fix(core): complete deepagent pipeline restructure downstream sync - #73
Merged
Conversation
Add nix/scripts/update-hashes.py and Platform: x86_64-linux Building node_modules_updater to compute correct hash ... Correct hash: sha256-2cnaDO1erE8kXD6iZtzAqd0hoLLl3IV1hwqXI9yoUhw= Updated x86_64-linux: sha256-2cnaDO1erE8kXD6iZtzAqd0hoLLl3IV1hwqXI9yoUhw= → sha256-2cnaDO1erE8kXD6iZtzAqd0hoLLl3IV1hwqXI9yoUhw= Done. You can now run: nix build .#deepagent-code app that builds node_modules_updater with fakeHash and extracts the correct hash automatically. No more manual hashes.json editing. The upstream nix-hashes CI already computes hashes for all 4 platforms; this adds local tooling for fork users and non-CI environments.
Wave 0 of CLI↔GUI parity plan. Replace the v2 half-surface (createRoutes from @deepagent-code/server) with the complete legacy server (Server.listen from deepagent-code/server/server), which is the only surface with a full session lifecycle — matching the GUI and the old 'deepagent serve'. - Add deepagent-code workspace dependency to cli package - Rewrite serve handler to call legacy Server.listen with mDNS/CORS flags; inject password via DEEPAGENT_CODE_SERVER_PASSWORD env (legacy auth reads env, not a listen argument) - Add session-client.ts migration seam centralizing session-execution calls for the future durable-engine swap - Add @/* path mapping and *.wasm ambient declaration so the deepagent-code source graph resolves during cli typecheck Verified: tsgo --noEmit passes, oxlint clean, serve responds on /api/health, /session, /provider, /mcp, /global/capabilities, /deepagent/packs/all (all HTTP 200); unauthenticated returns 401.
Wave 1 of CLI↔GUI parity plan. The default command (TUI) now accepts
-c/--continue, -s/--session, --fork, -m/--model, --agent, --prompt
and a project positional argument, matching the old 'deepagent' CLI.
- commands.ts: add params to root command with Flag.withAlias for
short names (-c, -s, -m)
- default.ts: assemble parsed flags into Args, resolve project to
working directory, validate --fork requires --continue/--session,
verify session existence via daemon client before entering TUI
- tui.ts: runTui accepts args and directory, replacing hardcoded {}
- daemon.ts: increase health-check retry from 5s to 60s to accommodate
legacy server startup time (full core initialization)
Verified: --help shows all flags with aliases, --fork without
--continue/--session errors, invalid session ID reports 'Session not
found', daemon auto-starts with legacy server.
Add 7 daemon-client commands for CLI↔GUI parity: - session list: SessionClient.list with table/JSON output and less paging - session delete <id>: SessionClient.deleteSession with NotFound handling - models [provider]: client.provider.list with --verbose JSON output - run [message..]: subscribe events → promptAsync → event loop to idle, supports --model/--agent/--format json/--continue/--session/--fork/ --variant/--dangerously-skip-permissions - export [sessionID]: SessionClient.get + messages reconstruction, --sanitize redacts sensitive transcript data - stats: aggregates cost/tokens from session list, table/JSON output - import <file>: /global/import endpoint (Codex/Claude format) with SSE progress stream consumption All commands follow dual-channel output contract (human-read TTY + machine-read --format json), use SessionClient migration seam for session execution calls, and pass typecheck + lint.
- auth login/list/logout: provider credentials via client.auth.set/remove - agent list: list agents via client.app.agents() - mcp list: MCP server status via client.mcp.status() - mcp add: persist MCP config via client.global.config.update() - All commands use daemon client (A mode) for consistency with Wave 0-2
Add 7 command groups for DeepAgent self-developed features: - packs list/pin/unpin (client.deepagent.packs*) - wiki list/get/search (raw HTTP, flag-gated: wiki) - oversight metrics/queue/approve/reject/ack/trace (raw HTTP) - review pending/approve/reject (client.deepagent.knowledge.*) - env-facts list/decide (client.deepagent.envFacts.*) - goal start/status/pause/resume/stop (raw HTTP, flag-gated: goalLoop) - panel status (raw HTTP, flag-gated: expertPanel) All commands use daemon client (mode A). Flag-gated commands check /global/capabilities and fail-close with a clear message when the feature is not enabled. Shared util.ts provides capability gating and raw HTTP escape hatch (client.client.request).
