Skip to content

CHANGELOG.md

Latest commit

 

History

History
793 lines (668 loc) · 38.7 KB

CHANGELOG.md

File metadata and controls

793 lines (668 loc) · 38.7 KB

Changelog

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

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

1.9.0 - 2026-04-26

Added

  • OpenCode CLI agent adapter: configure with agent.kind: opencode for autonomous issue-to-code workflows using the OpenCode CLI via opencode run --format json. Supports fork-per-turn execution, JSON event normalization, SSH remote dispatch, permission policy synthesis, token accounting, and companion deployment examples via examples/docker/opencode.Dockerfile and examples/WORKFLOW.opencode.md. (#476, #478, #479)

Fixed

  • SSH worker command assembly: multi-token remote commands are now appended verbatim instead of shell-quoted as a single word, so remote agent launches such as codex app-server and other pre-formed shell fragments no longer fail with command not found. (#493)
  • GitHub and Jira tracker adapters no longer return partially populated issues, slices, or state maps when tracked fetch operations fail; the nil-on-error contract is preserved for issue detail and state lookups. (PR #496)

1.8.0 - 2026-04-17

Added

  • Codex CLI agent adapter: configure with agent.kind: codex for autonomous issue-to-code workflows using OpenAI Codex CLI via the codex app-server JSON-RPC 2.0 protocol. Supports the same structured lifecycle as Claude Code and Copilot CLI adapters — event normalization, token tracking, timeout enforcement, graceful SIGTERM→SIGKILL shutdown, and session resume via ResumeSessionID. Tool calls are serialized through a channel to prevent concurrent stdin corruption. Handshake reads are cancellable via context, preventing stalled subprocesses from hanging the worker indefinitely. (#238)

1.7.1 - 2026-04-15

Changed

  • CLI: --version now outputs a single diagnostic line including commit SHA, build date, Go version, and OS/architecture — e.g. sortie 1.7.0 (commit: a1b2c3d, built: 2026-04-15, go1.26.1, linux/amd64). The previous GNU-style copyright/warranty block is removed. Build tooling (Makefile, Dockerfile, .goreleaser.yaml, and the release workflow) now injects Commit and Date via -ldflags at all build sites.

Fixed

  • Orchestrator: sortie_ci_escalations_total over-counted during CI escalation. escalateCIFailure incremented the metric unconditionally before calling the tracker API, then incremented again on error — producing two increments for one failed operation. It also incremented when TrackerAdapter was nil, recording a phantom escalation that was never performed. Both defects are fixed; the metric now increments exactly once per operation outcome, matching the pattern in escalateReviewFailure. (#449)

1.7.0 - 2026-04-13

Added

  • Cross-retry session resume: continuation retries now propagate the session ID from the exiting worker through the retry entry so the next worker can resume the agent conversation instead of starting a fresh session. The session ID is persisted to SQLite and restored on startup recovery. (#207)
  • Token usage cost estimation on the dashboard and JSON API: operators configure per-adapter token rates in WORKFLOW.md front matter (token_rates block); the dashboard surfaces per-session and aggregate USD cost estimates computed from running sessions. The JSON API includes active_estimated_cost_usd when token rates are configured. (#436)
  • Dashboard accordion tables: Running Sessions, Retry Queue, and Run History tables use an expand/collapse accordion pattern. Primary status columns are visible at a glance; secondary detail expands on click. Expansion state survives the 5-second auto-refresh via sessionStorage. (#432, #444)
  • StderrCollector buffer hardening: agent stderr collection now uses a 10 MiB scanner buffer cap and a head/tail ring-buffer retention strategy with a configurable byte budget, preventing silent line truncation and unbounded memory growth during long agent turns. (#387)

Changed

  • Retry timer tracker validation: HandleRetryTimer now calls FetchIssueByID instead of scanning all candidate issues via FetchCandidateIssues, reducing each retry timer fire from O(pages) tracker API calls to exactly one. (#206)

