CHANGELOG.md

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]

Added

• Bash-only SWE-bench agent loop (experiments/swebench_lite/run_in_loop.sh) — Pure-bash agent runner using curl + jq for tool dispatch. Zero Python framework dependency for the agent loop itself.
• Enhanced TUI welcome banner — Session start now displays project directory, registered tool count, and deny rule count alongside the existing version, model, and permission mode. The banner gives users immediate visibility into the session configuration without needing to run /tools or /permissions.

Fixed

• CI security scan — Added --ignore-vuln CVE-2026-3219 and --ignore-vuln CVE-2026-42561 to pip-audit invocations in ci.yml. Both are transitive dependencies pinned by upstream libraries with no satisfiable upgrade path.

[0.5.0] — 2026-05-08

Added

• Skills system overhaul — new skill management infrastructure with four new modules:
  • Security scanning (skills/security.py) — static analysis of skill files before installation. Detects 12+ threat categories including obfuscated eval/exec, base64-encoded payloads, crypto-miner patterns, and dangerous shell commands. Risk classification (clean/suspicious/dangerous).
  • Skill evolution (skills/evolution.py) — lesson tracking and automatic skill rewriting (AutoContext pattern). Accumulates corrections per skill, merges duplicates, and rewrites SKILL.md when enough high-confidence lessons accumulate. Timestamped backups before mutation.
  • Dream consolidation (skills/dream.py) — cross-session pruning, dedup, and date normalization modeled on Claude Code's Auto-Dream. Runs periodically (24h) to scan skill directories, merge duplicates, remove stale entries, and convert relative dates to absolute.
  • Wiki bridge (skills/wiki_bridge.py) — auto-generate SKILL.md files from llm-wiki knowledge base pages. Frontmatter extraction, tag filtering, topic matching, and reference preservation.
• Retrieval sub-agent (agent/retrieval_agent.py) — focused code exploration tool that delegates deep searches to a read-only sub-agent. Returns structured file:line-range results without polluting the main agent's context. READ_ONLY risk level; registered as retrieval tool.
• Speculative decoding benchmarks (scripts/benchmark_specdec.py, scripts/start_draft_server.py) — GPU-accelerated draft model server and benchmark suite for measuring tokens-per-second with/without speculative decoding on RTX 5070 Ti and compatible hardware.
• Enhanced codebase index (context/codebase_index.py) — improved semantic code search with better chunking, symbol extraction, and index freshness detection. 118 lines of enhancements.
• Model presets in config: fast, balanced, quality, local, cloud, frontier shortcuts in MODEL_PRESETS dict. Default model is now openai/qwen2.5-coder-14b with llama.cpp GPU spec decoding.
• Expanded driver catalog (llm/driver_catalog.yaml) — new entries for DeepSeek V4, Kimi K2.6, and updated NVIDIA NIM endpoints.
• datasets dependency added for training pipeline integration.
• Removed unused textual dependency — cleanup from prior framework experimentation.
• Enhanced llamacpp_manager tool — health checks, model lifecycle management, and server status reporting for llama.cpp inference.
• Expanded LLM client tests — 324 new assertions covering edge cases, retry logic, and model fallback chains.

Changed

• Default model changed from ollama/qwen3:4b to openai/qwen2.5-coder-14b with llama.cpp GPU spec decoding (~750 tok/s on RTX 5070 Ti with 1.5B draft model).
• Documentation audit fixes: standardized tool naming from Bash/Shell to shell(...) across docs/permissions.md, src/godspeed/security/rules.py, src/godspeed/security/permissions.py, and src/godspeed/config.py.
• Test reorganization: moved tests/test_skills_commands.py to tests/test_skills/test_commands.py; removed redundant tests/test_evolution/test_skill_gen.py.

[0.4.0] — 2026-05-01

Added

• Auto-load ~/.godspeed/.env.local on startup (cli._load_env_files). API keys in ~/.godspeed/.env / .env.local (and the project's .godspeed/.env / .env.local) are now injected into os.environ before LiteLLM runs, so NVIDIA_NIM_API_KEY=... godspeed one-offs AND long-lived ~/.godspeed/.env.local files both work without extra shell plumbing. Precedence: project/.env.local > project/.env > global/.env.local > global/.env; shell env wins over all of them so per-run overrides keep working. Matches dotenv/Vite/Next convention: .env.local is the gitignored local override. Log messages contain key names only, never values. Closes the "first-run auth error" UX paper cut that required users to remember to $env:NVIDIA_NIM_API_KEY = ... every shell.
• Troubleshooting doc: new "AuthenticationError: api_key client option must be set" section in docs/troubleshooting.md with the fix + the pre-v3.5.0 workaround.

Changed

• Performance optimizations across core modules: reduced fsync frequency in audit trail (1→10 with guaranteed first-write sync), pre-computed lowercase keyword set in dangerous command detection, cached model string in LLM client, single-pass post-processing in agent loop, output caching in background process registry, async file/folder I/O in mention resolution, compiled mention regex in completions, and module-level imports in TUI app.
• Background process auto-cleanup: completed processes are now automatically removed from the registry to prevent memory leaks during long sessions.
• Removed competitive product comparisons from README, CHANGELOG, and audit docs. Project positioning is now strictly feature-driven.
• Renamed PyPI package from godspeed to godspeed-coding-agent to resolve naming conflict with an existing project. Install command updated throughout documentation.

Fixed

• Background test compatibility: test_output_after_completion and test_kill_already_exited updated to handle auto-cleanup behavior where completed processes are removed from the registry.
• Home directory expansion: ~ now correctly expands in glob_search, grep_search, and resolve_tool_path.
• Inaccessible file handling: glob_search and grep_search gracefully skip files with permission errors on Windows.

[0.4.0] — 2026-04-22

Phase 2 of the world-class UX push. Two independent additions on top of v3.3.0, both landed with full test coverage and CI-green across Python 3.11 / 3.12 / 3.13. Zero breaking changes — every default behaves identically to v3.3.0 until you opt in.

Added

• /remember slash command for persisting permission rules without leaving the TUI (tui.commands._cmd_remember). Writes to ~/.godspeed/settings.yaml (or .godspeed/settings.yaml with --project) and adds the rule to the live PermissionEngine so it takes effect immediately — no restart. Accepts approve / allow / deny / ask as the action word; approve is an alias for allow that reads more naturally. Pattern syntax is validated up-front (Tool(argument) form) so typos fail loudly instead of saving an unmatchable rule. Duplicates are silently idempotent. New companion doc: docs/permissions.md covering the 4-tier model end-to-end.
• PermissionEngine.add_rule(pattern, action) for runtime rule injection — used by /remember but available for any code path that needs to widen (or narrow) the permission set mid-session.
• config.append_permission_rule(pattern, action, project_dir=) — generic YAML write-back helper covering all three tiers. Replaces the old allow-only append_allow_rule, which is kept as a back-compat wrapper.
• Task-aware model routing (llm/router.py, GodspeedSettings.cheap_model / strong_model). Each LLM turn is classified from conversation state (no extra LLM call) into one of plan / edit / read / shell; the existing ModelRouter swaps self.model for the duration of the call and restores on cleanup. cheap_model populates routing[edit/read/shell] and strong_model populates routing[plan] so users get the cost-vs-quality split without writing the full dict. Explicit routing: entries in YAML still win. Streaming and non-streaming paths both honor the task_type parameter; mid-turn cancel still closes the routed stream cleanly via the existing aclose() finally-block. Headless mode and configs without shortcuts behave identically to v3.3.0 (no routing → default model for every call).

