main
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
|
76f37d928f |
fix(post-suspension): vanity import path go.moleculesai.app/cli (closes molecule-ai/internal#71 phase 2)
Migrates go.mod + 22 Go imports + README + comments + generated config templates off the dead github.com/Molecule-AI/ identity onto the vanity host go.moleculesai.app, owned by us. Surfaces touched: - go.mod module declaration: github.com/Molecule-AI/molecule-cli -> go.moleculesai.app/cli - Every Go import statement under cmd/ + internal/ - README install section: rewritten to lead with the vanity install command (the previous text was migration-in-progress hedging) - Comment URLs in internal/backends/backend.go + internal/cmd/connect.go (https://github.com/Molecule-AI/molecule-cli/issues/10) -> point at git.moleculesai.app/molecule-ai/molecule-cli - Generated config templates in internal/cmd/init.go + internal/cmd/config.go: header URL updated so new users land on the live SCM - Adds internal/lint/import_path_lint_test.go — structural test that walks every *.go / *.mod / Dockerfile / *.md / *.sh / *.yml in the module and rejects future references to github.com/Molecule-AI/ or Molecule-AI/molecule-monorepo. Mutation-tested before commit. Test plan - go build ./... clean - go test ./... green (cmd/molecule + 5 internal packages + new lint gate, all pass) - TestNoLegacyGitHubImportPaths fails on injected canary, passes on clean tree (no tautology) Open dependency - go.moleculesai.app responder must be deployed before 'go install go.moleculesai.app/cli/cmd/molecule@latest' works externally. Internal builds + 'go build ./cmd/molecule' from a fresh clone work today (self-referential module path). - Responder code prepared (worker.js, vendor-portable for CF Workers / Vercel Edge); deploy tracked separately under internal#71 phase 1. Pairs with parallel migrations of plugin-gh-identity (#3) + molecule-controlplane + molecule-core under the same internal#71 sweep. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
|
b7a3c887c0 |
feat(connect): M1.3 — exec + claude-code backends
After this PR `molecule connect <id>` actually does something useful:
the default `--backend claude-code` runs `claude -p` as a subprocess
per inbound message, capturing stdout as the reply. The general
`--backend exec --backend-opt cmd="..."` lets operators bridge to any
program that reads stdin and writes stdout (Ollama, custom Python,
shell pipelines).
What's in:
- `internal/backends/exec/exec.go` — generic exec backend.
- cmd (required): shell command, run via /bin/sh -c (or cmd /c).
- timeout (default 60s): per-message wall clock; subprocess is
killed on timeout and the dispatcher keeps the message queued.
- pass_meta (default false): when true, populates env with
MOLECULE_WORKSPACE_ID / CALLER_ID / MESSAGE_ID / TASK_ID / METHOD.
- Stdin = joined text parts; stdout = reply text.
- Stderr is captured + surfaced in error messages so operators can
see what their command printed.
- Honours ctx cancellation — SIGTERM kills the subprocess immediately.
- `internal/backends/claudecode/claudecode.go` — thin shorthand.
Translates {bin, args, timeout, pass_meta} into the equivalent exec
config. Defaults: bin=claude, timeout=5m, pass_meta=true.
Reusing exec means timeout/stdin/env handling stays in one place.
- `internal/cmd/connect.go` — registers both backends so `--help`
shows them and the default `--backend claude-code` works without
flag tuning.
Test plan:
- exec: cmd-required, bad-timeout, zero-timeout, echo-stdin-to-stdout,
text-only-part-concat, non-zero-exit-surfaces-stderr, timeout-kills-
runaway, pass_meta-injects-env, no-pass_meta-leaves-env-untouched,
parent-env-inherited-when-pass_meta-off, ctx-cancel-kills-command.
- claude-code: registered, bin/args translation produces "echo -p hello",
bad-timeout-surfaces, default-config-builds.
- All Unix tests gated on `requireUnix(t)` — Windows-shell semantics
diverge; cross-platform coverage is M3 work (brew/scoop/winget).
- Full suite green under -race.
