Develop - #100
Merged
Merged
Conversation
Add LedgerCell, CapabilityLedger, LEVEL_RANK, CAPABILITY_FEATURES, and CAPABILITY_SCOPES to src/core/capabilities/ledger-types.ts. Suppress knip unused-export warnings for this file via ignoreIssues since consumers arrive in subsequent tasks.
warp/mcp/global, copilot/mcp/global, gemini-cli/hooks (proj+global) confirmed raise-candidates; windsurf/mcp/global confirmed native; windsurf project, copilot/hooks, kiro/hooks rejected with evidence.
…ns to native - agents (project/global): emit required groups + non-empty roleDefinition in .roomodes / custom_modes.yaml (Roo's modeConfigSchema has no default for either, so omitting them silently dropped ALL modes on load) - agents (global): fix custom_modes.yaml path to ~/.roo/settings/... (was missing the .roo/ prefix) - rules (global): stop redirecting the root rule to ~/.roo/AGENTS.md (never read by Roo Code); keep it at ~/.roo/rules/00-root.md like loadRuleFiles() actually reads, with the old AGENTS.md path kept as a legacy import fallback - mcp (global): downgrade native -> partial; ~/mcp_settings.json is not the real read path (McpHub resolves via non-deterministic globalStorageUri) - ignore (global): downgrade native -> none; RooIgnoreController only reads workspace-relative .rooignore, no home-directory concept exists - permissions (project): raise partial -> native; write/merge/import roo-cline.allowedCommands/deniedCommands via .vscode/settings.json Split index.ts into layout.ts/merge.ts to stay under the 200-line limit. Core: generatePermissionsFeature now passes scope context to generators. Recorded ledger provenance for all 6 changed cells and synced README/docs.
- permissions (project+global): none -> native via .codex/rules/*.rules using the same Starlark prefix_rule(pattern, decision, justification) DSL Codex already documents for command-execution permission decisions - hooks (project+global): drop canonical Notification hooks (no matching Codex lifecycle event exists) instead of silently emitting a dead hooks.Notification key; add a lint warning - additionalRules (project): write real nested <dir>/AGENTS.md / AGENTS.override.md files for scoped rules (Codex's actual directory-walk instruction hierarchy) instead of the unread .codex/instructions/ mirror plus a link-only root index entry - mcp: support remote/Streamable HTTP servers (url/bearer_token_env_var/ http_headers) on both generate and import; the importer previously hardcoded type:'stdio' and dropped url-based servers on import Recorded ledger provenance for all changed cells and synced README/docs.
… commands
- MCP (global): none -> native. ~/.copilot/mcp-config.json (mcpServers key),
wired via globalSupport.scopeExtras with generator + importer round-trip.
Corrects a prior ledger cell that pointed at the VS Code extension's
OS-specific user-profile mcp.json instead of Copilot CLI's own path.
- Hooks (global): none -> native. ~/.copilot/hooks/agentsmesh.json + wrapper
scripts, same {version, hooks} schema as project scope, wired via
scopeExtras so the project-shaped generator never leaks into global paths.
- Permissions (global): none -> partial. ~/.copilot/permissions-config.json
only records saved approvals (no deny/ask/default-mode/URL/tool-filter/
repo-local policy support), so this is a lint warning, not a generator.
- Commands (global): native -> none. Copilot CLI has no prompt-file/
slash-command mechanism at all (github/copilot-cli#618, maintainer,
closed); the phantom .copilot/prompts/ path is no longer generated or
imported. Project-scope commands (.github/prompts/, a real VS Code
Copilot Chat surface) are unaffected.
- Hooks (project, broken-fix): generateHooks now emits the real top-level
matcher regex field instead of a non-functional comment field, so hooks
actually filter by tool. The canonical '*' wildcard sentinel is omitted
(Copilot compiles matcher as ^(?:PATTERN)$, and a bare '*' is an invalid
regex that would skip the whole entry). Importer prefers the real field,
falling back to the legacy comment for previously-generated files.
Re-verification note: the research pass also flagged canonical Notification
hooks as unsupported -- current docs contradict that ('notification' is a
real, documented Copilot hook event), so no change was made there.
…tive)
- additionalRules (project + global): .opencode/rules/<slug>.md files are
now also declared in opencode.json's instructions array (project glob
.opencode/rules/*.md; global absolute ~/.config/opencode/rules/*.md).
