Skip to content

v3.8.1

Choose a tag to compare

View all tags
@fabioc-aloha fabioc-aloha released this 07 Jul 15:24
v3.8.1
b202442

Changelog

All notable changes to Alex ACT Edition.

[Unreleased]

[3.8.1] - 2026-07-07

Patch [behaviour] — VS Code 1.126–1.128 platform-awareness refresh (docs-only tool-awareness + terminal-safety table extensions, welcome-baseline $comment refresh) plus a forward-compatible heir_owned manifest field for next-gen Extension installs.

Two coordinated changes queued since v3.8.0, released together as a docs/manifest patch:

  1. VS Code 1.126–1.128 platform-awareness (Tier A of Alex_ACT_Supervisor/docs/proposals/vscode-1.126-1.128-adoption-2026-07-07.md). Extends the always-on tool-awareness table with 9 rows spanning three release cycles (Edit-mode deprecation, macOS/Linux terminal sandboxing, browser tools GA, BYOK Claude via own Anthropic key, and more), adds a 1.127 macOS/Linux terminal sandboxed by default row to terminal-command-safety, and refreshes welcome-baseline $comment to acknowledge 1.126 Edit-mode deprecation reinforcing the chat.agent.enabled: true pin.
  2. heir_owned manifest field (from v3.8.0 post-release commit cd32ece). Additive contract addition so next-gen static-fetch Extension installs can read the HEIR_OWNED policy as manifest data instead of executing fetched Edition code.

Both are additive; heirs that ignore the new tool-awareness rows still have correct rules, and existing Extensions still fall back to their legacy HEIR_OWNED parser.

Added

  • .github/instructions/tool-awareness.instructions.md — H2 renamed VS Code 1.122–1.125 conveniences → VS Code 1.122–1.128 conveniences; +9 rows covering 1.126 Edit-mode deprecation (reinforces chat.agent.enabled: true pin), 1.127 macOS/Linux terminal sandboxing / /troubleshoot on agent host / browser tools GA / file-based managed Copilot settings / built-in Ollama provider deprecated, 1.128 BYOK Claude via own Anthropic key / custom endpoint model options for BYOK / Claude agent → integrated browser DOM. Byte-identical Supervisor↔Edition mirror per shared-core-coherence.
  • .github/instructions/terminal-command-safety.instructions.md — H2 renamed VS Code platform changes (1.117–1.121) → (1.117–1.127); +1 row for 1.127 macOS/Linux terminal sandboxed by default (Windows unaffected — Backtick Hazard + Output Capture rules still apply); Backtick Hazard falsifier watermark bumped through 1.121 → through 1.128 (microsoft/vscode#295620 re-verified open, milestone On Deck).
  • .github/config/edition-manifest.json — new heir_owned field generated from _registry.cjs, forward-compatible for next-gen Extension installs.

Changed

  • .github/config/welcome-baseline.json ($comment only, settings block untouched) — audit watermark 2026-06-23 → 2026-07-07, VS Code coverage 1.121-1.125 → 1.121-1.128, +3 lines under Category (2) documenting VS Code 1.126 Edit-mode deprecation and reinforcing the chat.agent.enabled: true pin as doubly-aligned with platform direction (user policy AND platform default converged). Category (4) reference synced (1.117-1.121 features) → (1.117-1.127 features).
  • .github/instructions/tool-awareness.instructions.md — lastReviewed bumped 2026-06-23 → 2026-07-07.
  • .github/instructions/terminal-command-safety.instructions.md — lastReviewed bumped 2026-05-29 → 2026-07-07.

Brain contract

No change to min_extension_version (stays at 9.5.1), brain_subtrees, marker_schema, vscode_assets, or bootstrap_templates. The heir_owned addition is a new optional field that older Extensions ignore.

Heir upgrade

Run /upgrade (Extension ACT: Upgrade Brain) to receive the tool-awareness expansion + terminal-safety sandbox row + welcome-baseline $comment refresh. No behavior gate changes for existing rules; the mac/Linux sandbox note is informational context an agent uses when reasoning about terminal-command approval frequency on non-Windows heirs. No --allow-major needed.

Falsifier

2026-08-07 (30 days) per proposal § Falsifiability:

  • (a) If any 1.128 row I shipped turns out to be Insiders-only past 2026-07-05 → prune that row (rows are informational; per-row revert, not whole-batch rollback).
  • (b) If the macOS/Linux sandbox note misleads an agent into skipping the file-redirect fallback on a genuinely stripped-output case → revert the sandbox row.
  • (c) If the 1.126 Edit-mode note in welcome-baseline $comment produces heir confusion (heirs who deliberately used Edit mode with local overrides) → revert $comment refresh; chat.agent.enabled: true pin stands independently.

[3.8.0] - 2026-07-01

Minor [behaviour] — Hardens Edition bootstrap/upgrade and converter script safety, removes curator automation from the heir-shipped source tree, and raises the Node runtime baseline to 24+.

This release keeps the Edition contract stable while tightening the scripts that install, repair, and convert content inside heirs. It removes curator-owned automation from the Edition repo, clarifies user-vs-workspace VS Code settings ownership, makes upgrade rollback safer, preserves heir files inside known skill folders, and adds regression coverage for the failure modes that caused prior bootstrap/upgrade drift.

Removed

  • Removed Edition-owned .github/workflows/brain-qa.yml, .github/workflows/release-gate.yml, and .github/dependabot.yml. CI/update automation belongs outside the heir-shipped Edition brain; bootstrap, upgrade, and Extension install paths already treat workflows and Dependabot config as heir-owned and skip them, but the source repo no longer carries curator automation at all.

Changed

  • Clarified /configure-vscode and /configure-vscode-verify: user-scope settings are machine-level safe defaults, while workspace .vscode/settings.json remains the project-specific override layer handled by /configure-workspace.
  • Clarified /configure-workspace and /configure-workspace-verify: workspace settings intentionally override user-scope settings for the current project while preserving unrelated heir keys.
  • Removed chat.permissions.default from Edition's shipped .vscode/settings.json template so fresh bootstraps receive the safe baseline value from heir-workspace-settings-baseline.json instead of inheriting the Edition repo's own workspace preference.
  • Raised the Edition Node runtime baseline to Node.js 24+: package.json now declares engines.node >=24, and converter skill/script prerequisite text no longer advertises the older baseline.

[3.7.1] - 2026-06-30

Patch [behaviour] — Add markdown.styles: [".vscode/markdown-light.css"] to heir-workspace-settings-baseline.json with set-if-absent mode, closing the gap for heirs whose .vscode/settings.json predated bootstrap and never received the key from the first-install template.

Fresh heirs already received markdown.styles because Edition's own .vscode/settings.json (seeded as a bootstrap_template) carries it. But the bootstrap-template path is first-install-only — heirs whose .vscode/settings.json was preserved as HEIR_OWNED before the template seed (e.g., the heir already had a settings.json when bootstrap ran) never received markdown.styles and got default VS Code markdown preview styling instead of Edition's CSS. The baseline merge layer didn't carry the key either, so the gap persisted across upgrades.

v3.7.1 adds markdown.styles to the baseline with set-if-absent mode: heirs without the key get it (markdown preview now styled by Edition's CSS); heirs with a customized markdown.styles array (e.g., extended with their own CSS files) keep their array unchanged. Same pattern as chat.permissions.default.

Added

  • .github/config/heir-workspace-settings-baseline.json — new entry markdown.styles: [".vscode/markdown-light.css"] in settings; matching entry markdown.styles: "set-if-absent" in mergeMode. spec_version bumped 1.1 → 1.2. $comment block documents the rationale (closes HEIR_OWNED-settings-predates-bootstrap-template gap; preserves heir-custom CSS arrays).

Changed

  • .github/config/edition-manifest.json regenerated; edition_version 3.7.0 → 3.7.1; no shape change otherwise.

Heir-visible behaviour delta

  • Heirs running ACT: Upgrade Brain or /configure-workspace or node .github/scripts/upgrade-self.cjs --apply who currently have no markdown.styles will get it added on next upgrade; markdown preview will use Edition's light CSS.
  • Heirs with a customized markdown.styles array (any value) are untouched (set-if-absent semantics).
  • Heirs who deliberately want default VS Code styling: remove the key after upgrade; future upgrades will not re-add it because their setting is now non-absent (would be null if explicitly removed, which counts as "key present" under set-if-absent).

Brain contract

Unchanged: min_extension_version: 9.5.1, brain_subtrees: [.github], marker_schema: v2, manifest spec 1.4. No Extension republish required — static-fetch (ADR-009) delivers v3.7.1 to heirs at next upgrade.

Falsifier

2026-09-30 (90 days, shared with v3.7.0 falsifier) or sooner if: (a) the merger's set-if-absent mode handles arrays differently than expected and heir-custom markdown.styles arrays get clobbered; (b) heirs report markdown.styles not being added when their .vscode/settings.json lacks the key; (c) the value [".vscode/markdown-light.css"] proves wrong relative to where the CSS actually lives.

[3.7.0] - 2026-06-30

Minor [behaviour] — New /configure-workspace and /configure-workspace-verify prompts close the workspace-scope gap: bootstrap + upgrade already install workspace .vscode/ correctly, but until now there was no prompt-driven recovery path for heirs whose .vscode/markdown-light.css or discovery-location keys went missing.

Until v3.7.0, /configure-vscode (user-scope) and /configure-vscode-verify (user-scope audit) had no workspace-scope siblings. Heirs that cloned a repo with .vscode/ git-ignored, deleted assets accidentally, or lost the discovery keys had to run ACT: Upgrade Brain (Extension) or node .github/scripts/upgrade-self.cjs --apply (script) to recover — a full upgrade cycle for what's really an asset-refresh problem. The new pair gives heirs a targeted recovery surface that matches the user-scope pair shape (three-shell payload, single source of truth from .github/config/*.json, idempotent, heir-only).

Added

  • .github/prompts/configure-workspace.prompt.md — heir-only prompt that refreshes EDITION_OWNED vscode_assets (e.g. markdown-light.css) on every invocation, seeds HEIR_OWNED bootstrap_templates with .vscode/ prefix if missing (extensions.json, settings.json), and per-key merges heir-workspace-settings-baseline.json into .vscode/settings.json (respects mergeMode — set-if-absent keys like chat.permissions.default are preserved if the heir already set them). Source files fetched from GitHub raw at the pinned v<edition_version> tag (post-ADR-009 static-fetch pattern). bash/zsh + PowerShell reference commands.
  • .github/prompts/configure-workspace-verify.prompt.md — read-only audit. Reports which vscode_assets are missing, which bootstrap_templates (.vscode/-prefixed) are missing, and which baseline keys are compliant / drifted / missing in .vscode/settings.json. Recommends /configure-workspace only if drift is found.

Changed

  • .github/prompts/configure-vscode.prompt.md — cross-links to /configure-workspace for workspace-scope; Guardrail clarified that workspace-scope is owned by the new prompt. lastReviewed bumped 2026-06-10 → 2026-06-30.
  • .github/prompts/configure-vscode-verify.prompt.md — cross-links to /configure-workspace-verify. lastReviewed bumped 2026-06-10 → 2026-06-30.
  • .github/prompts/welcome.prompt.md step 5 — inserted /configure-workspace-verify → /configure-workspace recovery step between /configure-vscode and "Start a real chat". Clarifies that bootstrap already installs the workspace assets and this pair is the recovery path.
  • .github/config/edition-manifest.json regenerated — prompts 27 → 29 (the two new prompts added); edition_version 3.6.3 → 3.7.0.

Heir-visible behaviour delta

  • Two new slash commands: /configure-workspace and /configure-workspace-verify. Both refuse outside a heir workspace (require .github/.act-heir.json).
  • /welcome next-steps now has 4 items instead of 3.
  • Zero change to bootstrap or upgrade scripts — they already install workspace .vscode/ correctly (verified empirically against v3.6.3 in the 2026-06-30 investigation).

Brain contract

Unchanged: min_extension_version: 9.5.1, brain_subtrees: [.github], marker_schema: v2, manifest spec stays 1.4.

Falsifier

2026-09-30 (90 days) or sooner if any of the following fires: (a) the new prompts produce results that diverge from bootstrap-heir.cjs / upgrade-self.cjs for the same heir; (b) heirs report the prompts adding files that don't match Edition at the pinned tag; (c) the per-key merge logic doesn't respect mergeMode: set-if-absent and overwrites heir-chosen permission levels; (d) the GitHub raw URL pattern for Edition releases changes and the prompts can't fetch sources.

[3.6.3] - 2026-06-29

Patch [behaviour] — Direct bootstrap/upgrade now apply the ownership contract consistently: BOOTSTRAP_TEMPLATES is the only seeded heir-owned subset, broad HEIR_OWNED paths are never seeded from Edition, and upgrades seed missing templates while refreshing EDITION_OWNED assets.

Follow-up to v3.6.2 and the same-day Extension v9.5.6 investigation. The earlier v3.6.2 patch fixed direct-script upgrades refreshing .vscode/markdown-light.css, but bootstrap-heir.cjs still seeded every HEIR_OWNED path from Edition. That meant a fresh direct bootstrap could copy Edition's curator-side .github/workflows/** and .github/dependabot.yml into a heir — the same leak class found in the Extension path. v3.6.3 makes the direct scripts obey the same policy as the manifest: BOOTSTRAP_TEMPLATES is the explicit safe subset; the rest of HEIR_OWNED is heir territory only.

Fixed

  • .github/scripts/bootstrap-heir.cjs — imports BOOTSTRAP_TEMPLATES instead of HEIR_OWNED for first-install seeding. Fresh direct bootstraps now seed only .github/config/cognitive-config.json, .vscode/extensions.json, and .vscode/settings.json; workflows, dependabot, ISSUE_TEMPLATE, episodic, and local namespaces are no longer copied from Edition.
  • .github/scripts/upgrade-self.cjs — .github/ subtree copy now skips any source file matching HEIR_OWNED, except entries explicitly listed in BOOTSTRAP_TEMPLATES. Missing templates are seeded after restore, and .vscode/markdown-light.css still refreshes via the v3.6.2 Step 3.5 path.

