Changelog

May 10, 2026 · View on GitHub

All notable changes to this project.

The format follows Keep a Changelog, and the project adheres to Semantic Versioning.

[0.18.0] — 2026-05-10

Minor release. Adds first-class integration with the external Brain CLI/MCP memory system while keeping agentic-stack's project-local .agent/ runtime independent and optional.

Added

  • Brain integration. Adds agentic-stack brain ... and .agent/tools/brain_bridge.py so projects can use the external codejunkie99/brain CLI/MCP server as a git-backed long-term memory layer without vendoring the Rust workspace into agentic-stack.
  • Brain seed skill. Adds a brain skill that teaches host agents when to query or write Brain memory and how to avoid storing secrets.

Release

[0.17.0] — 2026-05-10

Minor release. Clears the open PR queue and ships new harness adapters, the Mission Control beta, semantic lesson retraction, and the consolidated pytest layout.

Added

  • GitHub Copilot CLI adapter. Installs AGENTS.md, .github/instructions/agentic-stack.instructions.md, .github/hooks/agentic-stack.json, and a .github/skills/ mirror so Copilot CLI can load the portable brain and run memory hooks.
  • Google Gemini CLI adapter. Installs gemini.md plus a .gemini/skills/ mirror for Gemini CLI projects.
  • agentic-stack mission-control. Adds a beta local web dashboard with a snapshot mode, local-only event store, static assets, collectors, renderer, and tests.
  • Semantic lesson retraction. Adds .agent/tools/retract_lesson.py so an accepted lesson can be marked status=retracted with a required rationale while preserving append-only lessons.jsonl history.

Changed

  • README, getting-started docs, POSIX installer help, and PowerShell installer help now list all 12 adapters and the Mission Control command.
  • Proactive recall now considers only the latest append-only state for each lesson id, excluding retracted lessons from future recall results.
  • LESSONS.md rendering deduplicates append-only lesson state by id so the rendered markdown reflects the current status without losing audit history.
  • Tests now live under tests/ with pytest.ini, including coverage for Copilot CLI, Gemini, Mission Control, upgrades, onboarding, and lesson retraction.

Release

[0.16.1] — 2026-05-10

Patch release. Ships the production-ready getting-started guide from PR #49 and fixes stale version text in the onboarding banner.

Changed

  • docs/getting-started.md now leads with the Homebrew install path while documenting source-checkout and PowerShell flows separately.
  • The guide now explains which commands are available through the global agentic-stack wrapper and which source-checkout users should run through ./install.sh.
  • The guide now documents current project-management verbs, including dashboard, status, doctor, upgrade, sync-manifest, add, remove, and manage.

Fixed

  • The onboarding banner now derives its displayed version from harness_manager.__version__ instead of showing stale v0.8.0 text.

Release

[0.16.0] — 2026-05-09

Minor release. Adds a safe installed-project upgrade path and repairs skill manifest drift so newly copied skills can trigger reliably after version upgrades.

Added

  • agentic-stack upgrade. Adds a safe project migration verb that refreshes skeleton-owned .agent infrastructure, copies new skills, supports --dry-run/--yes, and leaves adapter configs plus user memory untouched.
  • agentic-stack sync-manifest. Rebuilds .agent/skills/_manifest.jsonl from installed SKILL.md frontmatter so copied skills can trigger correctly.

Fixed

  • Install/add now re-sync the skill manifest when a project already has .agent/skills, preventing _index.md / _manifest.jsonl drift.
  • doctor now warns when Claude Code hook commands reference missing .agent Python files or when hook scripts are present but not wired in .claude/settings.json.

Migration

After upgrading the CLI, run agentic-stack upgrade --dry-run in installed projects to preview safe .agent infrastructure updates, then agentic-stack upgrade --yes to apply them. Run agentic-stack sync-manifest if a project's skill manifest needs repair without copying files.

Release

[0.15.0] — 2026-05-06

Minor release. Adds a production dashboard TUI for installed agentic-stack projects, with the trust-console inspection surface folded into the same user-facing entrypoint.

