Skip to content

Changelog

Every released version of the SwarmKit runtime, newest first, generated from the annotated git tags (v1.0.0 – present). Each version is a runtime + schema bump; see the GitHub releases and tags for the full diff.

The earlier per-series notes (v1.2, v1.1) carry hand-written detail for those ranges.

September 2026

  • v1.256.0 (2026-09-19) — runtime 1.256.0 — A2A budget forwarding (a2a-federation.md slice 2).
  • v1.255.0 (2026-09-19) — runtime 1.255.0 — worker fairness (worker-fairness.md).
  • v1.254.0 (2026-09-19) — runtime 1.254.0 — serve --profile production, a fail-closed startup preflight (production-profile.md).
  • v1.253.0 (2026-09-19) — runtime 1.253.0, schema 1.46.0 — topology input_schema (input-schema.md).
  • v1.252.0 (2026-09-19) — runtime 1.252.0 — fence every jobs-row write to worker ownership (worker-execution.md).
  • v1.251.0 (2026-09-19) — runtime 1.251.0 — queue observability (queue-observability.md).
  • v1.250.0 (2026-09-19) — runtime 1.250.0 — worker-execution hardening from an external scaling review.
  • v1.249.0 (2026-09-19) — runtime 1.249.0 — the API tier's SSE stream follows a worker's job to completion.
  • v1.248.0 (2026-09-19) — runtime 1.248.0 — the API/worker split for worker execution (worker-execution.md, slice 2).
  • v1.247.0 (2026-09-19) — durable job queue for worker execution (slice 1)
  • v1.246.0 (2026-09-19) — build the checkpointer once under concurrency
  • v1.245.0 (2026-09-19) — the storage service owns the connection pool; SQLite audit stays on the loop
  • v1.244.0 (2026-09-18) — configurable store connection pool + off-loop audit write
  • v1.243.0 (2026-09-18) — the compiled graph is cached per topology
  • v1.242.0 (2026-09-18) — load-test harness and the first published NFR numbers
  • v1.241.0 (2026-09-18) — swarmkit upgrade, with a breaking-change gate
  • v1.240.0 (2026-09-18) — A2A federation: a SwarmKit callee hands back its record
  • v1.239.0 (2026-09-18) — the audit is a write-through journal
  • v1.237.0 (2026-09-18) — the diagnostic commands report an unresolvable store instead of dying on it
  • v1.236.0 (2026-09-18) — reload keeps serve up with MCP servers configured; the catalogue in the portal
  • v1.235.0 (2026-09-18) — swarmkit skill: the catalogue and Agent Skills files from the command line
  • v1.234.0 (2026-09-18) — the fleet audit's last leftovers
  • v1.233.0 (2026-09-18) — memory on by default, and a fleet panel that approves as the person who clicked
  • v1.231.0 (2026-09-18) — serve needs no extra
  • v1.230.0 (2026-09-18) — the fleet control plane audited against the runtime and brought back into step
  • v1.227.0 (2026-09-18) — tutorials 1–16 rewritten from real runs; the bugs they found, fixed
  • v1.226.0 (2026-09-17) — the docs audited against the code: six new tutorials, a capstone demo, the knowledge pack carries the guides
  • v1.225.0 (2026-09-17) — agent skills offered to harness nodes through the governed gateway
  • v1.224.0 (2026-09-17) — add a remote agent by card URL from the portal
  • v1.223.0 (2026-09-17) — pack:workspace: every topology as a synthesized agent skill
  • v1.222.0 (2026-09-17) — another agent as a skill: implementation.type: agent
  • v1.221.0 (2026-09-17) — an A2A server: the Agent Card and task API as a transport onto jobs
  • v1.220.0 (2026-09-16) — the knowledge pack tells an LLM what shipped; the docs are generated where they rot
  • v1.219.0 (2026-09-13) — declarative model providers: a provider is YAML over a wire-format family
  • v1.218.0 (2026-09-11) — attachments on a run: a caller holding the file passes it beside the input