Heir-visible behaviour delta

  • Fresh direct bootstrap: no Edition workflows or dependabot.yml land in heirs.
  • Direct upgrade: no new Edition workflows or dependabot.yml land in heirs. Existing heir-owned workflows remain preserved.
  • Missing bootstrap templates are repaired on upgrade when absent; existing heir-owned settings/extensions files are not clobbered.

Brain contract

Unchanged: min_extension_version: 9.5.1, brain_subtrees: [.github], marker_schema: v2, manifest spec stays 1.4.

Falsifier

2026-09-29 (90 days) or sooner if a direct bootstrap/upgrade seeds any broad HEIR_OWNED file from Edition (.github/workflows/**, .github/dependabot.yml, .github/ISSUE_TEMPLATE/**, .github/episodic/**, or local/**) or fails to seed a listed BOOTSTRAP_TEMPLATES entry.

[3.6.2] - 2026-06-29

Patch [behaviour] — upgrade-self.cjs now refreshes EDITION_OWNED files outside .github/. Closes heir feedback 2026-06-27-machine-reset-vscode-assets-not-copied-on-bootstrap-or-upgrade.md.

The upgrade path's Step 3 has always done copyDirRecursive(editionGh, .github) — refreshing only the .github/ subtree. EDITION_OWNED has carried .vscode/markdown-light.css since the entry was added, but it was never refreshed on upgrade because the script's design treated anything outside .github/ as invisible. Heirs running upgrade-self.cjs carried whatever CSS they had at bootstrap forever; verified machine-reset heir had a 16,628-byte stale copy vs Edition's current 15,856-byte canonical. Correction from later 2026-06-29 investigation: the Extension was also affected, but for a different reason — its production cmdBootstrap / cmdUpgrade loops did not call lib/edition-install.js installFromTarball, so the manifest logic was tested but dead in production. See Extension v9.5.6 and Edition v3.6.3.

Fixed

  • .github/scripts/upgrade-self.cjs — new Step 3.5 inserted between brain install and heir-owned restore. Iterates EDITION_OWNED.filter(p => !p.startsWith('.github/')), expands each pattern against the fetched Edition, copies matching files into the heir root. Today filters to exactly one entry (.vscode/markdown-light.css); future-proof for any new non-.github/ EDITION_OWNED entries. HEIR_OWNED .vscode/settings.json and .vscode/extensions.json are preserved by the existing backup/restore flow (they are not in EDITION_OWNED). Summary line "Edition assets refreshed outside .github/: N" reports the count.

Heir-visible behaviour delta

  • Next /upgrade from any 3.x release: heirs whose .vscode/markdown-light.css differs from the canonical Edition copy will see the file refreshed. No other .vscode/ file is touched. No .github/ behavior change.
  • bootstrap-heir.cjs is unchanged — it already copies the CSS correctly via the EDITION_OWNED literal-path branch (verified empirically 2026-06-29).

Brain contract

Unchanged: min_extension_version: 9.5.1, brain_subtrees: [.github], marker_schema: v2, manifest spec stays 1.4.

Falsifier

2026-09-29 (90 days) or sooner if (a) a heir reports the upgrade refreshes a HEIR_OWNED .vscode/ file (filter is too broad); (b) a heir reports upgrade fails because Step 3.5 cannot find a declared file; (c) the deeper contract convergence (manifest vscode_assets vs registry EDITION_OWNED) surfaces a defect this fix masked.

[3.6.1] - 2026-06-23

Patch [behaviour] — Brain contract: min_extension_version 9.4.0 → 9.5.1. Closes 2 unresolved heir feedback items by enforcing the Extension floor that already carries the fixes.

The 2026-06-09 (vscode_assets handling) and 2026-06-10 (HEIR_OWNED source filter) install-contract fixes were shipped in Extension v9.5.0 and v9.5.1 respectively, but min_extension_version was left at 9.4.0. Heirs running v9.4.x silently missed both protections — confirmed by 2 feedback files in AI-Memory/feedback/ from 2026-06-12 (anychat heir): (a) .github/workflows/{brain-qa,release-gate}.yml + dependabot.yml shipped to heir repos and failed CI by construction; (b) .vscode/markdown-light.css was EDITION_OWNED but never installed on bootstrap. Bumping the floor to 9.5.1 turns these silent failures into explicit refusals ("update the Extension first") and the next /upgrade from a v9.5.1+ heir lands both fixes automatically.

Brain contract

  • min_extension_version: 9.4.0 → 9.5.1 — breaking floor for heirs whose Extension is older than v9.5.1; they see a clear refusal message on next upgrade per ADR-009 install validation. No other contract field changed: brain_subtrees: [.github], marker_schema: v2, manifest spec stays 1.4.

Heir-visible behaviour delta

  • Heirs on Extension v9.5.1 or newer (the common case post-2026-06-10): zero behavior change; the next /upgrade lands current Edition normally. The .github/workflows/** files inside Edition's .github/ are filtered out by the install-side HEIR_OWNED source filter; .vscode/markdown-light.css ships via vscode_assets manifest field.
  • Heirs on Extension v9.4.0 – v9.5.0: next /upgrade refuses with a typed error pointing at the new floor; user updates the Extension via VS Code Marketplace, then re-runs /upgrade. No silent breakage.

Changed (carried from prior unreleased work)

  • .github/config/welcome-baseline.json — extensions.autoUpdate: true → "on" per the VS Code 1.125 canonical shape. Heirs running 1.125+ would auto-migrate the literal true anyway; pre-1.125 heirs still accept true. The $comment Categories header refreshed 2026-06-12 audit ... 1.121-1.124 → 2026-06-23 audit ... 1.121-1.125 and Category (1) commentary rewritten to acknowledge 1.123 2-hour delay + 1.125 making extensions.autoUpdateDelay configurable (Edition does not pin a value — heirs ride the platform default) + 1.125 making "update only enabled extensions" the default behaviour (the explicit extensions.autoUpdateOnlyEnabledExtensions: false pin now overrides that to also update disabled extensions; retained as a deliberate heir-facing choice).
  • .github/instructions/tool-awareness.instructions.md — ## VS Code 1.122–1.124 conveniences → ## VS Code 1.122–1.125 conveniences. Table gains 4 new rows: (a) 1.124 Enterprise Copilot plugin policies (chat.plugins.enabledPlugins / extraMarketplaces / strictMarketplaces may block Mall installs in regulated orgs; heirs should surface a clear message rather than retry); (b) 1.125 extensions.autoUpdate value migration with backwards-compat note; (c) 1.125 extensions.autoUpdateDelay configurable, Edition unpinned; (d) 1.125 forwarded-port URL rewriting reduces remote browser-tool failures in Codespaces/Remote-SSH heir setups; (e) 1.125 Native MDM delivery may policy-lock baseline keys for MDM-managed heirs. lastReviewed 2026-06-12 → 2026-06-23. Byte-identical to Supervisor.

Falsifier

2026-09-23 (90 days) or sooner if (a) a heir reports the refused-upgrade UX is confusing or destructive; (b) bumping the floor strands a heir cluster (≥2 heirs) where Marketplace auto-update is disabled; (c) ≥1 new install-contract change ships without a corresponding min_extension_version bump (we already failed to ship this bump twice — once for v9.5.0, once for v9.5.1; root cause is missing release-preflight rule that release-preflight skill should grow next).

[3.6.0] - 2026-06-12

Minor [behaviour] — VS Code 1.124 settings drift correction. Removes schema-rejected values from baselines; corrects chat.permissions.default enum.

Full-repo audit 2026-06-12 against the live VS Code 1.124 schema surfaced three settings the brain shipped to heirs that VS Code now rejects, plus four documentation-drift items. The live JSON schema validator was the ground truth (not the cached release-notes prose) — several settings the brain copied from VS Code 1.122 release notes had different schema-accepted values when the build finally landed.

Changed

  • .github/config/heir-workspace-settings-baseline.json — chat.permissions.default value corrected from "defaultApprovals" (schema-rejected) to "default". Valid enum is default / autoApprove / autopilot. $comment block updated with the corrected enum names + a note that VS Code 1.124 enabled Autopilot Preview by default (the brain's pin to default is the deliberate opt-out for ACT's permission discipline). Merge mode stays set-if-absent so heirs with deliberate autoApprove / autopilot overrides are preserved.
  • .github/config/welcome-baseline.json — audit date bumped 2026-05-26 → 2026-06-12; refresh range 1.121-1.123 → 1.121-1.124.
    • Category 4 $comment documents removal of chat.tools.terminal.backgroundNotifications (deprecated by VS Code; notifications are now always-on).
    • Category 5 $comment documents removal of chat.utilityModel + chat.utilitySmallModel (schema rejected "gpt-4o-mini"; only "" is accepted now; users should set utility-model routing via the Chat: Manage Language Models UI introduced in 1.106+).
    • Category 8 $comment adds the heir-workspace chat.permissions.default: default companion + the 1.124 Autopilot-default context.
  • .github/instructions/tool-awareness.instructions.md — "VS Code 1.122 conveniences" section becomes "VS Code 1.122–1.124 conveniences". New rows: 1.123 session sync + /chronicle (clarifies the brain's local chronicle skill is adjacent, not replaced), 1.123 sandbox network-retry, 1.124 Autopilot enabled by default, 1.124 Advanced Autopilot opt-in. Table gains a Release column.
  • README.md — Utility-slot row in the model legend and the chat.utilityModel / chat.utilitySmallModel row in the Practical Recommendation table rewritten: Edition no longer pins a value in welcome-baseline.json; users route utility models via the Chat: Manage Language Models UI.
  • test/workspace-settings-merger.test.js — test fixtures rewritten to use the current VS Code enum (default / autoApprove / autopilot) instead of the invalid defaultApprovals / bypassApprovals strings. The merger tests test merge LOGIC, not VS Code SCHEMA, so this is a documentation-quality fix — future readers don't see invalid enum values in test fixtures. All 69 tests pass.

Removed

  • chat.tools.terminal.backgroundNotifications from welcome-baseline.json (VS Code-deprecated).
  • chat.utilityModel from welcome-baseline.json (schema-rejected for hardcoded model names).
  • chat.utilitySmallModel from welcome-baseline.json (same).

Brain contract: no change

min_extension_version: 9.4.0, brain_subtrees: [.github], marker_schema: v2. Manifest spec stays at 1.4. Edits are settings-baseline corrections + documentation refreshes; no install-contract surface changes.

Heir impact

Heirs running /upgrade from any 3.x release:

  • Fresh heirs (no prior chat.permissions.default): get the corrected "default" value pinned.
  • Heirs that received "defaultApprovals" from any prior baseline: set-if-absent preserves their value, so the schema warning in their .vscode/settings.json persists. Manual fix is the cheapest cleanup: edit "defaultApprovals" → "default" (or pick autoApprove / autopilot if a deliberate override is desired). VS Code's fallback behaviour means the warning is cosmetic, not breaking.
  • Heirs that received chat.tools.terminal.backgroundNotifications / chat.utilityModel / chat.utilitySmallModel via /configure-vscode: those keys remain in user-scope settings.json until manually removed. Running /configure-vscode-verify will surface them as drift; they are not removed automatically because welcome-baseline is additive (the prompts apply baseline; they do not unset keys absent from baseline).

Falsifier

If VS Code 1.125+ further changes any of the three corrected settings (chat.permissions.default enum values, chat.utilityModel schema, or deprecates more terminal-chat keys), re-audit. Re-evaluate this release's documentation by 2026-09-12 (90 days). Earlier triggers: any heir feedback that the schema warning on defaultApprovals is actually breaking VS Code behaviour, or first observed contradiction between the live schema and the brain's documented enum values.

[3.5.0] - 2026-06-10

Minor [behaviour] — Mac/Linux command parity for /configure-vscode, /configure-vscode-verify, /mall-refresh.

Audit 2026-06-10 (post-AP-01/AP-11 fleet adoption) surfaced that several lifecycle prompts shipped Windows PowerShell as the only working reference command. Heir agents on Mac/Linux were lifting $env:APPDATA and ConvertFrom-Json -AsHashtable verbatim, then failing.

Changed

  • .github/prompts/configure-vscode.prompt.md — "Windows Reference Command" section becomes "Reference Commands", adds working bash/zsh implementation (uses $OSTYPE to pick macOS ~/Library/Application Support/Code/User/settings.json vs Linux ~/.config/Code/User/settings.json, then Node one-liner for the merge). PowerShell block preserved verbatim. Both shells perform the same non-destructive merge.
  • .github/prompts/configure-vscode-verify.prompt.md — adds "Reference Commands (read-only audit)" section with bash/zsh + PowerShell implementations. Both compute Compliance/Drift/Missing counts and print a drift table; neither writes to settings.json.
  • .github/prompts/mall-refresh.prompt.md — ```pwsh fences on node .github/scripts/audit-mall-drift.cjs invocations changed to ```bash. The commands are platform-neutral; the misleading shell label was the bug.
  • .github/scripts/audit-mall-drift.cjs — catalog discovery candidate list gains ~/Development/Alex_Skill_Mall/ on Mac/Linux (symmetric to the existing Windows C:\Development\Alex_Skill_Mall\ fallback). Discovery order unchanged: sibling first, then home, then Development root, then home root, then GitHub raw HTTPS.

Brain contract: no change

min_extension_version: 9.4.0, brain_subtrees: [.github], marker_schema: v2. Manifest spec stays at 1.4. All edits are additive (more shells supported, more discovery paths tried).

Heir impact

Heirs on /upgrade from any 3.x release receive the refreshed prompts and the extra catalog discovery candidate. Heirs who had to translate PowerShell to bash by hand on every fresh-Mac install no longer need to. Zero breaking change to any existing workflow.

Falsifier

The three-shell reference is decorative if 90 days pass (re-evaluate 2026-09-10) and no Mac/Linux heir invokes the new bash block (heir feedback shows agents still defaulting to PowerShell). If so, narrow to a single Node one-liner that works on all three.

[3.4.1] - 2026-06-10

Patch [behaviour] — close bootstrap_templates leak + fix cross-platform test bugs.

