feat(daemon): install, uninstall and catalog hub agents from the daemon - #2477
Merged
Conversation
The TUI can only talk to agents over a stdin/stdout pipe, but the email agent has been a daemon-managed HTTP sidecar since #2191 — so it shows up in the hub greyed out and unusable. This plan scopes what it takes to make it a first-class TUI agent: install/run/uninstall from the hub and the command line, interactive-and-autonomous by default, and a settings screen. Also records two findings that shape the design: the CLI console auto-approves every gated tool (send, forward, permanent delete), and the canonical /query SSE contract cannot complete a confirmation-gated action.
Once the TUI can install, run, and configure the email agent, the gaia email subcommand is a second front door to the same sidecar with none of the UI. Scopes its removal, and pins the sequencing risk: it must land last, after the replacement is verified, or we delete the only working entry point first. Only the email subcommand is in scope — the rest of the legacy Python CLI is retired separately.
The user-journey design challenged ten points in the port plan; seven are accepted here. The biggest: preflight was scoped as late polish, but without it every transport failure reaches the user as a raw HTTP status — so it becomes Phase 1's exit test instead. Also drops 'always allow' on irreversible actions, replaces the three-level autonomy enum with fixed approval behaviour plus a background-work checklist, and moves background scanning from a TUI timer to launch time. Settings shrink to what actually persists, with the config and jobs APIs filed as #2470 and #2471.
…rt plan The plan told the decommission phase to keep src/gaia/daemon/agent_query.py because it was 'the generic daemon thin-client used by more than email'. That is false: run_query and ConsoleRenderer have exactly two callers, both inside the email handler being deleted, so removing the subcommand orphans all 346 lines. Left uncorrected, that phase would have kept dead code on a premise nobody would re-check. Reframed as an explicit choice — keep it as the reference implementation the Go transport mirrors (and say so in the module), or delete it with the subcommand — rather than a false statement of fact. Also ignores .claudia-worktrees/, which held six full checkouts one 'git add -A' away from being committed, and fixes two drifted line references.
The sidecar's QueryRequest.context is a required List with extras forbidden, so a null is a 422. A Go nil slice marshals to null, which means the obvious implementation fails on the very first turn — the one case with no history — and works on every turn after. A mocked 200 hides it entirely. Calls for initializing the slice and asserting the serialized body contains "context":[] rather than asserting the mock was called.
The TUI could only talk to agents it spawned itself as a subprocess, so the email agent — a long-lived HTTP sidecar the GAIA daemon starts, supervises and reverse-proxies — was visible in the hub but unreachable. The TUI is now a second thin client of the same relay `gaia email` uses: it discovers the daemon, ensures the sidecar, and streams the canonical seven-event `/query` contract. The subprocess transport is untouched, so the bash agent keeps working. Daemon discovery mirrors src/gaia/daemon: instance.json is trusted only after BOTH the recorded pid is alive AND a token-authed /daemon/v1/status probe answers with our service id and that same pid (a recycled port after a crash is a real case). Start-or-attach takes the same flock so two callers yield one daemon. The client token is re-read on any 401 — it rotates on every daemon restart — and never cached, logged, or printed; the sidecar's own bearer is discarded with the ensure response and never leaves the daemon. A stream that ends without a terminal final/error is reported as a failure, not an empty answer, and a turn is appended to the pushed transcript only when it terminated in `final` so a failed turn cannot poison the next one. Unknown event types and unparseable frames are surfaced instead of dropped. Also fixes three latent bugs found while in here: - an agent path containing a space was re-split on whitespace and never found; - cancelling a turn abandoned the read without stopping the child, so the tail of that turn's output surfaced as the next turn's events; - ChatModel held its streaming buffer in a strings.Builder, which panics the moment a copied non-zero Builder is written again — that would have crashed the TUI on the second streamed token of every answer.
