1.7.0 — OpenCode dialect fix, 132-pair migration sweep

Features

• Full-migration sweep (npm run sweep) — seeds every source agent's real files in its own dialect (TOML for codex, command-arrays for kilo, required transport tags for crush, commented JSONC), migrates all 12×11 = 132 pairs through the real CLI into fresh dirs, and validates each result: target detected, doctor clean, instructions at the documented path, both MCP servers (command+env and url+headers) preserved through every dialect, source files untouched. 132/132 passing.
• README reflects opencode.json[c] (both config names supported)

Fixed

• OpenCode adapter now matches OpenCode's documented schema — the scanner reads the documented mcp key (legacy mcpServers configs still read) and normalizes the local/remote dialect (array command + environment; url + headers); the writer emits that documented dialect instead of mcpServers with stdio tags, and remote servers are no longer dropped (they were previously written as an invalid {type: "stdio"} with neither command nor url); the detector also accepts opencode.json — previously the tool could not detect its own writer output; doctor validates the documented key with the legacy key as an alias
• Copilot's instruction file no longer leaks into other agents — migrating copilot → claude/codex/opencode/kilo/cursor wrote .github/copilot-instructions.md, which none of those agents read; instruction-name normalization now maps it to AGENTS.md like GEMINI.md/MUSE_CODE.md

Tests & docs

• 203 tests (was 195); opencode fixture, writer/scanner/detection/pipeline tests pinned to the documented dialect; new opencode detection regression tests

1.6.0 — fix --dry-run and honest migrate feedback

Note on the npm tarball: npm install -g @superdev2347832/agent-bridge ships 1.6.0 built from b2ccc94, one commit ahead of this tag. That commit adds two things not listed below: the OpenCode dialect fix (scanner/writer/doctor now match opencode.ai/docs/mcp-servers — mcp key, local command arrays with environment, remote url servers no longer dropped) and the full-migration sweep (npm run sweep), which validates all 132 agent-pair migrations end-to-end and passed 132/132. 203/203 tests green.

Features

• fix --dry-run — previews every planned auto-fix with before/after content and writes nothing; the same backup-transaction apply path as before when run without the flag
• Honest migrate feedback — migrate/apply/migrate-all surface doctor's config errors (file:line) instead of silently dropping an unparseable source config's servers; healthy projects stay warning-free
• plan and diff validate both agent ids against the registry — diff no longer exits 0 claiming "nothing to do" for an unknown target, and plan's stale 3-agent list is replaced by the full registry-derived list
• scan prints the agent id next to the display name (Detected: Claude Code (id: claude-code)) — the id every other command requires

Fixes

• CLI no longer crashes with EPIPE when output is piped (agent-migrate scan | head)
• Positional-arg parsing is flag-position transparent (migrate s t --dry-run no longer treats the flag as a path)
• Self-migration is rejected with exit 2; nonexistent project paths are an error, not a silent empty migration
• .agentbridge/ backup dirs are gitignored so users' git status stays clean

Docs

• Playtest findings documented; README's command table covers fix --dry-run

v1.5.0 — instruction sync in fix, compiled-CLI CI, tag-driven publishing

[1.5.0] - 2026-09-15

Features

• fix syncs divergent project instruction files — when AGENTS.md / GEMINI.md / MUSE_CODE.md carry different content, fix copies AGENTS.md (which every supported agent reads) over the stale copies, inside the standard backup transaction
• doctor now reports every divergent instruction file instead of stopping at the first, so fixing one no longer hides another agent acting on stale rules
• CI builds dist/ and smoke-tests the compiled CLI (scan, doctor's exit-1 rejection, fix's in-place repair) on Node 22 and 24
• Release workflow: pushing a v* tag runs tests → build → smoke → npm publish via the NPM_TOKEN secret, so a broken artifact can never ship