  • v1.217.1 (2026-09-08) — a human can put a fact into governed memory. swarmkit memory add and POST /memory write through the same governed path an agent writes through, so a seeded fact gets the same reconcile decision, the same contradiction quarantine and the same audit line: the command reports new/update/reinforce/refine/contradict rather than a bare "added", and a contradiction exits non-zero with the trusted memory untouched. Until now the only write path was the persistence-skill hook, so a fact entered memory only when an agent chose to emit it mid-run — while the facts most worth keeping are the ones discovered incidentally. Every write is now audited as memory.written, emitted by the store so the agent path carries the same line.
  • v1.216.3 (2026-09-08) — raises the [ui] floor to swarmkit-webui 0.18.0. 1.216.2 fixed the Connections page but left the floor at 0.15.0, so pip install "swarmkit-runtime[ui]" still resolved the 0.17.0 portal whose page throws on load — the fix shipped and the constraint let the broken build in anyway. A test now asserts the floor is at least the portal version in the tree, not merely that a floor exists.
  • v1.216.2 (2026-09-08) — fixes the Connections page, which crashed on every workspace in 1.216.0. Removing the channel blocks took channels out of the workspace-config API without updating the portal, which still read it — so anyone installing swarmkit-runtime[ui] got a page that threw on load. The page now shows event sinks, which is what the API returns. Also adds a showcase workspace that runs with no API keys and no network (just demo-showcase), and generates the portal screenshots and tour from a live serve rather than filming them once — the script that does it refuses to photograph a 404, which is how the crash was found.
  • v1.216.0 (2026-09-07) — communication leaves the runtime, and credentials get one service. An application now taps events instead of the runtime shipping chat integrations: GET /events?after=<cursor> is a durable log to reconcile from, events: sinks push best-effort, and a resolved gate resumes its run on its own. The Telegram, Discord and Slack providers are gone (~1,600 lines), with examples/event-consumer/ showing an application that receives an event, asks a human on its own platform, and resolves the gate — importing no runtime module. OAuth for remote MCP servers ships end to end: log in from the portal, tokens stored encrypted per owner and refreshed at the point of use, revoked upstream on disconnect. Every credential now resolves through one CredentialService, so run, chat, serve and mcp-serve all refresh alike rather than only the one path that remembered to. The portal gains a Connections page — the first surface that edits workspace infrastructure rather than artifacts.
  • v1.205.0 (2026-09-03) — MCP credentials a server declares now actually reach it: a credentials_ref or headers block resolves through the workspace credentials store instead of being validated, displayed and loaded by nothing. Skills can declare provenance.requires_runtime, so one that outgrows your runtime is refused at workspace load naming both versions rather than failing mid-run, and swarmkit skills check asks each skill's server whether its tool still exists. The runtime's dependency on swarmkit-schema is floored, so a runtime reading a new schema field can no longer install beside a schema that rejects it.
  • v1.201.0 (2026-09-02) — a harness is visible while it runs, on both chat surfaces
  • v1.200.0 (2026-09-01) — publish the web portal, and fix the guard that let it freeze