OpenCode does not auto-scan any rules directory (opencode.ai/docs/rules)
-- only AGENTS.md/CLAUDE.md auto-discover -- so generated rule files were
previously invisible to stock OpenCode.
- agents (project + global): emit a real permission object (e.g.
permission: { edit: deny }) mapped from canonical tools/disallowedTools
via a new permission-map.ts, instead of tools/disallowedTools frontmatter
keys that OpenCode's schema does not support (disallowedTools does not
exist; tools is deprecated and boolean-map shaped, not a string array).
The importer translates permission back into tools/disallowedTools so
restrictions round-trip.
- fix mergeOpenCodeSettings: it delegated to the generic Claude-shaped
settings merger, which only ever carried over permissions(plural)/hooks
from fresh content -- silently freezing mcp/permission/instructions at
whatever the first generate wrote. It now merges OpenCode's own
mcp/permission/instructions keys directly.
- add capability-ledger cells for opencode additionalRules/agents
(project + global), verdict confirmed, sourced from opencode.ai docs.
- target-contract.matrix.test.ts: a target's own native-directory prefix
legitimately appearing in its own generated settings file is not a
cross-target leak; exclude the current target's own prefix from the
no-target-specific-prefix-leak check.
…and frontmatter keys - additionalRules (global): embedded -> native. Non-root rules now write real files to ~/.qwen/rules/<slug>.md (loadRules() reads .qwen/rules/ from both global and project dirs identically) instead of being folded into ~/.qwen/QWEN.md's body. Adds QWEN_GLOBAL_RULES_DIR, global layout rewrite/rulePath wiring, and an importer-spec global source; removes the now-dead renderQwenGlobalInstructions embedding helper. - additionalRules (project + global): rename emitted frontmatter key globs: -> paths: in .qwen/rules/<slug>.md - Qwen's parseRuleFile() only recognizes paths: for conditional rule injection, so globs: was always silently ignored. Importer gains a frontmatterRemap to translate paths: back onto the canonical globs field for round-trip. - commands (project + global): stop emitting allowed-tools frontmatter in .qwen/commands/<name>.md - MarkdownCommandDefSchema has no tool- restriction field, so it was dead. Adds a lintCommands warning. - Ledger: add/update qwen-code cells (additionalRules project+global, commands project+global) as confirmed with source URLs. - Docs: regenerate README.md / supported-tools.mdx support matrices; hand-update the global-mode paths table row for the new rules dir.
- commands (project+global): native -> embedded. Amp has no declarative slash-command file format (ampcode.com/manual); commands only exist via amp.registerCommand(...) in a TS plugin. Generated output is unchanged (still .agents/skills/<name>/SKILL.md via the native skills surface) -- only the declared level was wrong. - hooks (project+global): native -> none. buildAmpScopedSettings() wrote a fabricated "amp.hooks" key into .amp/settings.json. Re-verified via live full-page render plus 14+ months of Wayback Machine snapshots (May 2025 - Jul 2026): ampcode.com/manual never mentions hooks. Only amp.on(...) plugin code exists, no declarative surface. Removed the amp.hooks branch from buildAmpScopedSettings and mergeAmpSettings, and added a lintHooks warning (mirrors the existing lintIgnore pattern). - Deprecated an existing lesson that had fabricated a detailed amp.hooks array schema from a hallucinated source, and added a corrected lesson with the verification trail. - Updated capability-ledger.json (commands confirmed embedded; hooks marked rejected for both scopes) and synced README/website support matrix via matrix:generate.
…ents/mcp/hooks/rules
Rebase the cline target on the standalone CLI's documented layout
(docs.cline.bot/cli/cli-reference) instead of the VS Code extension's
IDE-era paths, per confirmed capability-drift research.
- rules (project+global, native): .clinerules/{slug}.md -> .cline/rules/;
~/Documents/Cline/Rules/ -> ~/.cline/data/settings/rules/. Drop the
undocumented flat-file .clinerules convention (directory only).
- hooks (project+global, native): .clinerules/hooks/*.sh -> .cline/hooks/*.sh;
both scopes now resolve to the identical relative path.
