ptc_runner

# Changelog

All notable changes to this project will be documented in this file.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## [0.14.0] - 2026-08-24

### Breaking rewrite

This is a breaking 0.x replacement release. Existing integrations should be
treated as new integrations against the current guides and references; removed
APIs and commands have no compatibility facades.

- The legacy SubAgent, MCP, upstream, mutable-prelude, and trace products have
  been replaced by the bounded, owner-based `PtcRunner.Kernel`. Applications
  now use strict manifests, immutable component bundles, separate workflow and
  named mission environments, and explicitly host-installed capabilities.
- The command surface is now `ptc` (or `mix ptc` in a source checkout) for
  initialization, validation, execution, REPL use, diagnostics, transcripts,
  and the packaged trace Viewer. The former `mix ptc.run` and `mix ptc.repl`
  entry points and implicit environment discovery were removed.
- PTC-Lisp now runs through neutral contexts and results with bounded mission
  evaluation, generic Lisp-authored agent libraries, closed Java interop, and
  manifest-relative input and result contracts. Legacy agent journal, budget,
  progress, discovery, and upstream-inference forms were removed.
- Tool and model access is assembled as explicit host capability sources,
  including bounded LLM, filesystem, trace, Streamable HTTP MCP, and owned
  stdio MCP providers. Provider ownership, cleanup, caching, limits, and
  failure envelopes are enforced by the Kernel.
- Canonical events, private inspection, run-analysis navigation, transcripts,
  and the Viewer now share one bounded evidence model. Generated components can
  be materialized, statically gated, validated against host-owned trials, and
  supplied to a later run without granting a running program mutation authority.

## [0.13.0] - 2026-06-24

### Breaking Changes

- Slimmed the MCP `lisp_task` activity ledger (removed internal turn and
  args-hash fields from the success overview).

### Added

- Capability Preludes (V1): define and deploy reusable, namespaced PTC-Lisp
  capability libraries — including tool-backed exports — that programs and
  SubAgents can call. Backed by a versioned prelude store with editing tools,
  configurable defaults, and per-session prelude selection (root, MCP, and
  SubAgent). See the authoring & deploying guide.
- Session turn log and introspection: record per-turn session activity and
  query it from PTC-Lisp via the `log/` prelude and `(source ...)` discovery.
- Paged access to large tool results so programs can fold over big payloads
  within the memory budget.
- New PTC-Lisp helpers: `describe`, `doc`, and JSON `parse-lines`.
- SubAgent-to-upstream bridge so SubAgents can call configured upstream tools
  with fail-closed prelude requirements.
- Opt-in session feedback (collection hints) and configurable session
  preview/result caps.

### Changed

- Re-baselined sandbox heap accounting so granted environment data no longer
  counts against a program's memory budget; programs that previously hit the
  limit purely from large grants now run.
- Upstream tool calls apply a default side-effect guard.

### Fixed

- Hardened memory bounds so retained buffers (credential redaction, debug ring,
  trace collector, prelude store) can no longer grow unbounded.
- PTC-Lisp formatting/conformance fixes for regex literals, reader-macro forms,
  and turn-history references (`*1`/`*2`/`*3`).
- Fixed broken cross-references in the published API documentation.

## [0.12.0] - 2026-06-03

### Breaking Changes

- Renamed the upstream tool Lisp surface from the MCP-specific
  `(tool/mcp-call ...)` and `(mcp/servers)` forms to transport-neutral
  `(tool/call ...)` and `(tool/servers)`.
- Tightened upstream configuration transport names. Use `"openapi"`,
  `"mcp_stdio"`, or `"mcp_http"`; older ambiguous names such as `"stdio"` and
  `"http"` are rejected.
- Corrected several PTC-Lisp edge-case semantics for Clojure conformance,
  including `find` as associative lookup, `range` type errors, negative `nth`,
  duplicate literal map/set keys, and arity handling for `assoc`, `juxt`, and
  bitwise helpers.

### Added

- Added the root upstream runtime for embedded Elixir callers and `mix ptc.repl`,
  allowing PTC-Lisp programs to call configured OpenAPI, MCP stdio, and MCP HTTP
  upstream tools without running the MCP server.
- Added read-only JSON OpenAPI upstream support with credential bindings,
  operation allow-lists, schema loading, response caps, and the same tagged
  `tool/call` result model used for MCP upstreams.
- Added `PtcRunner.Session` for embedding stateful REPL-style PTC-Lisp
  evaluation with persistent `(def ...)` memory and bounded `*1`/`*2`/`*3`
  result history.
- Added upstream-aware `mix ptc.repl` options for loading upstream configs,
  listing tools, discovery, call budgets, response caps, and catalog snapshot
  modes.
- Added `mission_log_in: :user_message` format support so agents can keep system
  prompts stable for provider prompt caching.

### Changed

- Moved MCP aggregation onto the shared root upstream runtime, so root callers
  and the MCP server use the same config format, discovery forms, credential
  redaction, and `tool/call` execution semantics.
- Expanded PTC-Lisp Clojure conformance across nil-tolerant sequence helpers,
  multi-collection `map`/`mapv`/`pmap`, variadic `interleave`, three-arity
  `nth`, seq-form `replace`, Java `Boolean/parseBoolean`, and Java Math special
  cases.
- Improved release readiness with deterministic smoke, performance, coverage,
  docs, and package-content checks.

### Fixed

- Authentication is enforced before MCP HTTP method dispatch, required on
  non-loopback binds, rate-limited on failed bearer attempts, and scrubbed from
  traces/log-facing output.
- Upstream MCP result normalization now returns the first text block rather than
  only inspecting the first content item.
- Closed-but-unpruned MCP HTTP sessions now resolve to tombstones instead of
  crashing through a missing process.
- Invalid integer HTTP/MCP configuration values now fail fast with clearer
  errors.

