Release Notes

Register named providers that wrap one of the three installed SDKs (openai, anthropic, xai) and point it at a different endpoint — Ollama, LM Studio, vLLM, OpenRouter, internal proxies, anything compatible. Multiple custom providers coexist with the built-ins.

• CLI: bernard add-provider <name> --sdk <openai|anthropic|xai> --base-url <url> --model <model>, bernard remove-provider <name>, bernard providers.
• REPL: /provider now ends with + Add custom provider… (interactive wizard); for a custom provider, /model ends with + Type a new model name… and remembers what you typed.
• Keys are stored in keys.json alongside built-in keys and are never injected into process.env.

A new ask_user tool lets the agent pause the turn and ask clarifying questions instead of writing them as prose (which previously left the turn idle and, in coordinator mode, tripped the plan-enforcement loop).

• Batches up to 10 questions in one call — useful for “title + body + labels” style prompts — with a pinned tab strip showing [1 ✓ ▸2◂ 3] progress.
• Each question can be free-form or constrained by a choices list with an optional “Other” escape hatch (custom label supported).
• Esc mid-batch returns whatever was already answered, so partial input isn’t lost. Headless sessions get {unavailable: true} so the agent can pick a default.

The dangerous-command confirmation now uses the arrow-key menu instead of the legacy rl.question prompt, fixing a hang where the spinner blanked the prompt and the agent waited indefinitely. rm calls scoped to Bernard’s own scratch-script directory (<os-tmpdir>/bernard-*, e.g. /tmp/bernard-* on Linux — an internal constant in src/tools/shell.ts, not a configurable env var) and free of shell metacharacters skip the prompt entirely — that’s just Bernard cleaning up after itself.

The main agent’s direct shell, web_read, file_read_lines, and file_edit_lines calls are now transparently routed through their bundled wrapper specialists — same tool name and schema, only the execute path changes. The model can’t tell the difference, but wrappers add OS-aware examples, schema validation, and result capping.

• A one-shot repair hook is wired into every generateText site (main, specialist, subagent, tool-wrapper, cron). On InvalidToolArgumentsError or NoSuchToolError the model gets re-prompted with the actual error.
• Argument-truncation errors (the classic 16 KB heredoc cut mid-string) trigger a specific hint to use file_write + shell instead of inlining the payload.
• tool_wrapper_run now strips the wrapper’s reasoning array and caps the result field before returning to the parent, keeping the parent context clean. Reasoning still lands in logs/tool-wrappers.jsonl for debugging.
• Wrapper errors are mapped back to the native tool’s error shape (e.g. shell → { output, is_error: true }), so detectToolError and tool-profile learning keep working under the shim.

Tasks now receive a proportional fraction of the available step budget instead of a fixed step count, and the final step forces text-only output to guarantee structured JSON results.

• Introduced TASK_STEP_RATIO, getTaskMaxSteps, and makeLastStepTextOnly utilities in task.ts
• Overhauled wrapTaskResult to robustly extract and validate JSON from arbitrary output, handling nested objects, escaped quotes, and multiple JSON blocks
• Updated REPL and task tool to pass correct maxSteps and experimental_prepareStep options
• Expanded test coverage for result extraction edge cases

The auto-create threshold now accepts both fractional (0–1) and percentage (1–100) values, and pending specialist candidates are re-evaluated automatically when settings change.

• Added normalizeThreshold utility in config.ts that clamps values to the [0, 1] range regardless of input format
• Updated /agent-options threshold command to display both normalized and percentage values
• Refactored auto-creation logic into a reusable promoteCandidate helper
• Enabling auto-create now re-evaluates and promotes pending candidates above the threshold

A reusable Plan-Act-Critic (PAC) loop now wraps sub-agents, specialists, and cron job executions when critic mode is enabled. The critic verifies each agent’s work and retries on failure with targeted feedback.

• Extracted pac.ts module with runPACLoop() for consistent critic integration
• Sub-agents, specialists, and cron jobs all use the same PAC pipeline automatically
• On a FAIL verdict, the task is retried with the critic’s feedback (up to 2 retries)
• Compact critic verdicts — PASS and WARN are now single-line inline, reducing terminal noise while keeping full detail for FAIL

When a response hits the max-tokens limit and is cut off, Bernard automatically continues where it left off — up to 3 continuations — then recommends an optimal token limit based on actual usage.

• Seamless auto-continuation with no user intervention required
• Post-completion recommendation of the ideal max-tokens value (1.25× actual usage)
• Warning with /options max-tokens instructions if still incomplete after 3 continuations
• New /debug command — prints a diagnostic report (runtime, config, API key status, MCP servers, RAG stats) with no secrets leaked

Smarter specialist suggestions powered by token-based similarity scoring, plus a new auto-creation mode that saves high-confidence candidates without manual review.

• Jaccard similarity across name, description, system prompt, and guidelines (weighted)
• Candidates with >60% overlap are suppressed; partial overlaps suggest enhancing the existing specialist
• New /agent-options command to toggle auto-creation and set confidence threshold at runtime
• BERNARD_AUTO_CREATE_SPECIALISTS and BERNARD_AUTO_CREATE_THRESHOLD environment variables

Control how many tool-call iterations the agent can take per request, and benefit from step-by-step planning in every conversation — not just critic mode.

• BERNARD_MAX_STEPS environment variable (default 25) with runtime override via /options max-steps
• Sub-agents automatically receive 50% of the main agent’s step budget
• When the step limit is hit, Bernard offers to double it for the current session
• System prompt now includes a planning section by default, improving multi-step task execution

Bernard now has full temporal context — every user message carries an ISO 8601 timestamp, and the system prompt includes the current date and time.