Added

  • agentic-stack dashboard / dash. Adds a terminal dashboard for project health, installed adapters, doctor checks, harness verification, memory, team brain, skills, managed instances, transfer, and local data exports.
  • Trust-console parity. The dashboard includes a per-harness verify matrix, team brain status/init, skills listing, active instance listing, accepted/rejected memory review, and memory_why() evidence lookup.
  • Plain renderer. agentic-stack dashboard --plain and non-TTY fallback produce a script-safe text dashboard for logs, agents, and tests.
  • Interactive dashboard coverage. Adds local tests for renderer output, CLI aliases, trust-console parity sections, non-TTY fallback, up/down navigation, refresh, quit, and Enter-open behavior.

Changed

  • Bare interactive agentic-stack / ./install.sh opens the dashboard when an existing .agent/install.json is present. Non-TTY shells keep printing command guidance instead of launching an interactive UI.
  • README, POSIX installer help, and PowerShell installer help now document the dashboard entrypoint.

Fixed

  • test_claude_code_hook.py now works both as a standalone validation script and as a pytest-collected module by providing a real mod fixture.

Migration

No migration required. Existing projects can run agentic-stack dashboard after upgrading. Re-run an adapter install only if you want the latest copied adapter guidance files in a project.

Release

[0.13.0] — 2026-05-02

Minor release. Adds an onboarding-style transfer wizard for moving a portable .agent brain into Codex, Cursor, Windsurf, or terminal-only projects with a generated curl/PowerShell import command.

Added

  • agentic-stack transfer wizard. Adds an onboarding-style TUI that parses natural-language requests such as move my memory into Codex, previews the target adapter files, asks for confirmation, and either generates an import command, applies locally, or both.
  • Portable transfer bundles. Adds canonical JSON + gzip + base64url bundles with SHA-256 verification. The importer merges preferences and accepted lessons idempotently, restores selected memory files, copies skills, records import metadata, and installs selected adapters through the existing harness manager.
  • Full memory intent. move my memory now means preferences, accepted lessons, skills, working memory, episodic/history logs, and candidate lessons. Data-layer exports, flywheel traces, runtime indexes, and caches stay out unless future scopes explicitly add them.
  • Curl and PowerShell bootstraps. Adds scripts/import-transfer.sh and scripts/import-transfer.ps1 so another terminal can import a transfer bundle without manually cloning the repo first.

Changed

  • Windsurf installs a modern .windsurf/rules/agentic-stack.md workspace rule and still writes legacy .windsurfrules for older Windsurf builds.
  • agentic-stack transfer export and agentic-stack transfer import provide non-interactive surfaces for scripts and CI-style handoff flows.

Fixed

  • Transfer export blocks secret-like content, including private keys and common API token patterns, before payload generation.
  • Fresh Codex imports now copy the full .agent brain before installing the Codex AGENTS.md and .agents/skills adapter wiring.

Migration

No migration required. Existing installs keep working. Run agentic-stack transfer from a project that already has .agent/ to create or import a transfer bundle.

Release

[0.12.0] — 2026-04-27

Minor release. Adds the opt-in tldraw seed skill for live canvas diagrams and a skill-local snapshot store. The feature stays beta and off by default.

Added

  • tldraw seed skill — live canvas diagrams. Adds .agent/skills/tldraw/SKILL.md with MCP tool guidance, shape constraints, and a self-rewrite hook for diagram, sketch, wireframe, flowchart, whiteboard, and architecture visualization prompts.
  • Skill-local snapshot store. Adds .agent/skills/tldraw/store.py with snapshot, list, load, and archive CLI/API support. Runtime output is local and gitignored under .agent/skills/tldraw/: snapshots.jsonl, snapshots/, and INDEX.md.
  • Opt-in feature flag. Onboarding now writes a tldraw beta feature flag, default off. The skill loader skips flagged skills unless .agent/memory/.features.json explicitly enables them.
  • Manual MCP config source. Adds adapters/_shared/tldraw-mcp.json as the canonical tldraw MCP config block users can merge into Claude Code, Cursor, or Antigravity after enabling the feature.

Changed

  • Seed skill count is now nine: skillforge, memory-manager, git-proxy, debug-investigator, deploy-checklist, design-md, data-layer, data-flywheel, and tldraw.
  • tldraw persistence is intentionally skill-local storage, not a fifth memory layer: it has no dream-cycle, clustering, recall, or semantic-memory lifecycle.
  • Default adapter installs no longer wire beta tldraw MCP config automatically.