## [0.11.0] - 2026-05-25

### Breaking Changes

- Removed the underscore-prefix context firewall convention. `_`-prefixed
  context keys are now visible like ordinary keys.
- `(tool/mcp-call ...)` now returns tagged result maps directly. Remove uses
  of `mcp/text` and `mcp/json`; check `:ok` and read `:value` instead.

### Added

- Added PTC-Lisp ISO-8601 date-time parsing through `parse`,
  `LocalDate/parse`, and `Instant/parse`.
- Added PTC-Lisp bitwise integer builtins, including `bit-and`, `bit-or`,
  shifts, bit set/clear/flip/test, and `bit-not`.
- Added `(list & args)` as a Clojure-friendly alias for `vector`.
- Added `json/parse-string` and `json/generate-string` to PTC-Lisp.
- Added runtime callable support and unified REPL discovery helpers.
- Added Java duration helpers and namespace conformance audits.
- Added `KeyNormalizer.canonical_cache_key/2` for stable tool-result cache
  keys across atom/string map keys, map order, and integer-equivalent floats.
- Added opt-in `compaction:` support for pressure-triggered multi-turn context
  trimming.

### Fixed

- Improved PTC-Lisp type errors so common argument swaps and builtin values no
  longer leak Elixir internals.
- Preserved `sort-by` key errors and added support for sorting maps and vector
  paths.
- Accepted common `format` width hints.
- Kept compilation clean on Elixir 1.20 release candidates.

## [0.10.1] - 2026-05-04

### Fixed

- `count`, `reduce`, and `assoc` now treat `nil` as an empty collection, matching the nil-tolerance of other collection helpers (#863)

### Changed

- Lockfile refresh: `req_llm` 1.8.0 → 1.10.0 (constraint stays at `~> 1.8`, no consumer action needed)
- Dev tooling bumps: `credo` 1.7.18, `usage_rules` 1.2.6

## [0.10.0] - 2026-03-26

### Breaking Changes

- **MetaPlanner, PlanExecutor, PlanRunner, PlanTracer, PlanCritic removed** — The autonomous planning system with JSON task graphs, verification predicates, and replanning has been removed. Put orchestration and agent policy in PTC-Lisp instead — see [Building agents](docs/guides/building-agents.md).
- Planning prompt templates removed (`planning-examples.md`, `verification-predicate-guide.md`, `verification-predicate-reminder.md`, `signature-guide.md`)
- PlanExecutor telemetry events removed
- Internal `llm_client` package removed — use `PtcRunner.LLM` behaviour directly
- **`plan:` no longer auto-enables `journaling: true`** — Plans are now display-only labels for progress visibility. To use journaled task caching, set `journaling: true` explicitly.

### Added

**Model String Shorthand**

- Accept model strings directly in `SubAgent.run` — e.g., `llm: "haiku"` instead of building an LLM struct

**Clojure Conformance Expansion (~30 new functions/forms)**

- Control flow: `case`, `condp`
- HOF combinators: `comp`, `partial`, `complement`, `constantly`, `every-pred`, `some-fn`
- Collection operations: `cons`, `disj`, `empty`, `merge-with`, `reduce-kv`, `zipmap`, `filterv`, `update-keys`, `peek`, `pop`, `subvec`
- Sequence operations: `split-at`, `split-with`, `partition-by`, `dedupe`, `keep`, `keep-indexed`
- String/coercion: `format`, `name`, `keyword`, `hash-map`
- Type predicates: `int?`, `integer?`, `double?`, `float?`, `fn?`, `false?`, `true?`, `symbol?`, `decimal?`, `ratio?`, `rational?`, `nat-int?`, `neg-int?`, `pos-int?`, `infinite?`, `NaN?`
- Collection capability predicates (`sequential?`, `associative?`, `counted?`, etc.)
- Named fn for self-recursion (`(fn name [x] ...)`)
- `%&` rest args in `#()` short function syntax
- Keyword args via rest destructuring (`[& {:keys [a]}]`)
- `:strs` map destructuring for string-keyed maps
- `def`/`defn`/`defonce` can shadow builtins (Clojure-compatible)

**Java Interop Methods**

- String: `.startsWith`, `.endsWith`, `.toLowerCase`, `.toUpperCase`, `.contains`
- Date/DateTime: `.isBefore`, `.isAfter`

**Prompt System**

- 2-axis composable prompt architecture for flexible prompt composition
- Language reference included in default prompt compositions
- Pluggable `progress_fn` for custom turn feedback rendering

**Tracing**

- JSONL trace format v2 with flat event envelope
- Typed trace headers (`trace_kind`, `producer`, `query`, `model`, `trace_label`)
- Trace analyzer agent for investigating execution traces
- Streaming `query_events` and `aggregate_events` tools

### Fixed

- `keys`/`vals` nil-tolerant (like other collection helpers)
- `concat` type error diagnostics for non-collection args
- Hyphen/underscore normalization in flex_access lookups
- `and` returns last truthy value instead of boolean (Clojure conformance)
- `some` with keyword pred returns extracted value, not boolean
- `defn` inside `let` now visible across program expressions
- Code fence parsing uses line-by-line parser instead of regex
- Runtime exceptions classified as `:runtime_error` not `:tool_error`
- Sandbox-safe `list_traces` with bounded head/tail reads

### Changed

- Progress checklist renders for any agent with a `plan:`, regardless of `journaling:` setting
- `step-done` instruction text updated to reflect it is optional
- Bumped `req_llm` to `~> 1.8`