Found by pointing the new client at a real running GAIA daemon rather than the test fake. Three fixes, none of which the fake server could surface: `Client.Do` defaulted to the status-probe HTTP client, whose 1.5s timeout is deliberately tiny because a live daemon answers loopback in milliseconds. Any relay call taking longer than that timed out: `GET /v1/email/init` does real work (probing Lemonade, the model and connectors) and failed with a misleading "context deadline exceeded" instead of returning its readiness report. Nothing shipped today relied on the default — every current caller passes its own client — but the Phase 2 install/catalog calls and the Phase 5 preflight screen would each have walked straight into it. Do() now defaults to a general 60s client and the probe timeout is confined to the probe. A cancel POST for a run that is no longer in flight answers 404 by design, which is exactly the case when we abandon a stream that already ended. That was being logged as a failed cancel; it is now reported as "already finished". Also: a `data` line with no colon carries an empty value per the SSE spec, which the reference implementation's str.partition yields but strings.Cut discarded; and Close() landing while a query request is in flight no longer registers an un-cancellable run.
…try point Two independent reviews of the Phase 1 transport, plus a run against a real daemon. Everything below was found after the first commit was already green on 106 tests, which is the argument for the review pass rather than against it. Cancelling a subprocess turn emitted a spurious "agent process exited with code -1" bubble under the "cancelled" line about half the time: killing the child made the reader see a non-zero exit and a closed pipe, and it reported both as agent failures. A deliberate kill is not an error. The `emit` helper also selected over a ready channel send and a ready ctx.Done() simultaneously, so Go picked at random — it is now deterministic. Reproduced at 5/12 runs before the fix; a regression test now runs the cancel path eight times. The same cancel path could panic the TUI. `Send` read the stdin pipe outside the mutex while the cancel goroutine nil'd it, so a turn started right after a cancel could write to a nil pipe or scan a nil scanner. Turn state is now captured under one lock acquisition. Relatedly, killing also reaped the child while the reader was still reading its stdout pipe, which os/exec explicitly forbids — reaping is now the reader's job alone. A second test drives ten cancel-then-send cycles. The turn after a cancel could also silently produce nothing: the early-return paths skipped the reset, so the next turn reused the child that had just been killed. Found by the new test, not by review. A late-arriving event from a cancelled turn could tear down whichever turn was running by then — Bubble Tea cannot cancel a dispatched Cmd, so the old waitForEvent goroutine delivers whenever its channel finally closes. Events now carry the channel they came from and a superseded turn is ignored. The SSE client no longer holds that channel open behind a 10s cancel POST either. Nothing bounded the wait for response headers, so a daemon that accepted the connection and never answered hung the caller forever with a spinning cursor — the Python reference gets this for free from its read timeout. A recycled pid whose port was answered by an unrelated process permanently blocked daemon start, even though the probe had already proved the record was garbage; stale records are now classified so only our own wedged daemon is left alone. A 401 from a relayed call now blames the sidecar instead of sending the user to restart the daemon. Finally, the transport had no user-reachable entry point at all — roughly 1,700 lines that a reviewer could not exercise, and that Phase 2 would have been the first thing to drive. `gaia tui chat --agent <id>` now runs any catalog agent over whatever transport it declares, and with `--query` it is a genuine one-shot: answer on stdout, progress on stderr, exit 0/1, no alt screen. Driving it against the live daemon showed a bare "HTTP 404", so that case now explains that the installed sidecar predates the canonical /query contract. Transport selection moved off the Bubble Tea model into client.ForAgent so the headless paths do not need a UI, and the cancel path moved out of the daemon control-plane client into the SSE transport that owns the agent wire contract. The legacy event vocabulary is now marked frozen so new UI features are not built twice. Correcting the first commit's message: it said the daemon token is "never cached", which is stronger than the code. The token is re-read on every 401, never logged and never printed, and the sidecar bearer is discarded — but the last verified instance is held in memory so a run can be cancelled.
Teaches the TUI to reach agents through the daemon relay instead of only spawning a local binary over a pipe — the prerequisite for the email agent, which has been a daemon-managed sidecar since #2191 and so could never work over the subprocess path.
There was no way to exercise the TUI without a human at the keyboard, and no
way for a user to watch an assistant use it. `gaia tui --control` now exposes a
loopback HTTP API that injects keys into the *running* Bubble Tea program and
serves back the frame it just rendered, and an MCP server wraps that as tools —
so Claude Code can navigate the hub, filter, type, and launch agents while the
user watches the same terminal update in real time.