• ISO 8601 timestamp prefixes on every user message for precise time context
• System prompt upgraded from “Today’s date” to “Current date and time”
• time tool output now includes seconds and timezone
• Bug fix: MCP URL transport correctly uses HTTP for non-SSE endpoints (was always defaulting to SSE)

Two new tools — file_read_lines and file_edit_lines — give Bernard precision file editing without shelling out to sed or awk.

• Line-based editing with replace, insert, delete, and append operations
• Atomic writes — all edits in a single call succeed or all fail
• Binary file detection and pagination for large files (50 MB cap)
• file_read_lines supports offset/limit for targeted reads across big files

Override the provider and model for any specialist, sub-agent, or task invocation. Set a default model on specialists at creation time, or pass an override at call time.

• Specialists store a default provider/model that is used every time they run
• Sub-agents and tasks accept per-invocation provider/model overrides
• Cross-provider safety — auto-resolves the default model when switching providers
• Model tags shown in /specialists listings

Tasks are now single-step: one LLM call with tool use, then structured output. Saved tasks can be invoked by ID, and the REPL shows tasks and routines separately.

• Reduced to maxSteps: 2 (one tool-use round + structured result)
• New taskId parameter — run saved task routines by ID or via /task-{id} in the REPL
• Separate display for tasks vs routines in the REPL
• Auto-context injection — working directory, available tools, memory, and RAG are included automatically

Tool-result summaries shown to the critic now use an adaptive 8000-char budget instead of a fixed 500-char slice, so the critic sees far more context before judging.

• Adaptive 8000-char budget with a 500-char floor per tool result
• Response truncation raised from 2000 to 4000 chars, args from 500 to 1000
• Truncation markers (...) added to tool results
• Critic prompt updated to not treat truncation as failure evidence

Agents no longer retry the exact same failed command, and external API mutations get a short delay before re-verification to account for eventual consistency.

• No retrying the exact same failed command across all agent types and cron jobs
• Wait 2–5s before re-verifying external API mutations

Candidates that have already been saved as specialists are automatically filtered out of future suggestions.

• Auto-marks candidates as accepted when a specialist is created by draft ID or name
• Defensive filtering at startup and in /candidates output

Critic mode now retries up to 2 times when it detects issues, feeding its verdict back to the agent for correction before giving up.

• Up to 2 automatic retries when critic returns WARN or FAIL
• Critic feedback is injected into conversation history so the agent can self-correct
• Improved verdict parsing using regex-based extraction
• REPL prompt now shows a diamond symbol instead of [CRITIC] label
• New parseCriticVerdict utility extracted to output module

Added /create-task guided creation command and refactored duplicate creation logic.

• New /create-task REPL command for guided task routine creation
• Refactored duplicate guided creation logic into a shared runGuidedCreation helper
• Help menu now lists /routines, /create-routine, and /create-task commands

Specialists are now automatically matched to user input using keyword scoring, with high-confidence matches auto-dispatched.

• Keyword-based scoring matches user input against saved specialists
• Scores ≥ 0.8 trigger automatic dispatch; 0.4–0.8 prompt for confirmation
• Match advisory injected into system prompt so the agent knows which specialists fit
• Stop-word filtering prevents common words from inflating scores

Tasks are isolated, focused executions that return structured JSON output. Unlike sub-agents (which return free-form text), tasks always produce a {status, output, details?} response — making them ideal for machine-readable results, routine chaining, and conditional branching.

• 5-step budget — tasks are meant to be quick and focused
• Structured JSON output — always returns {status: "success"|"error", output, details?}
• Completely isolated from the current session — no conversation history
• Available as both a tool and a command via /task
• Shared concurrency pool with sub-agents (4 slots max)

Specialists are reusable expert profiles — persistent personas with custom system prompts and behavioral guidelines that shape how a sub-agent approaches work. Unlike routines (which define what steps to follow), specialists define how to work.

• Each specialist run gets its own generateText loop with a 10-step budget
• One JSON file per specialist in ~/.local/share/bernard/specialists/
• Manage via /specialists or invoke directly with /{id}
• Up to 50 specialists with lowercase kebab-case IDs

Bernard automatically detects recurring delegation patterns in your conversations and suggests new specialists. Detection runs in the background when you exit a session or use /clear --save.

• Review pending suggestions with /candidates
• Each candidate includes a name, description, confidence score, and reasoning
• Accept or reject candidates conversationally
• Auto-dismissed after 30 days if not reviewed; up to 10 pending at a time

Critic mode adds planning, proactive scratch/memory usage, and post-response verification. Recommended for high-stakes work like deployments, git operations, and multi-file edits.

• Planning — writes a plan to scratch before multi-step tasks
• Proactive scratch — accumulates findings during complex work
• Verification — a critic agent reviews the work and prints a verdict (PASS / WARN / FAIL)

A fourth specialized domain in Bernard's RAG memory system, joining Tool Usage Patterns, User Preferences, and General Knowledge. Conversation summaries capture what was discussed, approaches taken, tools/specialists/routines used, and outcomes.

• Extracts high-level summaries of what was accomplished each session
• Helps identify recurring patterns across conversations
• All four domain extractors run in parallel — no extra latency
• Search returns up to 5 results per domain (15 total max), preventing any single domain from crowding out others

Earlier releases — Routines, the multi-provider switch, and the first sub-agent loop — predate this changelog format and aren't reproduced here. See the git history for anything before v0.6.0.