## [0.9.0] - 2026-02-27

### Added

- `SubAgent.chat/3` for multi-turn chat with conversation history threading
- `on_chunk` streaming callback for real-time token-by-token output in text mode
- `PtcRunner.LLM` behaviour with `call/2` and optional `stream/2` callbacks
- `PtcRunner.LLM.callback/2` convenience API with built-in ReqLLM adapter
- Graceful streaming degradation — `on_chunk` fires once with full content when adapter doesn't support streaming

### Documentation

- LLM Setup guide with provider configuration, streaming, custom adapters, and framework integration
- Phoenix Streaming guide for LiveView integration with `chat/3` and `on_chunk`
- Structured Output Callbacks guide for implementing LLM callbacks
- Added phoenix-streaming and structured-output-callbacks guides to ExDoc
- Updated getting-started guide with chat, streaming, and LLM adapter sections

## [0.8.0] - 2026-02-25

### Breaking Changes

- Renamed JSON mode to text mode — `:json` and `:tool_calling` unified into single `:text` output mode
- Removed backward compatibility shims for old mode names
- Renamed builtin tool names to match new text mode conventions
- Migrated system prompts from markdown headings to XML tags
- Removed `gpt-nano` and `gpt-mini` model entries from LLM registry

### Added

**Unified Text Mode**

- `TextMode` module replacing `JsonMode`, with separate `JsonHandler` for structured output
- Native tool calling mode for smaller LLMs that support API-level tool use
- `ToolSchema` module for generating tool schemas from signatures
- Guard against nil `assistant_content` in tool-call messages

**PTC-Lisp Enhancements**

- `defonce` special form for idempotent variable initialization
- `pr-str` function for readable string representation
- `#"..."` regex literal support
- `CoreToSource` module for Core AST to PTC-Lisp source serialization
- MapSet support for `some`, `every?`, `not-any?`, `join`, `split`, `replace`
- Preserved `tool_calls` and `prints` from inside HOF closures
- Preserved `tool_calls` and `tool_cache` across `loop`/`recur` iterations
- Handle `:var` nodes in `SymbolCounter`
- Handle `#'name` var reader syntax in analyzer
- Handle `defonce` in `collect_undefined_vars` static analysis
- Skip bare vars in `or` during static undefined-var analysis
- Treat unbound vars as nil in `or` for safe memory defaults
- `str` fixed to use Clojure syntax for collections
- Return nil for keyword lookup on non-map types (Clojure conformance)

**SubAgent Improvements**

- `max_tool_calls` limit to prevent runaway tool loops
- `pmap_max_concurrency` config to control parallel task limits
- SubAgent `name` propagated to `Step` for TraceTree and Debug display
- Journal/step-done prompt sections gated behind `journaling: true`
- Moved `defonce` docs from base prompt to multi-turn addon

**LLM Client**

- `embed/2,3` and `embed!/2,3` for embedding API support
- Groq provider support
- Bedrock inference profile support
- Migrated to ReqLLM pricing (removed `LLMClient.calculate_cost`)

**Tracing & Viewer**

- Plan progress display in `Debug` and `TraceTree`
- ptc_viewer: multi-run span tree layout styles
- ptc_viewer: collapse span tree groups by default with count badges
- ptc_viewer: draggable sidebar resizer and preserved scroll position
- Trace sanitize `max_map_size` limit to prevent heap overflow

**Examples**

- ALMA: evolutionary memory design for GraphWorld and ALFWorld environments
- ALMA: domain-blind `Environment` behaviour with multi-env support
- ALMA: vector store with cosine similarity and real embeddings
- ALMA: grep-based DebugAgent with ptc_viewer drill-in
- RLM recursive: `pmap_max_concurrency` tuning and LCM-inspired directions

### Changed

- Consolidated `gpt-oss` registry entries to 120B only
- Simplified planner livebook to two-role pattern, dropped reviewer
- Simplified README Calculator example to use auto-extracted signatures
- Bumped `credo` 1.7.15 → 1.7.16
- Bumped `req_llm` 1.2.0 → 1.5.1

### Fixed

- Skip structs in `KeyNormalizer.normalize_keys`
- Include `raw_response` in `turn.stop` telemetry for parse errors
- Add agent names to joke workflow livebook for better trace display

## [0.7.0] - 2026-02-12

### Breaking Changes

- Removed PTC-JSON language entirely (PTC-Lisp only)
- Removed `CapabilityRegistry` module (no proven use case yet)
- Removed redundant JSON CLI and `LispAgent` shim
- Renamed `return_retries` to `retry_turns`

### Added

**Plan System & Multi-Agent Orchestration**

- Plan system with `PlanRunner` and `PlanExecutor` for multi-agent workflows
- `MetaPlanner` with trial & error replanning and replan-on-failure
- Per-task quality gates with evidence-based verification and telemetry
- Direct agent for LLM-free task execution
- Upstream dependency result injection into task prompts
- `--plan-only` / `--plan` CLI flags and Lisp syntax validation for plans

**Journaled Task System**

- `(task "id" expr)` — idempotent journaled execution with journal-based caching
- Dynamic expressions as task IDs
- `step-done` and `task-reset` forms with plan progress tracking
- Journal preserved on error paths

**PTC-Lisp Enhancements**

- Tree traversal functions: `walk`, `prewalk`, `postwalk`, `tree-seq`
- `boolean` and `type` built-in functions
- `:when`, `:let`, `:while` modifiers for `for` and `doseq`
- Map support for `take`, `drop`, and `distinct` family
- Extended keyword chars for operator keywords (Clojure conformance)
- `index-of` and `last-index-of` string builtins