Fixed

  • Updated the tldraw validation suite to target .agent/skills/tldraw/store.py instead of the removed .agent/memory/visual/visual_memory.py.
  • store.py renders INDEX.md while holding the JSONL lock so concurrent snapshots cannot leave the index stale.
  • Added coverage for feature-gated skill loading, no default MCP install, path traversal rejection, malformed JSONL recovery, same-second snapshot ids, and concurrent snapshots.

Migration

No migration required. Existing installs keep tldraw disabled until users opt in through onboarding reconfiguration or .agent/memory/.features.json.

Credits

  • PR #11 by @Siddharth11Roy, with review fixes and release prep by Codex.

Release

[0.11.2] — 2026-04-26

Patch release. Makes the data-layer dashboard easier to access from coding tools by turning the injected data-layer skill into the natural-language dashboard surface.

Added

  • Injected dashboard behavior. The data-layer skill now triggers on plain phrases such as "show me the dashboard", "what did my agents do", "agent status", and "usage report". When the model decides the user wants local agent activity, the skill tells it to render the terminal dashboard directly instead of making the user remember flags.
  • Natural-language exporter requests. The existing exporter now accepts requests such as python3 .agent/tools/data_layer_export.py show me last 7 days by hour. It maps common phrases to --window and --bucket; explicit flags still override the natural-language text for scripts and automation.

Changed

  • The terminal dashboard now uses the same rail/marker visual language as the onboarding flow while keeping dashboard.tui.txt plain text with no ANSI escape codes.
  • dashboard-summary.json records the natural-language request that produced the export when one was provided.

Migration

No migration required. Existing flag-based commands still work.

Release

[0.11.1] — 2026-04-26

Patch release. Makes the data-layer dashboard visible directly in coding-tool terminals and adds a visual SVG explainer for the data-layer flow.

Added

  • Terminal dashboard by default. The existing python3 .agent/tools/data_layer_export.py --window 30d --bucket day command now prints a compact TUI-style dashboard after writing exports. It shows resource numbers, latest bucket activity, top harnesses, top workflows, top categories, and artifact paths without requiring a browser.
  • dashboard.tui.txt. The same terminal dashboard is saved next to dashboard.html, CSV/JSON exports, and daily-report.md for agents and users who want to inspect or attach a plain-text report.
  • docs/data-layer.svg. README and docs/data-layer.md now include a visual of the local data-layer flow: input streams, exporter, browser dashboard, terminal dashboard, CSV/JSON, and approved handoff.

Changed

  • dashboard-report.json now advertises the terminal dashboard artifact.
  • Data-layer docs no longer require a separate command to see the terminal view; the normal export command prints it.

Migration

No migration required. Existing data-layer commands still work; they now print the terminal dashboard in addition to the previous status lines.

Release

[0.11.0] — 2026-04-26

Minor release. Adds two local-first data capabilities: a cross-harness monitoring/data layer and an approved-run data flywheel. Both stay private by default, write regenerated runtime artifacts under ignored .agent/ subdirectories, and avoid remote telemetry or model training.

Added

  • data-layer seed skill — local cross-harness monitoring. Adds .agent/skills/data-layer/SKILL.md and .agent/tools/data_layer_export.py to export dashboard-ready local data across Claude Code, Hermes, OpenClaw, Codex, Cursor, OpenCode, and custom loops sharing the same .agent/ brain. Outputs include agent events, cron timelines, KPI summaries, token/cost estimates, category breakdowns, dashboard.html, and daily-report.md. Runtime exports live under .agent/data-layer/, which is gitignored. Thanks to @danielfoch for PR #25.
  • data-flywheel seed skill — approved runs to reusable artifacts. Adds .agent/skills/data-flywheel/SKILL.md and .agent/tools/data_flywheel_export.py to turn human-approved, redacted runs into trace records, context cards, eval cases, training-ready JSONL, and flywheel metrics. Runtime exports live under .agent/flywheel/, which is gitignored. The tool is local-only, model-agnostic, and does not train models or call external APIs. Thanks to @danielfoch for PR #26.
  • Schemas, sanitized examples, tests, README sections, and architecture docs for both local data features.

