Repository navigation
Conversation
`omnigent devin` wraps the resident Devin CLI TUI in a runner-owned tmux pane and
mirrors it into an Omnigent conversation. It sits ALONGSIDE Devin's existing ACP
harness rather than replacing it:
* `devin-native` — this wrap. Omnigent policy enforcement, approval cards,
model + effort, resume, interrupt, cost tracking.
* `devin-acp` — the ACP path (`omnigent/inner/devin`), unchanged in behavior and
still the only one that surfaces Devin's sub-agents as child sessions. It moves
from the row id `devin` to `devin-acp` so the bare vendor spelling can
canonicalize to the native wrap, the way `opencode` resolves to
`opencode-native`.
Devin exposes a Claude-Code-shaped lifecycle-hook system carrying two ids the
other native harnesses have to reconstruct — `prompt_id` (used directly as the
Omnigent turn id) and `tool_use_id` (pairs a tool call with its output) — plus
`last_assistant_message` on `Stop`. So the mirror is hook-driven rather than
screen-scraped, and the forwarder is a fraction of the size of the
transcript-parsing ones.
Verified live against devin 3000.10.21:
* A `PreToolUse` hook returning `hookSpecificOutput.permissionDecision: "deny"`
blocks the tool even under `--permission-mode bypass`, and the reason reaches
the model — so the shared `native_policy_hook` seam drives Devin unmodified
(elicitation=HOOK, as claude-native). `PermissionRequest` accepts
`{"decision": "approve"}`.
* Interrupt needs Escape *twice* ("esc twice to interrupt"); one only clears the
composer draft.
* The composer stays writable mid-turn, so a queued message steers the running
turn (live_queue/steering).
* Effort is a model-variant suffix, not a flag, so a (model, effort) pair is
composed into one id at launch and validated against `devin models list`.
Hooks are registered through a session-scoped `--config` file merged from the
user's own config, so the user's settings survive and nothing is written into
their repository.
Co-authored-by: Isaac <no-reply@databricks.com>
Signed-off-by: Dhruv Gupta <dhruv0811@gmail.com>
…ist pinned Three CI-surfaced fixes on top of the devin-native harness: * `pickerEffortOptions` returned `[]` for devin-native, so the New Chat composer rendered no Effort section at all. Devin has no `--effort` flag — effort is a model-variant suffix Omnigent composes at launch — so without the ladder there was no way to choose effort when starting a chat, which is the whole point of carrying it as a separate axis. Devin now shares the Anthropic rung set, which is exactly its ladder. * Dropped `fullySupported` from the Devin row. That flag pins the picker's primary list to Claude Code + Codex and `nativeCodingAgents.test.ts` asserts exactly those two, so promoting a brand-new harness there is a maintainer's product call rather than a side effect of adding it. Devin folds into "More" with the other nine natives; flipping it back is one line. * `configured_harness_map()` gained the three devin readiness spellings, so the hello-frame coverage test lists them. Adds `tests/e2e_ui/chat/test_devin_native_picker.py`, which the E2E UI judge asked for: it stubs the host catalog and asserts Devin is offered, that its model list comes from the devin-native probe (every other harness's stub is empty, so a wrong catalog shows nothing rather than passing by luck), and that the Effort ladder renders — the regression the first bullet fixes. Not run locally: the e2e_ui rig's runner will not come online here, and the existing sibling `test_harness_support_level_split` fails identically at the same fixture. Co-authored-by: Isaac <no-reply@databricks.com> Signed-off-by: Dhruv Gupta <dhruv0811@gmail.com>
Like hermes-native, devin-native is a terminal-first TUI launched via `omni devin` (tmux pane + bridge dir) rather than `omnigent run --harness devin-native`, AND it wraps the own-auth `devin` CLI — so its spawn env carries no gateway/profile probe vars for that matrix to drive. Devin's ACP row is already excluded as an ACP_CLI_HARNESSES member. Co-authored-by: Isaac <no-reply@databricks.com> Signed-off-by: Dhruv Gupta <dhruv0811@gmail.com>
The first version asserted Devin sits behind the Other submenu and that Pi is not in the primary list — i.e. it pinned the `fullySupported` product decision, which is not what the test is for. It also failed in CI: the row was not found where the assertion insisted it must be, and the e2e_ui rig will not boot locally (the existing sibling test_harness_support_level_split fails at the same fixture), so the placement could not be checked before pushing. `_reveal_devin` now looks in the primary list and expands Other only if it has to, so the test fails on a Devin regression rather than on a placement change. The substance is unchanged: Devin's model list comes from the devin-native probe (every other harness's stub is empty, so a wrong catalog shows nothing rather than passing by luck), and the Effort ladder renders. Co-authored-by: Isaac <no-reply@databricks.com> Signed-off-by: Dhruv Gupta <dhruv0811@gmail.com>