Migrations

  • 009: Add nullable session_id TEXT column to retry_entries

1.6.1 - 2026-04-11

Added

  • Runtime terminal workspace sweep: workspace directories for issues that reach a terminal tracker state after their worker has exited are now cleaned up periodically by the event loop. This closes the gap between the startup sweep and in-flight reconciliation, preventing unbounded disk accumulation on long-running instances. (#428)

Fixed

  • Orchestrator: needs-human-review soft-stop now correctly triggers the handoff transition. Previously, the SoftStop branch in HandleWorkerExit matched all soft-stop reasons before the handoff case was reached, leaving the issue active and causing an infinite re-dispatch loop. (#426)

1.6.0 - 2026-04-10

Added

  • Self-review loop before PR creation: the orchestrator generates a workspace diff, executes configurable verification commands (tests, linters), assembles a structured review prompt, and iterates with the agent up to a configurable cap before proceeding. Opt-in via the self_review: block in WORKFLOW.md front matter (max_iterations, verify_commands, diff_max_bytes). (#312)
  • PR review comment routing: when a reviewer requests changes on an agent-created PR, the orchestrator detects CHANGES_REQUESTED reviews, extracts the review comments, and dispatches a continuation turn so the agent can address feedback automatically. Configurable via reactions.review_comments in WORKFLOW.md (max_retries, debounce_ms, escalation, escalation_label). Includes SCMAdapter domain interface for PR and review operations with a GitHub Checks/Reviews API implementation. (#305)
  • Unified reactions config block in WORKFLOW.md for event-driven continuation triggers. reactions.ci_failure replaces the top-level ci_feedback key (which remains supported for backward compatibility). Each reaction type shares provider, max_retries, escalation, and escalation_label fields. (#418)
  • Windows process lifecycle support for agent adapters and workspace hooks. Agent subprocesses are now placed in Windows Job Objects with JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE, enabling full process tree cleanup on timeout or cancellation. Graceful shutdown sends CTRL_BREAK_EVENT to the process group; force-terminate uses TerminateJobObject. Workspace hooks execute via cmd.exe /C on Windows with their own Job Object for timeout enforcement. The procutil package exposes cross-platform functions — SignalGraceful, AssignProcess, CleanupProcess, SetProcessGroup, KillProcessGroup — and WasSignaled is now platform-aware. Adapters no longer reference syscall.SIGTERM directly. (#390, #391)

Changed

  • CI pending backoff base now derives from the operator-configured poll_interval instead of a hardcoded 10 s default. A 30 s poll interval produces a (60 s, 120 s, 240 s, 300 s…) backoff schedule. Falls back to 10 s when poll_interval is zero or negative. (#385)

Deprecated

  • ci_feedback top-level config key in WORKFLOW.md. Use reactions.ci_failure instead. The legacy key continues to work; when both are present, reactions.ci_failure takes precedence. (#418)

Fixed

  • Workflow Manager continued using the pre-reconfiguration logger after logging.level was applied from WORKFLOW.md extensions, causing reload diagnostics to use the wrong log level. The Manager now updates its logger after every reconfiguration. (#394)
  • Reaction dispatch fingerprinting: MarkReactionDispatched was called at schedule time rather than actual dispatch time. If the process restarted between scheduling and dispatch, the fingerprint was permanently marked dispatched while the retry entry was lost, silencing future CI-fix reactions for that SHA. (#420)
  • Config: non-string YAML values in reactions fields (provider, escalation, escalation_label) now produce a ConfigError instead of silently coercing to an empty string. (#423)
  • install.sh: detect Rosetta 2 on macOS and prefer the native arm64 binary over amd64. (#391)

Migrations

  • 007: Add nullable review_metadata TEXT column to run_history
  • 008: Add reaction_fingerprints table for cross-restart reaction deduplication

1.5.1 - 2026-04-08

Added

  • GNU-style CLI help output with grouped sections, column-aligned descriptions, usage examples, and a "Learn more" link. Short aliases -h (help) and -V (version) are now recognized. Help text prints to stdout instead of stderr. Covers all three commands: sortie, sortie validate, sortie mcp-server. (#398)