Changed

  • Seed skill count is now eight: skillforge, memory-manager, git-proxy, debug-investigator, deploy-checklist, design-md, data-layer, and data-flywheel.
  • docs/architecture.md now describes five local modules: memory, skills, protocols, data layer, and data flywheel.

Migration

brew upgrade agentic-stack is enough. There are no on-disk schema changes. The new runtime directories, .agent/data-layer/ and .agent/flywheel/, are private and regenerated; both are gitignored.

Release

Credits

  • PR #25 and PR #26 by @danielfoch.
  • Merge conflict resolution, verification, and release prep by Codex.

[0.10.0] — 2026-04-26

Minor release. Adds the design-md seed skill (sixth seed skill in the brain), and fixes a hard crash on macOS-default Python 3.9 that hit every brew user on first run.

Added

  • design-md seed skill — DESIGN.md / Google Stitch support. A new portable skill that points coding agents at a root DESIGN.md (Google Stitch format) as the visual-system source of truth. Loads only when DESIGN.md exists at the project root; default behavior is read-only on the contract file, and validation prefers npx @google/design.md lint DESIGN.md over hand-checks. Brings the seed skill count from five to six. Thanks to @danielfoch for the contribution (PR #21). Decision recorded in .agent/memory/semantic/DECISIONS.md.

Fixed

  • Python 3.9 crash on first run (#27). Every brew user on macOS-default Python 3.9 hit TypeError: unsupported operand type(s) for |: 'type' and 'type' immediately after brew install agentic-stack because harness_manager/ used PEP 604 union syntax (Path | str) that requires Python 3.10+ at runtime. Added from __future__ import annotations (PEP 563) to the eight affected files so all annotations are stored as strings and never evaluated at import time. Works on Python 3.7+, which covers every macOS-shipped Python in the wild. Zero-cost vs. the reporter's suggested python@3.10 brew dep (~150 MB pull). Thanks to @WhoLsJohnGalt for the precise repro and the suggested workaround.

Migration

brew upgrade agentic-stack is enough — no on-disk schema changes. Users on Python 3.9 who hit #27 on v0.9.x can upgrade and the crash goes away. Users with existing .pi/extensions/memory-hook.ts from v0.9.0 already got the fixes via ./install.sh pi after the v0.9.1 upgrade; this release adds the design-md skill on top.

Release

  • Tag v0.10.0 cut from master.
  • GitHub release: https://github.com/codejunkie99/agentic-stack/releases/tag/v0.10.0
  • Formula/agentic-stack.rb bumped to v0.10.0 in a follow-up commit (same flow as v0.8.0 → v0.9.0 → v0.9.1): tag first, compute sha256, then bump url + sha256 + version together so brew always points at a real installable artifact.

Credits

  • PR #21 by @danielfoch (design-md skill), with cross-model review fixes applied as a follow-up commit.
  • Issue #27 by @WhoLsJohnGalt — clean repro of the macOS Python 3.9 crash including a working sed-based workaround.

[0.9.1] — 2026-04-26

Patch release that closes the gap between v0.9.0 and a working pi adapter. Every brew user on v0.9.0 hit the first bug; the rest are quieter but make the dream cycle and the cross-harness episodic log actually correct.