**Tracing & Observability**

- `ptc_viewer` web UI with interactive DAG graph visualization
- Gantt timeline for `pmap` parallel execution in trace viewer
- Expandable execution tree for recursive traces
- Turn pill badges and timeline overview
- Cross-process trace propagation via `TraceContext` module
- `PlanTracer` for plan-layer telemetry (phases, inputs, replan events)
- Trace result preview increased to 64KB
- `trace.stop` event with total duration on collector stop

**SubAgent Improvements**

- `thinking` option for SubAgent and demo CLI
- Configurable sandbox timeout and heap limits via application env
- Unified `builtin_tools` option (replaced `grep_tools`)
- Tool result caching and `child_steps` accumulation
- Pre-execution checks for undefined vars and unknown tools
- Prompt caching support for Anthropic, OpenRouter, and Bedrock

**Examples & Documentation**

- `page_index` example for hierarchical document retrieval with benchmarks
- `supply_watchdog` example project
- Meta Planner guide, Navigator guide, observability guide
- Plan-and-execute livebook and capability registry livebook

### Changed

- Centralized builtin tool injection into `SubAgent.effective_tools/1`
- Replaced `find_undefined_vars` with `Lisp.validate/1`
- Extracted `TraceContext` module for centralized trace propagation
- Refactored to use `Prompts` module for SubAgent loop
- Simplified `LanguageSpec` to use `Prompts` module
- Explicitly set `output: :ptc_lisp` in `PlanRunner`
- Bumped `req_llm` 1.2.0 → 1.5.1

### Fixed

- Prevented LLM thinking text from polluting message history
- Fixed XML-style `</clojure>` closers in code block parsing
- Made `grep` always-regex with BRE-to-PCRE auto-translation
- Hardened tracing reliability and removed duplicate tool telemetry
- Prevented `println`+`return` same turn conflict
- Fixed `(str x)` to convert single non-string arg to string
- Fixed false positive on `#"` check inside string literals
- Routed task failures to replan when `max_total_replans > 0`
- Prevented `TraceLog.Collector` crash when parent task is killed
- Collected child Steps on parent Step for TraceTree hierarchy
- Added named-arg usage example to tool signatures to prevent positional arg errors
- Scoped turn lookup by `span_id` to prevent cross-agent collisions in viewer
- Resolved dialyzer errors in linker and plan_executor

## [0.6.0] - 2026-01-30

### Added

**Language**

- `for` (minimal) list comprehension
- String functions: `.indexOf`, `.lastIndexOf`
- `builtin_tools` option for injecting builtin tools (e.g., `grep`, `grep-n`) instead of hardcoded builtins
- Collection functions: `extract`, `extract-int`, `pairs`, `combinations`, `mapcat`, `butlast`, `take-last`, `drop-last`, `partition-all`
- Aggregators: `sum`, `avg`, `quot`
- Reader literals: `##Inf`, `##-Inf`, `##NaN`

**SubAgent**

- `return_retries` for validation recovery with compression support
- `:self` sentinel for recursive agents
- `memory_strategy :rollback` for recoverable memory limit errors
- Budget introspection and callback for RLM patterns
- Last expression as return value on budget exhaustion
- `llm_query` builtin integrated into system prompts and tool normalization
- Auto-set `return_retries` for agents with tools during compile

**LLM-as-Tool Composition**

- `LLMTool` with `response_template` mode for typed LLM output
- Transparent tool unwrapping and input validation

**Tracing & Observability**

- `TraceLog` + `Analyzer` for structured SubAgent tracing
- Hierarchical tracing for nested SubAgents
- Chrome DevTools trace export
- HTML trace viewer
- Post-sandbox tool telemetry with span correlation

**Utilities**

- `PtcRunner.Chunker` for text chunking
- Configurable `pmap_timeout` for LLM-backed tools

### Changed

- Refactored SubAgent loop from recursive to iterative driver loop
- Extracted chaining, validation, and prompt modules into focused files

### Fixed

- Propagate `max_heap` option to Lisp.run and child agents
- Handle tool call positional args error gracefully
- Support `apply` with maps and variadic `max-by`/`min-by`

## [0.5.2] - 2026-01-23

### Added

- **Mustache Templates** - Standalone `PtcRunner.Mustache` module for template rendering (#719)
- **Unified SubAgent API** - CompiledAgent support with `then/3` for chaining (#709)
- Support `timeout` and `max_heap` options in compiled agent execution
- Allow SubAgentTools in compiled agents
- JSON reports with failure traces for demo benchmarks
- Signature naming convention documentation (underscores vs hyphens)
- Improved signature documentation and error messages (#715)

### Fixed

- Normalize hyphenated keys to underscores at tool boundary (#706)
- Enforce named args and string keys at tool boundary
- Normalize keys in `has_keys` constraint for better prompt clarity
- Return error for non-scalar Mustache variable expansion
- Use string keys for JSON mode and add `max_turns` for compile
- Allow `timeout` option in string convenience form
- Fix report filename extraction for Bedrock model IDs

## [0.5.1] - 2026-01-18

### Added

- **JSON Output Mode** - SubAgents can now return structured JSON instead of PTC-Lisp
  - Add `output:` field to SubAgent struct for declaring JSON schema
  - Add `Signature.to_json_schema/1` for JSON schema generation
  - Add `LLMClient.generate_object/4` for structured output generation
  - Add `LLMClient.callback/1` for SubAgent integration
  - Support array types and improved validation UX
- Add `re-seq` regex function to PTC-Lisp for extracting all matches
- Add debug mission display and tool call statistics with Clojure format output

### Fixed

- Convert keyword-style tool args to map in Lisp interpreter

## [0.5.0] - 2026-01-16

### Breaking Changes

- Replace `ctx/` namespace with `data/` and `tool/` namespaces for clearer separation
- Remove `tool_catalog` field from SubAgent (use `tools` directly)

### Added

**Observability & Message History (v0.5 theme)**

- Add `Turn` struct for immutable per-turn execution history with tool calls, prints, and memory snapshots
- Add `SingleUserCoalesced` compression strategy for token-efficient multi-turn conversations
- Add `compression: true` option to enable message compression in SubAgent
- Add `collect_messages: true` option to capture full conversation history
- Enhance `print_trace/2` with new options: `view: :compressed`, `messages: true`, `raw: true`, `usage: true`
- Add compression statistics to debug output
- Add prompt caching support by splitting static/dynamic sections

**New Functions**

- Add `distinct-by` for unique items by key function
- Add `re-split` for regex-based string splitting
- Add `rem` function and fix `mod` to match Clojure semantics
- Add multi-arity `map` and `partition` functions
- Add list index support to `get-in`, `assoc`, `update`, and related functions
- Add context filtering via static analysis to reduce memory pressure

**Other**

- Add configurable println truncation limit (`max_print_length` option)
- Add hidden fields filtering from LLM-visible output (fields starting with `_`)
- Add configurable sample limits and smart println for char lists
- Improve float support in PTC-Lisp

### Fixed

- Multi-arity map with variadic builtins
- Propagate `max_print_length` into closures and pcalls
- Show map field names in tool signatures for LLM
- Handle nil values in `Debug.print_trace` options
- Support builtin tuples in `fnil` for Clojure compatibility
- Show explicit "No tools available" message in prompt

## [0.4.1] - 2026-01-09

### Added

- Add `juxt` function combinator for multi-criteria operations
- Add variadic function support with rest parameters `[a & rest]`
- Add `max-key` and `min-key` for variadic comparisons
- Add IEEE 754 special values: `##Inf`, `##-Inf`, `##NaN`
- Add `float_precision` option to SubAgent (default: 2 decimal places)
- Add `context_descriptions` for automatic data inventory in prompts
- Extend `reduce` to work on maps, sets, and strings
- Add variadic `update` and `update-in` (match Clojure semantics)
- Add `java.time.LocalDate/parse` for date handling

### Fixed

- Preserve memory state on parse/analysis errors (multi-turn recovery)
- Handle `return`/`fail` correctly in threading macros (`->`, `->>`)
- Make `return`/`fail` terminate execution immediately
- Restore caller environment after closure execution
- Improve error messages with actionable suggestions

## [0.4.0] - 2026-01-06

### Added

- Add SubAgent API for high-level agent definition with type-safe signatures, auto-chaining, and resource limits
- Add Tracer system for immutable recording and visualization of agent execution
- Implement loop and recur support for iterative computation in PTC-Lisp
- Add character literals and string-as-sequence support for more flexible data handling
- Add `pcalls` for parallel execution of heterogeneous thunks
- Add `pmap` for parallel map evaluation
- Support vector paths in collection extraction functions for nested data access
- Add Clojure namespace normalization to improve LLM resilience

### Fixed

- Correct argument order for sort-by function to match Clojure semantics
- Fix update-vals argument order to match Clojure 1.11
- Update supported functions list (add frequencies, add float and for)
- Improve multi-turn agent guidance and system prompts
- Add specific error messages for predicate functions
- Fix Clojure compatibility for destructuring, count, and empty?

## [0.3.4] - 2025-12-25

### Added

- Add seqable map support to filter, remove, and sort-by operations
- Add entries and identity functions to PTC-Lisp
- Add sandbox support to PtcRunner.Lisp for resource limits

### Fixed

- Replace length() comparisons with Enum.empty? alternative
- Update error handling to use error tuples instead of raised exceptions

## [0.3.3] - 2025-12-22

### Added

- Add `update` and `update-in` map bindings for transforming values with functions
- Add function-based key support to `*-by` operations for custom sorting and grouping
- Add spec validation system for PTC-Lisp with multi-line examples and section reporting
- Improve JSON DSL prompts for better LLM accuracy

### Fixed

- Fix JSON agent to retry on empty LLM responses
- Improve deterministic ordering in keys/vals output
- Align `assoc-in` and `update-in` with Clojure semantics for intermediate path creation
- Correct `update/3` semantics to pass nil to function for missing keys
- Fix zip and into operations to return vectors instead of tuples
- Handle empty and nil LLM responses gracefully in agent loop

## [0.3.2] - 2025-12-20

### Added

- Add format_error/1 for human-readable error messages

### Fixed

- Include ptc-lisp-llm-guide.md in hex package

## [0.3.1] - 2025-12-13

### Added

- Improve PTC-JSON system prompt for better LLM accuracy
- Add object operation to construct maps with evaluated values (#253) (#254) ([#254](https://github.com/andreasronge/ptc_runner/pull/254))
- Enhance Clojure validation to execute and compare results
- Add auto-generated report filenames and reports directory
- Add cross-dataset join test case and clean up old reports
- Add --show-prompt option to display system prompts
- Add arithmetic operations (add, sub, mul, div, round, pct) #255
- Add membership operations (in, filter_in) (#257) (#259) ([#259](https://github.com/andreasronge/ptc_runner/pull/259))
- Add implicit object literals for memory storage (#256) (#261) ([#261](https://github.com/andreasronge/ptc_runner/pull/261))

### Fixed

- Handle Map values in constraint errors and fix GenServer timeout
- Correct round operation documentation for precision constraints
- Improve LLM prompt with arithmetic ops and better examples
- Evaluate filter_in value when it's a DSL expression
- Add sort_by order:desc to LLM prompt