test_native_seed_ids_are_byte_stable pins every native built-in's name-derived agent id, because a drifting id orphans persisted conversation.agent_id rows on redeploy. devin-native-ui adds one entry. Verified this is an addition, not a drift: recomputing builtin_agent_id over NATIVE_CODING_AGENTS leaves all eleven existing ids byte-identical, with devin-native-ui as the only new key. So no migration is owed and no session is orphaned — which is exactly the distinction the table's comment asks for before touching it. Co-authored-by: Isaac <no-reply@databricks.com> Signed-off-by: Dhruv Gupta <dhruv0811@gmail.com>
…lity The devin-native row declares only `devinMode`, but two render gates keyed off `modelPicker`/`permissionMode`/codex only: `selectedAgentHasKnobs` (which wraps the whole config dropdown) and the models-section gate. So selecting Devin in the New Chat picker showed neither a model list nor an effort ladder. `pickerModelOptions` and `pickerEffortOptions` already resolve devin-native; only the two gates missed it. OR `supportsDevinMode` into both. Adds a NewChatDialog vitest that drives the Devin agent through the picker and asserts its model families + effort rungs render. It fails with the exact e2e_ui signature (models section absent) without the gate fix. Co-authored-by: Isaac <no-reply@databricks.com>
Resolves the composer refactor (#7243, #7260) against the devin-native picker: - pickerEffortOptions: keep Devin's Anthropic-ladder branch, adopt main's normalizeEffortLabel rename. - models section: main moved it into <ComposerConfigSections>; thread supportsDevinMode into the new models gate so Devin still renders its model list (selectedAgentHasKnobs auto-merged with the same flag). - ChatPage effectiveModel: keep main's sessionModelSeeded branch, keep devin in the vendor-owns-model group. Co-authored-by: Isaac <no-reply@databricks.com>
…elper The picker e2e never opened Devin's config submenu — it only revealed the row, so `new-chat-landing-agent-models` (which lives in that submenu) was never visible and the test failed regardless of the gate fix. The vitest passed because it opens the submenu; the e2e drove the UI wrong. Rewrite it on the shared `_open_entry_models` helper (same navigation the passing Pi/Codex picker tests use), which clicks the `agent-config-*` Edit entry. That entry exists only when `selectedAgentHasKnobs` honours `devinMode`, so the corrected test still fails without the gate fix and passes with it. Co-authored-by: Isaac <no-reply@databricks.com>
…tore Devin persists each run_subagent delegate's full transcript as its own chain in the parent session's message_nodes forest (sessions.db) — not in the hook stream or the ATIF export. This is the direct analogue of claude-native reading Claude Code's per-subagent transcript files; Devin keeps the same data in SQLite. Adds omnigent/harnesses/devin_native/subagents.py: - reconstruct_transcript_nodes: task-keyed longest leaf→root walk that recovers the executed chain and sidesteps the forest's compaction/streaming snapshots (a leaf has one ancestor path, so no snapshot guessing); - transcript_items: converts a chain to the forwarder's own conversation-item shapes (user/assistant messages, function_call + paired output, reasoning); - parsers for the spawned/completed agent_id (which ride only free text) and a read-only, fail-soft message_nodes reader. Pinned against a trimmed real capture (devin 3000.10.21, two parallel sub-agents) in tests/data/devin_subagent_nodes.json; 16 tests. Not yet wired into the forwarder — that lands next. Co-authored-by: Isaac <no-reply@databricks.com>
…l transcripts Wires the sub-agent transcript reconstruction into the live path so Devin's run_subagent delegates appear as child sessions in the Agents rail, each carrying its complete internal transcript (prompt -> tool calls + outputs -> final report) — parity with claude/codex/antigravity native sub-agents. Forwarder (harnesses/devin_native/forwarder.py): - record each run_subagent spawn (agent_id parsed from the tool result, task + title from the input), persisted in the forward state; - on each turn-end, read Devin's session store, and for every newly-completed sub-agent reconstruct its chain and post it into a minted child session, then close the child idle. Best-effort per sub-agent, marked done to avoid re-posts. Server (mirrors the codex/antigravity sub-agent-start family): - external_devin_subagent_start event type + _persist_external_devin_subagent_start, keyed idempotently on