Fixed

  • Copilot CLI adapter: prefix --additional-mcp-config file paths with @ to match the documented Copilot CLI syntax. Bare file paths worked in Copilot CLI ≤1.0.18 via an undocumented fallback that was removed in v1.0.21, causing Invalid JSON in --additional-mcp-config on every turn. Operator-provided copilot-cli.mcp_config values are now auto-detected: inline JSON is passed unchanged, @-prefixed paths are preserved, and bare file paths receive the @ prefix automatically. (#404)
  • Agent adapters: reclassify exit-code-0 turns with zero output tokens and no result event as turn_failed instead of turn_completed. Previously, when an agent subprocess crashed immediately (e.g., MCP config parse error) but exited 0, all turns were counted as successful, causing the orchestrator to exhaust max_turns and trigger a false-positive handoff transition. Failed turns now retry with exponential backoff. Applies to both Claude Code and Copilot CLI adapters. (#404)

1.5.0 - 2026-04-07

Added

  • Always-on HTTP server with default port 7678 (mnemonic: SORT on T9). The server now starts unconditionally; the --port flag overrides the default but no longer acts as an activation trigger. Pass --port=0 to disable. Prometheus metrics, health probes, and dashboard are available out of the box without flags.
  • --host CLI flag and server.host workflow config field for configurable bind address. Default 127.0.0.1; container deployments override with --host 0.0.0.0. Resolution order: CLI flag > config > default.
  • --log-format CLI flag with values text (default) and json. JSON format emits one JSON object per log line with time, level, msg, and structured fields (issue_id, session_id, etc.) for integration with Loki, Datadog, CloudWatch, and ELK.
  • Dockerfile: multi-stage build producing a distroless container image with only /usr/bin/sortie. Users consume via COPY --from=ghcr.io/sortie-ai/sortie:latest /usr/bin/sortie /usr/bin/sortie in their own Dockerfile. Build flags match .goreleaser.yaml: CGO_ENABLED=0, -trimpath, -s -w, tags osusergo,netgo.
  • .dockerignore excluding build artifacts, .git, test fixtures, and documentation.
  • Agent-specific example Dockerfiles: docker/claude-code.Dockerfile and docker/copilot.Dockerfile with non-root user, health checks, and volume mounts.
  • Kubernetes deployment examples: k8s/deployment.yaml (Recreate strategy, liveness/readiness probes on /livez and /readyz), k8s/configmap.yaml, k8s/service.yaml, k8s/pvc.yaml.
  • Grafana dashboard template at grafana-dashboard.json covering all 22 Prometheus metrics. Panels for dispatch_transitions_total, tracker_comments_total, ci_status_checks_total, and ci_escalations_total added in dedicated CI Feedback and Integration rows. Uses __inputs/DS_PROMETHEUS pattern for portable data source selection.

Fixed

  • Agent stderr is now surfaced at WARN level when a turn fails instead of DEBUG. Both Claude Code and Copilot CLI adapters buffer stderr during a turn and log at WARN on non-zero exit, making startup rejections (e.g., --dangerously-skip-permissions under root) visible at default log level. Successful turns continue to log stderr at DEBUG.