Fleet-adoption batch shipped 2026-06-10 (AP-01/03/07/11/12) added .github/dependabot.yml to Edition. That file lives under HEIR_OWNED in _registry.cjs (heirs own their own dependency policy), but build-edition-manifest.cjs listBootstrapTemplates() inferred its bootstrap-template list from HEIR_OWNED literal-path entries that exist in Edition. Curator-only .github/dependabot.yml therefore joined bootstrap_templates, meaning the Extension would install Edition's curator-side dependency policy file into every heir on first install.

Fixed

  • .github/scripts/_registry.cjs now exports an explicit BOOTSTRAP_TEMPLATES array — the subset of HEIR_OWNED that's genuinely heir-starter-template (cognitive-config, .vscode/extensions, .vscode/settings). Adding a row here is now an explicit curator decision, not an inferred side-effect of touching HEIR_OWNED.
  • .github/scripts/build-edition-manifest.cjs listBootstrapTemplates() reads BOOTSTRAP_TEMPLATES directly instead of filtering HEIR_OWNED. The regenerated manifest contains the same 3 entries as before today's leak (cognitive-config.json + extensions.json + settings.json); .github/dependabot.yml correctly excluded.
  • test/registry.test.js readProfile/writeProfile tests now manage both USER (Unix) and USERNAME (Windows) env vars. Pre-fix, the tests set only USERNAME and passed only on Windows where USER is typically unset; on macOS / Linux USER always wins, so the tests returned the default profile instead of the user-specific one (3 failures on npm test).

Heir impact

Heirs running /upgrade against Edition v3.4.1 (via Extension v9.4.0+) will no longer receive Edition's curator dependabot.yml as a bootstrap template. Heirs that already received it on a v3.4.0 install can safely rm .github/dependabot.yml — the file is heir-owned per _registry.cjs. Extension v9.5.1 adds a defense-in-depth source-side HEIR_OWNED filter on the subtree copy that catches this same leak class for any future Edition release.

Brain contract: no change

min_extension_version: 9.4.0, brain_subtrees: [.github], marker_schema: v2. Manifest spec stays at 1.4. BOOTSTRAP_TEMPLATES is a build-side artifact, not exposed in the manifest contract.

Falsifier

The explicit BOOTSTRAP_TEMPLATES list is decorative if 90 days pass (re-evaluate 2026-09-10) without any new curator-only HEIR_OWNED file joining Edition that would have been incorrectly inferred. If no such pressure surfaces, the explicit list is still load-bearing (it documents intent), but the falsifier is the gap between intent-as-policy and intent-as-mechanism.

[3.4.0] - 2026-06-07

Minor [behaviour] — Adopt 4 authoring + prose-quality skills from Hermes Agent.

Closes named gaps in Edition baseline: no pre-write planning discipline (plan), no feasibility validation (spike), no test-first enforcement (test-driven-development), no deep on-demand prose humanization beyond the always-on markdown-author agent's 15-word filter (humanizer). All four load on description match — token cost is on-demand, not always-on.

Adapted from Hermes Agent (Nous Research, MIT) which itself ports from obra/superpowers, GSD, and blader/humanizer. All upstream sources MIT or MIT-compatible; attribution preserved in each SKILL.md footer.

Added

  • [behaviour] .github/skills/plan/SKILL.md — new skill. Plan-mode authoring discipline: write a concrete actionable markdown plan with bite-sized tasks (2-5 min each), exact file paths, complete code, verification steps, before any non-trivial implementation. Saves to docs/plans/. Adapted from Hermes Agent / obra/superpowers (MIT). Falsifier 2026-09-07. Loaded on description match — /plan, "design before building", multi-file work.
  • [behaviour] .github/skills/spike/SKILL.md — new skill. Throwaway feasibility experiments: decompose an idea into 2-5 questions, build minimal observable prototypes under spikes/<NNN>-<name>/, return VALIDATED / PARTIAL / INVALIDATED verdict. Adapted from Hermes Agent / GSD (MIT). Falsifier 2026-09-07. Composes with problem-framing-audit (frame-first) and plan (post-validation build).
  • [behaviour] .github/skills/test-driven-development/SKILL.md — new skill. RED-GREEN-REFACTOR enforcement for any production code change. Iron Law: no production code without a failing test first. Carves out throwaway prototypes, generated code, configuration files. Adapted from Hermes Agent / obra/superpowers (MIT). Falsifier 2026-09-07. Composes with systematic-debugging (test reproduces bug) and plan (every plan task starts with RED).
  • [behaviour] .github/skills/humanizer/SKILL.md + .github/skills/humanizer/examples/full-example.md — new skill. Strips 29 documented AI-writing patterns (Wikipedia's "Signs of AI writing" via blader/humanizer) from any prose. Fires on "humanize", "de-AI", "un-ChatGPT", or voice-match requests. Iterative draft → self-audit → final loop. Optional voice calibration from a user-provided writing sample. Composes with markdown-author agent's always-on 15-word banned-vocabulary filter (humanizer is the deeper on-demand pass) and Cardinal Rule 2 (em-dash ban; humanizer Pattern 14 documents the underlying reason). Adapted from Hermes Agent / blader/humanizer (MIT). Falsifier 2026-09-07 (sinks to Mall rather than removing entirely).

Changed

  • [clarification] .github/config/welcome-baseline.json — $comment refreshed against VS Code 1.123 release notes (2026-06-03). Categories header now reads VS Code 1.121-1.123. New sub-note under category (1) documents the 1.123 supply-chain feature: auto-updates apply a 2-hour delay after a new extension version is published (trusted publishers — Microsoft / GitHub / OpenAI — are exempt). extensions.autoUpdate: true semantics unchanged; the delay is an implicit platform behaviour, not a setting we can configure. No settings object change, no heir-visible behaviour delta.

Brain contract

min_extension_version: 9.4.0 (no change). brain_subtrees: [.github] (no change). marker_schema: .act-heir.json v2 (no change). Brain contract: no change.

Heir upgrade

/upgrade from any 3.x release fetches the 4 new skill folders automatically; no --allow-major needed. Skills load on description match — heirs that never invoke /plan, /spike, TDD workflows, or "humanize this" requests see zero behaviour change.

Falsifier

2026-09-07 (90 days) — per-skill event-based falsifiers in each new SKILL.md ## Would Revise If. If by then any of the 4 skills has zero observed invocations across the fleet, sunset that skill (TDD + humanizer sink to Mall rather than removing entirely). Track in Alex_ACT_Supervisor/docs/ledgers/curation-log.md tagged [HERMES-TIER-A-ADOPTION].

[3.3.0] - 2026-06-03

Minor [behaviour] — Per-key merge mode for the workspace-settings baseline + pin chat.permissions.default on fresh installs.

Heir feedback flagged that Edition holds policy locks for the Claude Agent permission surfaces (claudeAgent.allowAutoPermissions: false and claudeAgent.allowDangerouslySkipPermissions: false in welcome-baseline.json) but is silent on VS Code 1.122's chat.permissions.default — the workspace-scope chat permission level. Heirs who flipped Bypass or Autopilot had no policy signal from the brain.

The fix adds the safe default to the workspace baseline, but workspace-scope clobber on every /upgrade is the wrong shape: permission level is a per-repo workflow choice, not a fleet-wide policy stance. So this release also extends the merger with a new set-if-absent mode — pin on fresh installs, respect heir overrides on upgrade.

Added

  • [behaviour] .github/scripts/shared/workspace-settings-merger.cjs — new mergeMode map (per-key, optional) in the baseline JSON. Two modes:
    • enforce (default — used when mergeMode omits the key) — current behaviour: object deep-merge or scalar overwrite. The three existing chat.*Locations discovery keys keep this mode by default (their sub-keys must always be present or local skills/prompts/agents stop loading).
    • set-if-absent — skip wholesale if heir already has the key (under any value, including object/scalar/null). Use when the brain wants a safe default on fresh installs but respects per-repo heir overrides on upgrade.
  • [behaviour] .github/config/heir-workspace-settings-baseline.json — new key chat.permissions.default: "defaultApprovals" with mergeMode: { "chat.permissions.default": "set-if-absent" }. Fresh heirs get the safe VS Code default pinned at bootstrap; heirs that deliberately flip to bypassApprovals (experimental sandbox) or autopilot (long-running solo task) are never overwritten on /upgrade.
  • [behaviour] Merger result now carries a skipped array alongside changes. formatChangeSummary surfaces respected overrides so bootstrap and upgrade scripts can tell heirs exactly which keys the merger did NOT touch and why.
  • [clarification] 7 new tests in test/workspace-settings-merger.test.js covering set-if-absent on fresh repo, heir-existing value (no change), heir explicit null counts as presence, mixed enforce + set-if-absent routing in same baseline, default enforce behaviour preserved when mergeMode absent, and formatChangeSummary skipped-overrides surfacing.

Changed

  • [clarification] .github/config/heir-workspace-settings-baseline.json spec_version bumped 1.0 → 1.1 (additive mergeMode field, backward-compatible — readers that ignore mergeMode default to enforce).
  • [clarification] .github/config/README.md — added missing heir-workspace-settings-baseline.json row to the ownership table; new "How the workspace-settings merger applies the heir baseline" section documents the two modes.
  • [clarification] .github/scripts/shared/workspace-settings-merger.cjs — header docstring expanded to document mergeMode semantics and link to the new proposal.
  • [clarification] Added intentional-divergence audit markers to shared-core brain artifacts that ship with Supervisor-curated heir-portable phrasing (skill-creator, skill-review, agent-creator, agent-review, instruction-creator, instruction-review, prompt-creator, prompt-review, code-review, meditation, brain-audit). No behavior change; the markers make the audit trail explicit per shared-core-coherence-audit.
  • [clarification] Mirrored markdown-mermaid and alex-banner-generation skills (and the markdown-mermaid/references/ reference set) from Supervisor to clear documentation drift.

Brain contract

min_extension_version: 9.4.0 (no change). brain_subtrees: [.github] (no change). marker_schema: .act-heir.json v2 (no change). Brain contract: no change.

Heir upgrade

/upgrade from any 3.x release applies the new key per its mode:

  • Heir who never set chat.permissions.default → gets defaultApprovals written to .vscode/settings.json (visible in upgrade output as one applied change).
  • Heir who already set it (any value) → no change; upgrade output reports respected 1 heir override.

No --allow-major needed. No manual action required for either case.

Falsifier

2026-09-03 (90 days) or any VS Code release that removes/renames chat.permissions.default or changes its valid value set — re-evaluate the baseline key and the heir-feedback channel that drove the change.

Patch [clarification] — documentation and tooling cleanup, plus dead-file purge. No observable behavior change for any working consumer.

Removed

  • [behaviour] .github/scripts/shared/index.mjs — orphaned ESM bridge that imported a ./svg-pipeline.cjs module which never existed in git history or on disk. Loading the bridge threw Cannot find module for any consumer that tried. Audit confirmed zero live consumers; converter-qa.cjs only stat-checked its presence, never imported it. Honest cleanup; an ESM bridge can be reintroduced when a real ESM consumer arrives.

Fixed

  • [clarification] .github/scripts/converter-qa.cjs — md-to-html suite now skips gracefully when pandoc is absent, matching the guard the md-to-word, html-to-md, and docx-to-md suites already had. Eliminates the 2 spurious failures that appeared on machines without pandoc. Also dropped the presence-check line for the deleted shared/index.mjs.
  • [clarification] README.md — corrected skill/prompt counts (33 skills, 27 prompts) to match the actual brain shape; was lagging at 32/26.
  • [clarification] .github/copilot-instructions.md — removed two dead skill references from the cluster table (mall-installation and converter skills do not exist; the /mall-* prompts and 6 format skills already cover those domains). Added alex-banner-generation skill and /convert prompt as accurate replacements.

[3.2.0] - 2026-05-31

Minor [behaviour] — manifest spec 1.4 lands the static-fetch Extension contract (per ADR-009).

Pure infrastructure release: brain content is unchanged from v3.1.0. This release exists to ship the edition-manifest.json contract fields that Alex_ACT_Extension v9.4.0+ reads at install time to validate which subtrees it may install, what minimum Extension version is required, and what marker shape to write. Heirs running Edition v3.1.0 or earlier continue to work; the new fields are additive and null on legacy reads.

Added (manifest spec 1.4 — static-fetch Extension contract)

  • .github/config/extension-contract.json — hand-authored sidecar declaring min_extension_version, brain_subtrees, marker_schema. Source of truth for the static-fetch Extension contract (ADR-009).
  • .github/scripts/build-edition-manifest.cjs bumped to spec 1.4; merges the sidecar fields into the generated .github/config/edition-manifest.json at release time. Fields are null on legacy reads (sidecar absent).

Changed

  • .github/config/edition-manifest.json carries min_extension_version: 9.4.0, brain_subtrees: [".github"], marker_schema: {file_name: ".act-heir.json", version: 2}. Read by Alex_ACT_Extension v9.4.0+ static-fetch path to validate the install contract before any destructive op. See ADR-009.

Brain contract: spec_version 1.3 → 1.4, min_extension_version 9.4.0, brain_subtrees [".github"], marker_schema v2.

CHANGELOG discipline note (added 2026-05-30 per ADR-009 Phase 1A.6): every release entry going forward includes a **Brain contract**: line. Empty (Brain contract: no change) is valid; non-empty entries call out one or more of: min_extension_version bumped to X.Y.Z, brain_subtrees changed (added X, removed Y), marker_schema bumped to vN. Source of truth for fleet upgrade impact analysis. Bumping min_extension_version is a breaking change for any heir whose Extension is older than the new floor.

[3.1.0] - 2026-05-29

Minor [behaviour] — Plugin Mall v3 catalog integration (Phase 5a) + shared-core audit fixes. Rewrites /mall-search and /mall-show prompts to read the new trust-scored Plugin Mall catalog (catalog/index.json + catalog/stores/*.json) per PLAN-mall-automation v3 / ADR-008. Replaces legacy CATALOG.json lookups. Also bundles today's shared-core audit fixes that brain-auditor surfaced against Edition.