Devin's agent_id, stamping the devin-native-ui-subagent wrapper + title/tool_use_id labels; - _is_devin_native_subagent + _devin_subagent_display_tool drive the rail label (run_subagent title, agent_id as session_name); - devin joins _ALLOWED_EVENT_TYPES, the parse_item_data bypass set, and the terminal-status-forward exclusion (a mirror has no runner work entry). Capabilities + web: subagents=True + subagent_wrapper_label on the devin-native row; subagentWrapperLabel on the web row so the child resolves the Devin icon. Tests: server integration (mint / idempotency / missing-id 400) and forwarder flow (spawn recorded, completed -> mirrored child with full transcript, not mirrored until completion). Co-authored-by: Isaac <no-reply@databricks.com>
…t launch Two gaps kept the in-chat (in-conversation composer) model/effort controls from working for devin-native, even though the new-session composer had them: 1. UI gate: supportsEffortControl covered claude/codex/pi but not devin, so the in-chat Effort dial never rendered for a Devin session (the model picker did, via modelPickerKind). Add devin-native; effortLevelsForConv already returns the Anthropic ladder for it. 2. Switching didn't apply: the executor typed /model with config.model alone (the bare family), while reasoning_effort rode config.extra and was ignored — so an in-chat effort change (or effort-on-a-model change) never reached the pane. DevinNativeExecutor now recombines (family, effort) into one variant via resolve_devin_launch_model (cached per pair) before /model, matching the launch path. resolve_devin_launch_model degrades to the family if the catalog is unreachable, so a probe failure costs effort, not the turn. Tests: executor composition (compose, mid-chat effort switch re-applies /model, per-pair cache) and web capability gates (effort + model picker visible for devin-native in-chat). Co-authored-by: Isaac <no-reply@databricks.com>
… fixture Pre-commit (ruff format + end-of-file-newline) flagged two files the earlier commits added: the subagent reconstruction test's formatting and the captured node fixture, which json.dump wrote without a trailing newline. Co-authored-by: Isaac <no-reply@databricks.com>
The composer's slash-command menu unions a session's bundled skills with the extra skills its harness exposes in its own terminal, resolved per-vendor by resolve_harness_skills. devin-native had no provider, so it fell through to the generic host walk and Devin's own skills never appeared. Devin loads user-invocable skills (slash commands, including plugin-registered ones) from several of its own dirs (~/.config/devin/skills, ~/.agents/skills, .devin/skills, …) plus .claude/skills for compatibility, with its own precedence — more than a single dir walk here would faithfully reproduce. So devin_host_skills sources the menu authoritatively from `devin skills list --json --trigger user`, matching exactly the / commands the Devin terminal offers. Best-effort: a missing CLI / non-zero exit / timeout / bad output yields no host skills rather than raising, so the bundled skills still show. The web plumbing (session.skills -> composer menu) is generic, so no web change is needed; only the runner-side discovery was harness-specific. Co-authored-by: Isaac <no-reply@databricks.com>
resolve_devin_launch_model is typed str | None (it returns None only for a None family, which the guard above already excludes), so assigning it into the str-valued variant cache tripped pyrefly's unsupported-operation. Guard the None case explicitly — defensively correct and type-clean. Co-authored-by: Isaac <no-reply@databricks.com>
devin-native declared instruction_delivery=NOT_DELIVERED: the launch path never even received the agent spec, so a custom Devin agent's AgentSpec.instructions reached neither the CLI nor the model. Devin has no --append-system-prompt and ignores config-level instructions (verified against devin 3000.10.21). Its only channel that reaches every turn's system prompt is a Windsurf always-on rule (.windsurf/rules/*.md with a `trigger: always_on` frontmatter — without the frontmatter Devin treats the file as manual and never loads it; verified the model obeys it with the frontmatter). So thread ctx.agent_spec into _auto_create_devin_terminal (like pi/cursor) and write the raw instructions to <workspace>/.windsurf/rules/omnigent-agent-instructions.md at launch via the new write_devin_agent_rule. Stable name = one overwritten file, and a plain agent (no instructions) removes any rule a prior custom-agent launch left, so the workspace never carries stale instructions. Capability flips to AGENT_STARTUP_ADDITIVE. ponytail: the rule writes into the user's workspace — the only always-on channel Devin exposes (rules are CWD-relative; no out-of-tree rules dir). Co-authored-by: Isaac <no-reply@databricks.com>