1.4.0 - 2026-04-04

Added

  • CI feedback loop: when a CI pipeline fails on an agent-created branch, the orchestrator detects the failure, injects the CI failure logs into the next agent turn, and dispatches a continuation session so the agent can diagnose and fix the issue automatically. Controlled by a new ci_feedback config section in WORKFLOW.md (kind, max_retries, max_log_lines, escalation). Feature activation follows kind-based convention: present ci_feedback.kind enables, absent disables.
  • CIStatusProvider domain interface: adapter contract for fetching CI check status from an SCM platform. Returns structured CIResult with overall status (pending/passing/failing), individual check runs, and an optional log excerpt from the first failing check.
  • GitHub CIStatusProvider implementation via the Checks API: fetches check runs for a git ref, computes aggregate status, and retrieves truncated log output from the first failing GitHub Actions job. Log fetching is controlled by max_log_lines (0 disables). ANSI escape sequences are stripped from log output.
  • CI failure escalation: when ci_feedback.max_retries is exceeded, the orchestrator applies a configurable escalation action -- add a label (default needs-human) or post a comment on the issue.
  • Exponential backoff for CI pending re-enqueue: reconcileCIStatus now applies base * 2^attempts backoff (capped at 5 minutes) when CI checks remain pending or on transient API errors, reducing GitHub Checks API request volume from ~120 to ~15 per 20-minute CI run per issue. Stale pending entries expire after a 30-minute TTL.
  • TrackerOpsWg shutdown drain: fire-and-forget tracker API goroutines (comment posting, label adding) are now tracked by a dedicated sync.WaitGroup and drained during graceful shutdown with a 35-second timeout, preventing orphaned goroutines on process exit.
  • Automatic credential merging: when ci_feedback.kind matches tracker.kind, the orchestrator merges tracker credentials (api_key, project, endpoint) into the CI provider adapter config at startup, eliminating the need to duplicate credentials in a pass-through block.

1.3.0 - 2026-04-03

Added

  • MCP tool execution channel: agents can now call registered tools at runtime via the Model Context Protocol. The worker generates .sortie/mcp.json per session and passes it to the agent runtime via --mcp-config (Claude Code) or --additional-mcp-config (Copilot CLI). The agent runtime spawns sortie mcp-server as a stdio sidecar; the orchestrator does not manage the sidecar lifecycle.
  • sortie mcp-server subcommand: MCP stdio JSON-RPC server that exposes registered AgentTool implementations via tools/list and tools/call. Constructs its own TrackerAdapter and ToolRegistry by re-reading WORKFLOW.md from an absolute path passed via --workflow.
  • sortie_status MCP tool (Tier 1): returns live session runtime metadata -- current turn number, remaining turns, attempt number, session duration, and cumulative token usage. Reads from .sortie/state.json, a worker-written file updated at session start, each turn start, and on token usage events.
  • workspace_history MCP tool (Tier 1): returns up to 10 most recent completed run attempts for the current issue from the run_history SQLite table. Opens the database in read-only mode (?mode=ro). Non-fatal on database open failure -- the MCP server continues with other tools available.
  • Agent-to-orchestrator file protocol: agents can write blocked or needs-human-review to .sortie/status to suppress continuation retries. The orchestrator reads the file after each turn and before the tracker state refresh. Absent, unrecognized, or unreadable files degrade to normal behavior. Symlinks on either path component are rejected via Lstat.
  • RuntimeStatusSuffix auto-injection: the orchestrator appends A2O protocol instructions to the first-turn prompt so agents know how to signal blocked status without workflow author intervention. Continuation turns omit the suffix.
  • Soft-stop exit path in HandleWorkerExit: when the worker exits with a recognized status file signal, the orchestrator releases the claim and suppresses continuation retry. The issue re-dispatches only on tracker state change.
  • Operator MCP config merging: if mcp_config is set in WORKFLOW.md, the worker merges the operator's config with the sortie-tools entry. Name collision on sortie-tools is a validation error.

Documentation

  • Agent-to-orchestrator file protocol specification (docs/agent-to-orchestrator-protocol.md): 9-section normative document covering file format, recognized values, read timing, cleanup lifecycle, symlink rejection, and conformance checklist.
  • ADR-0009: MCP stdio sidecar for tool execution -- documents the chosen transport mechanism, process model, credential handling, adapter integration, and alternatives analysis.
  • Architecture Section 10.4 rewrite: AgentTool interface contract, ToolRegistry invariants, tier classification framework, and tracker_api tool specification aligned with implementation.