Added (Phase 5a — Plugin Mall v3)

  • /mall-search now reads catalog/index.json (trust-ranked, ~1.4 MB), surfaces Mall-curated entries (🏆) at the top by trust score (provenance bonus +50), shows store + version + truncated description per result. Falls back to GitHub raw URL when local Mall clone unavailable.
  • /mall-show is new — drills into one plugin from catalog/stores/<store>.json and displays full metadata: trust score + every signal that fed it, adapted_from (for Mall-curated entries), frontmatter.standard + extended + raw layers, available_refs (default branch + SHA + tags), source_url.
  • /mall-install placeholder: deferred to Phase 5b once heir feedback validates which install workflows actually matter (shape-handling, pinning strategy, multi-file with references). Today heirs install manually from source_url per the prompt's instructions; Phase 5b automates with a mall-install.cjs script.

Fixed (audit findings from brain-auditor dispatch 2026-05-29)

  • system-prompt-skepticism.instructions.md (high severity) — replaced broken reference to nonexistent worldview-integration.instructions.md with the actual file worldview.instructions.md. The drift had been shipping descriptive prose pointing at a file that doesn't exist; heirs reading the always-on instruction saw a dangling reference.
  • problem-framing-audit/SKILL.md (medium severity) — replaced 2 references to nonexistent /reframe prompt with the actual prompt name /problem-framing-audit. Pure mirror gap (Supervisor was already correct).
  • brain-auditor.agent.md + brain-audit/SKILL.md (medium severity) — mirrored Supervisor's Phase 7b stale-architecture row + Mall sibling-repo handling (added when the Supervisor brain was stripped of Mall operational artifacts after Mall self-curation went live).
  • docx-to-md/SKILL.md (medium severity) — removed dead Related-skill reference to nonexistent md-scaffold.
  • md-to-word/SKILL.md (medium severity) — removed 5 dead Related-skill references: pptx-generation, md-scaffold, book-publishing, svg-graphics, brand-asset-management. None of these skills ship in Edition.

Changed (shared-core architecture)

  • critical-thinking/SKILL.md — rewrote the documented 3-leg epistemic triad to 2 legs (anti-hallucination + critical-thinking). The awareness skill body referenced in the prior triad doesn't ship in either Supervisor or Edition; documentation now matches shipped reality. Error-detection during reasoning is currently distributed across epistemic-calibration (self-correction triggers table) + reliance-nudges (repeated-same-error nudge). If a heir reports the gap, restore awareness as a dedicated skill — tracked for 2026-08-29 retrospective.

Heir-visible behaviour delta

  • A heir running /mall-search code-review on Edition v3.1.0 sees Mall-curated entries at the top with trust scores (vs legacy CATALOG.json keyword match with no trust signal).
  • The new /mall-show <name> command surfaces the full signal breakdown — heirs can audit why a plugin scored what it scored before installing.
  • Always-on instructions no longer reference nonexistent files (worldview-integration).
  • Critical-thinking docs no longer describe a triad leg that doesn't ship.

[3.0.1] - 2026-05-29

Patch [behaviour] — additive brain refresh + terminology alignment + protection contract. Bundles four post-v3.0.0 commits that were not yet released, adds the protected-repo contract, adds a doc-hygiene routing instruction, and renames the heir-facing AI-Memory references to "shared memory bus" for terminology consistency with the sibling-repo architecture. No removals, no schema changes.

Added

  • .act-protected.json at repo root declaring kind: edition, bootstrap_allowed: false. Tells the Extension's AlexMaster-detection modal and migrateFromAlexMaster command to refuse Edition itself as a migration target. Pairs with the Extension-side guard (Extension v8.x patch — see Extension CHANGELOG).
  • .github/instructions/doc-hygiene.instructions.md — pattern-applied routing pointer (applyTo: "**/*doc*audit*,**/*doc*quality*,**/*drift*,**/*hygiene*") that fires on doc-audit / drift / hygiene file patterns and delegates to the existing doc-hygiene skill body.
  • anti-hallucination skill mirrored from Supervisor brain-auditor findings; sibling markdown-author agent folds in the prior ai-writing-avoidance content (commit 45dbeae).
  • Always-on rationale paragraphs mirrored across shared-core instructions from Supervisor (commit 2fcf31d) — names why each always-on rule earns its per-turn budget.
  • brain-auditor refreshed for the AI-Memory sibling-repo architecture and frontmatter-spec / applyTo-calibration gap-close (commits d87eea9, 189daac).

Changed

  • Terminology: heir-facing prompts and instructions now say "shared memory bus" (or "shared memory") instead of "AI-Memory" — cross-project-isolation.instructions.md, greeting-checkin.instructions.md, meditation.instructions.md, checkin.prompt.md, feedback.prompt.md, mall-contribute.prompt.md, note.prompt.md, save-session-note.prompt.md, .github/config/README.md. The underlying sibling repo ../Alex_ACT_Memory/ is unchanged; only the prose terminology updated.
  • meditation.instructions.md intentional-divergence marker added: Edition omits the Supervisor-only "Quarterly Retraining Integration" section; meditation frontmatter guidance updated to match the spec-aligned shape (description, applyTo for instructions; name + description for skills; lastReviewed always).
  • cross-project-isolation.instructions.md intentional-divergence marker added: Edition includes feedback-channel paths in applyTo and a "Fleet feedback" channel row — heir-side surfaces Supervisor doesn't write to.
  • mall-contribute.prompt.md frontmatter guidance for new Mall SKILL.md files updated from the legacy 9-field shape to the spec-aligned name + description + lastReviewed.
  • .github/VERSION → 3.0.1.

Heir-visible behaviour delta

  • A heir on v3.0.0 upgrading to v3.0.1 receives the new .act-protected.json, the new doc-hygiene.instructions.md routing pointer, the anti-hallucination skill, and the terminology rename across 9 instructions/prompts. No local/ overlay changes. _registry.cjs ownership unchanged. Marker edition_version rolls 3.0.0 → 3.0.1.

[3.0.0] - 2026-05-27

Major [behaviour] — structural consolidation and VERSION single-sourcing.

Changed

  • VERSION consolidation — removed stale root VERSION file; canonical version is now solely .github/VERSION, read by build-edition-manifest.cjs, bootstrap-heir.cjs, and upgrade-self.cjs.
  • Manifest edition_version updated from 2.6.0 to 3.0.0 (aligning with the v3.0.0 tag).
  • HANDOFF.md removed (stale 2026-05-26 session context, superseded).

Breaking

  • Heirs or scripts that read a root-level VERSION file must update to read .github/VERSION instead.

[2.6.0] - 2026-05-27

Minor [behaviour] — heir discovery setup automation. Fixes a silent-failure mode where Mall plugins installed under .github/skills/local/<name>/ (the official path per mall-installation.instructions.md) were invisible to VS Code Copilot's chat surface because skill / prompt / agent discovery walks each registered root one level only, while instruction discovery recurses. Empirically verified on the job heir 2026-05-27 — adding the three chat.*FilesLocations workspace settings surfaced 18 previously-invisible local skills.

Added

  • .github/config/heir-workspace-settings-baseline.json — declarative baseline of the three VS Code workspace settings (chat.agentSkillsLocations, chat.promptFilesLocations, chat.agentFilesLocations) that register .github/<type>/local/ as a second discovery root alongside .github/<type>/. Carries $comment block with full verification context and a falsifier deadline (2026-11-27 or sooner on any VS Code release that changes the settings contract). Registered as EDITION_OWNED in _registry.cjs.
  • .github/scripts/shared/workspace-settings-merger.cjs — shared merger module that performs idempotent per-key deep-merge of the baseline into the heir's .vscode/settings.json. Strips JSONC // and /* */ comments before parsing. Exports mergeWorkspaceSettings(repoRoot, baselinePath), writeMerged(result), formatChangeSummary(result, verb). Separates compute from write so callers can dry-run.

Changed

  • bootstrap-heir.cjs — invokes the merger after the marker write (inside the existing APPLY-gated execute block). Fresh heirs receive the three discovery settings merged on top of the existing .vscode/settings.json template; all unrelated keys are preserved verbatim.
  • upgrade-self.cjs — invokes the merger after step 5 (marker update). Existing heirs upgrading from any prior Edition version receive the three keys backfilled into their .vscode/settings.json on next /upgrade. Idempotent — subsequent upgrades report "Workspace settings: already current" with no changes.
  • mall-installation.instructions.md — new ### Discovery setup section at the top of ## Installation documents the one-level discovery walk, the three settings keys with two-root maps, references the baseline + automatic merge, and provides a manual fallback for heirs on Edition < 2.6.0 (paste the keys directly, or run scripts/apply-skill-discovery-settings.cjs --repo <heir-path> from a sibling Supervisor checkout). lastReviewed bumped to 2026-05-27.
  • build-edition-manifest.cjs — EDITION_CONFIG_FILES allowlist extended to include heir-workspace-settings-baseline.json so the manifest's configs array stays in sync with _registry.cjs. Manifest regenerated; configs count 3 → 4.

Heir-visible behaviour delta

Per-key merge semantics: if a heir has previously set one of the three chat.*FilesLocations keys to a different shape (e.g. a scalar value or an array instead of an object), the merger will scalar-replace that value with the baseline object. Heirs who have customized these keys for non-default discovery roots should re-apply their customization after upgrade. Heirs who have not touched these keys (the expected majority) see only the three new keys appearing in .vscode/settings.json with no other change.

Compatibility

Older VS Code (pre-1.118) silently ignores the three settings keys — the additions are backwards-compatible. Heirs on Edition < 2.6.0 can adopt the same fix manually using the documented fallback in mall-installation.instructions.md; no breakage if they do not upgrade.

[2.5.0] - 2026-05-27

Minor [behaviour] — baseline expansion + shared-core coherence. Two new baseline skills adopted from MALL (debugging + security disciplines), VS Code 1.122 conveniences surfaced across the always-on instruction set, repository-wide EOL normalization (.gitattributes), and three shared-core instructions mirrored from Supervisor to close the gaps identified in the 2026-05-26 audit.

Added

  • systematic-debugging skill adopted from MALL/obra-superpowers — four-phase root-cause-first method (investigate → pattern-analyze → hypothesize → implement) with iron law "NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST". Includes 3 supporting references (root-cause-tracing.md, defense-in-depth.md, condition-based-waiting.md). Closes a confirmed baseline gap — Edition had no debugging discipline; heirs defaulted to symptom-fixing.
  • security-and-hardening skill adopted from MALL/addyosmani-agent-skills — OWASP-aware security-first practices with three-tier boundary system (Always Do / Ask First / Never Do), input validation patterns, secrets management, dependency-audit triage decision tree, security review checklist. Examples in TypeScript but principles are language-agnostic. Closes a confirmed baseline gap — Edition had zero security coverage; heirs handling input/auth/storage started without OWASP awareness.

Changed

  • tool-awareness + lint-discipline instructions: VS Code 1.122 conveniences (/models slash, BYOK air-gapped, local agent host watchpoint; "Search only in changed files" toggle for lint scope) surfaced as additive shortcuts. No rule changes.
  • pii-memory-filter.instructions.md: added cross-link paragraph to memory-triggers.instructions.md § Memory Tier Selection clarifying that the filter constrains what may be written while MT constrains where. lastReviewed bumped to 2026-05-26. Mirrored from Supervisor (closes audit divergence).
  • falsifiability-deadlines.instructions.md: dropped stale / muscle from description after the .github/muscles/ folder was collapsed into .github/scripts/ in v2.4.0. lastReviewed bumped to 2026-05-26. Mirrored from Supervisor.
  • severity-tagged-commits.instructions.md: added in-file <!-- intentional divergence --> marker documenting why the Edition variant differs from Supervisor's (heirs don't ship ACT/**; trigger list includes .github/scripts/**; brain-qa-changelog cross-ref is generalized). lastReviewed bumped to 2026-05-27. Closes audit Recommendation #3.
  • README: skill count refreshed 30 → 32 after the two MALL adoptions.

Infrastructure

  • .gitattributes added with * text=auto eol=lf to normalize line endings across the repo. Required for stable EOL-normalized hashing in Supervisor's shared-core-coherence-audit skill. Pairs with Supervisor's matching .gitattributes (commit a0e7e8e).

[2.4.0] - 2026-05-26

Minor [behaviour] — substantial release: per-type review/creator pairs (ADR-007), full Edition audit closing 39 findings, .github/muscles/ folder collapsed into .github/scripts/, the 2026-05-26 cleanup pass (heir-facing docs validated, plugin guide removed, model table refreshed against GitHub Docs, .github/config/ pruned, .vscode/ workspace template established with curated CSS + 14-key settings), and edition-manifest.json extended to a complete file-level bill-of-materials (spec_version 1.0 → 1.3).

Added

Skills + prompts

  • deep-review skill mirrored from Supervisor — adversarial code review with three parallel perspectives (Advocate, Skeptic, Architect). Complements single-pass code-review for high-stakes PRs.
  • 3 baseline skills mirrored from Supervisor: code-review, git-workflow, status-reporting. Each with always-on routing instruction. Closes a real heir-baseline gap.
  • 6 per-type review/creator pair skills: instruction-review/instruction-creator, prompt-review/prompt-creator, agent-review/agent-creator (per ADR-007). Heirs gain Supervisor's curation surface.
  • 3 new slash-prompts: /review-instruction, /review-prompt, /review-agent — complete four-way symmetry with the existing /review-skill.
  • doc-hygiene skill mirrored from Supervisor (anti-drift rules, link integrity, count elimination, living-document maintenance).
  • markdown-mermaid/references/mermaid-reference.md (1339 lines) — deep-dive content extracted from SKILL.md (diagram tool selection, palette + theming, parser pitfalls, audit methodology, quality checklist).

Workspace environment + user settings

  • .vscode/markdown-light.css (edition-owned) — Mermaid-friendly markdown preview theme shipped directly. Replaces the prior copy-via-prompt workflow.
  • .vscode/settings.json (heir-owned bootstrap template, 14 keys) — wires markdown.styles at the new CSS, ships recommended Mermaid theme defaults (matches /polish-mermaid-setup Step 3), markdown-scoped editor settings (preserves two-space line breaks, wordWrap on, prose suggestions on), file hygiene (insertFinalNewline, trimFinalNewlines), sensible markdown preview defaults. Preserved across /upgrade.
  • welcome-baseline.json gains 3 user-scope keys:
    • github.copilot.chat.skillTool.enabled: true — load-bearing per tool-awareness.instructions.md; default true, explicit lock prevents drift across VS Code versions.
    • claudeAgent.allowAutoPermissions: false and claudeAgent.allowDangerouslySkipPermissions: false — defensive locks against VS Code 1.121's Claude Agent preview "Auto" mode and "Bypass all permissions" / YOLO mode. ACT's permission discipline is non-negotiable.
    • $comment refreshed to 8 categories with REQUIRED markers so future audits can distinguish load-bearing from UX-only settings.