# Conflicts: # web/src/shell/NewChatDialog.tsx
Devin's `--permission-mode` was reachable only via `omnigent devin` on the CLI; the composer had no picker. Add one, stacked on the devin-native PR. Devin uses the same `--permission-mode <mode>` launch flag as Claude Code, so the pick rides `terminal_launch_args` exactly like Claude's — but Devin has its own vocabulary (normal / accept-edits / smart / dangerous) and no running-session switch. To avoid entangling Claude's mid-chat permission machinery, this is an isolated `devinPermissionMode` capability mirroring Codex's separate `approvalMode` pattern rather than reusing `permissionMode`: - nativeHarnessModes.ts: DEVIN_NATIVE_PERMISSION_MODES (+ default "normal", which sends no flag so Devin keeps its own default) with args = ["--permission-mode", v]. - nativeCodingAgents.ts: `devinPermissionMode` capability on the devin row. - NewChatDialog: the permission hand-menu, chip summary, create-time terminal_launch_args emission, draft persistence, and harness-switch reconcile all gain a devin branch. Create-time only: Devin has no `/permissions` popup to drive a live switch. Server needs no change — terminal_launch_args validation is shape-only and the runner already appends passthrough to the Devin launch. Co-authored-by: Isaac <no-reply@databricks.com>
Contributor
|
…ve-harness # Conflicts: # web/src/shell/NewChatDialog.tsx
Deprecate the `devin-acp` ACP harness in favor of native `devin-native`, the backwards-compatible way: keep it registered and resolvable (module, capability, `--harness devin-acp`, existing sessions), but stop offering it as a fresh pick so native Devin is the sole "Devin". - cli_config.py: skip `devin-acp` in the `omnigent config` setup list (the one fresh-pick surface that still showed a second "Devin (ACP)" row). Everything else in the ACP catalog is untouched. - acp_cli_harnesses.py / docs/devin-native.md: note the deprecation. - harness_plugins.py: document the one accepted backwards-compat break — a session persisted before this cutover with harness `devin` (which meant ACP then) resumes on the native wrap. Not fixable by a static alias (bare `devin` must mean native for a fresh launch yet ACP for an old session), and the ACP harness shipped only recently, so pre-cutover ACP sessions are thin. - tests: the setup overview no longer lists "Devin (ACP)"; row-dispatch indices shift down by one after native Devin. Co-authored-by: Isaac <no-reply@databricks.com>
Contributor
|
Closed. If you want to pick this back up, comment |
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
1 participant
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.
Related issue
N/A — feature stacked on #7254 (devin-native harness). No separate issue.
Summary
Stacked on #7254. Adds a create-time permission-mode picker for Devin in the composer's permissions hand-menu — previously Devin's
--permission-modewas reachable only viaomnigent devinon the CLI.--permission-mode <mode>launch flag as Claude Code, so the pick ridesterminal_launch_argsexactly like Claude's. But Devin has its own vocabulary (normal/accept-edits/smart/dangerous, verified againstdevin 3000.10.21) and no running-session switch.devinPermissionModecapability mirroring Codex's separateapprovalModepattern rather than reusingpermissionMode.normal(default) sends no flag, so Devin keeps its own configured default.No server change:
terminal_launch_argsvalidation is shape-only and the runner already appends passthrough to the Devin launch, so the picked flag reaches the CLI as-is.ELI5: Devin already understood
--permission-mode; we just added the dropdown that sends it, with Devin's own list of modes.Test Plan
web:vitest run src/shell/NewChatDialog.test.tsx src/lib/nativeCodingAgents.test.ts src/pages/ChatPage.capabilities.test.ts— 443 pass, incl. two new tests asserting a picked mode emits["--permission-mode", <mode>]and the default emits nothing.tsc --noEmitclean;oxlintclean;prettierclean.Demo
UI change, but the composer menu could not be captured headlessly in this environment (no live browser). By-eye verification of the rendered permission chip/menu is pending a manual check: start
omnigent devinon this branch, open the composer permissions menu, and confirm the Normal/Accept edits/Smart/Dangerous options appear and a non-default pick starts Devin with that--permission-mode.Type of change
Test coverage
Coverage notes
Manual verification is limited to CLI + automated tests; the composer menu's visual rendering is unverified by eye in this environment (see Demo).
Changelog
Devin sessions can pick a permission mode (Normal / Accept edits / Smart / Dangerous) from the composer.
This pull request and its description were written by Isaac.