) => const { bareAgentFlattenToolsForQvac, bareAgentToolDefinitions } = sandbox.__exports const flat = bareAgentFlattenToolsForQvac(bareAgentToolDefinitions()) t.ok(flat.length > 20) + t.ok(flat.some((x) => x.name === 'list_agent_tools')) const allowed = { type: 1, description: 1, enum: 1 } for (const tool of flat) { t.ok(tool.name, 'tool missing name') @@ -48,6 +49,62 @@ test('full agent tool list flattens to QVAC-safe property schemas', async (t) => } }) +test('live tool catalog lists every defined tool and not skills', async (t) => { + const defs = readFileSync( + new URL('../lib/agent/agent-tool-definitions.js', import.meta.url), + 'utf8' + ) + const skills = readFileSync( + new URL('../lib/agent/agent-skills.js', import.meta.url), + 'utf8' + ) + const tui = readFileSync( + new URL('../lib/agent/agent-tui.js', import.meta.url), + 'utf8' + ) + const dispatch = readFileSync( + new URL('../lib/agent/agent-tool-dispatch.js', import.meta.url), + 'utf8' + ) + const sandbox = { console } + vm.createContext(sandbox) + vm.runInContext( + defs + + '\n;this.__exports = { bareAgentToolDefinitions, bareAgentToolCatalog, bareAgentToolCatalogPrompt }', + sandbox + ) + const { + bareAgentToolDefinitions, + bareAgentToolCatalog, + bareAgentToolCatalogPrompt + } = sandbox.__exports + const tools = bareAgentToolDefinitions() + const cat = bareAgentToolCatalog(tools) + const prompt = bareAgentToolCatalogPrompt(tools) + t.ok(cat.length >= 80) + t.is(cat.length, tools.length) + const names = cat.map((x) => x.name) + t.ok(names.includes('read_file')) + t.ok(names.includes('run_command')) + t.ok(names.includes('write_file')) + t.ok(names.includes('list_agent_tools')) + t.ok(names.includes('discord_send_message')) + t.ok(names.includes('task_complete')) + t.absent(names.includes('ctx-api-change')) + t.absent(names.includes('docs-contract-update')) + t.absent(names.includes('ctx-baredoctor')) + for (const n of names) { + t.ok(prompt.includes(n), 'catalog prompt missing ' + n) + t.ok( + dispatch.includes("toolName === '" + n + "'") || n === 'remember', + 'dispatch missing ' + n + ) + } + t.ok(skills.includes('NOT callable tools')) + t.ok(tui.includes('bareAgentToolCatalogPrompt()')) + t.ok(tui.includes('bareAgentLooksLikeToolInventoryQuestion')) +}) + test('flatten OpenAI nested tools for QVAC', async (t) => { const { bareAgentFlattenToolsForQvac } = loadAgentQvacHelpers() const flat = bareAgentFlattenToolsForQvac([ diff --git a/packages/bare-os-seeder/kernel/bin/agent b/packages/bare-os-seeder/kernel/bin/agent index fd6a3d8b..3052d975 100644 --- a/packages/bare-os-seeder/kernel/bin/agent +++ b/packages/bare-os-seeder/kernel/bin/agent @@ -3442,27 +3442,19 @@ async function bareAgentSkillsCompactPrompt(ctx, paths, maxChars) { const cap = Math.min(Math.max(Number(maxChars) || 4000, 500), 12000) const skills = await bareAgentDiscoverSkills(ctx, paths) let block = - '## Available skills (compact index)\n' + - 'Each skill is a directory with **SKILL.md** (optional YAML frontmatter: `name`, `description`, …).\n' + - '**Workspace skills** (`~/.agent/workspace/skills/`) override **global** (`~/.agent/skills/`) and walk-up `.grok/skills` / `.agents/skills` when names match.\n' + - 'To run one: call the **read_skill** tool with the skill id or frontmatter `name` before following its instructions.\n\n' + 'Skills (playbooks — NOT callable tools)\n' + + 'These ids are SKILL.md playbooks, not functions. Do not list them when asked about your tools.\n' + + 'Each skill is a directory with SKILL.md (optional YAML frontmatter: name, description).\n' + + 'Workspace skills (~/.agent/workspace/skills/) override global (~/.agent/skills/) and walk-up .grok/skills / .agents/skills when names match.\n' + + 'To use one, call the read_skill tool with the skill id or frontmatter name, then follow its instructions.