August 2026

  • v1.199.0 (2026-08-30) — command packs: local commands as a skill implementation type
  • v1.196.0 (2026-08-23) — supersedes 1.195.0: the schema stays in the prompt, since the grammar does not carry descriptions
  • v1.195.0 (2026-08-23) — structured-output prompts stop carrying another agent's schema, and max_tokens reaches the provider
  • v1.194.0 (2026-08-22) — an agent can set Ollama's think via model.options, instead of a model-family guess
  • v1.193.0 (2026-08-15) — swarmkit stop: ask a run to stop at its next agent boundary, keep its work, resume it later
  • v1.192.0 (2026-08-15) — docs rewritten around the pipeline removal, plus the run-supersedes chain a caller can finally write
  • v1.191.0 (2026-08-15) — a topology can declare that one skill must be called before another, enforced at the MCP permission seam on both executors
  • v1.190.0 (2026-08-14) — swarmkit validate reports which topology roots produce an output nothing checks, and --require-verified gates CI on it
  • v1.189.0 (2026-08-14) — the bundled pipeline is removed; sequencing belongs to the application
  • v1.188.0 (2026-08-14) — the bundled pipeline orchestrator is deprecated; sequencing moves to the application
  • v1.187.0 (2026-08-14) — a reference pipeline orchestrator outside the runtime, and correlation on the HTTP run path
  • v1.186.0 (2026-08-13) — resume a parked run over HTTP, record what a re-run supersedes
  • v1.185.0 (2026-08-13) — a field the durable row can answer is never reported absent
  • v1.184.0 (2026-08-13) — a gate names the run that produced it, and artifacts are fetchable over HTTP
  • v1.183.0 (2026-08-13) — a harness agent's diff survives the run and can be read back
  • v1.182.0 (2026-08-13) — a funnel's approve layer parks the run on a human, and resuming does not re-draft
  • v1.181.0 (2026-08-13) — read a gate's state with its approval policy applied
  • v1.180.0 (2026-08-10) — an escalated artifact says it was escalated, where the reviewer is looking
  • v1.179.0 (2026-08-10) — a correlated one-shot run can save and fetch its output
  • v1.178.0 (2026-08-10) — judge.rubric is read at last, and the reachability report names every binding
  • v1.177.0 (2026-08-10) — a harness run records which tools it called, with what and to what effect
  • v1.176.0 (2026-08-10) — correlate and label a one-shot run; audit events belong to the run that emitted them
  • v1.174.0 (2026-08-09) — a saga blocked by reconciliation can actually be released: the remedy bug 24's fix printed now works
  • v1.173.0 (2026-08-08) — report configuration no code path can reach: a wiring ledger plus an inert-binding report, after five bugs of the same anatomy
  • v1.172.0 (2026-08-08) — a bound funnel actually runs its layers: validate and judge were inert on every run serve or the CLI ever made
  • v1.171.0 (2026-08-07) — reconciliation no longer drives past a human gate: a stranded stage that declares one is refused rather than absorbed
  • v1.170.0 (2026-08-07) — a decision skill actually fires
  • v1.169.0 (2026-08-07) — advisory means advisory, and the memory that was unreachable is documented
  • v1.168.0 (2026-08-07) — curated facts reach the agent, tool calls open, and the workspace reloads
  • v1.164.0 (2026-08-07) — a restart no longer leaves work that looks like it is still happening
  • v1.163.0 (2026-08-07) — a run's detail page says what the run was, not only what it returned
  • v1.162.0 (2026-08-07) — a process's gateways all work, and a harness's artifact is not thrown away
  • v1.161.0 (2026-08-06) — stop correcting a broken executor, and never hand a harness a dead gateway
  • v1.160.0 (2026-08-06) — a good artifact is not discarded, and a toolless run is not a success
  • v1.159.0 (2026-08-06) — a reworked stage is its own run, not an overwrite of the one before
  • v1.158.0 (2026-08-06) — a harness agent receives its context files
  • v1.157.0 (2026-08-06) — a harness tool grant names the tools the gateway actually advertises
  • v1.156.0 (2026-08-06) — the dashboard shows real work, and a harness artifact is the agent's own
  • v1.155.0 (2026-08-06) — a chat turn is recorded, and its audits are findable
  • v1.154.0 (2026-08-05) — a job listed in history can actually be opened
  • v1.153.0 (2026-08-05) — runs and audit entries say what actually happened
  • v1.150.0 (2026-08-05) — a CLI run is part of the record
  • v1.149.1 (2026-08-05) — an output_schema no longer silently disables an agent's tools
  • v1.148.0 (2026-08-05) — a schema can live in a file, and a retry stops looking like an attack
  • v1.145.1 (2026-08-05) — the schema catches up, and a release fails rather than skipping a package
  • v1.145.0 (2026-08-05) — a stalled pipeline recovers itself, and the jobs page remembers
  • v1.144.0 (2026-08-04) — a harness node is finally a node, and two processes can start together
  • v1.141.0 (2026-08-04) — the portal can reach an identity provider, gates are approvable locally, and a failed stage stops the run
  • v1.137.0 (2026-08-03) — harness runs are observable, migrations leave a writable store, and rework carries the reviewer's comment
  • v1.133.0 (2026-08-03) — pipeline stage traces stop overwriting each other
  • v1.132.0 (2026-08-03) — checkpoints are actually durable
  • v1.131.3 (2026-08-03) — the MCP SDK 2.0 port, finished
  • v1.131.2 (2026-08-02) — MCP SDK 2.0 support
  • v1.131.1 (2026-08-02) — a missing checkpointer extra degrades, it does not kill serve
  • v1.131.0 (2026-08-02) — decision skill config is honoured, and a verdict's case is not its meaning
  • v1.130.0 (2026-08-02) — one storage service decides where every store lives
  • v1.129.2 (2026-08-02) — three reported bugs, two of them security, plus honest version reporting
  • v1.129.0 (2026-08-01) — human decisions carry comments, and agents read them
  • v1.128.0 (2026-08-01) — a bundled parked stage carries a real approval policy
  • v1.127.3 (2026-08-01) — launch substitution no longer deletes brace-bearing prompts
  • v1.127.2 (2026-08-01) — restore swarmkit orchestrator
  • v1.127.1 (2026-08-01) — forward the store URL with its password intact
  • v1.127.0 (2026-08-01) — approve a parked gate from the run view; fix the reported silent failures