Changed

Brain artifacts

  • 22 prompts stripped deprecated mode: agent frontmatter — deprecated and ignored per current Microsoft Learn prompt-files spec.
  • 23 prompts gain ## Would Revise If + lastReviewed: 2026-05-26 (20 from the broad sweep + 3 thin alias prompts via inline **Would revise if** text: checkin, meditate, note).
  • 3 converter skills gain ## Would Revise If — html-to-md, markdown-sanitization-chain, md-to-txt. 90-day windows with specific trigger conditions.
  • 8 always-on shared-core instructions mirrored byte-identical with Supervisor: act-foundations, act-pass, critical-thinking, epistemic-calibration, privacy-responsible-ai, proactive-awareness, system-prompt-skepticism, falsifiability-deadlines. Vague "would revise if" sections replaced with concrete 90-day falsifier windows (2026-08-26).
  • 4 worker agents tightened against Gate 6 (Tool Allowlist Minimality) — byte-identical mirror with Supervisor. document-assembler drops unused search/codebase; illustrator trims to read-only; markdown-author drops search/* + search/usages. brain-auditor retains edit.
  • markdown-mermaid SKILL.md trimmed 1648 → 327 lines (80% reduction) by extracting reference content to references/mermaid-reference.md. SKILL.md keeps the operational core (init template, ATACCU workflow, mode fragility). Description rewritten from slogan to what+when discovery aid.

Heir-facing docs + workspace assets

  • markdown-light.css moved to .vscode/ (was .github/skills/markdown-mermaid/markdown-light.css). Edition-owned per sync-policy.json (18 → 19 paths). Replaces the manual copy-via-prompt workflow; /polish-mermaid-setup Step 4 simplified accordingly.
  • README.md revised — instructions list aligned with actual brain (36 instructions across 6 cleaner categories; dropped 6 phantoms alternatives-and-tradeoffs/partnership-charter/technical-writing/creative-loop/debugging/scope-management; added 8 missing real ones agent-delegation/code-review/falsifiability-deadlines/git-workflow/no-deferred-debt/severity-tagged-commits/status-reporting/tool-awareness-categories). Path-count contradiction fixed ("Three entry paths"). Two redundant Mall subsections consolidated. Practical Recommendation aligned with the model table.
  • Model table validated against GitHub Docs Supported AI models — removed 5 (GPT-4o utility-only; 4 Microsoft-internal Claude variants); added 4 (Claude Opus 4.6 fast-mode preview, Gemini 3.5 Flash GA, Raptor mini, Goldeneye). Multiplier callouts inlined (30x, 14x, 7.5x). Snapshot date refreshed 2026-05-19 → validated 2026-05-26. GPT-5.4 nano deliberately not added (Codex VS Code only per GitHub Docs footnote).
  • .github/config/README.md rewritten — ownership table shows only files that actually exist + a "Read by" column for at-a-glance consumer visibility.

Fixed

  • init-edition.cjs heir-doctor path — bootstrap step 3 referenced .github/muscles/heir-doctor.cjs (no longer exists after the muscles → scripts collapse). Corrected to .github/skills/greeting-checkin/scripts/heir-doctor.cjs. Fresh inits now actually run heir-doctor instead of silently skipping it.

Removed

Brain artifacts

  • .github/muscles/ folder — brain-executable tier collapsed into .github/scripts/ after per-skill consolidation drained it of skill-bound content. Cross-cutting executables (converter-qa.cjs, audit-mall-drift.cjs) and shared library moved to scripts/. Heir Mall-install destination migrated from .github/muscles/local/** to .github/scripts/local/**.
  • academic-paper-drafting skill — semantic review found 683 lines (over skill-review Gate 2's 500-line cap). User chose removal over trim or accept-exception (reversed the earlier "keep" decision after the size violation surfaced). Mall plugin or .github/local/skills/ for heirs that need it.
  • /migrate-from-alex-master slash-prompt — fleet completed AlexMaster → ACT Edition migration months ago. Historical tooling whose continued presence implied active migration support.
  • audit-apis workflow — .github/EXTERNAL-API-REGISTRY.md + .github/prompts/audit-apis.prompt.md + (already-removed) .github/muscles/audit-api-drift.cjs. Low-yield: registry stayed stale; drift detector flagged dates not source-of-truth changes; semantic audit fell back on model knowledge anyway. Heirs that want it can install a Mall plugin or maintain their own under .github/local/.

Heir-facing docs + dead config

  • PLUGINS.md + assets/banner-plugins.svg — Fabio-specific plugin-registration guide hardcoding paths into Microsoft-internal .github-private stores most heirs cannot clone (9 of 12 plugins). Operator content, not heir baseline. Banner was orphan after the doc deletion.
  • scripts/ folder (root) — contained only cleanup-frontmatter.cjs, a one-shot from today's frontmatter sweep (997c6d6). Violates the .github/-namespace rule (c2764df).
  • .github/config/markdown-light.css (21KB fossil) — duplicate of the canonical skill-bundled copy with divergent content. Zero functional consumers; superseded by the .vscode/markdown-light.css shipment.
  • .github/config/mcp.json.template — zero functional consumers; documented "copy to .vscode/mcp.json" workflow was never wired. Heirs that want MCP servers can author .vscode/mcp.json directly.
  • mermaid-chat.enabled: true from welcome-baseline.json — dead setting; no marketplace extension publishes it. VS Code 1.121+ ships Mermaid Markdown Features built-in with no setting required.

Architecture

  • 4-artifact-type brain (was 5) — with muscles/ gone, the brain lexicon now describes 4 artifact types (skills/instructions/prompts/agents) plus an executable tier (.github/scripts/). Reflected in copilot-instructions.md, severity-tagged-commits.instructions.md, meditation.instructions.md + skill, falsifiability-deadlines.instructions.md, skill-creator/SKILL.md routing tables, welcome.prompt.md brain-inventory line, and README.md (header counts, What Else Ships, Customization Slots, Vocabulary).
  • Architecture-table refresh in copilot-instructions.md — cluster table now accurately lists only artifacts that actually ship (was naming 7 instructions deleted in the 2026-05-18 consolidation, violating Tenet III). 11 clusters covering the actual brain shape.
  • 6 converter skills strip legacy muscle: frontmatter field (no consumer; documentation-only field that drifted).
  • creative-writing skill description rewritten from generic prose to what+when format.
  • md-to-eml SKILL.md drops "graveyard" prose section that fossilized prior design discussion.
  • md-to-txt SKILL.md fixes double-dash typo in section header.

Internal

  • Audit method: 4 parallel subagent audits (skill-review, instruction-review, prompt-review, agent-review per ADR-007) surfaced 39 Revise findings across 27 skills / 33 instructions / 28 prompts / 4 agents. All 39 resolved in 5 fix batches.
  • Tier C decisions (judgment per Supervisor precedent): act-foundations + memory-triggers Gate 6 overages ACCEPTED as framework exception; falsifiability-deadlines Edition scope KEPT (heirs DO author brain artifacts when extending baseline).
  • brain-qa Edition findings: 0 after every batch. SHA-256 byte-identity verified on all mirrored artifacts.
  • edition-manifest.json spec_version bumped 1.0 → 1.3 — manifest is now the file-level bill-of-materials. 1.1 added instructions (36), scripts (15 incl. shared/), copilot_instructions, configs (3), version_file, vscode_assets (1). 1.2 added bootstrap_templates (2: .vscode/settings.json + .github/config/cognitive-config.json) — HEIR_OWNED files Edition ships as first-install templates, sourced live from HEIR_OWNED in _registry.cjs so no drift. 1.3 added skill_files (50) — recursive file enumeration under every skill directory (SKILL.md + scripts/ + references/ + assets/ + sub-prompts). skills retained at unit granularity for heir-doctor drift detection. Additive at every step: existing consumers reading older fields work unchanged; heir-doctor.cjs verified still healthy. Strict file-by-file audit: 139 files in .github/ + .vscode/, 139 enumerated by manifest, 0 untracked, 0 phantom.

[2.3.0] - 2026-05-25

Minor [behaviour] — init UX unified across CLI, README, and Extension; /welcome rebuilt as a true orientation tour. Three previously divergent post-bootstrap paths now hand the user the same next-step checklist, and the /welcome slash-command finally matches its name. Heirs auto-update via node .github/scripts/upgrade-self.cjs; no --allow-major required.

Changed

  • /welcome rebuilt as a read-only orientation tour. Previously, /welcome applied baseline VS Code user-scope settings — useful, but mis-named and surprising on first run. The settings logic moved to a new /configure-vscode command (verbatim port, no behavior change). The new /welcome is read-only: identity summary from .act-heir.json + copilot-instructions.local.md, three always-on capabilities worth knowing, Mall discovery, three good first-prompt examples, and where to go next. No writes.
  • Post-bootstrap next-step block is now shared across all three init paths (CLI init-edition.cjs, README Quick Start, Extension toast — Extension portion lands in the next Extension release). All three render the same checklist: edit copilot-instructions.local.md first, then /welcome, then /configure-vscode, then start a real chat. Removes the previous mismatch where the CLI suggested customizing identity after /welcome but the README didn't mention it at all.
  • README Quick Start expanded. Top-of-file gains a 3-line happy-path block for skimmers; the main Quick Start section now enumerates four entry paths (CLI, /initialize, Marketplace extension, after-bootstrap checklist) instead of one. Adds an explicit pointer to fill in ## Project Context before the first chat — identity grounding from session 1 beats session 10.
  • heir-doctor.cjs warning text for missing copilot-instructions.local.md updated to point at the new /welcome for guided orientation rather than the renamed settings command.

Added

  • .github/prompts/configure-vscode.prompt.md — applies baseline VS Code user-scope settings for fleet policy compliance. Identical behavior to the prior /welcome (loads from .github/config/welcome-baseline.json, which keeps its filename for backward compatibility).
  • .github/prompts/configure-vscode-verify.prompt.md — read-only audit of user-scope VS Code settings against the central baseline. Identical behavior to the prior /welcome-verify.

Removed

  • .github/prompts/welcome-verify.prompt.md — renamed to configure-vscode-verify.prompt.md. Heirs that bookmarked /welcome-verify should switch to /configure-vscode-verify; no aliasing layer ships (Edition slash-commands are read directly by the VS Code chat surface, so deprecation aliases would require shipping a stub-prompt which is itself friction).

Migration notes for heirs

Run node .github/scripts/upgrade-self.cjs. The upgrade copies the two new configure-vscode*.prompt.md files, overwrites welcome.prompt.md with the new orientation tour content (Edition-owned, safe), and deletes the old welcome-verify.prompt.md. If you wrote local overrides of welcome.prompt.md, move them to .github/prompts/local/ before upgrading or they'll be replaced.

Upgrade command

node .github/scripts/upgrade-self.cjs

Bump: minor · Breaks: none (one command renamed, behaviour preserved) · Deprecated: /welcome-verify (use /configure-vscode-verify) · Removed: welcome-verify.prompt.md

[2.2.1] - 2026-05-25

Patch — version bump only, no artifact changes. Cuts a new Edition release so the heir-side upgrade-self.cjs flow can be exercised end-to-end against a real version delta. No skills, instructions, prompts, agents, muscles, or config schemas changed since 2.2.0. Extension/overall surface remains at v9.0.0 (unchanged).

Upgrade command

node .github/scripts/upgrade-self.cjs

Bump: patch · Breaks: none · Deprecated: none · Removed: none

[2.2.0] - 2026-05-24

Minor — adds one always-on rule (no-deferred-debt), removes 6 always-on instructions that were not earning their tokens, retires heir migration tooling, and documents VS Code 1.118+ skill picker surfacing in tool-awareness. Net effect: leaner always-on set with sharper firing patterns. Heirs auto-update via node .github/scripts/upgrade-self.cjs; no --allow-major required.

Added

  • .github/instructions/no-deferred-debt.instructions.md — always-on rule: when a turn surfaces tech debt, dead links, stale references, or outdated content, fix it in the same turn. Composes with lint-discipline (which covers files I touched) by covering debt I surfaced regardless of authorship. Lifted from Alyva_Master heir-side discipline and adopted as Supervisor + Edition shared-core per FOUR-REPOS-COMPARISON.md Tier A §0.1 row 3. Lifecycle provisional; self-falsified at 10 opportunity-turns or 2026-08-23.

Changed

  • .github/instructions/tool-awareness.instructions.md — new section documenting VS Code 1.118+ skill picker surfacing (.github/skills/<name>/SKILL.md files with non-empty description now appear in the chat slash-command picker alongside prompts). Explains the verb-prompt / noun-skill pairing is intentional and the lever (github.copilot.chat.skillTool.enabled) is user-scoped, not a brain defect. Falsifier: revise by 2026-08-24 or sooner on first observed contradiction.

Removed

  • 6 always-on instructions that were not earning their tokens — debugging.instructions.md, creative-loop.instructions.md, partnership-charter.instructions.md, alternatives-and-tradeoffs.instructions.md, scope-management.instructions.md, technical-writing.instructions.md. Removal rationale: zero or near-zero cross-references in heir-facing brain; broad applyTo patterns firing on common conversational words (debug/build/scope/option/doc) for content already covered by stronger files (critical-thinking + Two-Hypothesis Floor, act-pass Step 3 Alternatives, problem-framing-audit root-cause reframe, no-deferred-debt, lint-discipline, communication-craft, markdown-mermaid, ai-writing-avoidance). Each removal was approved file-by-file by user after audit; mirror-deleted from Supervisor as canonical source.
  • migrate-to-edition.cjs + MIGRATION.md + .github/prompts/finalize-migration.prompt.md — heir migration tooling for moving pre-Edition Alex heirs to Edition. Fleet inventory shows all heirs on Edition v1.x+ for months; the migration path is historical. init-edition.cjs remains as the path for new heirs. Cross-refs cleaned up in README.md, init-edition.cjs, .github/muscles/heir-doctor.cjs (template-detector signal simplified to require only init-edition.cjs), .github/config/edition-manifest.json, .github/prompts/checkin.prompt.md, .github/prompts/initialize.prompt.md.