## [0.3.0] - 2025-12-11

### Added

- Add PTC-Lisp LLM generation benchmark (Phase 1)
- Improve generation and judge prompts for PTC-Lisp benchmark
- Improve benchmark with edge cases, better judge, and dry run output
- Add autonomous issue creation and GitHub Project integration to PM workflow
- Enhance PM workflow with tech debt priority and efficiency fixes
- Auto-trigger implementation on ready-for-implementation label
- Auto-trigger code review for PRs from claude/* branches
- Install git pre-commit hook in Claude workflow
- Create PtcRunner.Json public API and deprecate PtcRunner (#103) ([#103](https://github.com/andreasronge/ptc_runner/pull/103))
- Allow full Bash access in claude.yml workflow
- Implement PTC-Lisp parser infrastructure (Phase 1) - Closes #106 (#107) ([#107](https://github.com/andreasronge/ptc_runner/pull/107))
- Implement PTC-Lisp analyzer infrastructure (Phase 2) - Closes #108 (#109) ([#109](https://github.com/andreasronge/ptc_runner/pull/109))
- Implement PTC-Lisp eval infrastructure (Phase 1) - Closes #111 (#112) ([#112](https://github.com/andreasronge/ptc_runner/pull/112))
- Implement PtcRunner.Lisp entry point with memory contract - Closes #115 (#116) ([#116](https://github.com/andreasronge/ptc_runner/pull/116))
- Add hourly schedule trigger to PM workflow
- Add pre-computed phase status to PM workflow prompt
- Implement LispGenerators module with StreamData generators (#130) (#132) ([#132](https://github.com/andreasronge/ptc_runner/pull/132))
- Add property tests for evaluation safety and determinism (#133) (#134) ([#134](https://github.com/andreasronge/ptc_runner/pull/134))
- Add domain property tests for arithmetic, collections, types, and logic (#135) (#136) ([#136](https://github.com/andreasronge/ptc_runner/pull/136))
- Support flexible key access in where clause field accessors (#137) (#138) ([#138](https://github.com/andreasronge/ptc_runner/pull/138))
- Add Lisp.Schema module and extend Runtime with flexible key access (#139) ([#139](https://github.com/andreasronge/ptc_runner/pull/139))
- Add truncation hints to guide LLM query refinement
- Add PTC-Lisp CLI and enhance demo infrastructure
- Refactor PM workflow to use Epic Issue pattern
- Add LispTestRunner and improve multi-turn support
- Add file size analysis to PR review workflow
- Add #{...} set literal syntax support (Phase 1 of #164) (#166) ([#166](https://github.com/andreasronge/ptc_runner/pull/166))
- Add {:set, [t()]} to AST type specifications (#167) (#168) ([#168](https://github.com/andreasronge/ptc_runner/pull/168))
- Add set analysis support (Phase 3 of #164) (#170) ([#170](https://github.com/andreasronge/ptc_runner/pull/170))
- Add set evaluation support (Phase 4 of #164) (#172) ([#172](https://github.com/andreasronge/ptc_runner/pull/172))
- Add .env support and model selection for e2e tests
- Add flex_fetch/2 and flex_get_in/2 to Runtime module (#188) ([#188](https://github.com/andreasronge/ptc_runner/pull/188))
- Add update-vals for map value transformation
- Create TestRunner.Base with shared constraint/formatting functions (#197) ([#197](https://github.com/andreasronge/ptc_runner/pull/197))
- Create TestRunner.Report with markdown generation (#199) ([#199](https://github.com/andreasronge/ptc_runner/pull/199))
- Create TestRunner.TestCase with shared test definitions (#201) ([#201](https://github.com/andreasronge/ptc_runner/pull/201))
- Create CLIBase with shared CLI utilities (#203) ([#203](https://github.com/andreasronge/ptc_runner/pull/203))
- Set up demo test infrastructure (MockAgent, test config) - Closes #205 (#206) ([#206](https://github.com/andreasronge/ptc_runner/pull/206))
- Create JsonTestRunner with shared modules support
- Create JsonCLI module with test mode support (#217) ([#217](https://github.com/andreasronge/ptc_runner/pull/217))
- Add memory support to JSON Agent (#220) (#221) ([#221](https://github.com/andreasronge/ptc_runner/pull/221))
- Add agent injection to test runners for MockAgent testing (#222) (#223) ([#223](https://github.com/andreasronge/ptc_runner/pull/223))
- Add ModelRegistry and unify test cases (#227) ([#227](https://github.com/andreasronge/ptc_runner/pull/227))
- Add --runs=N option for running tests multiple times
- Add keyword/string type coercion to where clause comparisons (#232) (#233) ([#233](https://github.com/andreasronge/ptc_runner/pull/233))
- Align JSON DSL memory model with Lisp (#234)
- Add take, drop, and distinct operations to JSON DSL (#236) (#243) ([#243](https://github.com/andreasronge/ptc_runner/pull/243))
- Add enhanced stats to demo test runner report (#246) (#249) ([#249](https://github.com/andreasronge/ptc_runner/pull/249))