1.2.1 - 2026-04-01

Fixed

  • Prevent re-dispatch loop when effort budget (agent.max_sessions) is exhausted for an issue; the orchestrator now records a durable budget-exhaustion guard so the dispatch loop does not restart the issue on the next tick
  • Persist and display turns_completed per run in the dashboard and run history table
  • Show fully qualified owner/repo#N display identifiers for GitHub issues in the dashboard and API instead of bare issue numbers
  • Pass handoff_state to findCurrentStateLabel, extractState, normalizeIssue, and normalizeBlockers in the GitHub adapter so TransitionIssue accepts the handoff state and stale labels are removed during transitions
  • Rename dashboard footer cache label from "Cache:" to "Cache Read:" and add explanatory tooltips in the footer and Running Sessions table
  • Defer token_usage event emission in the Claude adapter when an assistant message carries zero output tokens (tool_use-only messages in Claude Code 2.x stream-json format); the adapter now accumulates input tokens and falls back to the result event which carries correct totals

Migrations

  • 004: Add index idx_run_history_issue_id on run_history(issue_id)
  • 005: Add turns_completed INTEGER NOT NULL DEFAULT 0 to run_history
  • 006: Add nullable display_identifier column to run_history

1.2.0 - 2026-03-31

Added

  • GitHub Copilot CLI adapter: configure with agent.kind: copilot for fully automated issue-to-code workflows using GitHub's headless Copilot CLI. Supports local execution and SSH remote dispatch via worker.ssh_hosts. Tool scope is controlled by allowed_tools, denied_tools, available_tools, and excluded_tools; --allow-all is the default when none are set. Session continuity across turns via --resume. Authentication uses token env vars when present, falling back to gh auth status.
  • worker.ssh_strict_host_key_checking: new optional worker config field controlling OpenSSH StrictHostKeyChecking for remote SSH agent sessions. Accepts accept-new (default, Trust On First Use), yes (strict verification, requires a pre-populated known_hosts), or no (disable host-key checking). Applies to both the Claude Code and Copilot CLI adapters.

1.1.0 - 2026-03-30

Added

  • GitHub Issues tracker adapter: configure with tracker.kind: github and tracker.project: OWNER/REPO. State management is label-based; TransitionIssue applies and removes GitHub labels with convergent retry on partial failure.
  • GitHub adapter: in-memory ETag cache for reconciliation polls — If-None-Match conditional requests return 304 Not Modified on unchanged issues, reducing GitHub API rate limit consumption during active runs.
  • sortie validate GitHub adapter config validation: emits diagnostics for tracker.project format (OWNER/REPO), GITHUB_TOKEN environment variable hint, empty state labels, and active/terminal state label overlap. Errors block dispatch; warnings are advisory.

Fixed

  • install.sh: checksum verification used substring matching (grep | awk) that could accept a checksum entry for the wrong archive entry; now uses exact field matching (awk '$2 == f').

1.0.0 - 2026-03-29

Added

  • SPDX JSON Software Bill of Materials (SBOM) included with every release archive, generated via syft in the GoReleaser pipeline for supply-chain auditing.

0.0.10 - 2026-03-28

Added

  • sortie validate template static analysis: three advisory warning classes — WarnDotContext (top-level key referenced inside {{ range }} or {{ with }} where dot is redefined), WarnUnknownVar (variable not in the {issue, attempt, run} contract), and WarnUnknownField (valid top-level key with an unknown sub-field, including depth-4+ field chains on known level-3 scalars). Warnings appear in both text and JSON output without blocking dispatch or changing the exit code.
  • sortie validate front matter schema validation: detects unknown top-level keys, unknown sub-keys within known sections, type mismatches, and semantic issues (non-positive hooks.timeout_ms, non-numeric max_concurrent_agents_by_state entries). Field paths are included in every diagnostic so operators can locate the offending key. Warnings are advisory only.
  • SORTIE_* environment variable config overrides: any workflow front matter key can be overridden via a SORTIE_-prefixed environment variable (e.g., SORTIE_POLLING_INTERVAL_MS=5000). Non-empty real env vars take precedence over .env file values. Raw line content is removed from .env parse errors and override values are excluded from debug logs to prevent secret leakage.
  • Orchestrator: tracker comments posted at session lifecycle points — session start, successful completion, and failure — with run duration and attempt metadata. Comments fire from a detached goroutine to avoid blocking the event loop.
  • Orchestrator: issues are transitioned to the configured in_progress_state on dispatch, with a no-op skip when the issue is already in the target state. sortie_dispatch_transitions_total Prometheus counter tracks success, error, and skipped outcomes.

