v0.1.13

Framework v0.1.13

Architectural Unification

• Guard evaluation: 5 entry points → 1 unified evaluate() with GuardBehavior enum
• RecordProcessor: 4 independent instantiations → single create() factory
• Lineage building: deduplicated across exhausted/passthrough builders
• Context builder: 235-line monolith → 4 focused functions (59-line orchestrator)
• Dead code deleted: HistoricalNodeDataLoader, BatchProcessor, apply_observe_filter (-2992 lines)

Data Model Fixes

• Fan-in merges all dependency branches via bus model (not just primary)
• Guard-skipped FILE-mode records get proper null namespace
• Guard-skip flat field leak fixed (pre-observe records used for tombstones)

UI Improvements

• Action Output shows only this action's fields, not the entire data bus
• Nested JSON renders as expandable tree nodes with syntax highlighting
• Guard-skipped records show clear indicator instead of dumping 25 namespaces
• Single unwrap path: scanner unwraps, frontend renders (no duplicate logic)

Skills

• Rewritten from ban-list to tool guide (positive teaching style)
• All 15 agentic patterns with copyable YAML configs
• Scaffolding templates for adding LLM, Record tool, and FILE tool actions

VS Code Extension v0.1.19

• Same UI improvements as Data Explorer
• Published to marketplace: https://marketplace.visualstudio.com/items?itemName=runagac.agent-actions

Examples

• All 6 example workflows validated on gemini-flash-lite-latest

v0.1.12

v0.1.12 — 2026-04-19

Enhancement or New Feature

• Self-reflection feedback strategy for reprompt retries
• LLM critique escalation for stubborn reprompt failures

Under the Hood

• Unified online and batch reprompt logic via shared helpers

Bug Fix

• Drop directives now correctly exclude fields from passthrough wildcards
• Pass compiled schemas to Groq, Cohere, and Mistral APIs
• Retry transient OpenAI 400 JSON parsing errors
• Cascade failures report as blocked instead of independent errors; batch-mode shows wall-clock duration
• Cross-workflow records deliver content and propagate source_guid
• Reprompt batch resolves source.* observe fields via storage backend; --downstream defers on batch pause
• FILE-mode output pipeline: fix lineage collision in _reattach_source_guid
• Batch resubmission loop: prevent resubmission when upstream guard filters records
• Guard-filtered upstream fields no longer crash downstream schema validation
• Prevent non-serializable data in OpenAI request payloads
• dispatch_task() resolves UDF functions instead of passing literal strings to LLM
• Warn when inline schema shadows schema_name

v0.1.11

v0.1.11

Enhancement or New Feature

• Cross-workflow chaining: --upstream and --downstream CLI flags, virtual action injection, and orchestrator support for multi-workflow pipelines
• FILE-mode record identity: tools receive full records with node_id, framework matches outputs to inputs by identity instead of heuristics (breaking change for FILE tools — use record["content"]["field"] instead of record["field"])
• Step-by-step progress display for sequential execution mode
• Add --example flag to agac init for scaffolding projects from bundled examples
• Add doc-vs-code audit framework for retry and reprompt modules

Bug Fix

• Fix FILE-mode lineage: preserve content from full records returned by FILE tools, and skip source data loading in cross-workflow runs
• Resolve .env files from project root instead of CWD
• Guard conditions can now reference output_field values directly

v0.1.10

Bug Fixes

• fix: HITL FILE granularity + guard pre-filter for FILE mode (#229)
• fix: HITL FILE mode truncates lineage to just its own node_id (#231)
• fix: resolve false-positive StaticTypeError for Jinja loop variables (#232)
• fix: accept cross-workflow dependency syntax, fix 4 crash sites
• fix: improve VS Code data card indentation, grammar, and rendering (#223)
• fix: improve docs data card indentation, contrast, and grammar (#224)

Tests

• test: deterministic lineage integrity tests (#221)
• test: context scope integration tests (#222)