UX after this PR:
# default — Claude Code on PATH
molecule connect ws_01HF... --token $T
# custom handler
molecule connect ws_01HF... --backend exec \
--backend-opt cmd="python myhandler.py"
# different model via Claude Code
molecule connect ws_01HF... \
--backend-opt args="--model claude-sonnet-4-6"
# CI smoke
molecule connect ws_01HF... --backend mock --backend-opt reply=ok
Out of scope (M1.4 + later):
- M1.4: GoReleaser tag-triggered release.yml workflow
- stdio backend (one persistent subprocess + line-delimited JSON-RPC)
— moved to M2 since exec covers the common case
- openai / mcp backends — M4 per RFC
RFC: https://github.com/Molecule-AI/molecule-cli/issues/10
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
|
db6b196631 |
feat(connect): M1.2 — heartbeat + activity poll loops
Implements the runtime side of `molecule connect <id>`. After this PR the CLI can actually attach to an external workspace and round-trip inter-agent messages through any registered backend. What's in: - `internal/connect/client.go` — platform-API client with bearer auth. Endpoints: POST /registry/register (delivery_mode=poll, no URL), POST /registry/heartbeat, GET /workspaces/:id/activity?type=a2a_receive, POST /workspaces/:id/a2a (reply target). Errors split into TransientError (network/5xx — retry with backoff) and PermanentError (4xx — abort with clear message). - `internal/connect/state.go` — atomic cursor persistence at ~/.config/molecule/state/<workspace-id>.json. Mode 0o600 (owner-only) from day 1 because future state additions may include rotated tokens. Atomic write-then-rename so a crash mid-write can never produce a half-written cursor. - `internal/connect/connect.go` — Run() orchestrator. Wires register- with-bounded-retry, then heartbeat goroutine + poll goroutine. Both respect ctx cancellation for clean SIGTERM. Robustness contract per RFC #10: * Cursor advances AFTER successful dispatch — crash mid-batch re-delivers, never drops. * 410 on cursor lookup → reset to "" and re-fetch (don't deadlock on a pruned cursor). * Heartbeat permanent error stops the heartbeat loop only; poll loop keeps running so the operator sees "stopped" + reason in logs and can SIGTERM. * Backend dispatch is sequential within a batch (avoids out-of- order replies for in-flight conversations). * Inter-agent reply path: POST envelope to /workspaces/<source>/a2a. * Canvas-origin reply (source_id == nil) logs + skips for now — M1.3 wires that via the task_update activity convention. - `internal/cmd/connect.go` — runConnect now actually calls connect.Run() (was a placeholder ctx-wait in M1.1). Test plan: - httptest workspace-server stub covers register / heartbeat / activity / a2a reply endpoints. - TestRun_RoundTrip_AgentReply: end-to-end ping → mock backend → pong reply lands at source, cursor saved. - TestRun_CanvasOriginMessageNotReplied: source_id=nil → backend fires but no reply post; cursor still advances. - TestRun_CursorPruned410ResetsAndContinues: server returns 410 once, cursor resets to "", next poll dispatches the fresh row. - TestRun_PermanentRegisterErrorAborts: 401 surfaces immediately. - TestRun_TransientRegisterErrorRetries: 503 then 200 → register succeeds on second attempt. - TestRun_OptionsValidation: missing Backend / WorkspaceID surface before any I/O. - State: round-trip, file mode 0o600, atomic-rename leaves no .tmp artifacts, corrupted file surfaces error. - All tests green under -race. Out of scope (next PRs in this stack): - M1.3: claude-code backend (canvas-origin reply convention rides with this) - M1.4: GoReleaser tag-triggered release.yml workflow - Push-mode (--mode push currently surfaces a clear "M4" error) RFC: https://github.com/Molecule-AI/molecule-cli/issues/10 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
|
3a6a7eb495 |
feat(connect): M1.1 — Backend interface + connect skeleton + mock backend
First step toward `molecule connect <id>` — the out-of-box external- runtime workspace connector specified in RFC #10. What's in this PR (foundational, ~300 LOC of code + matching tests): - `internal/backends.Backend` — the seam every concrete handler implements: HandleA2A(ctx, Request) → Response, Close(). Two methods, no inheritance, no surprise side effects. Concurrency-safe by contract (poll dispatch may parallelise). - Request/Response/Part/Config types — lossless JSON-RPC mirror so backends can re-issue downstream without re-parsing. - Compile-time registry — `Register("name", factory)` from each backend's init(); `Build(name, cfg)` selects at runtime. Panics on duplicate registration so drift fails loudly at startup, not on first message. - `mock` backend — single-template echo for CI smoke + tests + demos. `--backend-opt reply="<template>"` with `%s` for inbound text. - `molecule connect <workspace-id>` cobra command — flag surface, validation, --dry-run for smoke. Loops (heartbeat, activity poll, dispatch) land in M1.2 in internal/connect/. Coverage: - Registry: duplicate-name panic, empty-name panic, nil-factory panic, Build unknown-name error includes registered list. - Mock: default template, custom template, text-part concatenation, Final=true on terminal response. - Connect: --backend-opt KEY=VALUE parser (incl. value with =), flag validation (missing token, bad mode, bad opt, unknown backend), --dry-run happy path. All tests pass under -race. Out of scope (subsequent M1 PRs): - M1.2: heartbeat + activity poll loops in internal/connect/ - M1.3: claude-code backend (wraps molecule-mcp-claude-channel) - M1.4: GoReleaser tag-triggered release.yml workflow RFC: https://github.com/Molecule-AI/molecule-cli/issues/10 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |