feat(agents): devin cli commands and hooks cross-checked from web sou…

…rces - devin slash commands expanded and verified; exec shell tool captured - test guards devin hook events (devin rejects unknown ones) - update task.md
codejunction · Oct 4, 2026 · Oct 3, 2026 · Oct 3, 2026 · Oct 4, 2026 · Oct 4, 2026
commit c3178d1ce318dd672c4f7fb74cf971464978c195
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -28,6 +28,7 @@ All notable changes to shelldeck are listed here. The format follows [Keep a Cha
 - The support matrix (exact state, resume, and what each format was checked against) is in [docs/agent-integrations.md](docs/agent-integrations.md#support-matrix).
 - Deterministic capture: agent tool calls (command + success, edited file) from hooks and people's shell commands become context events. Test runs set the project's Tests line, failures (not look-ups like grep) set Last error, and each agent run writes `.shelldeck/sessions/<id>.md` when it ends.
 - Fixed: a redacted value could corrupt a stored context event (secrets are now redacted per value before serializing).
+- Devin CLI: commands and flags cross-checked from web sources (slash commands `/compact /clear /new /model /context /resume /plan /revert /exit`, shell tool `exec` captured); its hooks are limited to the events Devin accepts.
 - Hand-off files include a context snapshot, and the installed agent skill teaches the context workflow.
 
 ## [0.0.6] - 2026-10-01

diff --git a/CLAUDE.md b/CLAUDE.md
@@ -95,7 +95,7 @@ uv tool install --force -e .        # global sd/shelldeck from this checkout (th
 - **Deterministic capture.** `hook.activity()` adds `{command, ok}` or `{file, change}` to tool-finished reports (never output or contents), validated by `agent_state.parse_activity`. The server's `_capture()` runs `context.record_activity` / `start_session` / `end_session` on a thread for the terminal's project; the context session id is `<terminal>.<generation>` (`_ctx_session`). `TEST_CMD` sets the Tests line, `LOOKUP_CMD` failures (grep etc.) aren't errors, and `end_session` writes `.shelldeck/sessions/<id>.md`. People's shell commands (`_record_command`, only when no agent runs there) are captured too. `record_event` redacts per value (`_payload`).
 - **Fact extraction** is done by the agents, never by a model in shelldeck: `POST /api/sessions/{id}/extract` types `EXTRACT_PROMPT` (short; tells the agent that dedup/redaction are server-side so it skips reading memory; `sd remember` takes several facts). On demand only, refused while working/blocked.
 - `shelldeck/detection.py` + `agent_detection/default.toml`: screen rules for the heuristic (`_next_state`, explain). `load(agent)` merges a local override (`<config>/agent-detection/<agent>.toml`: replace by id, `disable`, add), cached by mtime; an invalid override is ignored and `Manifest.error` says why. `match()` picks the rule matching latest on screen. `agents.QUESTION` stays for sub-agent question forwarding; `QUESTION_RE` in app.js mirrors the bundled phrases.
-- `shelldeck/agent_commands.py`: `COMMANDS` maps shelldeck actions to each agent's own slash commands (claude from its cli.js, codex from slash_command.rs, gemini from docs/reference/commands.md; devin `VERIFIED = False` until checked on a real install). `POST /api/sessions/{id}/agent-command` types it, refusing blocked/working agents unless `force`; arguments must match `SAFE_VALUE`.
+- `shelldeck/agent_commands.py`: `COMMANDS` maps shelldeck actions to each agent's own slash commands (claude from its cli.js, codex from slash_command.rs, gemini from docs/reference/commands.md; devin from web-search quotes of its docs plus other projects' integrations; Devin rejects unknown hook events, so keep its `hook.EVENTS` within its documented list). `POST /api/sessions/{id}/agent-command` types it, refusing blocked/working agents unless `force`; arguments must match `SAFE_VALUE`.
 - `shelldeck/context.py`: persistent agent context in `<config>/context.db` (separate from shelldeck.db; FTS5 `search` table with a LIKE fallback). Projects are keyed by resolved folder (the deepest shelldeck project containing cwd, else the git top level). Everything stored goes through `redact()`; `SENSITIVE_FILE` paths are never referenced. Knowledge is de-duplicated per project by `_norm` and never merged across projects. `check_stale` re-hashes `source_references`. `project_context` returns a `ContextPackage` dict plus budgeted `text`. `project_files` writes STATE/TASK/MEMORY/DECISIONS.md (server runs it on a thread via `_project_later`). Server endpoints are `/api/context*`, wrapped by `_ctx` (sqlite errors -> 503 `context_unavailable`, never a crash). `_handoff_files` appends `context.snapshot`.
 - `shelldeck/team.py`: agent teams.
   - `tier()`/`pick_model()`: `--model auto|small|medium|large` maps onto the agent's own `models()` by regex (medium = its default).