July 2026

  • v1.123.0 (2026-07-29) — fix: webhook forwards its body as the pipeline input
  • v1.122.0 (2026-07-29) — show each pipeline node's input in the run view
  • v1.121.0 (2026-07-29) — fix: bundled orchestrator starts on named webhook events
  • v1.120.0 (2026-07-28) — fix: thread the pipeline input payload
  • v1.119.0 (2026-07-28) — bundled durable pipeline orchestrator + run-stage execution + dispatch/inspect surface
  • v1.118.0 (2026-07-27) — governed-memory UI + provenance (portal republished with swarmkit-webui 0.2.0)
  • v1.117.0 (2026-07-27) — governed memory
  • v1.111.0 (2026-07-27) — funnel deterministic validate (slice_budget + cited_change, harness diff threading) + gate coverage & comprehension-debt signals
  • v1.103.0 (2026-07-25) — orchestration & pipeline-triggering runtime
  • v1.98.0 (2026-07-20) — artifact env-substitution: ${VAR}/${VAR:-default}/$${VAR} resolved across all artifact YAML, with or without a workspace env file
  • v1.97.0 (2026-07-19) — multi-party approval (governance): role registry + approval-policy schemas, evaluation engine, and runtime gate resolution
  • v1.96.0 (2026-07-13) — serve hosts the web portal (swarmkit-runtime[ui] + swarmkit-webui)
  • v1.95.0 (2026-07-13) — topology canvas examine mode (run overlaid on the graph); get_job returns topology
  • v1.94.0 (2026-07-13) — gatewayed MCP for harnesses (container-reachable); harness feature complete
  • v1.93.0 (2026-07-13) — ephemeral governed MCP gateway for harnesses (native)
  • v1.92.0 (2026-07-13) — extract the one governed MCP-call path
  • v1.91.1 (2026-07-13) — container sandbox demo + close-out (feature complete)
  • v1.91.0 (2026-07-13) — container sandbox mounts + MCP reachability
  • v1.90.0 (2026-07-13) — build-in-sandbox (no local harness install)
  • v1.89.0 (2026-07-13) — enforced container egress (deny + allowlist proxy)
  • v1.88.0 (2026-07-13) — container sandbox provisioner (docker|podman)
  • v1.87.0 (2026-07-12) — harness executor + container sandbox docs; knowledge pack includes guides
  • v1.85.0 (2026-07-12) — trust accrual: repeated relay approvals propose an allowlist changeset (P3.5)
  • v1.84.0 (2026-07-12) — harness-gate review surface: CLI + HTTP API + serve UI + fleet UI
  • v1.83.0 (2026-07-12) — §6.3 input-request escalation (classifier + human inbox + memoize)
  • v1.81.0 (2026-07-12) — executor relay: park-resume mid-run permission approvals (verified e2e)
  • v1.78.0 (2026-07-12) — executor P3 complete: launch review gate + adapter authoring guide
  • v1.77.0 (2026-07-12) — executor P3: declarative harness adapters (harnesses are data)
  • v1.69.0 (2026-07-11) — executor abstraction P2 complete (harness executors, end to end)
  • v1.60.1 (2026-07-11) — Publish swarmkit-schema 1.11.0 (x-swarmkit-ref reference hints).
  • v1.60.0 (2026-07-11) — runtime 1.60.0
  • v1.59.0 (2026-07-11) — runtime 1.59.0
  • v1.58.0 (2026-07-10) — runtime 1.58.0
  • v1.57.0 (2026-07-10) — runtime 1.57.0
  • v1.56.0 (2026-07-10) — runtime 1.56.0
  • v1.55.0 (2026-07-10) — runtime 1.55.0
  • v1.54.0 (2026-07-10) — runtime 1.54.0, schema 1.10.0