Go side (tui/internal/control/): a Recorder wraps the root model so every
View() is cached, and named keys map to real tea.KeyMsg values. The existing
smoke tests send tea.KeyMsg{Type: KeyRunes, Runes: []rune("tab")} — three
letters, not the Tab key — so the mapping is contract-tested by asserting
KeyMsgFor(name).String() round-trips to the name the handlers switch on.
POST /wait is what makes automation reliable: it blocks on a broadcast channel
until the screen or the reported model state matches, and a timeout reports
what the screen actually contained. Callers no longer busy-poll.
Injection is synchronous end-to-end: tea.Program.Send only queues, so a
trailing sentinel message is used to answer /keys only once every key in the
batch has been handled and drawn. Without it a four-key batch returned with one
key applied.
Python side (gaia.mcp.servers.tui_mcp, `gaia mcp tui`): discovery reads
~/.gaia/tui/control.json (0600) and trusts it only when the recorded pid is
alive and a token-authed probe answers with a matching pid — a recycled port
after a crash is a real case. Loopback-only hosts are enforced client-side too,
since the bearer token rides every request.
Test plan:
- cd tui && make lint && go test ./... -race
- pytest tests/unit/test_tui_mcp.py
- pytest tests/integration/test_tui_control_e2e.py (boots the real Go server
and drives it with the real Python client; skips without a Go toolchain)
Rebasing onto the transport work collided two same-named helpers in package ui — the one-shot test's local run() and the control server's run(). Renamed the test helper to captureOneShot; the production name stays. Also corrects which server owns port 8766 (the MCP bridge, not agent_ui_mcp).
Lets an assistant drive the live TUI the user is watching — read the rendered screen, send keys and text, wait on conditions — so the terminal UI can finally be exercised and regression-tested without a human at the keyboard. Its own review pass caught six real bugs, including a key batch that was acknowledged before it was applied and stdio logging that corrupted the JSON-RPC stream (the latter also affects the shipped Agent UI MCP — #2472).
Before: an agent driven from a CLI or a subprocess executed every
confirmation-gated tool with no prompt at all. `OutputHandler.confirm_tool_execution`
returned True unconditionally and no CLI console overrode it, so the entire
mechanism was inert outside the Agent UI. In practice that meant `gaia email -i`
could send, forward, RSVP to, and permanently delete mail on the model's say-so,
and any agent could run shell commands and overwrite files, with no prompt and no
event the user could see.
After: gated tools prompt on the terminal — the tool name, its literal arguments,
and a yes/no answer that defaults to no (Ctrl-C and EOF also mean no). "Always
allow" is per tool and per console, and is not offered for run_shell_command /
run_cli_command, where the name says nothing about the next call. With no terminal
to ask on — piped stdin, redirected stdout, CI, a subprocess host — the tool is
denied with an actionable error naming both ways forward, instead of running
silently. The agent loop still returns {"status": "denied"}, so the model can
explain itself rather than the run crashing.
Details worth knowing:
- Silence covers narration, not consent. `gaia chat` builds a SilentConsole even
for a user sitting at a terminal, so the prompt lives in a mixin shared by both
CLI consoles; otherwise the flagship path would have gone from "runs gated tools
unasked" to "refuses them unasked".
- The gate itself never raises. A malformed tool_args from the model, a detached
tty that errors on read, or a closed stdout all resolve to a logged denial —
because the tool used to just run, an exception escaping here would be a new way
to kill a working session, and `Agent._execute_tool`'s callers do not catch.
- The unattended opt-in (GAIA_AUTO_APPROVE_TOOLS=1, or auto_approve_gated_tools on
the console) is read from the environment as it was at startup, so a
project-local .env cannot switch off a user's prompts. Every approval taken that
way is logged with the tool name.
- The Agent UI path is untouched: it still emits permission_request, blocks on the
modal, and denies on timeout; a new test pins that end to end through the gate.
Background/autonomous ticks keep their existing auto-deny.
- The OpenAI-compatible API server and scheduled UI runs have no channel to ask a
human, so they now deny rather than approve. Documented in the CLI reference,
troubleshooting, the console spec, and the API-server and output-handler pages.