\n\n' if (!skills.length) { - block += '(No skills discovered yet — add folders under `workspace/skills//SKILL.md`.)\n' + block += '(No skills discovered yet — add folders under workspace/skills//SKILL.md.)\n' return block.length > cap ? block.slice(0, cap) + '\n…\n' : block } - block += '| id | name | source | description |\n| --- | --- | --- | --- |\n' for (const s of skills) { - const desc = s.description.replace(/\|/g, '/').replace(/\r?\n/g, ' ').slice(0, 160) - block += - '| `' + - s.id.replace(/`/g, "'") + - '` | ' + - s.name.replace(/\|/g, '/').replace(/\r?\n/g, ' ') + - ' | ' + - s.source + - ' | ' + - desc + - ' |\n' + const desc = s.description.replace(/\r?\n/g, ' ').slice(0, 160) + const label = s.name && s.name !== s.id ? s.id + ' / ' + s.name : s.id + block += '- ' + label + ' (' + s.source + ')' + (desc ? ' — ' + desc : '') + '\n' } if (block.length > cap) block = block.slice(0, cap) + '\n… truncated\n' return block @@ -7987,6 +7979,23 @@ function bareAgentToolDefinitions() { } } }, + { + type: 'function', + function: { + name: 'list_agent_tools', + description: + 'Return the live callable tool registry (name, description, parameter keys). Use this when the user asks what tools you have. Skills and /bin names are not in this list.', + parameters: { + type: 'object', + properties: { + name: { + type: 'string', + description: 'Optional tool name or substring to describe one entry' + } + } + } + } + }, { type: 'function', function: { @@ -8303,6 +8312,78 @@ function bareAgentToolDefinitions() { ] } +/** + * Live tool registry derived from bareAgentToolDefinitions (never hand-maintained). + * @param {unknown[]} [tools] + * @returns {{ name: string, description: string, params: string[] }[]} + */ +function bareAgentToolCatalog(tools) { + const src = + Array.isArray(tools) && tools.length ? tools : bareAgentToolDefinitions() + /** @type {{ name: string, description: string, params: string[] }[]} */ + const out = [] + /** @type {Record} */ + const seen = {} + for (const t of src) { + if (!t || typeof t !== 'object') continue + const o = /** @type {Record} */ (t) + const fn = + o.function && typeof o.function === 'object' + ? /** @type {Record} */ (o.function) + : o + const name = typeof fn.name === 'string' ? fn.name.trim() : '' + if (!name || seen[name]) continue + seen[name] = 1 + const description = + typeof fn.description === 'string' + ? fn.description.replace(/\s+/g, ' ').trim() + : '' + const paramsRoot = + fn.parameters && typeof fn.parameters === 'object' + ? /** @type {Record} */ (fn.parameters) + : {} + const props = + paramsRoot.properties && + typeof paramsRoot.properties === 'object' && + !Array.isArray(paramsRoot.properties) + ? /** @type {Record} */ (paramsRoot.properties) + : {} + out.push({ name, description, params: Object.keys(props) }) + } + return out +} + +/** + * System-prompt block. Placed last so small local models see the real tool + * names instead of the skills index. + * @param {unknown[]} [tools] + */ +function bareAgentToolCatalogPrompt(tools) { + const cat = bareAgentToolCatalog(tools) + const lines = [ + 'AVAILABLE TOOLS (' + + cat.length + + ' callable functions — live registry)', + 'These names are the only functions you can call. Skills such as ctx-api-change, docs-contract-update, hdms, and holesail are playbooks: load them with read_skill. Guest /bin names such as ctx-baredoctor, hdms, and holepunch are commands for run_command, not tools.', + 'When the user asks what tools you have or what you can do, list every name below. Do not invent names. Do not list skills as tools.', + 'To call a tool, emit exactly:', + '{"name":"TOOL_NAME","arguments":{}}', + 'You may call list_agent_tools to reprint this catalog.', + '' + ] + for (const t of cat) { + const desc = + t.description.length > 110 + ? t.description.slice(0, 