### Fixed

- Move PM prompt to command file to fix expression length limit
- Use Bash(gh:*) pattern for PM workflow
- Trigger PM workflow on claude-approved label too
- Re-trigger code review on sync for claude/* branches
- Use --force in precommit to catch stale .beam files
- Add spec document verification to code review prompt
- Include PR comments and review comments in claude.yml
- Add mkdir permission to claude.yml workflow
- Add explicit Claude CLI install to workaround action bug
- Add safety net to push unpushed commits in PR fix workflow
- Mark PTC-Lisp implementation checklist items as complete (#123) ([#123](https://github.com/andreasronge/ptc_runner/pull/123))
- Update README with PTC-Lisp announcement and API migration guidance
- Complete API migration in Integration with LLMs section
- Implement compile-time extraction for PTC-Lisp schema prompt (#144) ([#144](https://github.com/andreasronge/ptc_runner/pull/144))
- Configure StreamData to run 300 iterations in CI (#146) ([#146](https://github.com/andreasronge/ptc_runner/pull/146))
- Make issue review always update the issue body
- Add sequential destructuring pattern type to CoreAST (#149) ([#149](https://github.com/andreasronge/ptc_runner/pull/149))
- Extend analyze_pattern for vector destructuring patterns
- Complete PR #151 - Add fn parameter destructuring documentation and tests
- Complete PR #151 - Remove stale documentation and add insufficient elements test
- Complete PR #151 - Remove stale documentation and add insufficient elements test
- Add E2E test for group-by with destructuring (#153) ([#153](https://github.com/andreasronge/ptc_runner/pull/153))
- Add analyzer unit tests for fn parameter destructuring patterns (#155) ([#155](https://github.com/andreasronge/ptc_runner/pull/155))
- Add evaluator unit tests for fn parameter destructuring patterns (#157) ([#157](https://github.com/andreasronge/ptc_runner/pull/157))
- Update LLM guide map example to use fn destructuring syntax (#159) ([#159](https://github.com/andreasronge/ptc_runner/pull/159))
- Enable sort-by with comparator and builtin HOF arguments (#160) ([#160](https://github.com/andreasronge/ptc_runner/pull/160))
- Extend multi-arity support to get and get-in (#163) ([#163](https://github.com/andreasronge/ptc_runner/pull/163))
- Unify concurrency groups for Claude issue workflows
- Add MapSet-safe collection operations and set runtime support (#175) ([#175](https://github.com/andreasronge/ptc_runner/pull/175))
- Add set literal formatting support to formatter (Phase 6 of #164) (#178) ([#178](https://github.com/andreasronge/ptc_runner/pull/178))
- Add test coverage for remove, mapv, empty?, and count on sets (#181) ([#181](https://github.com/andreasronge/ptc_runner/pull/181))
- Split eval_test.exs into multiple focused test files (#182) ([#182](https://github.com/andreasronge/ptc_runner/pull/182))
- Extract shared dummy_tool test helper (#183) (#184) ([#184](https://github.com/andreasronge/ptc_runner/pull/184))
- Support string key parameters in Lisp runtime functions (#185) ([#185](https://github.com/andreasronge/ptc_runner/pull/185))
- Standardize OpenAI model to gpt-5.1-codex-mini
- Rename duplicate module name in integration_test.exs
- Wire all call sites to use flex_fetch/flex_get_in for string/atom key interop
- Add integration tests and update docs for flexible key access (Phase 3)
- Update docs for flexible key access implementation
- Add @doc annotation to flex_get for API consistency
- Update ptc-lisp-overview.md to reflect completed flex key access (#192) ([#192](https://github.com/andreasronge/ptc_runner/pull/192))
- Update format_error references to PtcRunner.Json.format_error
- Update CHANGELOG format_error reference
- Change update-vals argument order to match Clojure 1.11
- Remove duplicate incorrect update-vals signature from LLM guide
- Handle FunctionClauseError in builtins with descriptive type errors
- Handle FunctionClauseError in multi-arity functions and complete type error messages
- Delete old TestRunner module and update README references (#219) ([#219](https://github.com/andreasronge/ptc_runner/pull/219))
- Require closing keyword in PR body for auto-close
- Add --report option to Lisp CLI Options table
- Update demo CLI to use ModelRegistry.resolve pattern (#229) ([#229](https://github.com/andreasronge/ptc_runner/pull/229))
- Update guide.md to reflect new JSON DSL API signature
- Update guide.md and demo to use new 4-tuple return format
- Handle invalid map destructuring syntax gracefully in analyzer
- Improve error message for update-vals with swapped arguments
- Update JSON agent to use new memory model API (#235) (#241) ([#241](https://github.com/andreasronge/ptc_runner/pull/241))
- Filter nil opts in CLI to allow Keyword.get defaults
- Split transformation_test.exs into access_test.exs and collection_test.exs (#244) (#247) ([#247](https://github.com/andreasronge/ptc_runner/pull/247))
- Align PTC-Lisp semantics with Clojure specification (#245) (#248) ([#248](https://github.com/andreasronge/ptc_runner/pull/248))
- Resolve remaining Clojure conformance test failures (#250) ([#250](https://github.com/andreasronge/ptc_runner/pull/250))

## [0.2.0] - 2025-12-05

### Added

- Add introspection operations (keys, typeof) to DSL (#92) ([#92](https://github.com/andreasronge/ptc_runner/pull/92))
- Improve DSL consistency for better LLM program generation (#94) ([#94](https://github.com/andreasronge/ptc_runner/pull/94))
- Add explore mode for schema discovery (#97) ([#97](https://github.com/andreasronge/ptc_runner/pull/97))
- Enable async execution for test modules (#98) ([#98](https://github.com/andreasronge/ptc_runner/pull/98))
## [0.1.0] - 2025-12-03