Denial reasons are bound to the tool they belong to, so a handler that denies
without recording one can never report another tool's explanation. Argument
shortening is display-only — an approved write receives the full content.
Confirmation-gated tools — send, forward, permanent delete, RSVP, shell, file write — executed with no prompt and no record whenever an agent was driven from a CLI or subprocess, because the base handler answered yes on the user's behalf. It now denies by default and the terminal console genuinely asks. Unattended hosts opt in with GAIA_AUTO_APPROVE_TOOLS=1, which deliberately cannot be granted by a .env file — a project file travelling with a directory is not an operator's decision. The SSE approval path is unchanged.
`gaia_agent_email/cli.py` was an in-process argparse entry point that nothing imported and no console script exposed. Its docstring still claimed `gaia.cli.handle_email_command` invoked it, which stopped being true at #2191 when that handler became a thin daemon relay — so the file was both dead and actively misleading about how a query reaches the agent. The sidecar (`gaia-agent-email serve` -> POST /query) is the only supported entry point; an in-process CLI would bypass its caller auth, trust gate, and SSE translation. Adds a guard test so the module can't creep back silently.
…orecard The npm package shipped two different scores for the same version: README and EVALUATION advertised 83.4/100 while the co-located SCORECARD.md — machine- generated from a real 3-run eval for v0.5.0 — reported 84.53. 83.4 is the 0.3.0-era number; the prose was never refreshed when the scorecard was regenerated, so an integrator comparing the two pages saw the package contradict itself about its own benchmark. SCORECARD.md's aggregate.value is the source of truth, so the prose moves to it. The CHANGELOG's 83.4 under 0.3.0 stays as-is: per-version entries are a historical record, and rewriting one would falsify the release history. Adds a guard that fails whenever a prose score claim drifts from the generated aggregate again, since prose does not regenerate with the eval.
…f hardcoding it The generated capability matrix contradicted itself inside one file: the Definitions section said the REST API has "16 functional verbs" while the matrix header five lines below computed 21 from the live OpenAPI spec. Anyone reading the doc to size the REST surface got the wrong number from the more prominent of the two. The 16 was a literal inside TOOLS_COUNT_DEFINITION, so the existing byte-identical drift guard could never catch it — regeneration reproduced the same stale literal. The blurb is now a template fed by the derived counts, and the two stale literals that had spread to a module comment and the gaia-agent.yaml comment are gone with it. Regenerated: 16 -> 21.
…nswer
Two swallowed failures on the same SSE path made an internal crash look like
the agent simply had nothing to say. The queue drain caught bare `Exception`
where it meant `queue.Empty`, so any non-Empty error read as "queue drained";
the stream then synthesized `{"type": "run_complete", "answer": ""}` — a
well-formed *successful* completion — whenever the worker thread had died
without emitting a terminal event. A client could not tell the two apart.
The drain now catches `queue.Empty` only, matching the sibling query route,
and a dead worker emits an `error` event naming the failure and the sidecar log
to check before the stream closes.
…profile Two degradations that left no trace at all. A malformed GAIA_EMAIL_TRIAGE_MAX_MESSAGES silently reverted to the 100-message default, and a negative or zero value was clamped to 1. The eval harness sets that variable to cover the whole labelled corpus, so a typo produced a quietly smaller scan and therefore a quietly different score with nothing to notice — exactly the regression-hiding the rule exists to stop. It now raises ConfigurationError naming the bad value and the way out, matching config._undo_window_seconds, which already handled the identical shape this way. An unreadable sender-profile row was reset to an empty record with no log, throwing away the reply-latency and interaction history that drives priority-sender promotion, while the sibling read paths log the same corruption class. The reset stays fail-soft — one bad row must not kill behavioral learning — but now says which sender lost what.