Fixed

  • agentic-stack pi crashed for every brew user with ModuleNotFoundError: No module named 'harness_manager'. The v0.9.0 Formula didn't include harness_manager/ in pkgshare.install, so install.sh couldn't dispatch. Adds it back.
  • Pi dream cycle never fired. The session_shutdown handler filtered on event.reason, but Pi's SessionShutdownEvent carries no reason field — verified against pi-coding-agent types.d.ts and the emit site at agent-session.js:1638. The filter rejected every event, so auto_dream.py never ran. Filter dropped; re-entrancy guard added.
  • Pi edit reflections lost the diff. Hook accessed event.input.edits[0] but Pi's EditToolInput is flat { path, oldText, newText } (no edits array — that's MultiEdit on Claude Code). Reflections silently degraded to Edited <path> with no old/new content. Read the flat fields.
  • decay.py crashed comparing aware-UTC entries to naive cutoff. The new session_shutdown hook surfaced this on every clean pi exit. Cutoff is now datetime.now(timezone.utc); entry timestamps are normalised to UTC before comparison.
  • Naive-local Python timestamps drifted against the UTC decay window. Decay's "naive == UTC" assumption was correct only if writers emitted UTC. They didn't: post_execution, on_failure, learn, graduate, promote, review_state, render_lessons all wrote naive-local. Every writer now emits aware UTC; every reader (salience, show._human_age / _daily_counts / failing_skills / last_dream_cycle, on_failure._count_recent_failures, review_state._age_factor, archive) normalises naive timestamps to UTC before comparison.
  • One bad regex in hook_patterns.json disabled every user pattern in the Pi hook. Pre-fix used a single combined RegExp per list, caught any error, returned null for both. Now per-fragment validation with incremental merge — same posture as claude_code_post_tool.py's _filter_valid / _build_with_fallback.
  • auto_dream lost entries that landed mid-cycle. Original _write_entries had a truncate-before-lock window — open(path, "w") truncates BEFORE a lock can be taken. The deeper bug: even after fixing the inner race, any append_jsonl() between _load_entries() and _write_entries(kept) would be truncated away by the rewrite. The cycle now holds a single exclusive flock on AGENT_LEARNINGS.jsonl across the entire read-modify-write window via _episodic_locked(). Mutually exclusive with _episodic_io.append_jsonl (same flock target). POSIX only; Windows falls back to historical best-effort behaviour.
  • _cachedSha went stale after git commit inside a pi session. TS hook cached the SHA per-process for performance, so every entry logged after a mid-session commit recorded the pre-commit SHA. Cache, but invalidate on bash commands matching git <subcmd> for HEAD-moving subcommands. Allows option flags between git and the subcommand (git -c key=val checkout main, git -C path switch dev).
  • salience over-scored future-skewed legacy rows. Legacy naive-local timestamps re-interpreted as UTC could read as a few hours in the future during the migration window. timedelta.days then went negative and recency exceeded the intended cap. Floor age at 0; clamp recency to ≤ 10.
  • decay archive filename used local date while cutoff was UTC. archive_{date}.jsonl now uses UTC date so a tz-jumping user gets a deterministic path.

Changed

  • adapters/pi/adapter.json no longer manages .agent/harness/hooks/pi_post_tool.py via from_stack. The TS hook is self-contained — all scoring + reflection inline, no Python subprocess per tool call. The .py still ships in the brain template under .agent/harness/hooks/ for standalone use.
  • tests/ is now untracked and listed in .gitignore. CI never ran these (.github/workflows/ci.yml runs test_claude_code_hook.py + verify_*.py at repo root). Pull from an older tag if you want the unittest suite locally.

Migration

brew upgrade agentic-stack is enough — there are no on-disk schema changes. Existing pi installs with the v0.9.0 hook get the new logic on the next ./install.sh pi. Existing .agent/memory/episodic/AGENT_LEARNINGS.jsonl files with naive-local timestamps continue to work — readers normalise them at compare time and writers emit UTC going forward.

Release

  • Tag v0.9.1 cut from master at d7b70b2.
  • GitHub release: https://github.com/codejunkie99/agentic-stack/releases/tag/v0.9.1
  • Formula/agentic-stack.rb bumped to v0.9.1 in a follow-up commit (same flow as v0.8.0 → v0.9.0): tag first, compute sha256, then bump url + sha256 + version together so brew always points at a real installable artifact.
  • Tarball sha256: 09cc3b8c9ec159cf8b85cf672fc15a29e6bfa9377cc9c59dd270acd441ced568 (verify locally with curl -sL https://github.com/codejunkie99/agentic-stack/archive/refs/tags/v0.9.1.tar.gz | shasum -a 256)

Credits

PR #24 by @aliirz; Codex CLI used for an independent second-opinion review that surfaced the auto_dream window race + the SHA-regex narrowness.