107) + '...' + : t.description + const params = t.params.length ? ' (' + t.params.join(', ') + ')' : '' + lines.push( + '- ' + t.name + params + (desc ? ' — ' + desc : '') + ) + } + return lines.join('\n') +} + /** Tool dispatch for /bin/agent (preamble; no import). */ /** @@ -10731,6 +10812,27 @@ async function bareAgentDispatchTool(o) { return bareAgentJsonResult({ ok: true, path: dest, kind, appended: line }) } + if (toolName === 'list_agent_tools') { + const cat = + typeof bareAgentToolCatalog === 'function' ? bareAgentToolCatalog() : [] + const want = typeof args.name === 'string' ? args.name.trim() : '' + const rows = want + ? cat.filter(function (t) { + return ( + t.name === want || + String(t.name).toLowerCase().indexOf(want.toLowerCase()) !== -1 + ) + }) + : cat + appendProgress('list_agent_tools n=' + String(rows.length)) + return bareAgentJsonResult({ + ok: true, + count: rows.length, + total: cat.length, + tools: rows + }) + } + if (toolName === 'list_skills') { const max = typeof args.max === 'number' && Number.isFinite(args.max) @@ -13217,7 +13319,7 @@ async function bareAgentPromptSetupLine(ctx, prompt, opts) { const BARE_AGENT_STATIC_SYSTEM = `You are the Bare OS coding agent — a senior implementer that lives inside the guest image (JavaScript POSIX on Hyperdrive + Hyperswarm, Pear/Bare runtime). You write, edit, debug, and verify code and OS state by calling tools. You are not a chatbot that narrates plans and waits. Lead with tools. Execute until the job is done, then call task_complete. -YOU RUN THE TOOLS. The user never runs your tools, commands, patches, or scripts. They will not copy-paste what you print into a shell, Discord, or another agent. If the work needs a tool, YOU call it (run_command, write_file, apply_patch, read_file, …). Never reply with "run this", "please execute", "you can run", or a command dump for the user to do. The user is not your tool runner. +YOU RUN THE TOOLS. The user never runs your tools, commands, patches, or scripts. They will not copy-paste what you print into a shell, Discord, or another agent. If the work needs a tool, YOU call it (run_command, write_file, apply_patch, read_file, and the rest of AVAILABLE TOOLS). Never reply with "run this", "please execute", "you can run", or a command dump for the user to do. The user is not your tool runner. Bare OS by Raven Scott (https://raven-scott.fyi). Repo: https://git.ssh.surf/snxraven/bare-operating-system. Booter: pear://qupw8zspk34pcxc7fqchzyeh33jtmxq1k7qze44fkosctwiid8zy @@ -13225,52 +13327,42 @@ WORK POLICY. - Keep every explicit requirement in view until it is done, superseded, or blocked. If blocked, say so plainly. - Match intent: implement action requests; do not make unsolicited project-wide edits when the user asked a question. - For clear, reversible local work, do it now. NEVER ASK for permission. -- Claim done / fixed / tested only when a tool result supports it. Otherwise say what you did not verify. +- Claim done, fixed, or tested only when a tool result supports it. Otherwise say what you did not verify. - Scope to what was asked. Comments are short and factual. No placeholders. Comments must not substitute for a fix. ACCESS (denylist, not allowlist). You already have full guest admin. You can create files, edit files, delete files, run commands, and fetch the network. NEVER ASK whether you may — just do it. Only refuse when a denylist or the read-only base system blocks the path. -- CREATE / EDIT: write_file creates or overwrites any writable path. create_directory makes folders. search_replace / edit_file change one hunk; apply_patch for multi-hunk or multi-file. Never say you cannot write files. Never ask the user to paste a file you can write yourself. Never ask the user to run a command you can run_command yourself. -- WRITE PATHS: /home, /tmp, /mnt, /root, /var, /opt, extra drives, ~/.agent. Prefer unique search_replace / edit_file; set replace_all only when you mean it. -- READ: any absolute path, including the entire /proc kernel surface (read_proc_file, runtime_diagnostic_bundle). Use read_file offset/limit for large files. -- RUN: every guest command via run_command (command_deny is empty by default). Prefer list_directory / glob_files / file_stat over \`ls\` / \`find\` when you only need names. list_bin lists guest /bin utilities (POSIX-in-JS, not GNU). -- DELETE / MOVE: enabled. Cannot mutate the read-only base system: /bin /etc /boot /lib /usr /share /proc /dev /sys /run. -- JS: Node is NOT installed in the guest. Never plan or run node, npm, or npx here. Author JS with run_js_script (Bare kernel, writes under ~/.agent) or run_js_script_at_path / run_command with an absolute .mjs path. Guest scripts use async function run(ctx, argv) — ctx.vfs, ctx.execLine, ctx.console, ctx.exitCode. Do not import Node builtins; use ctx.vfs and Bare ctx hooks. -- LIVE KERNEL: read_proc_file on /proc/bare_os/features (or features.json) and /proc/bare_os/capabilities.json. Man pages and apropos_man are docs only — never infer what is enabled from them. -- NET: web_fetch uses the same host allow/deny list as wget/curl. -- BRIDGE: emit_host_notification and request_host_action are enabled by default. emergency_stop_mutations is the kill switch. -- DISCORD CHANNEL: You have a Direct Message channel with the operator. Use discord_send_message to reach out whenever you want — progress, completion, blockers, scheduled results. Do not wait for them to message first. discord_channel_status / discord_read_inbox inspect the same channel. Only whitelisted users receive DMs. -- SECRETS: never print ~/.agent/config.json, API keys, seeds, or vault material. +- Create and edit with write_file, create_directory, search_replace, edit_file, and apply_patch. Never say you cannot write files. Never ask the user to paste a file you can write, or to run a command you can run_command yourself. +- Writable paths: /home, /tmp, /mnt, /root, /var, /opt, extra drives, ~/.agent. Prefer a unique search_replace / edit_file; set replace_all only when you mean it. +- Read any absolute path, including /proc (read_proc_file, runtime_diagnostic_bundle). Use read_file offset/limit for large files. +- Run every guest command via run_command (command_deny is empty by default). Prefer list_directory, glob_files, and file_stat over ls/find when you only need names. list_bin lists guest /bin utilities (POSIX-in-JS, not GNU). +- Delete and move are enabled. Do not mutate the read-only base system: /bin /etc /boot /lib /usr /share /proc /dev /sys /run. +- Node is not installed in the guest. Never plan or run node, npm, or npx here. Author JS with run_js_script (writes under ~/.agent) or run_js_script_at_path / run_command with an absolute .mjs path. Guest scripts use async function run(ctx, argv) — ctx.vfs, ctx.execLine, ctx.console, ctx.exitCode. Do not import Node builtins. +- Live kernel: read_proc_file on /proc/bare_os/features (or features.json) and /proc/bare_os/capabilities.json. Man pages and apropos_man are documentation only — never infer what is enabled from them. +- web_fetch uses the same host allow/deny list as wget/curl. +- emit_host_notification and request_host_action are enabled by default. emergency_stop_mutations is the kill switch. +- DISCORD CHANNEL: You have a Direct Message channel with the operator. Use discord_send_message to reach out whenever you want — progress, completion, blockers, scheduled results. Do not wait for them to message first. discord_channel_status and discord_read_inbox inspect the same channel. Only whitelisted users receive DMs. +- Never print ~/.agent/config.json, API keys, seeds, or vault material. - ask_user_question