Why this matters

Token economy and signal-to-noise. The 6 removed instructions averaged ~4KB each with broad applyTo patterns that fired on common project-conversation words. Their content was generic ("be hypothesis-driven", "consider alternatives", "manage scope") and covered by load-bearing rules elsewhere in the brain. Removing them sharpens what loads when. The migration tooling removal cuts ~40KB of heir-template payload that no longer corresponds to a live use case.

Upgrade command

node .github/scripts/upgrade-self.cjs

Gates verified

Gate Result
Supervisor brain-qa.cjs exit 0, 0 hard failures, 0 stale of 132 files
Edition test-edition-applyto-coverage.cjs 18/18 PASS, 0 capability gaps

[2.1.0] - 2026-05-24

Minor — two new always-on instructions for severity-tagged brain commits and falsifiability deadlines. Lifted from Karpathy_Loop's Phase 3 operational discipline (2026-05-23) and adopted as fleet-wide always-on per Supervisor brain-qa proposal 2026-05-24-02. These two rules sit underneath the existing brain-curation flow as additional gates — they don't replace the proposal-first protocol, they tighten it.

[constitutional] change: both new instructions are rules other rules now depend on. brain-curation-rules (Supervisor only) gained a Severity + Falsifiability Gate subsection that references both.

Added

  • .github/instructions/severity-tagged-commits.instructions.md — every brain-touching commit must carry one of [typo | clarification | behaviour | constitutional] in the subject line. [behaviour] requires the trimmed ACT pass; [constitutional] requires the full pass. Mixed commits get the highest tier. Self-falsified at 30 brain-touching commits or 2026-08-23, whichever fires first.
  • .github/instructions/falsifiability-deadlines.instructions.md — every new or materially edited brain artefact (instruction / skill / prompt / muscle / agent) must declare a specific falsification deadline: date, event, or count+time. "After N passes" without a date is not sufficient. Two-step lifecycle transition provisional → sinking → archived. Self-falsified at 5 new artefacts adopting deadlines or 2026-08-23, whichever fires first.

Why this matters

Severity tags restore credit-assignment fidelity that flat commit lists destroy — a typo fix and a Cardinal Rule change had identical weight in git log before. Falsifiability deadlines prevent provisional artefacts from accumulating in the brain indefinitely while their authors hope future evidence will materialise. Both disciplines proved operational in Karpathy_Loop (a live daughter heir) over the prior 24 hours; this release promotes them to the fleet baseline.

Upgrade command

node .github/scripts/upgrade-self.cjs

No --allow-major needed. The new instructions activate automatically — they apply to all commits and all new artefacts going forward, not retroactively.

Gates verified

Gate Result
Supervisor brain-qa.cjs exit 0, 0 stale of 139 files
Supervisor test-applyto-coverage.cjs 15/15 PASS, 0 capability gaps
Supervisor always-on set 17 files, 16,760 tokens (was 16 files, ~16,000 tokens — +760 net)
Edition test-edition-applyto-coverage.cjs 18/18 PASS, 0 capability gaps

References

  • Karpathy_Loop sources: local/severity-tagged-commits.instructions.md, local/add-falsifiability-deadlines.instructions.md
  • Supervisor commit: [constitutional] tag on the shipping commit
  • Tracker: FOUR-REPOS-COMPARISON.md §0.1 Tier A rows 1 + 2 → status "shipped (Edition v2.1.0)"

[2.0.5] - 2026-05-21

Patch — 25 shared-core brain files gain ## Would Revise If falsifier sections. Mirror of Supervisor D2(a) commit c6327bb. Each WRI names specific failure modes that would invalidate the file's advice — not boilerplate. Brain epistemic-qa coverage rises 45.5% → ~91% in Edition. No behavioral change for heirs: the files still direct the same actions; the WRI is an epistemic addition that names the conditions under which each rule should be revisited.

Closes the C1 falsifiability gap identified in Supervisor brain-qa 2026-05-21 findings (decision D2 option a, both phases). Satisfies the Cardinal Rule 3 quarterly CT-trifecta refinement requirement for Q2 2026 — the four CT-trifecta files (critical-thinking instruction + skill, problem-framing-audit, system-prompt-skepticism) are among the 25 with file-specific WRIs.

Added

  • ## Would Revise If section in 23 always-on instructions — act-foundations, ai-writing-avoidance, alternatives-and-tradeoffs, brain-audit, communication-craft, creative-loop, critical-thinking, emotional-intelligence, epistemic-calibration, knowledge-coverage, lint-discipline, markdown-mermaid, partnership-charter, pii-memory-filter, privacy-responsible-ai, proactive-awareness, problem-framing-audit, reliance-nudges, scope-management, session-health-monitoring, system-prompt-skepticism, tool-awareness, tool-awareness-categories
  • ## Would Revise If section in skills/critical-thinking/SKILL.md — covers the 7 disciplines + Discipline -1 frame audit
  • ## Would Revise If section in skills/markdown-mermaid/polish-mermaid-setup.prompt.md — falsifier for the workflow prompt

Verification

  • brain-qa.cjs: exit 0, 0 stale of 137 files
  • epistemic-qa.cjs (Edition): 100/100 score, warns 4 → 3 (the critical-thinking/SKILL.md mirror brought along Supervisor's D1 OVR01 reword fix)
  • test-edition-applyto-coverage.cjs: 18/18 PASS, 0 capability gaps
  • coherence-check.cjs: 0 hard, 0 soft

Upgrade

node .github/scripts/upgrade-self.cjs

No --allow-major needed. No /welcome re-run needed. WRI sections are additive content; existing heir behavior unchanged.

[2.0.4] - 2026-05-19

Patch — README Model Compatibility section gains the Copilot Language Models spec snapshot. Adds the factual model surface (context window, capability flags, in/out/cache costs) visible in VS Code 1.121's Language Models view (Settings → GitHub Copilot → Language Models). Documentation-only patch; no brain behavior change.

Added

  • README.md — Model Compatibility § "Snapshot: Copilot Language Models (2026-05-19)" — 22-row table covering every Copilot model in the VS Code 1.121 picker:
    • Context window (range: 68K → 1M)
    • Tools and Vision capability flags (universal across the lineup — not differentiators)
    • Input / output / cache cost in credits per 1M tokens (range: In 25–500, Out 200–3000, Cache 2.5–125)
    • Retirement warnings for GPT-4.1, GPT-5.2, GPT-5.2-Codex (all closing 2026-06-01)
  • One-paragraph framing below the table noting what is universal (Tools + Vision present everywhere) vs what is variable (context, in/out/cache cost), and pointing at the capability-floor benchmark (MAN.8.3) as the deferred work that will turn this spec sheet into an ACT-fit recommendation.

Notes for heirs

  • Verify against your own Language Models view before depending on these values. Model availability and pricing can change between releases.
  • Costs are credits per 1M tokens (Copilot internal accounting) — different from the premium request multiplier surface documented at docs.github.com/copilot/reference/ai-models/supported-models. Both surfaces matter; this snapshot covers the credits view.
  • The table is factual spec data, not a recommendation. The v2.0.3 architectural-needs framing in the same section is still the active recommendation. Measured ACT-discipline floor remains the open MAN.8.3 question.

Heir impact

Documentation-only. No /upgrade reconfiguration needed.

Audit trail

  • Companion to v2.0.3 README guidance (10bbe2e)
  • Supervisor README.md updated in parallel with identical table (single source of truth across both repos)
  • Source: VS Code 1.121 Language Models settings view, screenshot dated 2026-05-19
  • Cross-referenced against GitHub Docs supported-models page for retirement dates and plan availability
  • test-edition-applyto-coverage: 18/18 PASS (unaffected — no brain-file changes)
  • brain-qa: exit 0 (79 + 58 files)

[2.0.3] - 2026-05-19

Patch — README gains an honest Model Compatibility section. Adds explicit guidance for heirs about which Copilot models the brain is known to work with, what architectural needs the brain has, and what we have not measured. Closes a documentation gap; does not change any brain behavior or settings.

Added

  • README.md — new top-level section Model Compatibility between the cognitive-architecture intro and the Commands table. Contents:
    • Explicit "we have not characterised the minimum model" disclaimer citing MAN.8.3 in the Claims Registry
    • What we tested with: Claude Opus 4.7 (1M context) for v1.5.0 reasoning baseline + v2.0.0 release benchmark
    • Architectural needs: tool calling, long context (≥ 64K, ideally ≥ 128K), instruction adherence, multi-step reasoning
    • Practical recommendation: reasoning-class models (Claude Sonnet 4+, Claude Opus, GPT-4.1, GPT-4o or equivalent) for primary agent work; gpt-4o-mini reserved for the chat.utilityModel / chat.utilitySmallModel slots per v2.0.2 baseline (NOT for primary agent work)
    • Open question: call for heir feedback with reports of "this worked on X" / "this failed on Y" routed to AI-Memory/feedback/alex-act/

Heir impact

Documentation-only. No behavior change. Heirs on v2.0.2 reading the new section will learn what model class is recommended; no /upgrade reconfiguration is needed beyond pulling the new README.

What this is not

This release does not establish a measured minimum model. The MAN.8.3 claim remains open. The architectural-needs framing is the honest current state. A planned capability-floor benchmark (tracked in Supervisor HANDOFF.md outstanding item #8) will replace this guidance with measured floor on a future Edition release.

Audit trail

  • Companion to v2.0.2 baseline (178cb76) and v2.0.1 brain rules (f9aaffd)
  • Supervisor README updated in parallel with curator-facing framing
  • Provenance: user directed 2026-05-19 evening "A and then another day we refine it with B" — this is Option A (ship honest architectural-needs guidance now); Option B (run capability-floor benchmark to close MAN.8.3) deferred to a future session
  • test-edition-applyto-coverage: 18/18 PASS (unaffected — no applyTo or brain-file changes)

[2.0.2] - 2026-05-19

Patch — welcome baseline gains three VS Code 1.120/1.121 settings. Three keys added to .github/config/welcome-baseline.json so all heirs get them on next /welcome (or fresh setup). Companion to v2.0.1 (which updated the brain rules describing these capabilities); v2.0.2 wires the matching settings into the heir bootstrap.

Changed

  • .github/config/welcome-baseline.json — added three keys to the settings object:
    • chat.tools.compressOutput.enabled: true (1.120 Preview) — enables terminal output compression for git diff / ls -l / npm install and the 1.121 expansion (test runners, build tools, linters, Docker, package managers). The brain's file-redirect fallback remains valid for cases where compression strips data the agent needs.
    • chat.utilityModel: "gpt-4o-mini" (1.121) — routes title generation, rename suggestions, and settings search to a smaller cheap model. Heirs can override locally if gpt-4o-mini isn't in their model surface (VS Code falls back to default).
    • chat.utilitySmallModel: "gpt-4o-mini" (1.121) — same rationale for the small-model slot.
  • Baseline $comment updated to acknowledge that preview/experimental toggles can be included when explicitly requested by user policy and noted in CHANGELOG (was: "Stable settings only — keep preview/experimental toggles off unless explicitly requested"). chat.tools.compressOutput.enabled is the first such inclusion.

Heir impact

Heirs running /welcome (first-session bootstrap or new-machine setup) get all three settings. Heirs running /welcome-verify will see the three keys flagged as missing until they re-run /welcome. Existing user-level overrides for these keys are preserved by the merge step in /welcome (it merges, doesn't overwrite values that already differ — well, actually it does overwrite to match baseline; heirs who want a different utility model should set it AFTER /welcome). No --allow-major needed; standard /upgrade covers the baseline file.

Override guidance

Heirs who want a different utility model (e.g. running on BYOK with a different small-model name) should set their override in personal settings.json AFTER running /welcome. The /welcome merge is overwrite-to-baseline, so the override needs to be re-applied if /welcome runs again.

Validation

Dogfooded the /welcome reference command (verbatim from welcome.prompt.md) against the curator's personal settings.json before shipping — all three keys landed correctly as JSON booleans/strings (lowercase true, quoted model names). No existing keys disturbed.

Audit trail

  • Companion to v2.0.1 (brain rules) — commits b6dafc3 (Supervisor) + f9aaffd (Edition)
  • Proposal: original vscode-1.120-1.121-adoption-2026-05-19.md recommended these as personal-settings-only; user directed (2026-05-19) to bake into the heir baseline instead
  • Brain-qa: exit 0 across 79 Supervisor + 58 Edition files (no brain file changes in v2.0.2)
  • test-edition-applyto-coverage: 18/18 PASS, 0 capability gaps (no applyTo changes)

[2.0.1] - 2026-05-19

Patch — VS Code 1.120/1.121 feature adoption. Three brain files updated to reflect VS Code capabilities that shipped between 2026-05-13 and 2026-05-19. Mirrored byte-for-byte from Supervisor per the shared-core direction-of-edit rule. Zero behavior change for heirs — additive informational text + one factual correction + one extension-recommendation update.

Changed

  • terminal-command-safety.instructions.md — documents two new VS Code mechanisms that work alongside the existing rules:
    • NEW section VS Code 1.120 + 1.121 Terminal Output Compression (Preview) names chat.tools.compressOutput.enabled and the 1.121 expansion to pytest / jest / cargo test / tsc / cargo build / make / linters / Docker / package managers, plus auto-dispose of background terminals.
    • Terminal Hanging rule #1 now notes that VS Code 1.121+ auto-promotes sync→background after a configurable idle-silence period; the agent-intent rule remains correct and is still required on older builds.
    • Falsifier — Backtick Hazard watermark bumped from "through 1.118" to "through 1.121" with note that 1.120/1.121 ship no fix for microsoft/vscode#295620. The temp-file pattern remains mandatory.
  • session-health-monitoring.instructions.md — Proxy Heuristics opener corrected: VS Code 1.120 made BYOK token counts visible in the Chat-view context-window control. Opener now distinguishes BYOK (ground truth available) from non-BYOK / older builds (proxy heuristics still apply). Table below unchanged.
  • markdown-mermaid/SKILL.md — VS Code Extension Setup updated: VS Code 1.121 ships built-in Mermaid rendering in Markdown previews per microsoft/vscode#293028. Recommendation list keeps mermaidchart (chart authoring), vstirbu (standalone preview tab), and non-Mermaid tools (PlantUML, Graphviz, D2). bierner.markdown-mermaid removed — the built-in renderer covers its use case.