[0.9.0] — 2026-04-23

Added

  • Harness manager: manifest-driven adapter system. Each adapter now ships an adapters/<name>/adapter.json declaring its files, collision policy, optional skills directory mirror, and named post-install actions. Adding a new adapter is now a JSON-only PR — no Python code, no test wiring, no class registration. Lives in the new harness_manager/ Python package.
  • ./install.sh add <adapter> — append an adapter to an existing project without re-running the onboarding wizard.
  • ./install.sh remove <adapter> — confirmation prompt lists every file before deletion. Hard delete (no quarantine, no undo — git is the safety net). Reverses post-install actions automatically (e.g., openclaw agents remove).
  • ./install.sh doctor — read-only audit of installed adapters. Verifies tracked files exist, post-install state is valid, .agent/ brain is intact. Exits 0 on green, 1 on red. First run on a pre-v0.9.0 project asks before synthesizing install.json — never silently mutates.
  • ./install.sh status — one-screen view of installed adapters, brain stats (skills/episodic/lessons), last-updated timestamp.
  • .agent/install.json — authoritative record of what's installed. Schema-versioned. Atomic write via tempfile + rename, fcntl-locked on POSIX.
  • PowerShell parity from day one. install.ps1 is now a 70-line thin dispatcher to the same Python backend install.sh uses. The new add/remove/doctor/status verbs behave identically across mac/Linux/Windows. Was 270+ lines of duplicated bash-shaped logic.
  • docs/per-harness/standalone-python.md — gap-fill for the only harness that didn't have a per-harness doc.

Fixed

  • #18 — Claude Code hook commands break when cwd is not the project root. adapters/claude-code/settings.json template now uses {{BRAIN_ROOT}} placeholder, which the manifest backend substitutes with $CLAUDE_PROJECT_DIR at install time. Hook commands resolve correctly regardless of which directory Claude Code's cwd points at. Thanks to @palamp for the report and the proposal that shaped the larger feature.

Security

  • Manifest path-safety hardening (harness_manager/schema.py). The pre-existing path-traversal guard only tokenized on / and only treated /-prefixed paths as absolute, so Windows-style inputs (..\..\outside, \\server\share, C:\temp\x, C:foo) bypassed validation and could let install/remove read or write outside the adapter/project roots when run on Windows. Also extended the same validation to skills_link.target and skills_link.dst, which were previously only checked for presence — a manifest could otherwise point the symlink/rsync into arbitrary filesystem locations on any platform. Both POSIX and Windows separators are now normalized before traversal detection, and every common absolute-path form (POSIX root, Windows root, UNC, drive-letter) is rejected.

Changed

  • install.sh shrinks from 175 lines of bash case-statements to 35 lines of dispatcher. All install logic moved to harness_manager/. Existing CLI surface preserved: ./install.sh <adapter> [target] [--yes|--reconfigure|--force] works identically.
  • install.ps1 shrinks from 270+ lines to 70.

Migration

Existing v0.8.x users: brew upgrade agentic-stack, then run ./install.sh doctor in your project. Doctor detects existing adapters from filesystem signals and asks before writing install.json. Subsequent doctor runs are read-only.

Release checklist (post-merge, pre-brew upgrade)

The Homebrew Formula (Formula/agentic-stack.rb) intentionally still points at the v0.8.0 tarball in this PR. The v0.9.0 release flow is:

  1. Merge this PR to master.
  2. Tag v0.9.0 on master and create the GitHub release.
  3. Run curl -L https://github.com/codejunkie99/agentic-stack/archive/refs/tags/v0.9.0.tar.gz | shasum -a 256 to compute the new sha256.
  4. Open a follow-up PR that updates url, sha256, version together, and adds harness_manager + install.ps1 to the pkgshare.install line. This is the same pattern as commit abaa352 (the v0.8.0 sha256 bump).

Reason for the split: a Formula change that adds harness_manager/ to pkgshare.install while still pointing at the v0.8.0 tarball would fail brew install (file-not-found in the staged tarball). Bumping all four fields together as a follow-up after the tag exists keeps the formula always pointing at a real, installable artifact.