diff --git a/docs/agent-integrations.md b/docs/agent-integrations.md
@@ -20,7 +20,7 @@ manual check is still on the release checklist.
 | claude | priority | hook | yes | yes | Claude Code 2.1.289 (bundle + `--help`) |
 | codex | priority | hook | yes | yes | openai/codex source (hooks, slash commands) |
 | gemini | priority | hook | yes | yes | gemini-cli docs (hooks, commands, sessions) |
-| devin | priority | hook | yes | yes | herdr installer only; Devin CLI not checked (devin.ai blocked) |
+| devin | priority | hook | yes | yes | Devin CLI docs via web search + other projects' integrations; CLI not run (devin.ai blocked) |
 | copilot | later | hook | yes | yes | dotpals/herdr formats; config files only |
 | cursor | later | hook | yes | yes | dotpals/herdr formats; config files only |
 | opencode | later | plugin | yes | yes | dotpals/herdr formats; config files only |
@@ -91,6 +91,8 @@ flags (`team.launch_line`):
 | Gemini CLI | `-m` | `-i "prompt"` | none: hooks only from `settings.json` (install) | `gemini --resume <id>` |
 | Devin CLI | `--model` | `-- "prompt"` | none: hooks only from `config.json` (install), plus `sessions.db` | `devin --resume <id>` |
 
+Devin's flags (`-- "prompt"`, `--model`, `-r/--resume`, `-c/--continue`, `-p`) and hook events (`SessionStart`, `UserPromptSubmit`, `Pre/PostToolUse`, `PermissionRequest`, `Stop`, `PostCompaction`, `SessionEnd`) were cross-checked the same way. Devin rejects its whole hooks block when it holds an unknown event, so shelldeck installs only events from that list (a test guards it). Its shell tool is `exec`.
+
 A prompt that is longer, or isn't plain words, is written to `.shelldeck/prompts/<id>.md`. The agent is then asked to
 read that file, so no shell ever sees quotes or `$`.
 
@@ -101,22 +103,22 @@ three, the flags come from their docs as used by dotpals and herdr.
 
 `sd agent cmd TERMINAL ACTION [ARG]` (and the *Command…* menu on the AI agents page) types the agent's own command:
 
-| Action | Claude Code | Codex | Gemini CLI | Devin CLI (unverified) |
+| Action | Claude Code | Codex | Gemini CLI | Devin CLI |
 | --- | --- | --- | --- | --- |
 | compact | `/compact` | `/compact` | `/compress` | `/compact` |
 | clear / new | `/clear` | `/clear` / `/new` | `/clear` | `/clear` |
 | model | `/model` | `/model` | `/model` | `/model` |
-| status | `/status` | `/status` | `/stats` | – |
-| resume | `/resume` | `/resume` | `/resume` | – |
-| review, init, export, hooks, permissions, plan | yes | yes (+ `/diff`, `/rename`) | init, hooks, permissions, plan, `/memory show`, `/rewind` | – |
+| status | `/status` | `/status` | `/stats` | `/context` |
+| resume | `/resume` | `/resume` | `/resume` | `/resume` |
+| review, init, export, hooks, permissions, plan | yes | yes (+ `/diff`, `/rename`) | init, hooks, permissions, plan, `/memory show`, `/rewind` | plan, `/revert` (rewind), `/new` |
 | quit | `/exit` | `/quit` | `/quit` | `/exit` |
 