- attach: TUI connects to remote server with auth headers - db: database tools (path + SQL query with json/tsv output) - web: start full server + open browser - upgrade: self-update via Installation module - uninstall: remove files/config with dry-run and keep flags - pr: fetch PR branch via gh CLI then launch TUI - acp: ACP stdio server via @agentclientprotocol/sdk - github: delegate to deepagent-code binary for install/run - Add @agentclientprotocol/sdk and drizzle-orm dependencies
Repair over-deletions from f578e4f that left kept consumers unloadable, and continue its test cleanup: - migration.gen.ts: register 16 V4 migration files that existed on disk but were never wired into the registry (fresh DBs missed deepagent_workspace_config, session_steer, event bus, etc.) - session/sql.ts: restore SessionSteerTable (steer.ts service + V2 session-core steering still depend on it) - goal-loop.ts: restore readGoalTickCursor (goal-tick cold-recovery chain still consumes it; persistState keeps writing its format) - code-indexer.ts: restore symbolsForFilePaths (v4-event-runtime conflict-arbiter feed) - goal-manager.ts: restore goalStoreRoot export (v4-event-runtime imports it) - goal-loop-wiring.ts: restore SubagentTurnInput workspaceID/directory/ correlationID/maxTurnDurationMs (event turn runner still reads them) - goal-tick-port.ts: drop pendingPlanEdit/markPlanEditConsumed ports (API removed from core goal-loop + GoalDriverPorts) - delete obsolete prompt-cache-stability tests (volatile-tail design abandoned by the restructure) and stale world-state/im/wiki tests - builtin-agents.ts: drop removed maxFilesChanged limit values Remaining known desync (in progress): v4 flag reads in event-dispatcher, v4-event-runtime, session-completed-publisher, agent-push, supervisor-notifier, panel-convene-consumer; goal-driver API rework sync in goal-tick-port + goal-steer/steer tests; oversight/webhook httpapi registry; ~135 typecheck errors in packages/deepagent-code.
Their registrations were already removed from api.ts and server.ts in f578e4f; the group/handler files and the CLI oversight command (which called the removed /oversight/* endpoints) were dead code.
The generated SSE client's generator finally only releases the reader lock; the underlying fetch connection is never cancelled, leaking a socket that keeps the event loop alive. Pass an AbortController to event.subscribe and abort before breaking out of the consume loop.
discoverProviderModels requires apiKey: string; the cached wrapper's input has it optional. Default to empty string instead of passing undefined.
Definitions for v4PanelAutoConvene, v4AgentPushEnabled, v4EventDrivenIm, v4MultiAgentRuntime, v4EventDrivenArchive and v4Steering, read by the app IM capabilities and exercised by session/panel/server tests.
Adapt the remaining desynced consumers to the restructured goal driver: - goal-tick-port: inline the single-tick body (pause/stop checks + one loop.tick + status publish); the goal-steer relay (makeGoalSteerRelay, GOAL_STEER_DELIVERY, pendingSteer/markSteerConsumed ports) was removed from goal-driver/goal-loop-wiring/core goal-loop, so drop its wiring here and in v4-event-runtime (steerBuffer no longer needed). - multi-agent-runtime: drop the §C1 maxFilesChanged and §E2 maxTokensPerHour gates plus the token-budget tracker — both limit fields were removed from the agent limits schema, making the gates dead code. - Delete test files pinning removed APIs: goal-steer (steer relay), overflow (soft-landing state machine), plan-status-cache (volatile-tail cache design abandoned by the restructure), httpapi-webhook (webhook API dropped in 1db089f); trim removed-API cases from steer, multi-agent-runtime and model-discovery tests; adapt goal-tick-cold-recovery to the single-tick port. typecheck in packages/deepagent-code goes from 75 errors to 0. Four runtime test failures in steer/multi-agent-runtime predate this change (baseline: 8 failures on the same two files) and remain as restructure follow-up.
|
Thanks for your contribution! This PR doesn't have a linked issue. All PRs must reference an existing issue. Please:
See CONTRIBUTING.md for details. |
|
This PR doesn't fully meet our contributing guidelines and PR template. What needs to be fixed:
Please edit this PR description to address the above within 2 hours, or it will be automatically closed. If you believe this was flagged incorrectly, please let a maintainer know. |
Take upstream v4.0.3 semantics in the 6 conflicted files (plan-gate warn-only downgrade, PlanStore authority, V4.1 dual-path resume, sessions.clear, consecutive_blocks backfill, app version 1.4.2).
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.
Summary
Completes the downstream sync left over from the deepagent context pipeline
restructure (f578e4f / eeeac7e), and fixes two regressions found while
verifying the Wave 2–5 CLI commands against a live daemon.
oversight/webhookHTTP API modules whose service layerswere removed by the restructure; the orphaned files broke typecheck and the
CLI
oversightcommand targeted endpoints that no longer exist.dacode runhanging after completion: the session event stream is nowaborted on
session.status idle, cancelling the underlying SSE fetchconnection (the generated SSE client never cancels the fetch in its
generator
finally, leaking the socket and pinning the event loop).apiKeyin the provider discovery cache fetch wrapper.v4PanelAutoConvene,v4AgentPushEnabled,v4EventDrivenIm,v4MultiAgentRuntime,v4EventDrivenArchive,v4Steering) still consumed by app capabilities and test RuntimeFlagsoverrides.
logic in
goal-tick-port(pause/stop check +loop.tick+ status publish),remove dead steer-relay wiring and dead
maxFilesChanged/maxTokensPerHourgates, and delete/prune tests that targeted removedAPIs (
goal-steer,overflow,plan-status-cache,httpapi-webhook,plus pruned blocks in
steer,multi-agent-runtime,model-discovery,goal-tick-cold-recovery).Test plan
bun typecheckacross the workspace: deepagent-code package 75 → 0 errors,turbo typecheck 15/15 green.
serve --registerstarts andregisters cleanly;
runexits normally (~36s, previously hung past 150s);debug agents,agent list,models,session list,packs list,service statusall pass.steer,goal-tick-cold-recovery,multi-agent-runtime,model-discovery): 41 pass / 4 fail — the 4 failures are pre-existingruntime-behavior leftovers from the restructure (baseline was 8 failures on
the same files), unrelated to typecheck errors and noted for follow-up.