[3.3.0] — 2026-04-22

World-class coding-agent UX push. Five additions on top of the security-first core, all born from a 6-task daily-use benchmark that surfaced two real production bugs (both fixed) and four UX gaps (all addressed). Every change shipped with tests, CI-green across Python 3.11 / 3.12 / 3.13.

Version jumps 3.1.0 → 3.3.0 to align with the research-track label (v3.2 covered the LLM-judge / best@k selector + four null results; artifacts are in experiments/swebench_lite/ and the tech-report draft on Desktop).

Added

• Mid-turn cancellation (agent.result.AgentCancelledError, agent.loop.agent_loop(cancel_event=...)). The agent now checks a cancel event at three checkpoints per iteration — top of loop, after pause release, between streaming chunks — and unwinds cleanly via AgentCancelledError when set. _streaming_call uses a finally-block aclose() so the underlying httpx stream shuts down promptly. TUI installs an asyncio SIGINT handler on Linux/macOS: first Ctrl+C cancels the current turn; second press within 1 s hard-interrupts (Jupyter pattern). Windows ProactorEventLoop degrades gracefully. Distinct from pause — pause stops at iteration boundary; cancel stops immediately.
• Diff approve-before-write gate (tools.base.DiffReviewer Protocol, Tool.produces_diff class attribute, ToolContext.diff_reviewer). Two independent axes of consent now: PermissionEvaluator answers "may this tool run?" (existing); DiffReviewer answers "should THIS specific diff be applied?" (new). file_edit / file_write / diff_apply opt in via produces_diff = True; each calls await context.diff_reviewer.review(tool_name, path, before, after) just before write. "accept" proceeds; any other value (including forward-compat "edit") rejects. TUI _InteractiveDiffReviewer renders a side-by-side diff via Rich's Syntax("diff") lexer with (y)es · (n)o · (a)lways keys; session-scoped "a" bypass. Headless/CI: diff_reviewer stays None, writes proceed as before.
• Per-turn cost/token HUD (tui.output.format_status_hud). Compact one-line summary after every agent turn: · 1,234 in + 567 out (1,801) · $0.0024 · model · 3 turns. When max_cost_usd > 0 shows $X / $Y and tints WARNING when remaining budget drops below 20%. Provider prefix stripped for readability. Pure function; reuses existing LLMClient accumulators.
• Troubleshooting + Windows quickstart docs (docs/troubleshooting.md, docs/quickstart_windows.md, docs/demo.md). Top-of-funnel issues from the production audit: UnicodeEncodeError on cp1252, NIM 40 RPM mitigations, WSL + Docker for swebench (with swebench 4.1.0 pvlib NumPy bug caveat), audit-verify workflow, Ollama first-run, post-edit syntax gate rejections, diff reviewer keys, mid-turn Ctrl+C semantics. Windows quickstart: 5-minute path from Miniconda → pip → godspeed init → first task. Demo recording is user-gated; docs/demo.md captures the asciinema + agg recipe.
• README "What's new in v3.3.0" section — 5-row summary table cross-linking every new feature + the production audit.

Fixed

• file_edit silent indentation corruption on multi-line replace (PR #78). The fuzzy matcher could find an indented region, but a poorly-indented new_string would write a syntactically broken file without warning. Now the tool runs ast.parse(new_content) on .py / .pyi and json.loads(new_content) on .json. If the pre-edit content parsed but the post-edit would not, the write is rejected with a clear error asking the agent to include more context. Pre-broken files pass through so the agent can fix them. Surfaced by the daily-use benchmark's T4 task (multi-file user_id → account_id rename); with the fix the same task now passes.
• Windows UnicodeEncodeError on CLI exit (PR #78). The agent's final summary often contains →, —, or smart quotes — none encodable in cp1252. cli.py now calls _force_utf8_stdio() at module load to re-wrap sys.stdout and sys.stderr as TextIOWrapper(encoding="utf-8", errors="replace"). Idempotent; skips already-utf8 streams.

Changed

• Dependabot now ignores pydantic and python-dotenv bumps (PR #73). Both are pinned strictly by litellm>=1.83.7 (pydantic==2.12.5, python-dotenv==1.0.1) — upstream-blocked so auto-opened PRs are unmergeable. Un-ignore when LiteLLM relaxes its pins.

Tested

• Daily-use benchmark on C:\Users\ttimm\Desktop\godspeed_benchmark\ runs 4/4 passes after the two fixes above; extended to 6/6 with harder refactor
  • bug-fix scenarios (T5, T6).
• Production audit adds path-traversal (T7a blocked correctly), audit-trail chain verify (21 records on a live session, tamper detection confirmed), and full-suite regression (1861+ pass; 10 flaky system_optimizer tests pre-exist and are environmental).
• CI 8/8 green across lint / security / type-check / CodeQL / Analyze / tests 3.11/3.12/3.13 on every feature PR (#78-#82).

[3.1.0] — 2026-04-21

Agent-in-loop Docker oracle tool + ensemble selector + SystemOptimizerTool

• driver registry + web-search hardening. Headline number is from an oracle-selector best-of-5 ensemble, disclosed transparently — see the Benchmarks subsection below and experiments/swebench_lite/findings_2026_04_21.md.

Added

• Agent-in-loop Docker oracle (experiments/swebench_lite/run_in_loop.py, docker_test_tool.py, run.py --agent-in-loop). The agent can call swebench_verify_patch mid-session to run the real SWE-Bench test harness against its current edits, then iterate on the failure output. Budget: 5 verify calls per instance (hard cap 8), with a working-tree SHA short-circuit to prevent re-verifying unchanged diffs.
• Oracle-guided best-of-N selector (experiments/swebench_lite/oracle_merge.py). Takes predictions.jsonl:report.json pairs and picks per instance the patch that actually resolved (preferring shortest among resolvers; falling back to shortest non-empty). Methodology is the same pattern other open-source agent projects publish — explicitly labeled as oracle_best_of_N in output model_name_or_path.
• Driver registry (src/godspeed/llm/driver_catalog.yaml, src/godspeed/agent/prompt_profiles.py, scripts/validate_driver.py, docs/adding_a_driver.md). Ten catalogued drivers (NVIDIA NIM free tier, Moonshot direct, Anthropic frontier, Ollama local + cloud). Three prompt profiles (default, thinking, minimal) with robust fallbacks. New-driver smoke runs 3 easy instances and gates on a ≤20% LLM-error rate before allowing benchmark use.
• SystemOptimizerTool (src/godspeed/tools/system_optimizer.py), inspect and recommend modes (both READ_ONLY). Cross-platform (Windows/Linux/macOS) using psutil + optional pynvml. Hard deny-list for system-critical processes (explorer.exe, systemd, launchd, kernel_task, etc.), documented up front even though act mode is not yet shipped. Recommends Ollama unload, memory/disk pressure remediation, and outlier-process investigation.
• Instance cooldown flag (run.py --instance-cooldown SECONDS) to smooth NVIDIA NIM free-tier rate-limit (40 RPM) when running agent-in-loop sessions sequentially.
• Web-search disk cache (src/godspeed/tools/web_fetch.py). 7-day TTL at ~/.godspeed/cache/web/<sha1>.json with 50 MB LRU-by-mtime eviction. New no_cache: true bypass flag. Transparent; no agent prompt change needed.
• --allow-web-search flag on experiments/swebench_lite/run.py (defaults to False). Benchmark runs now register the tool registry with tool_set="local" so the agent cannot leak the ground-truth fix via web_search / web_fetch during evaluation.
• reproduce_v3_1.sh one-command SWE-Bench Lite reproduction script.
• SWE-Bench research memo at ~/Desktop/SWE-Bench-Research-Memo/ (MD + PDF) documenting current SOTA, the dataset noise floor, the five compounding levers, and honest positioning for a solo-engineer OSS agent.