…CLI claim CONTRACT.md — the doc that calls itself the contract reference — declared schema 2.3 while contract.py has been on 2.5 since #2154. Both additive bumps (2.4's /v1/email/query agent loop, 2.5's /v1/connections OAuth intake) shipped without a history row, and every copy-pasteable JSON example told integrators to send the stale version. The same file still explained that the interactive CLI loads the model at 32K via an `agent_context_sizes` registry in src/gaia/cli.py; that registry no longer exists and no CLI loads a model at all since #2191 made every front-door relay to the sidecar. npm/SPEC.md stated the contract version as 2.4 without saying whose — true of the client, wrong for the sidecar it documents. Now says both and why 2.4 requests still work. Adds guards tying CONTRACT.md's headline, examples, and history table to contract.SCHEMA_VERSION, so the next bump cannot land doc-less.
…tops startup Follow-up to the fail-loudly pass: raising from _inbox_scan_ceiling() achieved nothing. Both scanning tools call it inside their own `except Exception -> _envelope_err`, so the ConfigurationError was caught two lines later and handed to the LLM as a per-call error string. An eval run with a typo'd GAIA_EMAIL_TRIAGE_MAX_MESSAGES would have failed every triage call, apologised, and still written a scorecard — worse than the silent default it replaced. The resolver moves to config.py beside default_undo_window_seconds and becomes an EmailAgentConfig field, so a bad value fails construction, where an operator sees it. Blank now means "unset" on both, matching that precedent. Also from review: the profile-reset warnings logged a raw sender address at WARNING, which gaia diagnostics bundles into user-attached bug reports — they log the store row id instead. CONTRACT.md's new version-policy sentence contradicted the 2.3 row it introduced (a breaking change shipped as a MINOR, which clients gate on MAJOR alone and never saw), and its ctx note overreached: gaia eval benchmark does still build the agent in-process. The docs site repeated the stale 2.3 schema claim and is now guarded with CONTRACT.md.
… environment The config field added in the previous commit was validated at construction and then never read: both scanning tools still called the env resolver on every invocation. `EmailAgentConfig(inbox_scan_ceiling=500)` was accepted and silently ignored, and because the resolution happened per call the ceiling could move underneath a live agent when the environment changed mid-run. The registrar now resolves it once, preferring the config and falling back to the same env resolution for the duck-typed configs some hosts pass (the existing `debug` flag reads the same way). Tests cover all three: explicit config wins, a config without the field still works, and the value is stable across calls. Also drops the ✅ the debug playground printed on an empty `run_complete` — a dead worker now emits an error event immediately before it, and a green check after that read as success. Reverts drive-by reformatting of 13 untouched files from a whole-tree black run.
Fixes a case where the agent could stream a blank answer as success after its worker had died, and makes a bad inbox scan ceiling stop startup rather than degrade quietly. Also settles two doc contradictions at the source: the package advertised an eval score its own scorecard disagreed with, and the capability matrix hardcoded a verb count that its own generated table contradicted. Deletes the dead in-package CLI module retired by #2191.
The email→TUI port had an engineering scope but no user experience: a list of capabilities, not a journey. This walks all eleven stages a person actually passes through — from `gaia tui` on a bare machine to a triaged inbox — with ASCII mockups for every screen state, a keybinding map checked against what the hub and chat models already bind, and an explicit build order. The finding it argues: the port's hard problem is the four-precondition readiness gate (local LLM, model, daemon, mailbox), not the transport. Every one of those is already reported by an existing endpoint with an actionable hint, so seven of the ten highest-value fixes need no backend change at all. Pushes back on the engineering plan in ten places, notably: preflight is Phase 1's acceptance criterion rather than later polish, "always allow" on irreversible actions should not ship, the three-level autonomy enum should be two plain controls, and background timers do not belong inside a foreground TUI.
Review pass against the code found four things the doc asserted that aren't true as written, each of which would have sent an implementer down a dead end: - `ctrl+b` and `ctrl+t` were proposed as free chat keybindings. Both are taken by bubbles/textarea's default keymap, which receives every key the chat model doesn't handle. Moved to `ctrl+g` / `ctrl+r` and documented the full set of chords the textarea reserves, since checking only chat/model.go misses them. - The quit message told users to stop the background service with `gaia kill`, which kills a process by port and cannot stop the daemon. It is `gaia daemon stop`. - `GET /daemon/v1/catalog` was referenced as if it exists; it is proposed in the engineering plan's Phase 2. - `gaia tui` itself does not exist — no `tui` subparser in the CLI, and the Go binary's cobra root is already named `gaia`, so installing both yields two different `gaia` commands. Adds a "what exists today vs. what this design assumes" table so every proposed surface is labelled in one place, and tightens the `diagnose()` line range.