0.8.0 — 2026-04-21

Added

  • Google Antigravity adapter. ./install.sh antigravity drops an ANTIGRAVITY.md into the project root so Antigravity agents pick up the portable brain in .agent/. Matches the pattern of the other root-instruction harnesses. Brings the supported-harness count to 9. Thanks to @smartsastram for the contribution (PR #9).
  • Rich PostToolUse episodic logging for Claude Code. New .agent/harness/hooks/claude_code_post_tool.py reads the JSON payload Claude Code sends via stdin and derives a real action label, importance score, and non-empty reflection per tool call. Replaces the old hardcoded post-tool ok that produced identical entries every session and left the dream cycle with nothing to cluster on. Ships with a 54-test validation suite (test_claude_code_hook.py). Thanks to @aliirz for the contribution (PR #8).
  • User-owned stack tuning via hook_patterns.json. Drop your own high-stakes and medium-stakes command patterns in .agent/protocols/hook_patterns.json so the hook scores vercel deploy, supabase migrate, etc. correctly for your stack. Ships with empty arrays and a _examples section; universal patterns stay hardcoded.
  • on_failure() severity overrides. New importance= and pain_score= parameters so a failed production deploy records its real severity instead of the flat importance=7 / pain_score=8 defaults. Lets the dream-cycle salience formula actually distinguish a failed migration from a failed ls.
  • Bash wrapper-aware failure detection. _is_success() now detects explicit exit-masking wrappers (|| true, || :, || exit 0, ; true) and falls through to stderr-based signal when they are present, so masked production failures are still captured. Quoted strings and set +e are excluded from masking detection to avoid false positives on patterns like echo '... || true ...' and set +e; grep X log; set -e.
  • 33-check regression verifier. verify_codex_fixes.py validates every classification path after 7 rounds of codex review. Named verify_*.py (not test_*.py) to avoid pytest collection side effects. Uses a TMPDIR / $HOME / repo-local / VERIFY_TMPDIR fallback chain so it runs in constrained sandboxes.

Fixed

  • Bash exit_code=0 no longer second-guessed via stdout. Commands like grep Error /var/log/app.log and cat failures.log used to be recorded as failures because the output contained error-looking strings. Exit code is now authoritative for Bash responses.
  • User regex fragments in hook_patterns.json can't crash the hook. Each fragment is validated standalone via re.compile; invalid ones are dropped with a stderr warning. Merged-compile failures (e.g., a fragment like (?i)foo that validates alone but breaks once embedded) fall back to an incremental build that drops only the offending fragments, preserving universals and good user fragments.
  • on_failure reflection no longer prefixes str: for string errors. Only Exception objects get a type-name prefix now.

Changed

  • Wizard version bumped to 0.8.0 in onboard_render.py.
  • Wizard outro now points users at .agent/protocols/hook_patterns.json so they know they can extend the importance scorer with their stack's service names.

0.7.2 — 2026-04-20

Changed

  • README repositioning. Leads with the actual buyer pain — switching coding-agent tools keeps resetting how your agent behaves — so the adapter list, wizard, and memory architecture read as proof instead of preamble. Follow / coded-using / article framing moved into the Credits section.

0.7.1 — 2026-04-20

Changed

  • Relicensed from MIT to Apache 2.0.

0.7.0 — 2026-04-20

Added

  • learn.py host-agent tool. Teach the agent a rule in one command: python3 .agent/tools/learn.py "Always serialize timestamps in UTC" --rationale "past cross-region bugs". Stages, graduates, and renders in one step. Idempotent. Cleans up staged files on heuristic reject; preserves on crashes so retries work.
  • recall.py host-agent tool. Surfaces graduated lessons relevant to what you're about to do. Ranked lexical-overlap hits with per-entry source labels. Merges lessons.jsonl and seed bullets in LESSONS.md so graduating your first lesson doesn't hide the seeds. Logs every recall to episodic memory for audit.
  • show.py host-agent tool. Colorful dashboard of brain state (episodes, candidates, lessons, failing skills, 14d activity sparkline). --json / --plain / NO_COLOR flags.
  • Adapter wiring for recall across all 8 harnesses. Every adapter (claude-code, cursor, windsurf, opencode, openclaw, hermes, pi, standalone-python) now instructs the model to run recall.py "<intent>" before deploy / migration / timestamp / debug / refactor work, and to surface results in a Consulted lessons before acting: block.
  • Pre-graduated seed UTC lesson so new installs see proactive recall return a real hit on first try. Stored at .agent/memory/semantic/lessons.jsonl.