- mcp (project, native): .cline/cline_mcp_settings.json -> .cline/mcp.json
(old name + .cline/mcp_settings.json still accepted on import).
- mcp (global): native -> none. No global MCP path is documented.
- skills (global, native): .cline/skills/ -> .cline/data/settings/skills/.
- agents (project, native): restructure from one file per agent
(.cline/agents/<name>.md, undocumented) to a single combined
.cline/agents.yaml (CLI docs: "Agent definitions"). agentPath() now
returns null (no per-name destination), matching roo-code's .roomodes.
- agents (global): native -> none. No agents.yaml/agents/ is documented
under ~/.cline/data/settings/.
- ignore (global): native -> none. docs.cline.bot/customization/clineignore
documents .clineignore as strictly per-workspace-root.
- permissions (project+global): none -> partial. No writable permissions
file exists; approval is --auto-approve/CLINE_COMMAND_PERMISSIONS/UI only.
A no-op generator stub plus a lint warning point users at these.
Commands (workflows) are unaffected — the CLI reference doesn't document
a workflows surface, so the existing .clinerules/workflows/ path is kept.
Records ledger provenance for every changed cell and syncs README +
website docs (support matrix, global-mode paths table, CLI reference
prose, hooks Windows-portability note).
…schema
- Commands (project + global, none -> native): .rovodev/prompts.yml
(+ ~/.rovodev/prompts.yml) saved-prompts manifest, with a
.rovodev/commands/<name>.md content file per canonical command.
Commands are no longer projected as skills for this target.
- MCP (project): native -> none. No project-level MCP file is documented;
only ~/.rovodev/mcp_config.json (global) exists.
- MCP (global): renamed the generated/imported file from
~/.rovodev/mcp.json to the real ~/.rovodev/mcp_config.json.
- Permissions (global): rewrote toolPermissions to the real nested
tools.<name> / tools.bash.{default,commands} schema instead of a flat
allow/deny/ask list.
- Updated capability ledger, README/website support matrix, contract
fixtures, and reference-rewrite test helpers to match.
…nds, downgrade project hooks
- hooks (project): native -> none; docs.langchain.com/oss/javascript/deepagents/code/hooks
documents only the global ~/.deepagents/hooks.json, no project-level surface exists
- hooks (global): broken-fix; generator wrote Claude-Code-style {EventName:[...]} shape,
Deep Agents actually reads a flat {hooks:[{command,events}]} array (hooks-format.ts);
wired via globalSupport.scopeExtras (gated scope==='global') plus a matching importer
- agents (project+global): none -> native; docs list a dedicated
.deepagents/agents/{name}/AGENTS.md (project) / ~/.deepagents/{agent}/agents/{name}/AGENTS.md
(global) subagent surface, distinct from skill-projection (agent-format.ts)
- commands (project+global): none -> embedded; no dedicated command format exists,
routed through the existing skills-projection mechanism (declaration-only fix)
- rules/skills/agents (global): broken-fix; global paths were identical to project paths
instead of the documented per-agent-instance dir ~/.deepagents/{agent}/... (default
instance name "agent" per docs) - added DEEPAGENTS_CLI_DEFAULT_AGENT_NAME
- updated capability-ledger.json (confirmed cells + one rejected for project hooks),
README/docs support matrices, and blast-radius references (import-maps, contract
paths, e2e reference-targets, target-contract-matrix projected-agents set)
…global ignore - Rules/AdditionalRules/Commands/Agents (global, native): move from the ~/.kilo/ mirror to the documented ~/.config/kilo/ unified config dir; non-root rules now register under kilo.jsonc's `instructions` key (a bare rules dir isn't auto-loaded at global scope). - MCP (global, native): fold into kilo.jsonc's `mcp` key with kilo's own type/command schema instead of a standalone mcp.json + mcpServers wrapper; add importGlobalKiloMcp() for the reverse direction. - mergeKiloConfig now overlays instructions/mcp alongside permission so the three writes to the shared kilo.jsonc compose correctly. - Ignore (global): native -> none; no documented global .kilocodeignore equivalent exists (workspace-root-only per kilo's own docs). - Skills (global) re-verified unchanged: ~/.kilo/skills/ is correct. - Update capability-ledger.json, README/website support matrix, and add a changeset.
