Skip to content

CHANGELOG.md

Latest commit

 

History

History
177 lines (159 loc) · 10.1 KB

CHANGELOG.md

File metadata and controls

177 lines (159 loc) · 10.1 KB
Raw

Changelog

All notable changes to Gas City will be documented in this file.

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

Unreleased

Fixed

  • Kiro provider launch behavior is now explicit in release notes and provider docs: the built-in Kiro provider starts kiro-cli with chat, --no-interactive, --agent gascity, and --trust-all-tools by default. Operators who do not want unrestricted tool trust can replace the full default argv with an explicit [providers.kiro].args list in city.toml.
  • Tmux and runtime provider-overlay staging now surface nonfatal preservation warnings on stderr, including the Kiro AGENTS.md preservation notice when project instructions already exist.
  • jsonl-export.sh no longer mis-classifies a bead database with an empty issues table as a failed export. dolt sql -r json returns {} (not {"rows":[]}) when a queried table is empty; validate_exported_issues now treats the bare-object form as zero rows so the database lands in the success path with an issues.jsonl committed to the archive instead of appearing in the failed: summary.
  • The built-in control-dispatcher trace now defaults to ${GC_CITY_RUNTIME_DIR}/control-dispatcher-trace.log (falling back to ${GC_CITY}/.gc/runtime/control-dispatcher-trace.log) instead of writing at city root. This keeps workflow-trace appends inside the controller's watcher-excluded runtime subtree, avoiding continuous config-changed reconciliations. After upgrading, operators tailing the default trace should switch to .gc/runtime/control-dispatcher-trace.log; the old ${GC_CITY}/control-dispatcher-trace.log file becomes stale and can be removed. After upgrading, restart or recycle existing control-dispatcher sessions so they pick up the new trace path; otherwise they keep their previous trace target and can continue retriggering reconciles. Validation currently covers watcher exclusion, dispatcher warning routing, and the graph-workflow integration shard; there is not yet a dedicated patrol-cadence stress test.
  • proxy_process services now receive a GC_SERVICE_URL_PREFIX that the supervisor's public listener actually routes. Previously the prefix was the per-city-relative /svc/<name>, so any service that composed CallbackURL = $GC_API_BASE_URL + $GC_SERVICE_URL_PREFIX (the documented shape for adapter self-registration) would 404 on inbound calls. The prefix is now the full /v0/city/<cityName>/svc/<svcName> path. The per-city router contract (config.Service.MountPathOrDefault) is unchanged.
  • gc session reset now documents its named-session circuit-breaker behavior: when the target is a named session, reset clears a tripped respawn breaker before requesting a fresh restart.

Changed

  • ACP, subprocess, and Kubernetes session staging now apply pack and agent overlays through the provider-aware per-provider/<provider>/ contract. Custom ACP overlays that previously expected a literal per-provider/ subtree in the session workdir should move provider-specific files under the matching provider slot so they are flattened at launch.
  • The review-quorum durable contract now documents that synthesized findings_count is deduplicated, top-level mutations_delta is reserved for synthesis-created changes, lane mutation deltas remain under their lane records, lane-scoped finalizer failures use lane=<lane_id> reason=<stable_reason> entries, and unknown lane verdict values are hard contract failures. Reviewer lane prompts now require durable lane_id, provider, and model fields, and the finalizer rejects blank lane IDs without merging contract-invalid lane findings, evidence, or usage into the synthesized summary.
  • [[orders.overrides]] rig matching is stricter and clearer. A rigless override (rig unset) still matches only city-level orders; if the named order exists only as per-rig instances, the error now names every matching rig so it's obvious what to type. rig = "*" is a new wildcard that targets every instance of the named order (city-level + per-rig). The literal "*" is reserved and rejected as a real rig name by config validation.
  • Managed Dolt config now emits listener backlog and connection-timeout keys. Existing managed cities may see a dolt-config doctor warning until gc dolt restart or the next managed server start regenerates dolt-config.yaml.
  • In bead-backed pool reconciliation, scale_check output is now documented and enforced as additive new-session demand. Assigned work is resumed separately; custom checks that previously returned total desired sessions should return only new unassigned demand.
  • Session bead reconciliation now stops suspended and orphaned runtimes before closing their beads; resuming one of those sessions starts a fresh lifecycle instead of continuing the previous runtime process.
  • gc hook --inject is now silent legacy compatibility for already-installed Stop/session-end hooks. Fresh managed hook configs no longer install it; routed work pickup should happen through the SessionStart claim protocol or an explicit non-inject gc hook call.
  • The built-in Claude provider's model = "opus" option now emits claude-opus-4-7. Cities that rely on the opus alias should expect the new model target after upgrading.