June 2026

  • v1.8.0 (2026-06-22) — context compression (lossless columnar + reversible headtail + plugin backends), per-surface policy, context_retrieve, observability, per-run contextvars isolation; eval harness slices 1-2. swarmkit-schema 1.5.0.
  • v1.3.6 (2026-06-13) — generic per-model options passthrough (schema 1.3.3, runtime 1.3.6)
  • v1.3.5 (2026-06-13) — schema-constrained output decoding from output_schema
  • v1.3.4 (2026-06-13) — surface structured MCP tool output when no text block
  • v1.3.3 (2026-06-11) — Gemma/Google model tool-calling fixes
  • v1.3.2 (2026-06-06) — GBrainMemory compiler wiring
  • v1.3.1 (2026-06-03) — tool call tracing + UI trace panel
  • v1.3.0 (2026-06-01) — M11 launch prep + expertise packages

May 2026

  • v1.2.70 (2026-05-31) — Topology Composer (3 views, CRUD, YAML editing), cost optimization (dual model, token tracking)
  • v1.2.69 (2026-05-31) — dual model, token tracking, vedanta optimizations
  • v1.2.68 (2026-05-31) — dual model (tool/synthesis split), accurate token tracking
  • v1.2.67 (2026-05-31) — schema version bump for PyPI
  • v1.2.66 (2026-05-31) — accurate token tracking, CRUD endpoints, composer scaffold, configurable store, vedanta scripture fixes
  • v1.2.63 (2026-05-28) — chat UI polish, real-time progress, markdown, retry, token tracking
  • v1.2.62 (2026-05-28) — UI dashboard + chat, SQLite persistence, workspace memory, canary deployments
  • v1.2.60 (2026-05-28) — schema version bump for PyPI publish
  • v1.2.59 (2026-05-28) — workspace memory (GBrain + local), canary deployments, UI dashboard, serve triggers
  • v1.2.49 (2026-05-19) — maintenance release
  • v1.2.47 (2026-05-19) — fix single-agent tool stripping + add all MCP tools to archetype
  • v1.2.46 (2026-05-19) — architect solution reasoning in scope + mermaid syntax rules
  • v1.2.45 (2026-05-19) — topology-level synthesis prompt config + Sterling HLD prompt
  • v1.2.44 (2026-05-19) — fix synthesizer file write: natural language template extraction + default output path
  • v1.2.43 (2026-05-19) — synthesizer trace + openrouter provider + version bump
  • v1.2.42 (2026-05-19) — synthesizer + writing tool gating + schema update
  • v1.2.41 (2026-05-19) — add synthesis field to workspace schema
  • v1.2.40 (2026-05-19) — automatic synthesizer: single large-context call for document generation
  • v1.2.39 (2026-05-19) — compiler-driven two-phase execution with tool gating
  • v1.2.38 (2026-05-19) — fix OpenAI tool_calls message format for Qwen/Groq/Together
  • v1.2.37 (2026-05-18) — timeout + retry with jitter on all model provider calls
  • v1.2.36 (2026-05-18) — fix resume fast-path: root skips to architect on task plan resume
  • v1.2.35 (2026-05-18) — fix resume routing to architect on crashed run recovery
  • v1.2.34 (2026-05-18) — fix trace recursion + differentiated configurable tool limits
  • v1.2.33 (2026-05-18) — fix forced synthesis dead code, tool loop guards, docs-researcher redesign
  • v1.2.32 (2026-05-18) — Wave 3 complete + gate-validator MCP server
  • v1.2.31 (2026-05-18) — Wave 2 complete: deterministic grounding + Rynko Flow integration