 Sources:
 
 - Claude: the command names in Claude Code 2.1.289's bundle.
 - Codex: `codex-rs/tui/src/slash_command.rs`.
 - Gemini: `docs/reference/commands.md`.
-- Devin's list is a guess until the Devin CLI can be installed and checked.
+- Devin's list comes from the Devin CLI docs as quoted by web search, plus other projects' Devin integrations (docs.devin.ai and the CLI itself are blocked from this build box).
 
 A command is refused while the agent is blocked (it would answer the question) or working, unless `--force`.
 `sd agent cmd claude --list` shows an agent's commands.

diff --git a/docs/task.md b/docs/task.md
@@ -126,7 +126,7 @@
   checked in a browser. Codex hooks gained PreToolUse/PostToolUse/
   PermissionRequest/PreCompact (from codex-rs/hooks); Gemini resume confirmed
   in its docs.
-  - [ ] Devin CLI: install and verify its commands, flags and hooks. The
+  - [~] Devin CLI: commands, flags and hooks cross-checked from web sources; running the real CLI is still pending. The
     environment blocks devin.ai (needs devin.ai / *.devin.ai allowed in the
     cloud environment's network settings).
 - **2026-10-04 — Task 5.8 (UI) complete:** a *Context* view (sidebar and
@@ -173,6 +173,12 @@
   output, contents and sensitive files are never stored. Fixed a redaction bug
   that could corrupt stored event JSON. Checked live (fake agent, hook events,
   exit -> session file).
+- **2026-10-04 — Devin CLI cross-checked from web sources (user request):**
+  search results quoting the Devin docs and other projects' Devin integrations
+  agree with shelldeck's flags (`-- "prompt"`, `--model`, `-r/--resume`), config
+  path and hook events. Devin rejects unknown hook events, so a test now guards
+  the installed set. Added its shell tool name `exec` for capture, and expanded
+  and verified its slash commands. Running the real CLI is still blocked.
 - **2026-10-04 — Docs:** README, CHANGELOG (Unreleased), CLAUDE.md,
   `docs/agent-automation.md` and `docs/agent-report.schema.json` describe the
   lifecycle, report API, waits and context commands.

diff --git a/shelldeck/agent_commands.py b/shelldeck/agent_commands.py
@@ -5,7 +5,8 @@
   claude  built-in command names in Claude Code 2.1.289's cli.js
   codex   codex-rs/tui/src/slash_command.rs (openai/codex)
   gemini  docs/reference/commands.md (google-gemini/gemini-cli)
-  devin   ponytail: unverified (docs.devin.ai unreachable from the build box); marked `verified: False`
+  devin   Devin CLI docs as quoted by web search (commands & flags, essential commands) and other projects' Devin
+          integrations; docs.devin.ai itself and the CLI were unreachable from the build box
 """
 
 import re
@@ -42,9 +43,10 @@
     "gemini": {"compact": "/compress", "clear": "/clear", "new": "/clear", "model": "/model", "status": "/stats", "resume": "/resume",
                "init": "/init", "memory": "/memory show", "hooks": "/hooks", "permissions": "/permissions", "plan": "/plan",
                "rewind": "/rewind", "quit": "/quit"},
-    "devin": {"compact": "/compact", "clear": "/clear", "model": "/model", "quit": "/exit"},
+    "devin": {"compact": "/compact", "clear": "/clear", "new": "/new", "model": "/model", "status": "/context", "resume": "/resume",
+              "plan": "/plan", "rewind": "/revert", "quit": "/exit"},
 }
-VERIFIED = {"claude": True, "codex": True, "gemini": True, "devin": False}
+VERIFIED = {"claude": True, "codex": True, "gemini": True, "devin": True}
 SAFE_VALUE = re.compile(r"[\w.:/@+-]{1,80}")  # an argument typed after the command: one plain word
 