Fixed

  • Linux systemd supervisor service restarts now preserve managed tmux sessions for re-adoption. Linux users should rerun gc supervisor install after upgrading so the user unit is regenerated with KillMode=process and the preserve-on-signal environment. If the currently active Linux supervisor predates the preserve-on-signal environment, gc supervisor install now refuses the warm refresh before sending a signal and tells operators to stop or drain agents intentionally with gc supervisor stop --wait, then rerun the install. Once the active supervisor already supports preserve mode, Linux warm refresh sends the main supervisor PID SIGTERM first so preserve-mode shutdown can close workspace services and flush traces, with a bounded SIGKILL fallback if the process does not exit. The Linux refresh also stops orphan-prone workspace service process groups owned by registered cities before starting the replacement supervisor; supervisor startup repeats the same owned-service cleanup after crashes. Service-managed SIGTERM preserves sessions for re-adoption, while SIGINT remains a destructive escalation path. Preserve mode intentionally leaves the beads provider running so preserved sessions can keep using the store; the bundled managed-Dolt start path is idempotent when it finds an already-running server, but custom exec providers must make start reattach or no-op safely after preserve-mode restarts. macOS launchd upgrades still use launchd unload/load rather than the Linux main-PID refresh path; macOS supervisor startup now warns that automatic orphaned workspace-service cleanup is Linux-only, lists the registered GC_SERVICE_STATE_ROOT roots to inspect, and tells operators to stop stale workspace-service processes before restarting affected cities after non-graceful exits.

1.0.0 - 2026-04-21

First stable release. Between v0.15.1 and v1.0.0 the project received 610 commits across 1,273 files (+303,902 / −46,437) from the core team and 12 community contributors. See the GitHub release page for the full narrative.

Added

  • gc reload [path] — structured live config reload. Failures keep the previous runtime config active instead of silently degrading.
  • gc prime --strict — turns silent prompt/agent fallback paths into explicit CLI failures for debugging.
  • rig adopt — adopt existing rigs without a full rebuild.
  • Provider-native MCP projection for Claude, Codex, and Gemini, with multi-layer catalog resolution and projected-only gc mcp list.
  • Per-agent append_fragments so prompt layering is configurable through the supported config and migration paths.
  • Wave 1 pass over orders and dispatch runtime — store resolution, dispatch surfaces, rig-aware execution, and verifier coverage.

Changed

  • Session model unified. Declarative [[agent]] policy/config is now cleanly separated from runtime session identity; session beads are the canonical runtime projection.
  • Pack V2 is the active layout. Bundled packs use [imports.<name>]; builtin formulas, prompts, hooks, and orders come from the builtin core pack. V1-era city-local seeding is retired.
  • gc init is back on the pack-first scaffold contract. Agent and named sessions belong in pack.toml; machine-local identity stays in .gc/site.toml; city.toml keeps workspace/provider state.
  • gc import install is now the explicit bootstrap path for importable packs.
  • gc session logs --tail N returns the last N entries (matches Unix tail convention) instead of the old compaction-oriented behavior.
  • Supervisor API migrated to Huma/OpenAPI; Go client regenerated; dashboard SPA restored.
  • Order "gates" renamed to triggers.

Fixed

  • Startup proofs for hook-enabled providers — correct startup prompt delivery, no duplicate SessionStart hook context, no replay of startup prompts on resumed sessions.
  • Managed Dolt hardening: recovery, transient failures, health probes, runtime-state validation, and late-cycle macOS portability fixes (start-lock FD inheritance, path canonicalization, lsof reachability, PID confirmation, portable sed parsing).
  • Pack V2 tmux startup regression where large prompt launches could silently fall back to the known-broken inline path.
  • Custom provider option defaults now fail early instead of silently degrading.
  • Beads storage core quality pass — cache recovery, close-all fallback semantics, watchdog reconciliation cadence, dirty-cache fallback reads.
  • Long tail of session lifecycle, wake-budget, and pool identity fixes.