Docs & tooling

• RELEASING.md documents the release flow, NPM_TOKEN setup, and the tag-pins-workflow gotcha
• 184 tests (was 182)

v1.4.0 — instruction-file validation in doctor

[1.4.0] - 2026-09-15

Features

• doctor validates instruction files: warns when a detected agent has none of its own (AGENTS.md / CLAUDE.md / GEMINI.md / MUSE_CODE.md / .github/copilot-instructions.md, mirrored from the scanners' lists), and warns when project-general instruction files (AGENTS.md / GEMINI.md / MUSE_CODE.md) carry divergent content — different agents would act on different rules. Agent-specific supplemental files (CLAUDE.md, copilot-instructions) are deliberately excluded from the comparison
• CI runs doctor over the repo root as an explicit step alongside tests and typecheck

Docs & tooling

• 182 tests (was 177)

v1.3.0 — doctor, fix, writer-derived plan/diff

[1.3.0] - 2026-09-15

Features

• doctor [path] — validates every detected agent's config against its documented schema: parseability by format (with a rename-to-.jsonc hint when a strict-JSON file carries comments), MCP key shape and wrong-key aliases, per-server command/url presence, and each agent's transport-tag policy (crush requires type on every server, cursor on stdio servers, gemini/codex forbid it). Exits 1 on errors, so it gates scripts and CI
• fix [path] — applies the safe auto-fixes doctor identifies, inside the standard backup transaction (rollback just works). Today that's one class: strict-JSON configs failing only from comments/trailing commas are rewritten comment-free in place — data preserved, agent keeps its documented filename (renaming to .jsonc would break agents like Claude Code, so fix never renames). Schema problems are listed for a human with exit 1
• CI now runs on every push and PR (previously main-only), with concurrency cancellation
• Test fixtures are doctor-guarded — a fixture that stops satisfying an agent's schema fails the build through the existing npm test CI step

Refactor

• Plan/diff statuses derive from each target's real writer via a writerSupports probe — the 132-entry rules table and compatibility engine are deleted, and what plan reports can no longer drift from what migration does
• Adapter + writer registered together in one list per agent; registerWriter throws on duplicates, making the "registered adapter, forgot writer" bug class structurally impossible (invariant test enforces exactly one adapter and one writer per agent)
• README compatibility matrix drift-guard now runs through the production planMigration code path

Docs & tooling

• 177 tests (was 163)

v1.2.0 — 12 agents, 132 bidirectional migrations

[1.2.0] - 2026-09-15

Features

• Seven new agents, both directions each — OpenAI Codex (.codex/config.toml, TOML), GitHub Copilot CLI (.copilot/mcp-config.json, instructions to .github/copilot-instructions.md), Crush (.crush.json, mcp key), Grok CLI (.mcp.json), omp/Oh My Pi (.pi/mcp.json), Muse Code (MUSE_CODE.md), Pi (AGENTS.md) — 132 ordered-pair migrations across 12 agents (Cursor CLI covered by the Cursor adapter)
• Shared JSON-adapter factory generates detector/scanner/writer triples from one per-agent spec; hand-rolled minimal TOML parser/serializer powers the Codex adapter (throws loudly on unsupported TOML features instead of mis-parsing)
• Merge-don't-clobber extended to TOML and commented-JSON targets — user keys in config.toml and hand-commented .json configs survive migration
• migrate-all works across all new agents with per-target rollback ids

Fixes

• migrate --dry-run was silently ignored (flag dropped in CLI arg parsing) — it now plans without writing, matching migrate-all
• Crush output now schema-compliant: type written on every server (stdio/http) per the required field in Crush's docs
• Commented .json configs no longer fall back to whole-file replace — comment-tolerant parsing applies at the merge boundary too
• Instructions normalize MUSE_CODE.md → AGENTS.md for all targets (not just GEMINI.md)
• OpenCode writer's obsolete opaque-config gate removed; parsing is now the test

Docs & tooling

• README compatibility matrix regenerated for all 132 directions from real writer behavior (drift-guard test enforces it)
• 163 tests (was 119)