…gnore-lint dispatch - reference/map.ts, output-source-map.ts: track combined-agent output sources so the rewriter resolves references inside merged agent files - import-maps/cline.ts, roo-code.ts, trae.ts: add missing import-map entries for agent and hook paths introduced by capability audit - generate/engine.ts: gate rules rewrite on descriptor layout presence, preventing empty rules blocks from leaking into embed-only targets - lint/linter.ts: dispatch descriptor.lint.ignore so partial-ignore targets (Jules, Pi Agent, Kiro, etc.) emit warnings - zed/lint.ts: remove unused lintHooks export (hooks=none; silent-drop guard covers it)
… loss, report over-declared cells - src/core/capabilities/merge.ts: extract mergeCell/hasNonEmptyFingerprint into pure unit-tested module; register pnpm capabilities:merge script - scripts/merge-capability-ledger.ts: fix two data-loss bugs — confirmed/rejected cells now keep their researched maxAchievable ceiling; fingerprint preservation checks all three arrays (topLevelKeys, requiredFrontmatter, keyChecks) - src/core/capabilities/audit.ts: report over-declared cells independently from unverified so both signals surface together - package.json/pnpm-lock.yaml: register capabilities:merge script - CONTRIBUTING.md: document audit/merge workflow
…ws path to ~/.gemini/antigravity/global_workflows/
…er and content contracts
…rtial, scope-gate lintMcp
…ith lint warnings
…dded, permissions/project to partial
…obal and permissions to partial
… full test coverage
… changeset with raises
…ions partial with lint coverage
…te all builtin descriptors
…ed, 21 rejected cells)
…ect cell formats and fingerprints
…, sync generated outputs
agentsmesh check verified only canonical sources against the lock, so direct edits to generated files passed CI despite documented behavior. - generate records checksums of all generated outputs in .agentsmesh/.lock (outputs map); full runs replace the map, --targets/--features runs merge per-path - check re-hashes recorded outputs and exits 1 on hand-edited or deleted files; JSON gains outputsModified/outputsRemoved/outputsChecked - new check --no-outputs flag skips output verification (gitignored outputs in CI); old-format locks skip with a hint until regenerated - hashes mirror hashFileForManifest (BOM strip + LF normalize) so line-ending-only rewrites are not drift - MCP check tool now performs output verification (passes projectRoot) - merge's lock rewrite intentionally omits outputs until regenerate - docs: check/generate/merge CLI pages, CI guide, generation pipeline, MCP reference, README lock description
The hono ^4.12.25 CORS-advisory override (6885150) was added only to package.json pnpm.overrides, leaving pnpm-workspace.yaml and the lockfile overrides snapshot stale. This broke CI's pnpm 10 --frozen-lockfile with ERR_PNPM_LOCKFILE_CONFIG_MISMATCH. Add hono to the workspace overrides and regenerate the lock so all three sources agree.
…act ignores The .cursor/, .kiro/, .warp/ (and sibling) ignore patterns were unanchored, so they matched the same-named dirs inside tests/e2e/fixtures/ too. Three fixture sources were thus silently gitignored and never committed: cursor-project/.cursor/cli.json -> permissions.yaml kiro-project/.kiro/hooks/review-on-save.json -> hooks.yaml warp-project/.warp/.mcp.json -> mcp.json filesystem server They exist locally (tests pass) but are absent on CI's fresh checkout, so import reads null and silently skips emission, failing import-capabilities e2e only on CI. Root-anchor every generated-artifact pattern (leading /, matching the existing /docs//tasks/ convention) and commit the three files.
generateRules built the execution-rule output path with
`rule.source.split('/').pop()`, which on Windows leaves the whole native
backslash absolute path as the "slug" — emitting a bogus
`.codex/rules/C:\...\.agentsmesh\rules\<name>.rules`. Its dirname mkdir
throws ENOENT (drive-colon mid-path), so `generate --targets codex-cli`
exits 1 on Windows CI while passing on POSIX (where absolute paths use '/').
Split on both separators (/[\\/]/) so the slug is the basename regardless
of platform. Adds a platform-independent regression test feeding a Windows
absolute source. Fixes codex-format-roundtrip, global-roundtrip-codex e2e
and generate unit test on Windows.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Type of change
Checklist
pnpm test,pnpm lint,pnpm typecheck)pnpm changeset) for user-visible changes