  • v1.2.29 (2026-05-18) — structured checkpoint instructions (JSON action specs)
  • v1.2.28 (2026-05-18) — auto-populate source fields from tool provenance
  • v1.2.27 (2026-05-18) — MCP provenance envelope on every tool call
  • v1.2.26 (2026-05-18) — default output_schema for workers (structured inter-agent communication)
  • v1.2.25 (2026-05-18) — planning config (scope_required + two_phase) as workspace/topology settings
  • v1.2.24 (2026-05-18) — create-scope + update-scope + read-scope tools
  • v1.2.23 (2026-05-17) — check for state-changing tools after nudge response
  • v1.2.22 (2026-05-16) — tool loop detects update-task-plan for two-phase execution
  • v1.2.21 (2026-05-16) — two-pass task plan handler for correct two-phase execution
  • v1.2.19 (2026-05-16) — handle update-task-plan + freeze-scope feedback in tool loop
  • v1.2.18 (2026-05-16) — Phase 1 checkpoint prompts freeze-scope + update-task-plan
  • v1.2.17 (2026-05-16) — fix auto-synthesize for single-task Phase 1 plans
  • v1.2.16 (2026-05-16) — fix freeze-scope tool loop handling
  • v1.2.15 (2026-05-16) — freeze-scope platform tool + spec-conformance decision skill
  • v1.2.14 (2026-05-16) — sync bundled schemas for decision_skills validation
  • v1.2.13 (2026-05-16) — governance decision skill evaluator, retry loop, Sterling grounding
  • v1.2.12 (2026-05-16) — governance decision skills with workspace/topology merge
  • v1.2.11 (2026-05-16) — forced synthesis forbids planning language
  • v1.2.10 (2026-05-16) — planning text filter, jira-researcher upgrade, strip narration from history
  • v1.2.9 (2026-05-16) — archive run-state after completion for history
  • v1.2.8 (2026-05-15) — swarmkit why reads trace files and task plans
  • v1.2.7 (2026-05-15) — trace hierarchy fix for task-executed children + Sterling README update
  • v1.2.6 (2026-05-15) — fix read-task-result re-delegation loop
  • v1.2.5 (2026-05-15) — nudge on empty response for delegation agents
  • v1.2.4 (2026-05-15) — auto-add synthesis task when model omits it
  • v1.2.3 (2026-05-15) — auto-fix missing depends_on for synthesis tasks
  • v1.2.2 (2026-05-15) — enforce depends_on in task plans, fix flat plan bug
  • v1.2.1 (2026-05-15) — structured delegation PRs 2-5: execution engine, checkpoint review, summaries, resume from disk
  • v1.2.0 (2026-05-15) — structured delegation (task plan + tool injection) + compiler split into 8 modules
  • v1.1.18 (2026-05-15) — checkpoint visibility, resume on error, checkpoints CLI
  • v1.1.17 (2026-05-15) — delegation cap per child, workers on deepseek-v4-flash
  • v1.1.16 (2026-05-14) — trim developer to 17 skills, fix pandoc tool name
  • v1.1.15 (2026-05-14) — partial child results, tighter archetype prompts
  • v1.1.14 (2026-05-14) — tool result details in progress output
  • v1.1.13 (2026-05-14) — fix parent re-invocation during child delegation
  • v1.1.12 (2026-05-14) — user-facing progress output during execution
  • v1.1.11 (2026-05-14) — fix re-delegation loop on tool limit, bump default to 50 turns
  • v1.1.10 (2026-05-14) — trace parent tracking
  • v1.1.9 (2026-05-14) — run trace with agent call graph and token tracking