Changed

• src/godspeed/tools/shell.py force-kills the process tree on timeout. Replaces subprocess.run(timeout=N) with subprocess.Popen + communicate(timeout=...) + psutil-based tree kill on TimeoutExpired. Addresses a Windows-specific failure mode where grandchildren holding pipes survived the parent's kill and blocked the runner for 60–100+ minutes. Now returns within timeout + ~5s cleanup window. Six new regression tests in tests/test_shell_tool.py guard the path.
• --verify-retry deprecated in favor of --agent-in-loop. The former is a post-hoc single-shot retry; the latter gives the agent live oracle feedback during the session. --verify-retry now logs a deprecation warning and no-ops when combined with --agent-in-loop.
• psutil>=5.9,<8.0 is now a required dependency (was optional) — SystemOptimizerTool registers by default and the test suite needs it.
• aiohttp>=3.13.4 pin-up to patch 10 CVEs (CVE-2026-34513..34525, CVE-2026-22815) transitively carried through litellm.
• Settings.yaml.example pre-registers moonshot/kimi-k2.6, moonshot/kimi-k2.5, ollama/kimi-k2.6:cloud, ollama/kimi-k2.5:cloud driver paths (commented) so swapping providers is a one-line edit.

Fixed

• CodeQL was flagging intentionally-broken fixture code (benchmarks/fixtures/easy-fix-syntax-01/app.py) as a syntax error. .github/codeql/codeql-config.yml now excludes benchmarks/fixtures/, experiments/, and tests/fixtures/ from scanning.
• python-dotenv CVE-2026-28684 is suppressed via pip-audit --ignore-vuln CVE-2026-28684 in CI. The fix (python-dotenv 1.2.2) is unsatisfiable because litellm pins python-dotenv==1.0.1 strictly; the suppression is temporary, pending a litellm release that relaxes the bound. Commented in .github/workflows/ci.yml.

Benchmarks

SWE-Bench Lite dev-23 (Python-only, 23-instance subset). All numbers are free-tier; $0 API spend, ~6 hours total wall-clock on a single consumer laptop (RTX 5070 Ti 16 GB). Method column is the important part — each row uses a different inference strategy over the same 23-instance split.

Version	Method	Resolved	Rate
v2.12.0 baseline	Kimi K2.5 single-shot	8 / 23	34.8%
v3.1 Phase 1 (null result)	Kimi K2.5 + agent-in-loop, single seed	7 / 23	30.4%
v3.1.0 headline	Oracle-selector best-of-5 (free-tier)	12 / 23	52.2%

Methodology disclosure — the 52.2% headline is an oracle_best_of_5: the 5 constituent runs (Kimi K2.5 single-shot, GPT-OSS-120B, Qwen3.5-397B iter1, Qwen3.5-397B seed3, Kimi K2.5 + agent-in-loop) were each submitted to sb-cli standalone and paid their own quota slot. The selector (oracle_merge.py) picks per instance the patch from whichever run resolved — preferring shortest among resolvers; falling back to shortest non-empty when no run resolved. This is the same pattern used in published open-source agent research; it is explicitly labeled as such.

Single-run result was a null result. Kimi K2.5 + agent-in-loop alone scored 7 / 23 = 30.4%, below the 34.8% single-shot baseline. Prompt tuning did not fix it. The instance-level analysis shows agent-in-loop is additive inside the ensemble (it contributed marshmallow-code__marshmallow-1359, a unique resolve no other run landed) — but on this driver, a verify-feedback loop alone over-edits on hard instances and stops early on easy ones.

Limitations. dev-23 is a 23-instance subset of SWE-Bench Lite's 300 dev instances. Single-seed runs for most constituent drivers (Kimi seed 2 was contamination-affected by NVIDIA NIM concurrent-job contention and excluded). Test-50 and dev-300 headline validation is pending for v3.1.x. Published SOTA on full SWE-Bench Lite (April 2026) is 62.7%; top open-source agents with paid frontier drivers sit in the 40-50% band on the same benchmark.

Reproducibility:

./experiments/swebench_lite/reproduce_v3_1.sh

Full findings, per-instance resolution map, run-by-run pre-merge numbers, and the research memo that framed this release:

• experiments/swebench_lite/findings_2026_04_21.md
• ~/Desktop/SWE-Bench-Research-Memo/ (local, not checked in)

[2.11.0] — 2026-04-19

Benchmark and test-infrastructure polish. No runtime behavior changes to the agent itself — this release strengthens the evidence that Godspeed works end-to-end before any fine-tuning work begins.

Added

• Real-LLM smoke tests under @pytest.mark.real_llm (tests/test_smoke_real_llm.py, tests/test_integration_real.py). Skipped by default; run with pytest -m real_llm and a running Ollama server on localhost:11434. Includes a multi-turn read-edit-verify scenario that proves the agent loop chains real tool calls with a real model — the foundation claim the prior test suite did not exercise.
• Benchmark fixture directories (benchmarks/fixtures/<task_id>/) for 19 of the 20 tasks in benchmarks/tasks.jsonl. Each is copied into an isolated temp workspace per run so scores are reproducible across models and invocations. Optional _setup.py hook handles tasks that need runtime state (git repos with dirty/staged changes). Optional verify.py hook runs post-agent and returns a machine-checkable mechanical-success signal for 13 tasks (syntax fix, SQL-injection remediation, requests→httpx migration, CI workflow creation, etc.).
• waste_penalty metric in BenchmarkScore and BenchmarkSuiteResult. Deducts up to 0.3 when an agent issues > 1.5× the expected tool calls, so honest efficiency shows up in the score instead of being hidden by the Jaccard + LCS primaries. Tie-breaker, not a dominant signal.
• test_parallel_file_reads_real_tools — complements the existing _TrackedTool synthetic test by running real FileReadTool in parallel on real files, locking in the concurrency guarantee for the production tool path.

Changed

• scripts/run_benchmark.py now uses benchmarks/fixtures/<task_id>/ as the per-task workspace when present (falls back to --project-dir otherwise). Records mechanical_success per task and mechanical_pass / mechanical_evaluated / mean_waste_penalty in summary.json.

Fixed

• Prior audit erroneously flagged speculative tool dispatch as scaffolded-but-inactive. Verified the cache is populated by _speculative_dispatch during streaming (src/godspeed/agent/loop.py) and exercised by 7 passing tests in tests/test_speculative.py.

Benchmarks

First honest end-to-end numbers for Godspeed on a real LLM, run against the polished fixtures:

Driver	Overall	Pass	Mech
nvidia_nim/qwen/qwen3.5-397b-a17b	0.608	11/20	7/13
nvidia_nim/moonshotai/kimi-k2.5	0.548	9/20	6/13
nvidia_nim/mistralai/devstral-2-123b-instruct-2512	0.446	5/20	2/13
nvidia_nim/qwen/qwen3-coder-480b-a35b-instruct	0.333	5/20	2/13
ollama/qwen3-coder:latest	0.107	1/20	1/13

Recommended production driver: nvidia_nim/qwen/qwen3.5-397b-a17b. Full table: experiments/benchmark_shootout_2026_04.md.