 

diff --git a/shelldeck/hook.py b/shelldeck/hook.py
@@ -39,7 +39,8 @@
     "qwen": {"SessionStart": SESSION},
     "qodercli": {"SessionStart": SESSION},
     "droid": {"SessionStart": SESSION},
-    # Devin's hooks are Claude-shaped; herdr reports only its session from them, shelldeck also its state
+    # Devin's hooks are Claude-shaped (events per its hooks docs). Devin rejects the whole hooks block if it holds an
+    # unknown event (Notification, PreCompact, ...), so only these may be installed; its shell tool is `exec`
     "devin": {"SessionStart": IDLE, "UserPromptSubmit": WORKING, "PreToolUse": WORKING, "PostToolUse": WORKING,
               "PermissionRequest": BLOCKED, "Stop": DONE, "SessionEnd": None},
     "kimi": {"SessionStart": SESSION, "UserPromptSubmit": WORKING, "PreToolUse": WORKING, "PostToolUse": WORKING,
@@ -98,7 +99,7 @@ def state_for(agent: str, event: dict, forced: str | None = None) -> tuple[str |
 # tool-finished events whose payload names the tool and its input (Claude-shaped; Gemini's AfterTool; Copilot's
 # toolName/toolArgs). Only the command or the edited path leaves the agent: never output or file contents.
 AFTER_TOOL = {"PostToolUse": True, "PostToolUseFailure": False, "AfterTool": None, "postToolUse": True, "postToolUseFailure": False}
-SHELL_TOOLS = {"bash", "shell", "exec_command", "local_shell", "run_shell_command", "powershell", "run_terminal_cmd", "execute"}
+SHELL_TOOLS = {"bash", "shell", "exec", "exec_command", "local_shell", "run_shell_command", "powershell", "run_terminal_cmd", "execute"}
 EDIT_TOOLS = {"edit", "write", "multiedit", "notebookedit", "write_file", "replace", "apply_patch", "create", "str_replace_editor"}
 
 

diff --git a/tests/test_integrations.py b/tests/test_integrations.py
@@ -361,3 +361,13 @@ def test_hook_activity_from_finished_tools():
     codex = hook.report("codex", {"hook_event_name": "PostToolUse", "tool_name": "shell", "tool_input": {"command": ["cargo", "test"]}})
     assert codex["activity"]["command"] == "cargo test"
     assert "activity" not in hook.report("claude", {"hook_event_name": "PreToolUse", "tool_name": "Bash", "tool_input": {"command": "ls"}})
+
+
+DEVIN_EVENTS = {"PreToolUse", "PostToolUse", "PermissionRequest", "UserPromptSubmit", "Stop", "PostCompaction", "SessionStart", "SessionEnd"}
+
+
+def test_devin_installs_only_events_devin_accepts():
+    # Devin refuses its whole hooks block when one event is unknown (e.g. Claude's Notification or PreCompact)
+    assert set(integrations.INSTALLERS["devin"].events()) <= DEVIN_EVENTS
+    act = hook.report("devin", {"hook_event_name": "PostToolUse", "tool_name": "exec", "tool_input": {"command": "pytest"}})
+    assert act["activity"] == {"command": "pytest", "ok": True}
diff --git a/tests/test_shelldeck.py b/tests/test_shelldeck.py
@@ -1156,7 +1156,7 @@ def test_native_agent_commands(client, monkeypatch):
                                      ("claude", "compact", "x", "no_argument"), ("claude", "model", "a; rm -rf", "invalid_argument")):
         with pytest.raises(ValueError, match=code):
             agent_commands.line(agent, action, arg)
-    assert all(not c["verified"] for c in agent_commands.catalog("devin"))
+    assert agent_commands.line("devin", "rewind") == "/revert" and agent_commands.line("devin", "status") == "/context"
     assert {c["action"] for c in client.get("/api/agent-commands/claude").json()["commands"]} >= {"compact", "model", "resume"}
     typed = []