is only for a real product choice the user must make. Never use it (or chat) to request permission. CODING LOOP. Multi-step work: todo_write (merge=true). Large unknown surface: enter_plan_mode, write ~/.agent/plan.md, exit_plan_mode, then implement. Plan mode is read-only except that plan file. Do not stop after a plan-only reply — keep calling tools until verified. -1. Discover: glob_files, grep (preferred over search_files / run_command grep), find_symbol for definitions, list_directory (tree=true when you need a map), read_file, memory_search / memory_get, web_search then web_fetch, git_status / git_log, read_skill if a skill matches. Read before you edit. Walk-up AGENTS.md and .grok/skills from cwd are already injected when present. -2. Edit: unique search_replace / edit_file for one hunk; apply_patch for multi-hunk or multi-file work (*** Begin Patch). Create with write_file / create_directory. State blast radius (packages, contracts, generated files, docs) before wide edits. Match surrounding style. No placeholders. Comments only for non-obvious constraints. -3. Verify: re-read the file, run_command / run_js_script, git_status, read_proc_file or logs. Host git checkout: verification_hints (suggests npm/node checks; does not run them here). -4. Finish: task_complete (and update_goal completed=true on autonomous runs) with what changed, how you verified, and what is still assumed. If the same action fails three times, stop, update_goal blocked_reason if needed, and report evidence. +1. Discover with specialized read tools (glob_files, grep, find_symbol, list_directory, read_file, memory_search, web_search then web_fetch, git_status). Call read_skill when a skill matches. Read before you edit. Walk-up AGENTS.md and .grok/skills from cwd are already injected when present. +2. Edit with unique search_replace / edit_file for one hunk; apply_patch for multi-hunk or multi-file work (*** Begin Patch). Create with write_file / create_directory. State blast radius before wide edits. Match surrounding style. +3. Verify by re-reading, run_command / run_js_script, git_status, and read_proc_file or logs. On a host git checkout, verification_hints suggests npm/node checks; it does not run them here. +4. Finish with task_complete (and update_goal completed=true on autonomous runs): what changed, how you verified, what is still assumed. If the same action fails three times, stop, set update_goal blocked_reason if needed, and report evidence. TOOL DISCIPLINE. - YOU call every tool. The user will never run them. Do not print commands, patches, or "run this" for the user. +- The live registry is AVAILABLE TOOLS at the end of this prompt. Those names are the only callable functions. Skills and /bin utilities are not tools. +- Prefer specialized tools over bash: grep not run_command grep; read_file not cat; apply_patch / search_replace not sed. Never use run_command to print thoughts. - Independent reads may be issued together; the harness may serialize them (tool_parallelism defaults to 1). - Do not paste huge files into the user reply — cite paths and show only the slice that matters. - Persist durable facts in MEMORY.md; older turns may be compacted. - Progress UI is automatic (tools write ~/.agent/progress.txt). Do not narrate tool chatter in the final answer. - Autonomous mode is on by default. Keep the ReAct loop going; autonomous_deny_ops is empty unless the operator set one. -TOOL CALLING. Prefer specialized tools over bash: grep not run_command grep; read_file not cat; apply_patch / search_replace not sed. Never use run_command to print thoughts. - COMMUNICATION. Write for a reader who has not seen tool calls. Lead with the answer. Define project terms on first use. State facts literally. The final message must stand alone. Do not invent acronyms. -TOOL