[2.10.0] — 2026-04-18

Minor release bundling two independent features: the Phase A1 synthetic tool-calling training-data pipeline, and weak-model robustness at the agent layer (tool-name aliasing + tool-set filtering).

Added — Phase A1 pipeline (experiments/phase_a1/)

• End-to-end synthetic data pipeline producing ~6,200 {messages, tools} OpenAI-format training samples on a $0 free-tier-only budget. Stages: specs (stratified 21-tool × 4-category gen plan) → blueprints (LLM-planned JSON with per-tool arg validation + retry) → executor (real tool dispatch against a seeded sandbox, with fixtures for network-touching tools) → narrator (LLM assistant-reasoning with anti-hallucination guards + retry) → emit (canonical OpenAI record) → judge (4-dim rubric scoring) → validate (schema + coverage gate) → assemble (merge anchor + augment + synthetic + distill, dedup by user-prompt hash, deterministic shuffle).
• Per-tool argument validation inside blueprint generation. The same invariants validate.py enforces post-execution now run pre-execution via the new public validate_tool_call_args(tool_name, args) helper. Malformed blueprints (e.g. github.action=None, grep_search.pattern="") are rejected before executor + narrator spend is incurred, then retried with a temperature bump. Closes the three failure modes that burned a prior 5-sample smoke run (0/5 yield).
• Anti-hallucination narrator guards. System prompt now forbids the narrator from inventing facts (PR state, tool output values, counts) not present in the executed transcript. A narrate_session retry loop catches structural failures like pre_call length != N and re-samples with bumped temperature before giving up.
• Observability. executor.py, judge.py, and four sites in orchestrate.py now call logger.warning(..., exc_info=True) on caught exceptions so failed samples are debuggable from the log stream rather than only via the truncated error field in run_metrics.jsonl.
• Multi-provider async router with SQLite-backed per-provider daily quota tracking (providers.py): Cerebras + Z.ai + Groq + Ollama cascade, UTC-midnight quota reset, rate-limit backoff.
• Makefile a1-* targets for each stage (a1-smoke, a1-run, a1-run-prod, a1-validate, a1-judge, a1-anchor, a1-distill, a1-augment, a1-assemble).
• Per-file ruff ignore for experiments/**/*.py — allow S311 (jitter via random), N818 (error naming), SIM103 style nits in research code.
• .gitignore rules excluding derived data artifacts: experiments/phase_a1/data/*.jsonl, *.db, sessions/, _*/, plus .env.local.

Added — Agent tool-name aliasing + tool-set filtering

• src/godspeed/tools/aliases.py: common hallucinated tool names (read_file, grep, glob, etc.) get canonicalized to their registered equivalents inside _parse_tool_call. Small open-source models express a correct intent with the wrong label — this closes the dead-end without a fine-tune.
• src/godspeed/tools/tool_sets.py: named capability surfaces (local / web / full). New --tool-set CLI flag constrains the registry so local-codebase runs hide web_search / web_fetch / github entirely. Weak models stop picking web_search over file_read by construction, not by prompting.
• QUALITY_PROMPT extension with explicit Tool Selection Defaults so the model's written guidance matches the runtime constraint.

Tests (+18 new)

• experiments/phase_a1/tests/test_blueprints.py (11 tests): validates the per-tool-arg rejection of github.action=None and grep_search.pattern="", verifies retry-on-bad-args, retry-on-invalid-JSON, temperature bump on each retry, failure after exhausted retries, and prompt content for error-prone tools.
• experiments/phase_a1/tests/test_narrator.py (7 tests): verifies retry on pre_call length mismatch, retry on invalid JSON, temperature bump, failure after exhausted retries, content-injection into the session JSONL, and anti-hallucination prompt content.
• Phase A1 test suite total: 125 passing (excludes test_swesmith_distill.py which has a pre-existing sklearn import gap unrelated to this release).
• Project-wide suite: 1706 passing / 20 skipped.

[2.9.1] — 2026-04-17

Patch release — compatibility shim and local-model documentation. Shipped alongside the Stage A findings from the Track B re-evaluation plan.

Added

• src/godspeed/llm/qwen3_coder_parser.py — regex extractor for Qwen3-Coder's <function=name>\n<parameter=key>\nvalue\n</parameter>\n</function> tool-call XML. Ollama 0.20.x doesn't recognize this format — calls come back in the content field instead of tool_calls. The parser synthesizes OpenAI-shaped tool_calls so the rest of Godspeed's pipeline sees a standard response. Hooked into LLMClient._call() as a no-op when tool_calls is already populated.
• Model table entries for ollama/qwen3:14b and ollama/qwen3-coder:latest in the models CLI command and settings.yaml.example — documented but not promoted to default.
• scripts/run_benchmark.py — measurement runner that shells out godspeed run per task and scores via training.benchmark. Not a shipped CLI command; a tool for Stage A / future benchmark runs.
• Per-file ignore for scripts/**/*.py in pyproject.toml — print() is intentional in CLI scripts.
• .godspeed/training/, .godspeed/checkpoints/, .godspeed/memory.db* added to .gitignore — session artifacts should not track into repos.

Tests (+20 new)

• tests/test_qwen3_coder_parser.py covers the parser's detector, the type-coercion (bool/int/float/JSON/string), multi-call responses, malformed input, and uniqueness of synthesized IDs.

[2.9.0] — 2026-04-17

Final entry in the v2.5.1 review follow-up chain. Auto-index on session start so the agent has semantic code search available by default instead of requiring a manual /reindex.

Added

• Auto-index on session start (src/godspeed/context/auto_index.py). New helper maybe_start_auto_index(project_dir, auto_index_enabled) that:

  • Returns None when disabled, when chromadb isn't installed (graceful degradation via the [index] extra), or when the index is already fresh (needs_reindex() returns False).
  • Otherwise schedules build_index_async() as an asyncio.Task so the session continues without blocking on index construction.
  • Swallows exceptions in the build coroutine so an indexing failure never crashes the agent.

  Wired into both the TUI and headless paths in cli.py, right after the ToolContext is constructed.

• GodspeedSettings.auto_index field (default True). Disable with auto_index: false in ~/.godspeed/settings.yaml or GODSPEED_AUTO_INDEX=false.

Changed

• Sessions started without a fresh codebase index now rebuild it in the background automatically (assuming [index] extra is installed). The previous behavior required a manual /reindex.

[2.8.0] — 2026-04-17

Closes the last of the v2.5.1 review follow-ups. New LLM-driven test generation tool, architectural affordance for tools that need an LLM, and the semantic fix to verify._verify_with_retry that was deferred from earlier patches.

Added

• generate_tests tool (src/godspeed/tools/generate_tests.py) — reads a source file, asks the LLM to produce a complete pytest module, writes it to tests/test_<basename>.py (or a caller-specified path). The agent can then run test_runner to confirm the generated tests pass. Strips markdown code fences if the LLM returns them. Closes the test-first discipline loop with one tool call.
• LLMInvoker protocol + ToolContext.llm_client field (src/godspeed/tools/base.py) — architectural affordance for tools that need to make LLM calls. Decouples tools/ from a concrete llm.LLMClient reference. Both headless and TUI paths now populate llm_client when constructing the ToolContext.