Out of scope (deliberate)

Three 1.120/1.121 features documented in the proposal but not adopted pending field data: chat.tools.riskAssessment.enabled (overlaps act-pass severity), Claude auto-permission mode (overlaps act-pass), and workspace-level forcing of chat.tools.compressOutput.enabled (still preview).

Heir impact

None for the contract. Heirs on v2.0.0 reading the updated rules gain awareness of upstream-handled mechanisms; the rules themselves continue to fire correctly. No --allow-major needed; standard /upgrade covers it.

Proposal + audit trail

  • Proposal: Alex_ACT_Supervisor/docs/proposals/vscode-1.120-1.121-adoption-2026-05-19.md
  • Supervisor commit: b6dafc3 (origin/main)
  • Brain-qa: exit 0 across 79 Supervisor + 58 Edition files
  • test-applyto-coverage: 15/15 PASS, 0 capability gaps
  • test-edition-applyto-coverage: 18/18 PASS, 0 capability gaps

[2.0.0] - 2026-05-19

Major — reasoning-quality release. Same brain shape (36 instructions, 18 skills, 23 prompts, 16 muscles, 4 agents), same heir API surface, same /upgrade mechanism. Behavior changes are improvements to always-on reasoning disciplines that close measured coverage gaps in benchmark scenarios while reducing total credits-per-solved-problem. Major version bump signals that heirs upgrade via --allow-major and acknowledges that v2 reasoning IS measurably different from v1.5.0 (sharper verify-before-report, frame audits on explain frames, output-discipline gates).

Validated by:

  • Compose verification benchmark (5 scenarios): 13/15 → 15/15 composite, -22.5% credits (228.5 → 177.0)
  • S360 real-world adoption: heir adopted on main 2026-05-19 and self-promoted .act-heir.json from 2.0.0-candidate → 2.0.0 after multi-commit follow-through validation
  • Tenet X demonstration in S360: v2 brain refused a stale templated instruction from Supervisor (exactly the failure mode the new rules were designed to catch)
  • Terminal-safety fix empirical validation: 3 post-fix commits in S360 with $env:TEMP pattern, zero .commit-msg.tmp leaks (verified via git show --stat)

Breaking

  • None for the heir contract. File inventory unchanged. Heir-side .act-heir.json, scripts, local/ skills, HANDOFF.md, episodic/, workflows/ all preserved on upgrade.
  • Behavior changes are intentional and visible: heirs will notice more verify-before-report patterns firing on search/summary work, more explicit frame-audit markers on "explain X" / "tell me how Y works" prompts, more by-name citations during disagreement-mode refusals. Net effect per benchmark + S360: better outcomes at lower cost, but takes a session to feel natural.
  • --allow-major required on node .github/scripts/upgrade-self.cjs for heirs upgrading from any v1.x to v2.0.0.

Added

  • epistemic-calibration.instructions.md Output-discipline subsection (Phase 3.3) — Anti-Hallucination Signals table split into Input-discipline (existing 5 rows: claims about generation) and Output-discipline (3 new rows: claims about reporting):
    • "No matches found" / "Verified clean" / "Nothing returned" → verify search scope before reporting absence; cite paths/globs/file-count
    • "The doc says X" / "Per README" / "According to spec" → cross-check doc against filesystem; cite both
    • "I checked and..." / "Verified that..." → name what was actually checked; unattributed verification is theatre
    • Plus 4th Core Principle: "A search that didn't run looks identical to a search that found nothing — verify the scope before reporting absence."
  • problem-framing-audit.instructions.md Explain/Summarize Frame subsection (Phase 5 Option C) — complementary discipline for summarization patterns the Output-discipline literal triggers don't catch:
    • Literal trigger phrases: "Explain X", "Tell me how Y works", "Describe Z", "Summarize <doc>", "Read <file> and...", "Walk me through...", "What does <doc> say about..."
    • Required action: name source file(s) read + cross-check ≥1 structural claim against filesystem; if doc and filesystem disagree, surface the gap and report both
    • Visible marker: **Verified against**: <doc path> + <filesystem check>
  • terminal-command-safety.instructions.md temp-file location guidance — closes a heir-reported defect (the git commit -F tmpfile + git add -A interaction silently committed temp message files into commits):
    • Warning paragraph: place temp files outside the working tree ($env:TEMP\<slug>.txt on Windows, /tmp/<slug>.txt on Unix) OR add the pattern to .gitignore before staging
    • Preferred PowerShell template using Join-Path $env:TEMP + Set-Content -NoNewline + git commit -F + Remove-Item

Changed

  • 3 instruction files updated (see Added above for the substantive changes):
    • .github/instructions/epistemic-calibration.instructions.md
    • .github/instructions/problem-framing-audit.instructions.md
    • .github/instructions/terminal-command-safety.instructions.md
  • .github/VERSION bumped 1.5.0 → 2.0.0.
  • File inventory unchanged at 36 instructions / 18 skills / 23 prompts / 16 muscles / 4 agents.
  • Always-on token growth: ~+908 tokens/session (~+3632 bytes across the 3 instruction files). Trade is net-positive per benchmark: -22.5% credits across the 5-scenario Compose set; breakeven at ~1 avoided corrective turn per 10 sessions; observed rate in benchmark + S360 is 2-3+ per 10.

Scope correction

A pre-flight diff during release prep reported 46 files differing between Edition v1.5.0 and the v2 candidate workspace. On verification (parent-commit checkout + spot-check) the 43 "accumulated v2-candidate dev" files were already byte-identical to Edition v1.5.0; the apparent difference was line-ending normalization (CRLF in Edition vs LF in the v2 candidate scratch workspace) that git smooths over at commit time. The actual v2.0.0 release scope is 3 instruction files (Phase 3.3 + Phase 5 + terminal-safety) plus VERSION + CHANGELOG. This entry was corrected before push.

Upgrade

node .github/scripts/upgrade-self.cjs --allow-major

Heirs preserve all of: skills/local/, .act-heir.json, HANDOFF.md, episodic/, workflows/, scripts/ (heir-specific), docs/, all non-.github/ content. Heir-doctor may surface cosmetic warnings on first run if the heir's edition-manifest.json is stale; they clear after upgrade completes.

Why

The v1 line optimized brain size; v2 optimizes brain outcomes. The Phase 1 baseline benchmark surfaced one real coverage gap (output-verification — verify before reporting search/doc claims, scored 2/3 not 3/3 in S4 and S7). Phase 2 dual audit of all 17 always-on rules produced 1 Grow / 8 Compose / 8 Unchanged / 0 Shrink / 0 Restructure — confirming the brain isn't bloated, it's a tightly composed system where every rule earns its cost. Phase 3 applied the Grow (epistemic-calibration Output-discipline). Phase 5 added the complementary Explain/Summarize frame for surface patterns the Phase 3 literal triggers miss. Terminal-safety added the temp-file location guidance to close a heir-reported defect that hit S360 twice.

The release reaches Edition because S360 adopted v2 candidate in real-world product work on 2026-05-19 and demonstrated all the predicted improvements PLUS a real Tenet X self-correction moment (v2 brain refused a stale Supervisor instruction). Real-world signal outweighs synthetic benchmark for ship/no-ship decisions.

Falsifiability

This release is wrong if any of the following occur within 14 days:

  • ≥2 heirs report behavior regressions traceable to Phase 3.3 or Phase 5 changes (triggers partial rollback or v2.0.1 fix-forward)
  • The terminal-safety fix doesn't prevent a .commit-msg.tmp-class leak in a heir that upgrades to v2.0.0
  • S360 reverts its .act-heir.json marker from 2.0.0 back to 2.0.0-candidate or below (the bellwether heir)
  • A fleet-cost regression appears that wasn't visible in the 5-scenario benchmark (triggers the Phase 6.2 light re-baseline that was deferred)

If any fire: cut v2.0.1 with fix; document in Supervisor's brain-qa-changelog.md tagged [V2-REGRESSION].

References

[1.5.0] - 2026-05-18

Minor — converter-qa harness restoration + complete-coverage tests for all converters (PNG + SVG image handling verified end-to-end).

Added

  • converter-qa suites for previously uncovered converters:
    • md-to-html.cjs: end-to-end + image handling — structural HTML (DOCTYPE, CSS, headings, lists, tables, links, blockquotes, code), PNG embedding (data URI or referenced), SVG handling (inline / data URI / <img> ref)
    • md-to-txt.cjs: strip formatting + preserve image alt text — formatting markers removed (**, *, backticks, #, |), content + image alt text preserved
    • html-to-md.cjs: structure + image preservation — headings (ATX or setext), bold/italic, bullet + numbered lists, tables (pipe / simple / grid form), links, blockquotes, code, PNG + SVG image refs preserved with alt text
    • docx-to-md.cjs: round-trip + image extraction — md → docx → md via md-to-word leg + docx-to-md leg; headings, lists, tables, and image extraction (or inline ref preservation) verified
  • md-to-word.cjs: [toc] marker warn-and-ignore suite — confirms v1.4.0 behavior (warning logged when [toc] marker found without --toc, marker stripped from body, output docx has no TOC field)
  • Font + margin value assertions added to md-to-word.cjs: table styling regression suite — header w:sz="18" (9pt), data w:sz="17" (8.5pt), cell margins w:w="20" (1pt T/B) + w:w="60" (3pt L/R). Catches the v1.4.0 numeric changes that previously flowed through silently
  • createImageFixtures helper in converter-qa — generates a minimal 1x1 PNG (67 bytes, embedded as binary literal) + a labeled SVG fixture for use across the new converter suites
  • SVG image section added to docs/testing/md-to-word-coverage.md regression corpus

Removed

  • 4 dead test suites referencing modules that no longer exist:
    • Shared: replicate-core.cjs (3 suites: base, batch retry & validation, negative-prompt & prompt-file)
    • Shared: svg-pipeline.cjs
    • These were pre-Edition artifacts from a Replicate AI image-gen + SVG pipeline that has been out of the brain for some time. The harness FATALed on the first missing module, blocking the entire test run
  • 2 obsolete callout assertions in Shared: markdown-preprocessor.cjs suite — ::: tip and > [!WARNING] callout transformation tests. The preprocessor never implemented these (always returned input unchanged), and the syntaxes aren't part of any active Edition workflow. If callout rendering is needed in the future, separate feature request
  • 2 file-inventory checks for shared/replicate-core.cjs and visual-memory.json (also no longer exist)

Changed

  • converter-qa.cjs internal version 1.2.0 → 1.3.0; JSDoc updated to reflect new suite list and assertion count (284 → 256 after dead-suite removal + new-suite additions; net: cleaner)

Why

User request: "make sure all other converters do a complete job" with explicit "mds can contain svg and png images, make sure they are supported by the converters." Before this release, only md-to-word and md-to-eml had dedicated suites; the other 4 converters had zero coverage, and the harness FATALed at startup so even existing suites couldn't run end-to-end. Result: 256 PASS, 0 FAIL, 4 SKIP (skips are pandoc-availability checks when pandoc is missing in CI, plus adm-zip on systems without it).

Falsifiability

If a converter regresses on PNG or SVG handling in a future change, the new suites will catch it. If callout syntax becomes a real requirement, add the feature to markdown-preprocessor.cjs and re-add the assertions (or new ones) — the prior version is in git history.

[1.4.0] - 2026-05-18

Minor — md-to-word table tightening + [toc] marker honors documented default.

Changed

  • muscles/md-to-word.cjs table styling (internal muscle version 5.4.0 → 5.5.0):
    • Header font: 10pt → 9pt (w:sz 20 → 18) — still bold white on Microsoft blue
    • Data cell font: 9pt → 8.5pt (w:sz 18 → 17)
    • Cell margins: T/B 40 twips (2pt) → 20 twips (1pt), L/R 80 twips (4pt) → 60 twips (3pt)
    • Visible effect: denser, more reference-document-style tables; same colors, borders, and zebra striping
  • muscles/md-to-word.cjs [toc] marker semantic: previously, a [toc] line in the source silently set args.toc = true. Now the marker line is still stripped from the source, but TOC is not auto-enabled. A warning is logged so the heir can either pass --toc explicitly or remove the marker. Aligns the skill with its documented default (--toc | off).
    • Version-management note: classified as minor rather than major because the [toc] auto-detect was never in the skill's documented Options table — it was a side-effect of the preprocessor. Removing an undocumented side-effect to honor the documented default is a quality fix, not a breaking change to the documented contract. Heirs whose source files use [toc] AND who don't pass --toc will see the warning in their next conversion run and can adjust trivially. If a heir reports regression within 14 days, fallback is v1.4.1 reverting just the [toc] semantic.
  • skills/md-to-word/SKILL.md — Options table notes new [toc] marker behavior; Table Formatting section reflects 9pt/8.5pt fonts and 1pt/3pt padding; Version History gains v5.5.0 row; currency + lastReviewed stamped 2026-05-18.

Added

  • docs/testing/md-to-word-coverage.md — regression corpus exercising every markdown feature the skill claims to support (H1-H6, inline formatting, all list types, tables with alignment and pagination, code blocks in multiple languages, blockquotes including nested, horizontal rules, footnotes, PNG image refs, Mermaid flowchart / sequence / state / class diagrams). Includes a 21-item verification checklist for post-conversion spot-check, and a TOC-marker behavior test procedure.

[1.3.4] - 2026-05-18

Patch — heir discoverability of the AI-Memory formal contract.

Changed

  • skills/ai-memory-setup/SKILL.md — added "Formal contract" subsection pointing heirs at AI-Memory/SCHEMA.md (the Supervisor-maintained contract document covering subfolder ownership, frontmatter, and lifecycle rules). Folder Structure tree updated to show SCHEMA.md at root.
  • skills/ai-memory-setup/SKILL.md § Write Feedback — aligned filename format and frontmatter description to the canonical shape in SCHEMA.md (filename now YYYY-MM-DD-<heir-id>-<short-slug>.md; frontmatter keys date, heir_id, severity, category — lowercase, matches what Supervisor expects when triaging).