0.0.9 - 2026-03-27

Added

  • sortie validate subcommand for one-shot workflow file validation without starting the orchestrator, opening the database, or spawning a filesystem watcher. Supports --format text (stderr diagnostics) and --format json (structured stdout output) for CI pipelines and pre-commit hooks. Flag-parse errors are routed through the diagnostics emitter.
  • --dry-run flag for a single read-only poll cycle that validates the full startup sequence (workflow load, preflight, database, adapter wiring) without dispatching work or persisting state.
  • --log-level flag and logging.level workflow extension key to set the minimum log severity at startup (debug, info, warn, error).
  • Dashboard: Workflow column in Active Sessions table and a new Run History table showing completed session outcomes, timing, and workflow file. SQL migration 003 adds a nullable workflow_file column to run_history.
  • Workspace root write-permission check in dispatch preflight — surfaces a clear diagnostic instead of failing mid-dispatch.
  • Homebrew tap distribution via GoReleaser-managed tap repository (brew install sortie-ai/tap/sortie).
  • Jira adapter: User-Agent header (sortie/<version>) sent on every HTTP request.
  • Claude Code adapter: per-request APIDurationMS on token_usage events for API-call-level latency visibility, clamped to a minimum of 1 ms.
  • Claude Code adapter: tool error text now included in EventToolResult.Message.

Fixed

  • CLI: adapters implementing io.Closer are now closed during graceful shutdown, preventing resource leaks.
  • Orchestrator: TurnCount increments on session_started instead of session finalization, correctly reflecting in-progress turns.
  • Claude Code adapter: tool error messages stripped of XML markup and tail-truncated to prevent oversized events.

0.0.8 - 2026-03-26

Added

  • JSON API server with GET /api/v1/state, GET /api/v1/<identifier>, and POST /api/v1/refresh endpoints for programmatic access to orchestrator state. Enabled via --port flag or server.port config.
  • HTML dashboard at / with auto-refreshing view of running sessions, retry queue, token totals, and runtime statistics when the HTTP server is enabled.
  • /livez and /readyz health endpoints following Kubernetes z-pages conventions. /readyz checks database accessibility, preflight validation, and workflow loading.
  • Prometheus /metrics endpoint exposing session gauges, dispatch/worker/retry counters, token counters, tracker request counters, tool call counters, poll and worker duration histograms, and sortie_build_info. Uses a dedicated prometheus.Registry — compatible with standard Prometheus scrape configs.
  • tracker_api client-side tool: agents can query the tracker during sessions to fetch issues and comments, scoped to the configured project.
  • SSH worker extension via worker.ssh_hosts config: dispatch agent runs to remote hosts over SSH with round-robin host selection and per-host concurrency limits (worker.max_concurrent_agents_per_host).
  • Per-session token breakdown in JSON API and dashboard: input_tokens, output_tokens, cache_creation_tokens, cache_read_tokens.
  • Per-session timing breakdown in JSON API and dashboard: elapsed, agent_time, idle_time, agent_pct.
  • Claude Code adapter: tool_result events now emitted, making agent tool invocations visible in the dashboard and API.
  • Worker failure logging in HandleWorkerExit: WARN with next_attempt and delay_ms for retryable errors, ERROR for non-retryable errors.
  • Structured logging: issue_id, issue_identifier, and session_id context fields now present on all orchestrator lifecycle log lines. Agent tool calls logged at INFO level.
  • POSIX-compatible install script (install.sh) for automated binary installation.