### Added

- Add CI check to verify STATUS.md is updated in PRs
- Implement Phase 1 core interpreter with JSON parsing and sandbox execution (#10) ([#10](https://github.com/andreasronge/ptc_runner/pull/10))
- Add pre-implementation check for blockers in PM workflow
- Implement get operation for nested path access (fixes #17) (#18) ([#18](https://github.com/andreasronge/ptc_runner/pull/18))
- Implement comparison operations (neq, gt, gte, lt, lte) (#22) ([#22](https://github.com/andreasronge/ptc_runner/pull/22))
- Implement collection operations (first, last, nth, reject) (#26) (#27) ([#27](https://github.com/andreasronge/ptc_runner/pull/27))
- Implement contains, avg, min, max operations (#28)
- Implement let variable bindings for Phase 3 (#30) (#31) ([#31](https://github.com/andreasronge/ptc_runner/pull/31))
- Implement if conditional operation for Phase 3 (#32) (#33) ([#33](https://github.com/andreasronge/ptc_runner/pull/33))
- Implement boolean logic operations (and, or, not) for Phase 3 (#34) (#35) ([#35](https://github.com/andreasronge/ptc_runner/pull/35))
- Implement combine operations (merge, concat, zip) for Phase 3 (#37) ([#37](https://github.com/andreasronge/ptc_runner/pull/37))
- Implement call operation for tool invocation (#41) ([#41](https://github.com/andreasronge/ptc_runner/pull/41))
- Add Jaro-Winkler typo suggestions for unknown operations (#44) ([#44](https://github.com/andreasronge/ptc_runner/pull/44))
- Add ExDoc and Hex package metadata (#45) (#46) ([#46](https://github.com/andreasronge/ptc_runner/pull/46))
- Implement declarative schema module for DSL operations (#52) ([#52](https://github.com/andreasronge/ptc_runner/pull/52))
- [Phase 5] JSON Schema Generation (#50) (#55) ([#55](https://github.com/andreasronge/ptc_runner/pull/55))
- [Phase 5] E2E LLM Testing Infrastructure (#51) (#57) ([#57](https://github.com/andreasronge/ptc_runner/pull/57))
- Adopt program wrapper as canonical PTC format - Update to_json_schema/0 (#63) ([#63](https://github.com/andreasronge/ptc_runner/pull/63))
- Adopt program wrapper as canonical PTC format in parser (#58) (#64) ([#64](https://github.com/andreasronge/ptc_runner/pull/64))
- Add structured output support with generate_program_structured! for E2E tests (#65) (#67) ([#67](https://github.com/andreasronge/ptc_runner/pull/67))
- Validate tool function arities at registration time (#42) (#68) ([#68](https://github.com/andreasronge/ptc_runner/pull/68))
- Add interactive demo CLI for PTC with ReqLLM integration (#75) ([#75](https://github.com/andreasronge/ptc_runner/pull/75))
- Add to_prompt/0 for token-efficient LLM text mode (#80) ([#80](https://github.com/andreasronge/ptc_runner/pull/80))
- Add security gates and hardening to Claude workflows

### Fixed

- Add safety improvements to GitHub workflows
- PM workflow commits STATUS.md directly to main
- Avoid parallel PRs by including STATUS.md in implementation PR
- Simplify STATUS.md update rules to prevent merge conflicts
- Improve PM workflow action handling
- Trigger PM workflow when issue becomes ready-for-implementation
- Ensure git push happens immediately after commit in Claude workflow
- Use PAT in issue-review workflow to trigger PM workflow
- Optimize min_list and max_list performance and update avg docs
- Correct documentation for sum vs avg behavior with non-numeric values
- Use anyOf for nested expressions in LLM schema (#71) ([#71](https://github.com/andreasronge/ptc_runner/pull/71))
- Improve LLM schema descriptions and use Haiku 4.5 (#73) ([#73](https://github.com/andreasronge/ptc_runner/pull/73))
- Store last_result in Agent state to avoid regenerating random data (#79) ([#79](https://github.com/andreasronge/ptc_runner/pull/79))
- Add test_coverage configuration to exclude test support modules (#89) ([#89](https://github.com/andreasronge/ptc_runner/pull/89))
[0.14.0]: https://github.com/andreasronge/ptc_runner/compare/v0.13.0...v0.14.0
[0.13.0]: https://github.com/andreasronge/ptc_runner/compare/v0.12.0...v0.13.0
[0.12.0]: https://github.com/andreasronge/ptc_runner/compare/v0.11.0...v0.12.0
[0.11.0]: https://github.com/andreasronge/ptc_runner/compare/v0.10.1...v0.11.0
[0.10.1]: https://github.com/andreasronge/ptc_runner/compare/v0.10.0...v0.10.1
[0.10.0]: https://github.com/andreasronge/ptc_runner/compare/v0.9.0...v0.10.0
[0.9.0]: https://github.com/andreasronge/ptc_runner/compare/v0.8.0...v0.9.0
[0.8.0]: https://github.com/andreasronge/ptc_runner/compare/v0.7.0...v0.8.0
[0.7.0]: https://github.com/andreasronge/ptc_runner/compare/v0.6.0...v0.7.0
[0.6.0]: https://github.com/andreasronge/ptc_runner/compare/v0.5.2...v0.6.0
[0.5.2]: https://github.com/andreasronge/ptc_runner/compare/v0.5.1...v0.5.2
[0.5.1]: https://github.com/andreasronge/ptc_runner/compare/v0.5.0...v0.5.1
[0.5.0]: https://github.com/andreasronge/ptc_runner/compare/v0.4.1...v0.5.0
[0.4.1]: https://github.com/andreasronge/ptc_runner/compare/v0.4.0...v0.4.1
[0.4.0]: https://github.com/andreasronge/ptc_runner/compare/v0.3.4...v0.4.0
[0.3.4]: https://github.com/andreasronge/ptc_runner/compare/v0.3.3...v0.3.4
[0.3.3]: https://github.com/andreasronge/ptc_runner/compare/v0.3.2...v0.3.3
[0.3.2]: https://github.com/andreasronge/ptc_runner/compare/v0.3.1...v0.3.2
[0.3.1]: https://github.com/andreasronge/ptc_runner/compare/v0.3.0...v0.3.1
[0.3.0]: https://github.com/andreasronge/ptc_runner/compare/v0.2.0...v0.3.0
[0.2.0]: https://github.com/andreasronge/ptc_runner/compare/v0.1.0...v0.2.0
[0.1.0]: https://github.com/andreasronge/ptc_runner/releases/tag/v0.1.0