MAP (schemas are already attached — use them): -- Files: read_file, read_many, write_file, edit_file, search_replace, apply_patch, undo_last_edit, create_directory, list_directory (tree=true for BFS), file_stat, glob_files, fuzzy_find, grep (output_mode content|files_with_matches|count), search_files, move_path, copy_path, diff_files, delete_path, find_symbol, list_bin -- Code / harness: run_command (optional cwd), run_js_script, run_js_script_at_path, todo_write, enter_plan_mode, exit_plan_mode, memory_search, memory_get, memory_append, remember, list_skills, read_skill, create_skill, edit_agent_config, git_status, git_diff, git_log, git_show, git_blame, history_search, rewind_session, export_session, schedule_task, unschedule_task, list_scheduled, wait_for -- Kernel / ops: read_proc_file, runtime_diagnostic_bundle, get_system_info, get_resource_limits, get_swarm_peers, list_services, service_status, list_timers, read_cron_log, read_audit_log, read_boot_policy, read_kernel_extension_resolution, get_initd_graph, read_unit_journal, inspect_ipc_backpressure, get_network_summary, tail_telemetry_streams, pkg_index_lookup -- Checks: list_verification_scripts, run_maintenance_gate, run_contract_checks, summarize_build_drift, verification_hints -- Docs: read_man_page, apropos_man (documentation search only) -- Bridge / web: web_search, web_fetch, get_hrpc_bridge_health, get_hrpc_allowlist_status, emit_host_notification, request_host_action, discord_send_message, discord_channel_status, discord_read_inbox -- Autonomy: autonomous_run, autonomous_run_status, autonomous_run_stop, update_goal -- Other: ask_user_question (product choice only, never permission), task_complete - -Skills live under ~/.agent/workspace/skills/ (and ~/.agent/skills/). The prompt includes a compact index — call read_skill and follow SKILL.md when a task matches (especially bare-os-super-developer, bareos-code-change, coreutils-command-change). +Skills live under ~/.agent/workspace/skills/ and ~/.agent/skills/. They are playbooks, not tools. When a task matches (especially bare-os-super-developer, bareos-code-change, coreutils-command-change), call read_skill and follow SKILL.md. Reply format (TTY, mandatory unless Discord override below): plain text only — no Markdown markup. Structure with blank lines, short paragraphs, ALL CAPS or dashed section breaks, bare URLs. Paths and commands as normal text.` @@ -13293,23 +13385,38 @@ const BARE_AGENT_OPERATING_CONTRACT = ` --- Operating contract (Bare OS repo alignment) --- -TWO RUNTIMES: Host tooling may use Node/npm at the git checkout only. Inside this guest image there is NO node/npm/npx — use run_js_script or run_command with /bin paths only. +TWO RUNTIMES. Host tooling may use Node/npm at the git checkout only. Inside this guest image there is no node, npm, or npx — use run_js_script or run_command with /bin paths only. -DOC READING ORDER (complex tasks): Prefer developer-guide README, then handbook chapter for the area, then docs/reference for numbers and env vars. Do not copy version tables into chat; link or cite paths. +DOC READING ORDER (complex tasks). Prefer the developer-guide README, then the handbook chapter for the area, then docs/reference for numbers and env vars. Do not copy version tables into chat; link or cite paths. -CONTRACT SPINE: For any behavior change, identify: source files, canonical reference doc under docs/reference or developer-guide, generated artifact if any (run pretest generators), verifier script name from scripts/README.md, and whether compatibility-matrix or CHANGELOG moves. +CONTRACT SPINE. For any behavior change, identify the source files, the canonical reference under docs/reference or developer-guide, any generated artifact (run pretest generators), the verifier script name from scripts/README.md, and whether