Changed

  • POST /api/v1/refresh returns 409 Conflict during graceful shutdown instead of accepting requests that cannot be fulfilled.

Fixed

  • Claude Code adapter: duplicate token_usage events no longer emitted when assistant-level usage is already reported in the result message.
  • HTTP server: 405 Method Not Allowed responses now include the Allow header per RFC 9110.
  • Jira adapter: sortie_tracker_requests_total counter no longer increments on no-op calls with empty ID lists.

0.0.7 - 2026-03-24

Added

  • Graceful shutdown: on SIGTERM/SIGINT the orchestrator now drains running workers (up to 30 s), persists final state to SQLite, flushes pending agent events, and cancels retry timers before exiting.
  • Issue handoff via tracker.handoff_state config field — when an agent session completes normally and the issue is still in an active state, the orchestrator transitions it to the configured handoff state (e.g., "In Review") and skips the continuation retry.
  • TransitionIssue operation on the TrackerAdapter interface — Jira adapter uses the workflow transitions API; file adapter uses an in-memory override map.
  • Per-issue effort budget via agent.max_sessions — limits total agent sessions dispatched per issue before releasing the claim. Default 0 (unlimited).
  • Documentation site at https://docs.sortie-ai.com/ with initial configuration reference.

Fixed

  • Orchestrator: continuation retry attempt counter now increments correctly across sessions instead of resetting to 1 on every normal exit.
  • CLI: orchestrator-only fields (max_turns, max_concurrent_agents, max_retry_backoff_ms, max_concurrent_agents_by_state) removed from the adapter config map, fixing silent shadowing of adapter extension keys such as claude-code.max_turns.
  • Jira adapter: extractStringSlice now handles []string from the config layer — previously only []any was handled, silently reverting to default states and causing configured active_states / terminal_states to be ignored.
  • Jira adapter: FetchIssueStatesByIDs now queries by numeric id instead of key, and results are keyed by issue ID, fixing reconciliation failures where state changes on running issues were never detected.
  • Jira adapter: non-numeric IDs are now rejected instead of silently mangled, and empty ID lists no longer produce invalid id IN () JQL.
  • File and Jira adapters now return ErrTrackerNotFound for missing issues in FetchIssueByID and FetchIssueComments.
  • Orchestrator: INFO-level tick summary log after each dispatch cycle with candidate, dispatched, running, and retrying counters to distinguish normal operation from a stall.

0.0.6 - 2026-03-23

Added

  • Orchestrator engine with state management, concurrency-limited dispatch, worker lifecycle, exponential-backoff retry scheduling, active-run reconciliation, and event-driven poll loop with graceful shutdown.
  • Full startup sequence: workflow load, preflight validation, database open, state reconciliation, and poll loop — in that order.
  • Dispatch preflight checks that validate adapter availability, required API keys, and agent configuration before dispatching work.
  • Adapter metadata via AdapterMeta and RegisterWithMeta so adapters can declare requirements (e.g., RequiresAPIKey) checked during preflight.
  • Retry classification on TrackerErrorKind and AgentErrorKind — errors are now classified as retryable or permanent for dispatch decisions.
  • ErrTrackerNotFound error kind for HTTP 404 responses from tracker adapters.
  • Configurable db_path field in workflow configuration with ~ and $VAR expansion.
  • Workflow validation callback (ValidateFunc) that guards config promotion during hot-reload.

Fixed

  • Workspace CleanupByPath now rejects non-canonical paths and uses the actual workspace path for pending cleanup instead of reconstructing it from config.
  • Startup: preflight checks now run before opening the database, preventing .sortie.db creation when configuration is invalid.
  • Startup: .sortie.db is now created adjacent to WORKFLOW.md instead of in the working directory.