  • v1.1.8 (2026-05-14) — invalid self-delegation fix + all-children-done stripping
  • v1.1.7 (2026-05-14) — coordinator nudge fix
  • v1.1.6 (2026-05-14) — dynamic recursion limit
  • v1.1.5 (2026-05-14) — sequential delegation + MCP timeout/retry
  • v1.1.4 (2026-05-14) — sequential delegation fix
  • v1.1.3 (2026-05-14) — forced synthesis fix
  • v1.1.2 (2026-05-13) — forced synthesis at tool limit
  • v1.1.1 (2026-05-13) — cache fix + MarkItDown CWD
  • v1.1.0 (2026-05-13) — M8 MCP integration layer
  • v1.0.36 (2026-05-08) — DAG dependency graph, Rynko content pipeline
  • v1.0.35 (2026-05-05) — parallel delegation, image fallback, path fix
  • v1.0.34 (2026-05-05) — image fallback + path sanitisation fix
  • v1.0.33 (2026-05-05) — cache error fix, Confluence page/image workflow
  • v1.0.32 (2026-05-05) — tool cache, history compaction, /clear
  • v1.0.31 (2026-05-03) — ASCII banner + suppress noisy MCP logs
  • v1.0.30 (2026-05-03) — DRY refactor, code quality fixes, authoring UX
  • v1.0.29 (2026-05-03) — prompt_toolkit for author, AIMessage fix, incomplete nudge
  • v1.0.28 (2026-05-03) — fix AIMessage/Message type mismatch in retry
  • v1.0.27 (2026-05-03) — nudge incomplete tool loop responses
  • v1.0.26 (2026-05-03) — conversation context for worker agents
  • v1.0.25 (2026-05-03) — enforce code-first workflow, 8-turn tool limit
  • v1.0.24 (2026-05-03) — multi-turn tool loop
  • v1.0.23 (2026-05-03) — tool result synthesis, smart path sanitisation, relative grep
  • v1.0.22 (2026-05-03) — smart path sanitisation, relative grep paths
  • v1.0.21 (2026-05-02) — fix embedded env var expansion in MCP servers, master ingestion script
  • v1.0.20 (2026-05-02) — agentic retry loop for tool use
  • v1.0.19 (2026-05-02) — v1.0.19
  • v1.0.18 (2026-05-02) — prompt_toolkit chat UX, CDT config server
  • v1.0.17 (2026-05-02) — MCP server cwd support with schema
  • v1.0.16 (2026-05-02) — MCP server cwd support
  • v1.0.15 (2026-05-02) — fix filesystem MCP server cwd
  • v1.0.14 (2026-05-02) — v1.0.14 debug
  • v1.0.13 (2026-05-02) — maintenance release
  • v1.0.12 (2026-05-02) — multi-tool execution, path sanitisation
  • v1.0.11 (2026-05-02) — multi-tool execution, /model command, DeepSeek workers
  • v1.0.10 (2026-05-02) — /model chat command, verbose logging, forced delegation
  • v1.0.9 (2026-05-01) — verbose agent logging, forced delegation, custom MCP servers
  • v1.0.8 (2026-05-01) — ChromaDB batch ingestion, FTS5 search, Graphify wrapper, PDF/DOCX/Excel support
  • v1.0.7 (2026-05-01) — MCP session persistence, command arg expansion, newline-delimited JSON protocol

April 2026

  • v1.0.6 (2026-04-29) — graceful MCP server failure handling
  • v1.0.5 (2026-04-29) — lazy provider imports, Sterling workspace improvements
  • v1.0.4 (2026-04-28) — multi-turn chat, Sterling workspace, observability prompts
  • v1.0.3 (2026-04-27) — fix swarmkit init prompt formatting
  • v1.0.2 (2026-04-27) — dry-run, markdown audit, annotations
  • v1.0.1 (2026-04-26) — observability
  • v1.0.0 (2026-04-26) — launch prep — skills catalogue, docs site, Docker, PyPI, CHANGELOG