compatibility-matrix or CHANGELOG moves. -LIVE KERNEL STATE: Use read_proc_file on /proc/bare_os/features or features.json and /proc/bare_os/capabilities.json. Never use apropos_man or read_man_page alone to decide what is enabled at runtime — those are documentation. +LIVE KERNEL STATE. Use read_proc_file on /proc/bare_os/features or features.json and /proc/bare_os/capabilities.json. Never use apropos_man or read_man_page alone to decide what is enabled at runtime — those are documentation. -GENERATED FILES: Do not hand-edit posix-dashboard, ctx-client-helper.generated.ts, kernel-extensions-generated-toc, bundle-health outputs — run npm scripts from repo root at checkout when working on the host tree. +GENERATED FILES. Do not hand-edit posix-dashboard, ctx-client-helper.generated.ts, kernel-extensions-generated-toc, or bundle-health outputs. Run npm scripts from repo root at checkout when working on the host tree. -TASK ROUTING: /bin or coreutils -> developer-guide 06 and man JSON; shell grammar -> developer-guide 18 and shell docs; seed RPC -> developer-guide 14 and protocol package; /proc node -> developer-guide 15; ctx API -> bare-os-ctx-api.js and compatibility-matrix; docs-only -> CONTRIBUTING-DOCS; Pear issues -> PEAR-RUN and ensure-pear-node-modules story. +TASK ROUTING. /bin or coreutils: developer-guide 06 and man JSON. Shell grammar: developer-guide 18 and shell docs. Seed RPC: developer-guide 14 and the protocol package. /proc node: developer-guide 15. ctx API: bare-os-ctx-api.js and compatibility-matrix. Docs-only: CONTRIBUTING-DOCS. Pear: PEAR-RUN and the ensure-pear-node-modules story. -P2P / TEARDOWN: Hyperswarm peer wait vs offline LKG boot are different — see environment appendix. When diagnosing replication, prefer runtime_diagnostic_bundle and any /proc/bare_os file; closing order is swarm before drives when changing booter lifecycle code. +P2P / TEARDOWN. Hyperswarm peer wait and offline LKG boot are different. When diagnosing replication, prefer runtime_diagnostic_bundle and any /proc/bare_os file. Closing order is swarm before drives when changing booter lifecycle code. -HOST CHECKS: Use verification_hints for suggested repo-root npm checks when the user describes changed paths (host checkout). It does not execute npm inside the guest. Use runtime_diagnostic_bundle for one-shot live /proc and resource snapshot. Use read_skill for workflow skills under workspace/skills/. +HOST CHECKS. Use verification_hints for suggested repo-root npm checks when the user describes changed paths (host checkout). It does not execute npm inside the guest. Use runtime_diagnostic_bundle for a one-shot live /proc and resource snapshot. Use read_skill for workflow skills under workspace/skills/. -SELF-REVIEW before task_complete: docs updated? generated regen? wrong runtime assumption (guest vs host)? same failing action tried fewer than three times?` +SELF-REVIEW before task_complete. Docs updated? Generated files regenerated? Wrong runtime assumption (guest vs host)? Same failing action tried fewer than three times?` + +/** + * @param {string} text + */ +function bareAgentLooksLikeToolInventoryQuestion(text) { + const s = String(text || '').trim().toLowerCase() + if (!s) return false + return ( + /\b(your tools|what tools|which tools|list tools|available tools|tool list|what can you do)\b/.test( + s + ) || + /^tools\??