0.0.5 - 2026-03-21

Added

  • Workspace manager: safe path computation from issue identifiers with containment validation and symlink rejection.
  • Workspace manager: atomic directory creation and reuse with CreatedNow flag for hook gating.
  • Workspace hook execution with configurable timeout, truncated output capture, and restricted subprocess environment (only PATH, HOME, SHELL, and SORTIE_* variables are inherited).
  • Workspace lifecycle orchestration: Prepare, Finish, and Cleanup functions that sequence after_create, before_run, after_run, and before_remove hooks with appropriate failure semantics (fatal vs best-effort) and context.WithoutCancel for teardown hooks.
  • Batch workspace cleanup (CleanupTerminal) for removing terminal-state issue workspaces with per-identifier error collection and best-effort before_remove hook execution.
  • ListWorkspaceKeys for enumerating workspace directory names under a root, skipping non-directories and symlinks.

0.0.4 - 2026-03-20

Added

  • AgentAdapter interface and normalized event model: 13 event types, TokenUsage, AgentConfig, Session, TurnResult, and AgentError with 9 error kinds.
  • Agent adapter registry (registry.Agents) for registration and lookup by kind.
  • Mock agent adapter (kind "mock") with configurable turn outcomes, delays, and cumulative token accumulation for orchestrator and integration testing.
  • Claude Code agent adapter (kind "claude-code") that launches the CLI as a subprocess, reads JSONL events from stdout, and normalizes them to domain event types. Supports graceful SIGTERM→SIGKILL shutdown on context cancellation and session resumption via ResumeSessionID.

Fixed

  • Claude Code adapter: double-wait race between RunTurn and StopSession — gracefulKill is now fire-and-forget with timer-based SIGKILL escalation.
  • Claude Code adapter: error on missing binary now includes the actual command name instead of a hardcoded string.

0.0.3 - 2026-03-20

Added

  • Normalized Issue model and TrackerAdapter interface for multi-tracker support.
  • Typed adapter registry with thread-safe registration and lookup.
  • File-based tracker adapter for local JSON task definitions.
  • Jira Cloud REST API v3 adapter with cursor-based paginated search, issue detail retrieval, state tracking, and comment fetching.
  • BFS flattener for Atlassian Document Format (ADF) descriptions to plain text.
  • JQL builder with string escaping and optional query_filter clause support.
  • query_filter field in tracker configuration for custom JQL expressions.
  • GoReleaser configuration for reproducible cross-platform binary releases (linux/darwin/windows, amd64/arm64).

Fixed

  • Jira search endpoint migrated from retired /rest/api/3/search to /rest/api/3/search/jql (Atlassian returns 410 Gone on the old endpoint).
  • Infinite loop guard in Jira comment pagination when the API returns inconsistent offsets.

0.0.2 - 2026-03-19

Added

  • SQLite persistence layer with WAL mode and single-writer enforcement.
  • Schema migration runner with versioned SQL files.
  • CRUD operations for retry entries, run history, session metadata, and aggregate metrics.
  • Startup recovery loader that resumes incomplete retry entries on restart.

Fixed

  • Deterministic ordering for session metadata queries via session_id tie-breaker.

0.0.1 - 2026-03-18

Added

  • WORKFLOW.md file loader with YAML front matter and prompt body parsing.
  • Typed configuration layer with $VAR environment variable resolution and ~ home directory expansion.
  • Prompt template engine using Go text/template in strict mode (unknown variables and filters cause hard errors).
  • Turn-based prompt builder for multi-turn agent conversations.
  • Filesystem watcher for live WORKFLOW.md reload via fsnotify.
  • CLI entry point (sortie) with graceful shutdown and signal handling.

Fixed

  • Environment variable expansion now preserves inline $VAR references inside URIs instead of silently dropping them.
  • Fractional float values no longer silently coerced to integers during config parsing.

0.0.0 - 2026-03-18

Added

  • Go module scaffold and project directory structure.
  • Structured logging built on log/slog with issue-aware and session-aware contextual fields.
  • CI pipeline with golangci-lint, gofmt enforcement, and test execution via GitHub Actions.
  • Architecture Decision Records (ADR-0001 through ADR-0005).