Changed

• verify._verify_with_retry now returns ToolResult.failure when unresolved lint errors remain after retries (previously it returned a success-typed result with "some remaining" in the output body). The MUST-FIX gate still fires on the same fingerprint — it now checks both result.error and result.output so it's robust across both the new and legacy shapes. Downstream callers get a clear is_error=True signal where before they had to substring-match the output.
• _build_tool_registry (cli.py) registers GenerateTestsTool alongside the other quality tools.
• The three tests in test_verify_fix_retry.py that previously asserted not result.is_error with remaining issues have been updated to assert result.is_error and read the fingerprint from result.error.

Deferred to v2.9.0

• Auto-index on session start (separate concern — session-init refactor).

[2.7.0] — 2026-04-17

Quality-tooling minor release, continuation of v2.6.0. Closes two more deferred items from the v2.5.1 review and adds an integration-level invariant check between the audit trail and training log.

Added

• complexity tool (src/godspeed/tools/complexity.py) — wraps radon cc (Python, with grade mapping from the McCabe CC threshold) and lizard (polyglot, when installed). Functions above max_cc (default 10) fail the call — lets the agent enforce a complexity budget on edits rather than catching it in code review.
• dep_audit tool (src/godspeed/tools/dep_audit.py) — auto-detects the project's package manager from manifests (pyproject.toml / requirements.txt → pip-audit; package.json → npm audit; Cargo.toml → cargo-audit) and runs the matching CVE scanner. Vulnerabilities produce an error result so the agent treats them as gating. Explicit manager argument overrides detection.
• E2E drift test (tests/test_audit_log_drift.py) — integration-level assertion that the audit trail's session_end detail and ConversationLogger.log_session_end record agree on every field (exit_reason, exit_code, all metrics). Catches regressions where the two streams could drift after independent edits.

Changed

• _build_tool_registry (cli.py) registers ComplexityTool and DepAuditTool alongside the existing quality tools.

[2.6.0] — 2026-04-17

Quality-tooling minor release. Two new tools the agent can call to enforce production-grade code; efficiency telemetry in training logs; follow-up cleanup from v2.5.1.

Added

• coverage tool (src/godspeed/tools/coverage.py) — wraps coverage run pytest + coverage report -m. Optional min_percent argument fails the call below the threshold, turning coverage into a quality gate the agent can enforce per-session. Auto-detects --source (./src if present, else cwd). Graceful error when coverage isn't installed.
• security_scan tool (src/godspeed/tools/security_scan.py) — wraps bandit (Python) and semgrep (polyglot, when installed). Both optional; falls back when neither is available. Non-zero findings produce an error result so the agent treats them as gating rather than advisory. Configurable minimum severity (low / medium / high).
• AgentMetrics.must_fix_injections counter — increments each time the MUST-FIX gate injects a fix-required message. Exposed in headless JSON output, audit trail session_end detail, and ConversationLogger log_session_end record. Training signal: agents that trigger many MUST-FIX injections are less efficient per unit of successful work; downstream RL (GRPO) can penalize against this counter.

Changed

• Fingerprint coupling reduced — verify.REMAINING_ERRORS_FINGERPRINT is now a module-level constant imported by loop._maybe_inject_must_fix. Eliminates the duplicated "some remaining" string literal identified in the v2.5.1 review follow-ups.
• ConversationLogger.log_session_end(...) accepts a new must_fix_injections keyword argument (defaulted to 0 for backwards-compatibility with existing callers).
• _build_tool_registry (cli.py) registers CoverageTool and SecurityScanTool alongside the existing quality tools.

Fixed

• tests/test_agent_result.py::test_duration_before_finalize — flaked on Windows where time.monotonic() granularity could return identical values across a 0.01s sleep. Assertion relaxed to >= 0.0 (the semantically correct invariant — duration is never negative).

[2.5.1] — 2026-04-17

Patch release — code-quality enforcement and training-record alignment. No new user-facing CLI flags or JSON schema changes.

Added

• ConversationLogger.log_session_end() emits a terminal record per headless session with exit_reason, exit_code, iterations_used, tool_call_count, tool_error_count, duration_seconds, cost_usd. Fields mirror the audit-trail session_end detail so the two streams stay comparable. Enables RL pipelines (GRPO) to shape rewards on exit_code without parsing the audit log.
• Quality defaults in the system prompt: a new QUALITY_PROMPT block (type hints, test-first, read-before-edit, comments-only-when-non-obvious, secrets policy, anti-premature-abstraction) appended after WORKFLOW_PROMPT.

Changed

• MUST-FIX gate on auto-verify failures. When _auto_verify_file leaves unresolved lint errors (fingerprint: "some remaining"), the agent loop injects a user-role message naming the file and errors, instructing the model to fix them before any other edits. Capped at 3 injections per session — after the cap, the gate logs a warning and fails open so the agent isn't deadlocked on a fundamentally unfixable error (e.g., broken project ruff config). Wired into both parallel and sequential dispatch paths. Structural enforcement closes a silent-error channel where the model could read a verify success marker despite persistent lint issues.

Fixed

• src/godspeed/__init__.py:5 — stale __version__ (was "2.2.0"; now tracks pyproject.toml).

[2.5.0] — 2026-04-16

MLOps-readiness release. Closes the headless/pipeline gaps found in the feature + logic audit on 2026-04-16. The godspeed run command is now suitable for unattended use in CI, W&B sweeps, and ML research pipelines.

Added

• Exit code contract (godspeed.agent.result.ExitCode). godspeed run now returns differentiated exit codes so orchestrators can switch on them: 0 success, 1 tool error, 2 max iterations, 3 budget exceeded, 4 LLM error, 5 invalid input, 6 timeout, 130 interrupted. Codes are stable across minor versions.
• JSON output schema extended with exit_reason, exit_code, iterations_used, tool_calls (list of {name, is_error}), tool_call_count, tool_error_count, duration_seconds, cost_usd, and audit_log_path. Existing fields preserved.
• --timeout N wall-clock session cap on godspeed run. Wraps the agent loop in asyncio.wait_for; exits with code 6 on timeout. Default 0 (no limit) preserves existing behavior.
• --prompt-file FILE flag on godspeed run reads the task from a file. Task input precedence: --prompt-file > positional > stdin. Passing - as the positional task reads from stdin; an empty positional with a piped stdin also reads from stdin.
• Rate-limit retry with exponential backoff + jitter in LLMClient. 429/quota errors now retry up to 4 times on the same model (delays 1s/2s/4s/8s ± 25% jitter, capped at 60s) before falling over to the next model. Honors Retry-After when the provider supplies it (upward-only jitter — we don't retry earlier than the provider asked).
• AgentMetrics accumulator on agent_loop() via the new optional metrics kwarg. Populates iterations_used, tool_calls, exit_reason, duration_seconds.

Changed

• Headless mode now has an audit trail by default. Previously godspeed run wired audit=None into the tool context — unattended sessions had no tamper-evident log, the opposite of the project's security posture. Audit now writes to ~/.godspeed/audit/{session}.audit.jsonl with session-start and session-end records bookending every run.
• agent_loop() gains the optional metrics: AgentMetrics | None kwarg. Omitting it preserves the prior behavior exactly — backwards-compatible.

Fixed

• Addresses G1–G7 from the 2026-04-16 feature audit.