Fixed

  • Canonical pattern_id. Conditions are casefolded, unicode whitespace collapsed, zero-widths stripped, deduped, sorted — the same logical set always yields the same id.
  • Stricter heuristic check. validate.heuristic_check now requires ≥3 content words in a claim (blocks junk like !!!abc that passed the raw-length gate).
  • Idempotent graduate.py retries. Re-renders LESSONS.md, honors original reviewer / rationale from lessons.jsonl to keep stores in sync, refuses retries against legacy rows missing metadata.
  • Advisory flock on lessons.jsonl. render_lessons and append_lesson now hold an exclusive flock during writes. Concurrent writers serialize; LESSONS.md can no longer go stale relative to lessons.jsonl. Atomic rewrite via temp file + rename.

0.6.0 — 2026-04-17

Added

  • Pi Coding Agent adapter. ./install.sh pi drops AGENTS.md and symlinks .pi/skills to .agent/skills so pi sees the full brain with zero duplication. Safe to install alongside hermes / opencode (all read AGENTS.md; the installer skips the overwrite if one exists).

Changed

  • BREAKING: openclient adapter renamed to openclaw. Installed file changed: .openclient-system.md → .openclaw-system.md. Existing OpenClient users: re-run ./install.sh openclaw.

0.5.0 — 2026-04-17

Added

  • Host-agent review protocol. Python handles filing (cluster, stage, heuristic prefilter, decay). The host agent handles reasoning via list_candidates.py / graduate.py / reject.py / reopen.py. Graduation requires --rationale so rubber-stamping is structurally impossible. Zero unattended reasoning, zero provider coupling.
  • Structured lessons.jsonl as source of truth. LESSONS.md is rendered from it. Hand-curated content above the sentinel is preserved across renders; legacy bullets auto-migrate on first run.
  • Proper single-linkage clustering with bridge-merge. Pattern IDs derived from canonical claim + conditions, stable across cluster- membership changes, distinct for generic-canonical collisions.
  • Query-aware retrieval. context_budget ranks episodes by salience × relevance and filters lessons to status=accepted only — provisional, legacy, and superseded entries never leak into the system prompt.
  • [BETA] FTS5 memory search (.agent/memory/memory_search.py). Opt-in via onboarding or .agent/memory/.features.json. Default off. Prefers ripgrep when FTS5 is not available, falls back to grep. Restricted to .md / .jsonl so source files never pollute results.
  • Windows-native installer. install.ps1 runs natively under PowerShell; install.sh continues to work under Git Bash / WSL.

Fixed

  • Batch-sound graduation gate.
  • Stable slugs across cluster drift.
  • Provisional re-review and supersession semantics.
  • REVIEW_QUEUE refreshes on every CLI action.
  • Heuristic-rejection stamping so unrelated LESSONS.md edits do not churn.
  • Atomic graduate.py (semantic write first, candidate move last).
  • .gitignore ordering so .agent/memory/.index/ is actually ignored.
  • Fallback search restricted to .md / .jsonl.
  • Feature toggle file and [BETA] label in onboarding.

0.4.0 — 2026-04-16

Added

  • Interactive onboarding wizard (onboard.py, clack-style UI) that auto-fills PREFERENCES.md after install. Flags: --yes (CI / defaults), --reconfigure (re-run). Apple-style redesign with 7 scenes (Memory, Skills, Dream Cycle added).

0.3.0 — 2026-04-16

Fixed

  • Cron-safe paths in auto_dream.py and Stop hook matcher.
  • Deny-glob syntax in settings.json.

0.2.0 — 2026-04-16

Added

  • Homebrew formula at Formula/agentic-stack.rb.

Fixed

  • standalone-python path detection.
  • Harness count in README.
  • brew tap URL in README.

0.1.0 — 2026-04-16

Added

  • Initial release. Portable .agent/ brain folder with adapters for Claude Code, Cursor, Windsurf, OpenCode, OpenClient (later OpenClaw), Hermes, and standalone Python. Homebrew installer (replaces the earlier npx-based flow).