Docs

• docs: update HITL, granularity, and guards docs for FILE-only HITL and guard pre-filter (#230)

v0.1.9

Bug Fixes

• fix: resolve FileUDFResult lineage by source_mapping index, not source_guid scan (#220)
  • FILE-mode tools with multiple input records sharing the same source_guid (e.g. after flatten) now produce correct per-record lineage instead of all inheriting the first match
  • Fixes silent downstream data duplication when using context_scope.observe on pre-flatten ancestors
  • source_mapping is now fully consumed by the runtime — no longer validated-then-discarded

v0.1.8

Features

• feat: enforce context_scope as sole data gate (#199)
• feat: namespace llm_context by action to prevent version data loss (#208)
• feat: redesign VS Code data card with 5-section tree view (#215, #216)

Bug Fixes

• fix: deterministic record matcher — exact node_id join, no fallbacks (#195)
• fix(vscode): safe JSON injection for card view records (#197)
• fix(cla): grant contents write permission for signature commits (#200, #201)
• fix: resolve context_scope field prefix patterns for version_consumption merge (#202)
• fix: template error hints reference context_scope.observe (#205)
• fix: guard null context_scope to prevent crash on YAML indentation errors (#211)
• fix: hint at action namespacing when schema validation fails on tool UDF output (#214)

Refactoring

• refactor: eliminate field prefix pattern from version resolution (#206)
• refactor: simplify SKILL.md with Anthropic skill patterns (#210)

Docs

• docs: align docs with recent PRs — namespaced context, record_limit, guards, prompt traces (#209)

Chores

• chore: add Contributor License Agreement and CLA bot (#198)

v0.1.7

Features

• feat(storage): prompt trace storage for compilation-level observability (#184)
• feat(workflow): support record_limit on downstream actions (#187)
• feat(ui): surface prompt traces in Data Explorer and VS Code extension (#189)

Bug Fixes

• fix(input): remove double-parse of JSON in staging pipeline (#185)
• fix: guard error taxonomy and circuit breaker for semantic errors (#186)
• fix: preflight guard condition AST validation for unquoted strings (#188)

Tests

• fix(tests): eliminate 14 theater unit tests (#190)

v0.1.6

Bug Fixes

• fix(docs): render images in workflow README tab (#170)
• fix(docs): logs badge green when count is zero (#171)
• fix(docs): use absolute URLs for README images on PyPI (#173)
• fix(validation): scope prompt validator to current workflow (#182)

Documentation

• Add Project Surface and Dependencies to all 17 module manifests
• Tighten AGENTS.md link to manifest system (#169)

v0.1.5: fix: rename "LLM workflow" to "agentic workflow" (#156)

What's New

• Rich execution summary — agac run now renders a structured post-execution tree showing per-action status, type badges, provider, and latency
• Configurable seed data path — set seed_data_path in agent_actions.yml to rename your seed data directory
• Unified agac docs command — replaces separate docs generate + docs serve with a single command
• Cleaner CLI output — demoted 14x "fallback heuristic" spam and other noise to debug level

Fixes

• Place example READMEs where docs scanner expects them (README tab now works)
• Correct pass rate calculation and surface failed runs in logs
• Rebuild frontend to remove stale qanalabs references
• Remove dead ephemeral field from codebase
• Path traversal validation on seed_data_path config values
• LSP resolver now respects custom seed_data_path

DX Improvements

• agac init now scaffolds seed_data/ directory with config defaults
• Proper singular/plural in CLI messages ("1 Tool" not "1 Tools")
• Homepage stats polish and redesigned run detail view

v0.1.4

What's New

• Runtime tab added to docs site logs screen
• Gantt-style execution timeline on runs detail screen
• Action metrics enriched with latency, provider, cache, and disposition
• Runtime warnings extracted from events.json into catalog
• Project name field added to agent_actions.yml
• Retry failed/skipped actions on workflow re-run
• Pre-flight guard checks for nullable fields flowing into tool schemas
• ActionStatus enum and DEFERRED records audit trail (ARCH-003, ARCH-004, ARCH-006)
• Shared pipeline context prevents batch/online drift (ARCH-005)
• Partial failure visibility with completed_with_failures status

Fixes

• LSP hover shows all metadata, not just name and line
• VS Code extension expands version ranges correctly
• Batch validation rejects synchronous tool/HITL actions
• WhereClauseParser silent operator mismapping resolved
• Tool-action failure surfacing and zero-success detection
• Guard-skipped actions marked correctly in tally
• Hardcoded references removed across codebase

Docs & DX

• .env.example files added to all example projects
• Examples tightened with schemas, flow diagrams, and improved READMEs
• Ollama batch support docs updated