Security

• Audit trail fails closed on I/O errors: AuditTrail.record() now raises AuditWriteError when a write or fsync fails, instead of silently logging and advancing the in-memory chain. Chain state (sequence, prev_hash) does not advance on failure, so a successful retry chains cleanly from the last persisted record. This closes a gap where disk-full or permission errors would poison the hash chain while the agent continued executing tools.
• Evolution safety gate blocks mutations to security-sensitive tool descriptions: Mutations whose artifact_id is in SECURITY_SENSITIVE_TOOL_IDS (shell, bash, file_write, file_edit, diff_apply, git, github, background) now require human review. Mutations whose text matches any SECURITY_BYPASS_PATTERNS regex (e.g. "always granted", "bypass permission", "ignore safety", "auto-approve") also require review regardless of artifact.

Changed

• README, SECURITY.md, and architecture doc updated: dangerous-pattern count corrected from "72+" to 71 (actual); tool count updated from "18+" to 25.
• CI matrix extended to Python 3.13 (previously 3.11, 3.12).
• Pre-commit adds mypy (src-scoped) and bandit (low-severity filter) hooks; ruff hook bumped to v0.14.1.
• make lint no longer auto-fixes or formats — matches CI exactly. New make fix target runs ruff check --fix && ruff format .. make test now runs with coverage gate to match CI.

Added

• CodeQL workflow (.github/workflows/codeql.yml): security-and-quality query set, runs on push/PR and weekly cron.
• Release workflow (.github/workflows/release.yml): builds wheel/sdist on v* tag push, generates CycloneDX SBOM, attests build provenance via sigstore, attaches to GitHub release.

[2.3.0] — 2026-04-13

Added

• Fine-tuning data pipeline (Units A-F): Full infrastructure for collecting training data and fine-tuning tool-calling LLMs on Godspeed conversations.
  • ConversationLogger: Persists every conversation message (user, assistant, tool results, compaction summaries) to per-session JSONL at ~/.godspeed/training/. Gated on log_conversations config.
  • TrainingExporter: Converts conversation logs to openai, chatml, and sharegpt fine-tuning formats with filtering (min_tool_calls, success_only, min_turns, tool whitelist). CLI: godspeed export-training.
  • Per-step reward annotations: Automatic reward signals for GRPO/DPO — success (+1.0), verify passed (+0.5), dangerous command (-1.0), efficient sequence bonus (+0.5). Session-level summarization.
  • Benchmark suite: 20 hand-crafted tasks (easy/medium/hard) with Jaccard tool selection scoring and LCS sequence quality scoring.
  • Tool description enhancement: All 10 core tools now include inline usage examples and JSON Schema examples fields for better training signal.
  • Common workflows in system prompt: 5 canonical multi-step patterns (fix bug, add feature, explore codebase, git workflow, research/debug).
• 110 new training pipeline tests (total: 1,557 passing)

[2.2.0] — 2026-04-12

Added

• Self-evolution system (Units 20-27): Learn from execution traces to improve prompts, tool descriptions, and permissions automatically. Runs entirely on Ollama for $0 — optional API acceleration.
  • Trace analyzer: Parse audit trail JSONL into failure patterns, latency stats, permission insights, and repeated tool sequences.
  • Evolution engine: GEPA-style LLM-guided mutations for tool descriptions, system prompt sections, compaction prompts, and auto-generated skills.
  • Fitness evaluator: A/B testing with LLM-as-judge scoring (0.5×correctness + 0.3×procedure + 0.2×conciseness). Length penalty for bloat.
  • Safety gate: Size limit (<2x growth), semantic drift (Jaccard ≥0.3), fitness threshold, confidence check, human review for high-impact artifacts.
  • Evolution registry: Append-only JSONL history with apply/revert/rollback. Originals backed up before mutation.
  • Runtime hot-swap: Update tool descriptions in-memory via ToolRegistry._description_overrides. Prompt section overrides loaded on startup.
  • Cross-session learning: Aggregate insights across sessions, model-specific analysis, regression detection with rollback/investigate/monitor recommendations.
  • Skill auto-generation: Detect repeated multi-tool patterns (≥3 occurrences) → generate reusable skill markdown with YAML frontmatter.
  • Permission pattern learning: Analyze denial/approval patterns → suggest allowlist optimizations with rationale.
  • Hardware-aware model selection: Auto-detect VRAM (nvidia-smi + Jetson /proc/meminfo), pick largest fitting model from tier list. Scales candidates and batch sizes by available memory. Jetson Orin Nano 8GB → qwen2.5:3b with 2 candidates.
  • /evolve command: status, run, history, rollback, review, approve, reject subcommands.
• 175 new evolution tests (total: 1,400+ passing)

Changed

• ToolRegistry gains _description_overrides dict, update_description(), clear_description_override(), get_description() methods
• GodspeedSettings includes evolution_enabled and evolution_model fields
• Trace analyzer uses streaming line-by-line reads instead of readlines() for low-memory devices

[2.1.0] - 2026-04-12

Added

• Extended thinking: Pass thinking parameter to Anthropic/Claude models with configurable token budget. /think [budget] slash command to toggle. Thinking blocks displayed in collapsed dim panel.
• Notebook (.ipynb) support: file_read detects .ipynb and renders cells as structured text with outputs. notebook_edit tool for cell-level operations (edit, add, delete, move).
• Image read tool: image_read reads PNG/JPG/GIF/WebP files, returns base64-encoded content for vision-capable LLMs. Size limits (20MB reject, 5MB warning). Path traversal protection.
• PDF read tool: pdf_read extracts text from PDF files with page range support (pages: "1-5"). 20-page-per-request limit. Graceful fallback when pymupdf not installed.
• Cost budget enforcement: Hard cost limit via max_cost_usd config. BudgetExceededError stops agent when exceeded. /budget [amount] slash command to set/show budget.
• GitHub PR/issue tool: github tool wraps gh CLI for create/list/view PRs and issues, add comments. 8 actions with structured JSON output. HIGH risk (modifies external state).
• Background command execution: shell tool gains background: true parameter. BackgroundRegistry singleton tracks processes. background_check tool for status/output/kill.
• Unified diff apply tool: diff_apply accepts unified diff format, applies hunks in reverse order. Multi-file support, fuzzy context matching (±3 lines), dry-run mode.
• Architect mode: /architect toggles two-phase plan-then-execute pipeline. Phase 1 uses read-only tools for planning, Phase 2 uses full tools guided by the plan.
• Speculative tool execution: During streaming, READ_ONLY tool calls are dispatched immediately before the full response completes. Results cached and consumed by the main loop — eliminates dispatch latency for reads.
• Risk-based parallel/serial dispatch: Phase 3 partitions tool calls by risk level. READ_ONLY tools run in parallel via asyncio.gather(), write tools run sequentially. Maintains ordering.
• Cheapest-model compaction: _compact_conversation() selects the cheapest model from the fallback chain for summarization. get_cheapest_model() in cost module.
• Typed event system: AgentEvent union type with 11 frozen dataclass event types (ThinkingEvent, TextChunkEvent, ToolCallEvent, etc.) for composable agent loop consumption.
• 250+ new tests (total: 1,200+ passing)

Changed

• agent_loop() signature expanded: on_thinking callback, speculative cache for streaming
• LLMClient tracks thinking_budget, max_cost_usd, total_cost_usd
• GodspeedSettings includes thinking_budget, max_cost_usd, architect_model, sandbox fields
• _streaming_call() accepts tool_registry and speculative_cache for speculative dispatch
• /help updated with new commands under "Agent Control" section