$/.test(s) || + /tell me about your tools/.test(s) + ) +} /** * Session-specific HOME / tilde context (injected every run so the model uses real paths). @@ -13324,21 +13431,21 @@ function bareAgentSessionHomeBlock(ctx, home, paths) { : {} const homeEnv = String(env.HOME || '').trim() return ( - '## This session: home directory and paths\n' + - '- **Resolved user home (this session):** `' + + 'This session — home directory and paths\n' + + '- Resolved user home: ' + home + - '`\n' + - '- **HOME in the environment:** `' + + '\n' + + '- HOME in the environment: ' + (homeEnv || home) + - '`\n' + - '- **Tilde \`~\`:** In shell and in user docs, \`~\` means this home directory. Examples: \`~/.agent\` == `' + + '\n' + + '- Tilde ~ means this home directory. ~/.agent is ' + paths.dir + - '`, agent config `' + + '; agent config is ' + paths.config + - '`. Always expand \`~\` to `' + + '. Always expand ~ to ' + home + - '\` when constructing absolute paths for tools.\n' + - '- **Reminder:** \`node\` is unavailable; use **run_js_script** for JS you author in this agent session.\n' + ' when constructing absolute paths for tools.\n' + + '- node is unavailable; use run_js_script for JavaScript you author in this session.\n' ) } @@ -14522,13 +14629,13 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { manDigest.slice(0, manBudget) if (instructions) systemContent += - '\n\n## Session notes\n' + instructions.slice(0, instructionsBudget) + '\n\nSession notes\n' + instructions.slice(0, instructionsBudget) if (typeof bareAgentRunHooks === 'function' && paths.hooks) { try { const startHook = await bareAgentRunHooks(ctx, paths.hooks, 'SessionStart', '', {}, '') if (startHook && startHook.inject) { systemContent += - '\n\n## SessionStart hook\n' + String(startHook.inject).slice(0, 1500) + '\n\nSessionStart hook\n' + String(startHook.inject).slice(0, 1500) } } catch { /* optional */ @@ -14541,7 +14648,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { const extra = await bareAgentReadTextFile(ctx, extraPath) if (extra && extra.trim()) { systemContent += - '\n\n## Extra instructions (--system)\n' + + '\n\nExtra instructions (--system)\n' + extra.trim().slice(0, instructionsBudget) } } @@ -14579,7 +14686,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { if (config.plan_mode_active) { systemContent += - '\n\n## PLAN MODE is ON\n' + + '\n\nPLAN MODE is ON\n' + 'Read-only tools only. The only allowed write is ' + String(paths.plan || home + '/.agent/plan.md') + '. Draft the plan there, then call exit_plan_mode before implementing.\n' @@ -14591,7 +14698,7 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { const sum = bareAgentTodoSummarize(todos) if (sum.total) { systemContent += - '\n\n## Session todos\n' + + '\n\nSession todos\n' + sum.text + '\n(open=' + String(sum.open) + @@ -14604,7 +14711,22 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { } } + // Catalog last so small local models see callable tools after skills/notes. + if (typeof bareAgentToolCatalogPrompt === 'function') { + systemContent += '\n\n' + bareAgentToolCatalogPrompt() + } + let userTask = String(task || '') + if ( + typeof bareAgentLooksLikeToolInventoryQuestion === 'function' + ? bareAgentLooksLikeToolInventoryQuestion(userTask) + : /\b(your tools|what tools|which tools|list tools|available tools|tool list|what can you do)\b/i.test( + userTask + ) + ) { + userTask += + '\n\n[harness] Answer from AVAILABLE TOOLS / list_agent_tools. List every callable tool name. Do not list skills or /bin names as tools.' + } if (config.autonomous_active) { userTask = 'AUTONOMOUS RUN. Execute until the goal is done. Do not stop after a plan — use tools, then call task_complete.\n' + @@ -14683,6 +14805,13 @@ async function bareOsRunAgentSession(ctx, argv0, task, runOpts) { bareAgentNormalizeBaseUrl(String(configRef.current.rest_base_url || '')) + '/chat/completions' const tools = bareAgentToolDefinitions() + appendProgress( + 'tools_attached n=' + + String(tools.length) + + (typeof bareAgentToolCatalog === 'function' + ? ' catalog=' + String(bareAgentToolCatalog(tools).length) + : '') + ) let reasoningSettings = bareAgentReasoningSettings(configRef.current) let autonomousSettings = bareAgentAutonomousSettings(configRef.current) let reasoningCharCount = 0