No behavior change. Purely heir-facing documentation alignment so the deployed spec and the Supervisor-side contract agree.

[1.3.3] - 2026-05-18

Final cleanup of the handoff-tier convention introduced in v1.2.2. Reading side now agrees with the writing side: prior-session context comes from repo-root HANDOFF.md first, session memory only as legacy fallback.

Fixed

  • instructions/proactive-awareness.instructions.md Cross-Session Context Recovery (PA1): Step 1 now checks repo-root HANDOFF.md first (the canonical cross-session handoff per memory-triggers.instructions.md). Session memory becomes Step 2, labeled as "legacy/secondary signal" with an explicit note that any handoff content there predates the v1.2.2 tier convention. Surface-context table updated to recognize HANDOFF.md present with recent content as the primary trigger.

Why

v1.2.2 + v1.3.1 + v1.3.2 fixed the write side (where handoff content goes). The read side (proactive-awareness Step 1) was still pointing at /memories/session/ as the primary source for prior-session context. Symmetry now holds: write to HANDOFF.md, read from HANDOFF.md.

Verification

  • brain-qa exit 0 (58 Edition files)
  • test-edition-applyto-coverage 18/18 PASS, 0 gaps

[1.3.2] - 2026-05-18

Internal consistency fix for the handoff-tier convention shipped in v1.2.2 + v1.3.1. Two files still routed handoff content to ephemeral session memory — the exact pattern s360 originally flagged.

Fixed

  • skills/meditation/SKILL.md Extract routing table: replaced Session continuity for next chat → /memories/session/<name>.md with Cross-session handoff (next session needs to know) → repo file (HANDOFF.md at repo root) — NOT session memory. The skill's Extract step now agrees with its Handoff step (Step 5) and with memory-triggers.instructions.md / save-session-note.prompt.md.
  • skills/meditation/SKILL.md Step 5 (Handoff): renamed SESSION-HANDOFF.md → HANDOFF.md. v1.3.1's rename had missed this SKILL.md file (caught in heir-aware spot-check).
  • instructions/session-health-monitoring.instructions.md Graceful Handoff section: was telling the agent to write state + completed work + next steps + pending decisions to /memories/session/[name].md (which clears at conversation end). Now routes to repo-root HANDOFF.md with an explicit "session memory is for in-conversation scratch only" clarifier.

Verification

  • brain-qa exit 0 (58 Edition files)
  • test-edition-applyto-coverage 18/18 PASS, 0 gaps
  • Zero write /memories/session/... handoff misroutings remaining in the live brain tree (verified via grep)

[1.3.1] - 2026-05-18

Naming clarity + zero-behavior-change refactor. Per proposal docs/proposals/prompt-overlap-audit-2026-05-18.md (Supervisor side).

Changed

  • Session-handoff artifact unified on HANDOFF.md (was SESSION-HANDOFF.md). Per user instruction: "saved in root... always use root, like we did here". Updated 4 files in lockstep:

    • instructions/meditation.instructions.md
    • prompts/meditate.prompt.md
    • prompts/note.prompt.md
    • prompts/save-session-note.prompt.md

    This unifies with the convention memory-triggers.instructions.md ships in v1.2.2 (cross-session continuity → repo file HANDOFF.md, not session memory). The two filenames were a naming inconsistency introduced earlier today; v1.3.1 fixes it.

    save-session-note.prompt.md includes a legacy-migration note: if a heir still has SESSION-HANDOFF.md at root from before this rename, the prompt mentions it during confirm — the heir manually reviews and either merges into HANDOFF.md or deletes the legacy file. Never silently discards content.

  • welcome baseline extracted to config — the user-scope VS Code settings payload that /welcome applies and /welcome-verify audits previously lived duplicated in both prompt files. Future drift trap removed by extracting to .github/config/welcome-baseline.json as the single source of truth. Both prompts now load from there. Updated sync-policy.json to mark the new config file as edition_owned (overwritten on upgrade).

Added

  • .github/config/welcome-baseline.json — the VS Code user settings baseline applied by /welcome and audited by /welcome-verify. Edit once, both prompts pick it up.

[1.3.0] - 2026-05-18

Tier 3 token rationalization: port of verified Supervisor trims plus load-bearing Tenet X discipline added to the always-on ACT pass. Always-on body tokens: 12,467 → 11,863 (-604, -4.8%). Zero capability regression — verified by scripts/test-edition-applyto-coverage.cjs (18/18 scenarios pass). Per proposal docs/proposals/edition-optimization-2026-05-18.md (Supervisor side, Batch 2).

Added

  • act-pass.instructions.md § Self-Application (Tenet X always-on hook): 6-row pattern/signal/correction table (reasoning theatre, hedge laundering, authority deference, symmetric balance, adversarial-probe skip, self-flattering meta-cognition) so the always-on pass enforces the Tenet X discipline. Previously this discipline lived only in act-foundations (which is now scoped on the Supervisor side; will be evaluated for Edition in a later batch).
  • tool-awareness-categories.instructions.md (NEW, scoped): Common deferred tool categories table moved out of tool-awareness.instructions.md to a scoped sibling that loads only on tool/MCP/GitHub/browser/notebook work. Always-on file shrinks; reference table still available where needed.

Changed

  • communication-craft.instructions.md trimmed (989 → 660 tokens): dropped Explaining Concepts §2, Tone Anti-Patterns, Integration table; kept SBI + stakes + voice + audience lead + needs/solutions tables. LLM-inherited communication behaviors no longer need always-on reinforcement.
  • session-health-monitoring.instructions.md trimmed (774 → 398 tokens): dropped High-Token-Cost Operations table (duplicates tool-awareness), Session Memory Template (operational on-demand), Integration section; kept proxy heuristics + warning signs + checkpoints + handoff.
  • emotional-intelligence.instructions.md trimmed (658 → 468 tokens): collapsed Adaptation Rules verbose sections to single 4-row table; kept signal detection table + mimicry prevention as one paragraph.
  • knowledge-coverage.instructions.md trimmed (425 → 244 tokens): compressed KS2 pre-response assessment + KS3 visible-badge logic to taxonomy table + brief rule.
  • tool-awareness.instructions.md trimmed (499 → 263 tokens): moved Common Deferred Tool Categories table to new scoped tool-awareness-categories.instructions.md; kept core Rules + External Ingest.

Falsifiability watch (2 weeks)

If reasoning quality degrades within 2 weeks (sycophancy returns, alternatives missed, instructions ignored), the most likely culprit is the act-pass Tenet X addition or the tool-awareness split. Revert in that order. Falsifier closes 2026-06-01.

[1.2.2] - 2026-05-18

Fixed

  • memory-triggers.instructions.md: Clarified that session-handoff documents live in repo HANDOFF.md, not in /memories/session/ (which is cleared at conversation end). Per heir feedback (s360, 2026-05-09 handoff-tier-confusion): the natural phrase "session handoff" reads like exactly what session memory is for, but session memory is by-design ephemeral. The Memory Tier Selection table now distinguishes "Cross-session handoff (next session needs to know)" from "In-conversation scratch (current session only)". New § "Cross-Session Continuity" makes the rule explicit. Trigger Conditions table now points the "Session > 30 min OR end-of-session" trigger at the repo file rather than a generic "Handoff" prompt.

[1.2.1] - 2026-05-13

Added

  • Brain audit trifecta: Added brain-audit.instructions.md, skills/brain-audit/SKILL.md, and /audit-brain workflow support.
  • Dedicated audit worker: Added agents/brain-auditor.agent.md for deterministic local Edition audits.
  • Mall refresh workflow: Added drift-aware /mall refresh support and docs wiring for curated-subset states.

Fixed

  • Audit robustness: Hardened Edition audit flows for template context and npm probe handling.

Changed

  • Propagated Supervisor-side audit remediations into Edition artifacts.

[1.2.0] - 2026-05-05

Added

  • Mall contribute prompt: New /mall contribute command for submitting plugins to the Plugin Mall.
  • Tool awareness instruction: New tool-awareness.instructions.md documenting deferred tools and external ingest.

Fixed

  • Instruction count: 33 → 34 across documentation (README, copilot-instructions).
  • converter-qa stale currency tag: Updated to current date.

Removed

  • /fleet prompt: Moved to Supervisor scope (not heir-relevant).

[1.1.0] - 2026-05-05

Added

  • Tool awareness instruction: New tool-awareness.instructions.md documents deferred tools (require tool_search before use) and external ingest for remote/virtual workspaces (VS Code 1.118/1.119).
  • VS Code 1.118 agentic execution sub-tool note in terminal-command-safety.instructions.md: documents output pre-filtering behavior and when redirect-to-file fallback is still needed.
  • AI-Memory knowledge index: Heirs now pointed to AI-Memory/knowledge/index.json for on-demand reference material.
  • Extension scaffold relocated: VS Code extension moved to independent repo Alex_ACT_Extension.

Fixed

  • 23 epistemic-integrity findings resolved across 19 brain files (ABS01x4, OVR01x1, REV01x18): added revision conditions, qualified absolute claims, and sourced authority claims. Epistemic integrity score: 91 to 100/100.
  • Stale references: /find-skill to /mall search, /install-from-mall to /mall install, corrected skill/instruction counts.
  • Cross-platform improvements: Path resolution and file handling.

Changed

  • Removed .github-v0 pre-refactor brain backup (no longer needed post-v1.0.0).
  • Project docs (PLAN, decisions/) relocated to Alex_ACT_Supervisor.

[1.0.0] - 2026-05-02

The v1 brain refactor: a complete restructure of the cognitive architecture for clarity, token efficiency, and maintainability. Every instruction, skill, prompt, and muscle was re-evaluated, clustered, trimmed, or consolidated.

Breaking

  • Architecture table replaced: copilot-instructions.md now describes 8 functional clusters (was 7 generic domains). Heir copilot-instructions.local.md that reference old domain names ("Reasoning", "Learning", "Growth") should update.
  • 6 converter instructions consolidated into 1: docx-to-md, html-to-md, md-to-html, md-to-txt, md-to-word, md-to-eml instructions replaced by single converter.instructions.md. Format-specific logic stays in skills. Heirs with custom converter references should update to /convert.
  • plugin-store-routing removed: Absorbed into mall-installation.instructions.md. Heirs referencing it by name should update.
  • 5 instructions demoted from always-on to conditional: alternatives-and-tradeoffs, agent-delegation, partnership-charter, worldview, creative-loop. They still fire on relevant file patterns and conversational context, but no longer consume always-on token budget.
  • upgrade-self.cjs major-version path now backs up and recreates: Major bumps trigger backup + fresh install + recovery of heir-owned content (was in-place overwrite). Requires --allow-major.
  • .github/episodic/** is now heir-owned: Was edition-owned (silently wiped on upgrade). Meditations, post-mortems, and calibration logs are preserved.

Added

  • AI-Memory setup: new standard skill (ai-memory-setup) with 8-provider cloud drive discovery (OneDrive, iCloud, Dropbox, Google Drive, Box, MEGA, pCloud, Nextcloud), auto-create, CLI (_registry.cjs --discover/--init/--resolve), and cognitive-config.json persistence (ai_memory_root, ai_memory_exclude).
  • bootstrap-heir.cjs --ai-memory flag: Explicit cloud drive selection during bootstrap. Auto-creates AI-Memory and persists the choice.
  • ACT.md onboarding note: Generated on bootstrap with project-aware plugin recommendations based on detected tech stack.
  • Heir-added artifact relocation: upgrade-self.cjs detects artifacts heirs placed in edition-owned paths and relocates them to local/ automatically (incremental and major paths).
  • Deprecated file cleanup: upgrade-self.cjs removes files that Edition no longer ships.
  • Pass 3.5 (Episodic Memory) in finalize-migration.prompt.md: Guides heirs to restore episodic files during migration.
  • Plugin Mall v2 integration: /mall search and /mall install prompts, plugin.json manifests, shape/engines/token_cost metadata, CATALOG.json v2.1.
  • Worker subagents: markdown-author, illustrator, document-assembler for delegated mechanical work.
  • edition-manifest.json: Machine-readable inventory of shipped skills, prompts, agents. Used by heir-doctor.cjs (replaces stale hardcoded allowlists).
  • 17 skills (was 11), 20 prompts (was 19), 3 agents (new), 20 muscles.

Fixed

  • Mall repo name: all gh api and git clone URLs updated from Alex_ACT_Plugin_Mall to Alex_Skill_Mall (the actual GitHub repo name).
  • Episodic wipe on upgrade: moved .github/episodic/** from edition_owned to heir_owned in sync-policy.json.
  • Episodic drop on migration: removed episodic/ from EXTENSION_ONLY patterns in migrate-to-edition.cjs.
  • heir-doctor.cjs false positives: now reads edition-manifest.json instead of hardcoded allowlists.
  • AI-Memory path resolution: standardized across all artifacts (was 6 different hardcoded candidate lists). Now all flow through cognitive-config.json override + auto-discovery.
  • Windows reparse points: cloud drive folders with ReparsePoint attribute (common for OneDrive) now detected by _registry.cjs.

Changed

  • Always-on token budget: 25,835 (v0.9.1) to 13,886 (v1.0.0). 46% reduction.
  • Context-loaded artifacts: 79 (v0.9.1) to 73 (v1.0.0). 10 converters consolidated, 1 absorbed.
  • Instructions: 37 (v0.9.1) to 33 (v1.0.0). DRY pass removed redundancy.
  • copilot-instructions.md Architecture table: 7 generic domains replaced with 8 functional clusters matching actual artifact organization.
  • README.md: Updated all artifact counts, removed stale references.

[0.9.9] - 2026-05-02

Phase 0-9b of the v1 brain refactor. All 59 capabilities migrated and verified.

[0.9.1] - 2026-04-30

Fleet pull-based architecture, heir self-update, Mall v2 design.

[0.9.0] - 2026-04-30

Edition brain reset from AlexMaster. Clean baseline for v1 refactor.

[0.7.0] - 2026-04-28

Initial Mall integration, converter improvements, mermaid fidelity.

[0.6.2] - 2026-04-29

Mermaid viewport fix, ZWSP checkbox fix, banner cross-reference.

Assets 2
Loading