[2.0.0] - 2026-04-11

Added

• Parallel tool execution: When the LLM returns multiple tool calls, they execute concurrently via asyncio.gather(). 3-phase dispatch: parse → permission check (sequential) → execute (parallel). Config: parallel_tool_calls: true (default). Falls back to sequential for single calls or when disabled.
• Multimodal user messages: add_user_message() accepts str or list[dict] content blocks. build_image_content_block() and build_multimodal_message() helpers for image URLs and base64 data URIs. Token counter estimates 765 tokens per image block.
• @-mention context injection: @file:path, @folder:path, @web:url syntax in user input. Parsed via regex, resolved to content blocks injected into conversation. File/folder mentions validate paths via resolve_tool_path() (traversal protection). Web mentions enforce HTTPS-only with 100KB size limit.
• @-mention tab completion: Typing @ suggests mention types; @file: and @folder: complete file paths relative to project directory.
• Auto-commit workflow: After configurable number of successful edits, generates a conventional commit message via LLM and commits with Co-Authored-By: Godspeed <noreply@godspeed.dev> attribution. Config: auto_commit: false (default off), auto_commit_threshold: 5.
• /autocommit slash command: Toggle auto-commit on/off, set threshold. Listed in /help and tab-completable.
• Lint-fix retry loop: Auto-verify now retries ruff check --fix up to N times for Python/JS files until clean. Config: auto_fix_retries: 3 (default). Languages without deterministic fixers skip retry.
• MCP SSE/HTTP transport: MCPSSEClient connects to remote MCP servers via HTTP/SSE alongside existing stdio transport. Config: transport: "sse", url, headers fields in MCP server config. Backward compatible — missing transport field defaults to "stdio".
• Parallel tool TUI output: format_parallel_tool_calls() shows grouped header with tool count and names. format_parallel_results() shows batch summary with success/error markers. Callbacks: on_parallel_start, on_parallel_complete.
• Deferred auto-verify/auto-stash for parallel mode: Auto-verify runs sequentially after parallel batch completes. Auto-stash counts writes across batch.
• 100+ new tests across 7 test files (total: 960+ passing)

Changed

• agent_loop() signature expanded: parallel_tool_calls, skip_user_message, auto_fix_retries, auto_commit, auto_commit_threshold, on_parallel_start, on_parallel_complete parameters
• Conversation.add_user_message() accepts multimodal content blocks (list[dict])
• Token counter handles image_url type blocks with flat 765-token estimate
• GodspeedSettings includes new config fields for all v2.0 features

[1.0.0] - 2026-04-11

Added

• Token cost tracking: llm/cost.py — model-aware pricing table for 20+ models (Claude, GPT, Gemini, DeepSeek). /stats command shows token usage and estimated cost. Quit screen includes session cost. Ollama/local models always show "free".
• Enhanced diff previews: Permission prompts now show unified diff format with @@ hunk headers, line change stats (+5 -3 lines), and up to 30 context lines. File write prompts show line count and overwrite warning.
• Multi-file project instructions: Loads GODSPEED.md, AGENTS.md (Linux Foundation AAIF standard), CLAUDE.md, and .cursorrules. Priority: GODSPEED.md > AGENTS.md > CLAUDE.md > .cursorrules. Zero-friction migration from other agents.
• Prompt caching: System prompt marked with cache_control: ephemeral for Anthropic/OpenAI models. ~50% cost reduction on repeated prefixes via LiteLLM.
• Conversation export: /export [name] command writes session as formatted markdown to .godspeed/exports/. Includes system prompt, messages, tool calls, and results.
• Multi-language verify: Auto-verify now supports Python (ruff), JS/TS (biome/eslint), Go (go vet), Rust (cargo check), and C/C++ (clang-tidy). Linters detected dynamically. Shared _run_linter() helper.
• Test runner tool: test_runner tool auto-detects project framework (pytest, jest, vitest, go test, cargo test). Runs targeted or full test suites. Available to the agent for edit-test-fix loops.
• Headless/CI mode: godspeed run "task" --headless for non-interactive execution. --auto-approve levels (reads/all/none), --json-output for structured results, --max-iterations control. Exit code reflects success/failure.
• Web search tool: web_search — DuckDuckGo HTML search, no API key required. Returns titles, URLs, snippets. Agent can look up docs and error messages.
• Web fetch tool: web_fetch — HTTP GET with HTML-to-text extraction. Blocks local/private network access. 10K char limit. Agent can read documentation pages.
• 12 new built-in tools (total: 12 built-in + MCP + sub-agents)
• 92 new tests (total: 806+ passing), all new features tested
• /stats and /export slash commands with tab completion

Changed

• Auto-verify triggers on .js, .jsx, .ts, .tsx, .go, .rs, .c, .cpp, .h, .hpp (was Python-only)
• Quit screen shows estimated session cost for paid models
• Permission prompt diff uses difflib.unified_diff (was basic -/+ prefix)
• File write permission prompt shows line count and create/overwrite indicator

[0.9.0] - 2026-04-11

Added

• Skill framework: Markdown .md files with YAML frontmatter define reusable prompt skills. discover_skills() scans ~/.godspeed/skills/ and .godspeed/skills/, project overrides global. /{trigger} commands inject skill content into conversation. /skills lists available skills. Tab-completion for skill triggers.
• Auto-permission learning: ApprovalTracker counts repeated user approvals per pattern. After 3 approvals, suggests adding as permanent allow rule. append_allow_rule() persists to .godspeed/settings.yaml. Thread-safe, session-scoped.
• Hook system: HookDefinition pydantic model with 4 event types (pre_tool_call, post_tool_call, pre_session, post_session). HookExecutor runs shell commands with template variables ({tool_name}, {session_id}, {cwd}). Pre-tool hooks can block execution. Configurable timeout (1-300s). Wired into agent loop and CLI lifecycle.
• Task tracking: TaskStore (in-memory, sequential IDs) + TaskTool (create/update/list/complete). /tasks command shows themed Rich table. Registered as built-in LOW-risk tool.
• Codebase index: Optional ChromaDB-backed semantic search ([index] extra). AST-based chunking for Python, sliding window for other languages. CodeSearchTool for natural language code queries. Background indexing, /reindex command, stale detection.
• Architecture document: GODSPEED_ARCHITECTURE.md — 6-part reference covering core loop, security model, tool system, intelligence, autonomy, and memory/TUI. Mermaid diagrams. HTML comment delimiters for chunk loading.
• hooks field in GodspeedSettings for YAML hook configuration
• chromadb optional dependency under [index] extra
• 734 tests, ~90% coverage

[0.6.0] - 2026-04-11

Added

• Midnight Gold visual identity: tui/theme.py — single source of truth for all colors, styles, and branded strings. Electric gold primary, steel blue structure, mint green success, warm red errors, amber warnings, slate gray muted.
• Branded prompt: lightning bolt (⚡) icon with icon_prompt() supporting normal, plan, and paused states via prompt-toolkit HTML
• Thinking spinner: Rich Status spinner shown while waiting for LLM response; auto-clears on first output callback
• Semantic color constants: CTX_OK/WARN/CRITICAL, PERM_ALLOW/DENY/ASK/SESSION, TABLE_KEY/VALUE/BORDER/HEADER — no hardcoded Rich color strings remain in src/godspeed/
• Theme test suite: 24 tests covering all helpers (styled(), brand(), icon_prompt()), Rich rendering compatibility, and constant validation
• 626 tests, ~90% coverage

Changed

• All TUI modules (output.py, commands.py, app.py) and CLI (cli.py) import colors from tui/theme.py instead of hardcoding Rich markup
• Permission prompt input uses themed BOLD_WARNING style
• godspeed version command uses brand() helper for consistent rendering
• godspeed models and godspeed init commands use themed table and text styles

[0.5.0] - 2026-04-11

Added

• User memory with SQLite: UserMemory class backed by SQLite with WAL mode; persistent preferences (key/value CRUD) and corrections table; safe concurrent access; auto-creates ~/.godspeed/memory.db
• Session memory: SessionMemory records session lifecycle events (start, end, tool calls, errors) to SQLite; cross-session history with event filtering and limits
• Correction tracker: CorrectionTracker with heuristic detection of user corrections (negation patterns like "no", "don't", "stop", "instead"); auto-records to UserMemory; format_for_system_prompt() surfaces top-N corrections as "User prefers X over Y" guidance
• memory_enabled config: toggle memory system via settings.yaml
• 602 tests, ~90% coverage

[0.4.0] - 2026-04-11

Added

• Sub-agent coordinator: AgentCoordinator spawns isolated sub-agents with separate conversations; depth limit 3, iteration limit 25; SpawnAgentTool (HIGH risk) enables multi-agent orchestration via agent_loop() reuse
• spawn_parallel(): run multiple sub-agents concurrently via asyncio.gather()
• MCP client: MCPClient connects to MCP servers via stdio transport; MCPToolAdapter maps MCP tool definitions to Godspeed Tool ABC (all HIGH risk); graceful when mcp package not installed
• MCP server discovery: mcp_servers config in settings.yaml auto-discovers and registers remote tools at startup
• Model routing: ModelRouter routes LLM calls by task type (plan/edit/chat) to different models; configurable via routing in settings
• Human-in-the-loop pause/resume: asyncio.Event shared between TUI and agent loop; /pause stops at next iteration, /resume continues, /guidance <msg> injects mid-conversation correction and resumes
• Rich permission prompts: contextual detail in permission dialogs -- file_edit shows mini-diff, file_write shows first 10 lines, shell shows syntax-highlighted command, file_read shows path
• 549 tests, ~90% coverage

Changed

• LLMClient.chat() accepts optional task_type parameter for model routing
• agent_loop() accepts pause_event parameter for human-in-the-loop control
• format_permission_prompt() accepts optional arguments dict for contextual display

[0.3.0] - 2026-04-11

Added

• Tree-sitter repo map tool: RepoMapTool extracts symbol outlines (functions, classes, methods) from Python/JS/TS/Go files using tree-sitter; graceful degradation when tree-sitter not installed
• Plan mode: /plan command toggles read-only mode -- permission engine blocks all non-READ_ONLY tools; system prompt updated to instruct explore-only behavior
• Git stash/stash_pop actions: GitTool supports stash (with auto-message) and stash_pop operations
• Auto-stash before risky operations: agent loop tracks consecutive file edits; after 3+ consecutive writes, auto-stashes working state as a safety net
• Model-aware compaction: compaction prompts adapt to model context window size -- small models (<=32K) get aggressive summarization, frontier models (>100K) get detailed preservation
• MODEL_CONTEXT_WINDOWS mapping: prefix-matched context window sizes for Claude, GPT, Gemini, Ollama models with get_model_context_window() utility
• /checkpoint [name] command: save conversation state snapshots to .godspeed/checkpoints/; list checkpoints with metadata (tokens, messages, model, timestamp)
• /restore <name> command: restore a saved checkpoint, rebuilding full conversation state
• Checkpoint management: save_checkpoint(), load_checkpoint(), list_checkpoints(), delete_checkpoint() in context/checkpoint.py
• 485 tests, ~90% coverage

Changed

• Compaction in agent loop now uses model-aware prompts via get_compaction_prompt() instead of hardcoded prompt
• Moved tests/test_context.py into tests/test_context/ package for better organization

[0.2.0] - 2026-04-11

Added

• Stuck-loop detection: after 3 identical tool errors, injects a replan message forcing the model to try a different approach
• Verification cascade: VerifyTool runs ruff check on Python files; auto-verifies after every file_edit/file_write so the agent self-corrects lint errors
• /extend N command: override max iterations per agent turn (default: 50)
• /context command: show context window usage — tokens, percentage, message count with color-coded thresholds
• Audit trail compression: compress_session() rotates .jsonl → .jsonl.gz; verify_chain() transparently handles compressed logs
• FileEdit confidence reporting: output includes [match=exact confidence=1.00] or [match=fuzzy confidence=0.87 line=42] so the agent can gauge match quality
• 26 new dangerous command patterns: iptables, mount/umount, fdisk, shutdown/reboot, docker rm -f, kubectl delete, env exfiltration, Windows destructive ops, supply-chain attacks
• Ollama auto-start: detects when Ollama is not running and starts ollama serve as a background process before first LLM call
• Lazy LiteLLM import: deferred import reduces cold startup from ~1.5s to ~300ms
• Smart retry for connection errors: skips retry+sleep when Ollama/server is down, returns actionable error immediately
• Non-TTY crash guard: graceful error message when launched from non-interactive shells
• ToolRegistry.has_tool() method
• agent_loop() accepts max_iterations parameter
• 411 tests, 90% coverage

Fixed

• Route logs to stderr and scope verbose mode to godspeed.* namespace only — eliminates LiteLLM/httpx/markdown_it debug noise in TUI
• Add debug logging on tiktoken encoding fallback instead of silent pass
• Add docstrings to PermissionEvaluator and AuditRecorder protocol methods

Changed

• godspeed init command — creates ~/.godspeed/ and default settings.yaml
• godspeed models command — shows popular model options with provider, cost, and API key info
• settings.yaml.example — full reference config with free/paid model examples and permission rules
• Audit trail retention cleanup — expired sessions are purged on startup based on retention_days setting; handles both .jsonl and .jsonl.gz
• Token counter model mappings for Claude, Gemini, DeepSeek, Ollama models
• Default model changed from paid claude-sonnet-4-20250514 to free ollama/qwen3:4b — zero-cost out of the box
• Expanded pyproject.toml classifiers and keywords for PyPI discoverability

[0.1.0] - 2026-04-10

Added

• Hand-rolled agent loop — model decides when to stop, no framework dependency
• 4-tier permission engine (deny > ask > allow) with deny-first evaluation, pattern matching, and 46 dangerous command patterns
• Hash-chained JSONL audit trail (SHA-256) with tamper detection and godspeed audit verify
• 4-layer secret protection: file deny rules, context cleaning, output filtering, audit redaction (27 regex patterns + Shannon entropy)
• LiteLLM integration for 200+ LLM providers (Claude, GPT, Gemini, Ollama, etc.) with fallback chains
• 7 built-in tools: file_read, file_write, file_edit (with fuzzy matching), shell, glob_search, grep_search, git
• Rich + prompt-toolkit TUI with syntax highlighting, diff rendering, and streaming
• Slash commands: /help, /model, /undo, /audit, /compact, /clear, /quit
• GODSPEED.md project instructions (walk-up-tree loading, like CLAUDE.md)
• Conversation compaction when approaching context limit
• Configuration cascade: global (~/.godspeed/settings.yaml) > project (.godspeed/settings.yaml) > CLI flags (deny rules are additive — project can't weaken global denies)
• CLI entry points: godspeed (TUI), godspeed version, godspeed audit verify
• 243 tests, 82% coverage
• Full repo-standards scaffolding: CI pipeline, pre-commit hooks, dependabot, issue/PR templates, CONTRIBUTING.md, SECURITY.md, LICENSE (MIT)