Defines the end-to-end experience the port is building toward, using email as the driving case: what a first-time user sees, where the walls are, and which fixes are worth the most per hour. Its pushback reshaped the engineering plan — most importantly, the readiness gate moved from late polish to Phase 1's exit test, because without it every transport failure surfaces as a raw HTTP status.
The plan's email paths predated the #2060 restructure, so every reference pointed at a directory that no longer exists. Also splits the bug list by status: two are fixed and merged, two remain open, and four are closed as moot because they are flag behaviour on the subcommand Phase 5b deletes — reading them as open would send someone to fix a command being removed. Adds the hub tab-cursor bug found while the control API drove the TUI.
Until now the only way to install a GAIA agent was the web UI's FastAPI
server on port 4200 — no CLI, no daemon API, so the TUI and any scripted
setup were blocked. The daemon now owns install, which means the TUI, the
CLI and the Agent UI share one installer, one SHA-256 check and one
install lock instead of three forks of the logic.
New routes (all behind the existing daemon client token):
GET /daemon/v1/catalog hub catalog + installed state
POST /daemon/v1/agents/{id}/install 202, queue an install
GET /daemon/v1/agents/{id}/install-status progress + terminal error
DELETE /daemon/v1/agents/{id} stop, verify, remove the dir
Safety semantics that are not negotiable and are covered by tests:
- The install dir IS the sidecar's binary cache, so install and uninstall
stop the sidecar first and hold it stopped for the whole mutation via a
new SidecarRegistry.hold_for_mutation(): an ensure arriving mid-download
waits instead of respawning the process from the directory being
rewritten. A pid that survives the tree-kill aborts the operation.
- Install and uninstall of one id share a slot, so a DELETE can never
report success while a download is still writing into that directory.
- Integrity is the hub manifest's server-computed SHA-256, verified before
anything is written; a mismatch leaves nothing installed and there is no
"use it anyway" path.
- Reserved built-ins are refused, and ids are validated before they become
a path segment under ~/.gaia/agents.
- The catalog only offers agents the daemon has a sidecar spec for, and
reports the ids it hid rather than dropping them silently.
Also drops the daemon's dead dependency on the committed
binaries.lock.json (placeholder SHAs, and not shipped in the wheel): a
missing lock now means "not installed yet" and names `gaia hub install`
as the remedy instead of blaming a broken install.
There was no way to install an agent from a terminal at all — setup meant opening the Agent UI. `gaia hub` drives the daemon's install routes, so a scripted or headless machine can now provision agents, and it shares the UI's installer rather than forking it. `list` shows the catalog with installed versions and an update marker; `--installed` answers from the local .installed sentinels and needs no network; `--refresh` bypasses the 5-minute catalog cache. `install` polls and prints each phase, exiting 1 with the reason on failure. `uninstall` stops the agent's sidecar and verifies its pid is gone before removing anything. None of these renderers can print the sidecar bearer token.
The registry only verifies the pid it still tracks, so a sidecar the daemon lost track of would have its binary replaced underneath it. That is not hypothetical: on this machine a daemon-spawned email sidecar (pid 31647, port 57920) has been serving /health for an hour while `GET /daemon/v1/agents` reports the agent as `stopped` — the daemon spawned a second process 12 seconds after the first without killing it and dropped the reference. Install and uninstall now also ask the OS what is actually running out of the agent's directory and abort with the surviving pids and a remedy if anything is. exe/argv[0]/argv[1] only, so a log path passed to another tool cannot block an install.
Upstream's installer gained a trust gate: any agent outside the `verified` security tier now needs `trusted=True`, because installing it runs third-party code on the user's machine. The daemon install path did not pass it, so `POST /daemon/v1/agents/email/install` and `gaia hub install email` failed for every user — email is `experimental`. The opt-in is threaded end to end: a `trusted` field on the install body, `--trust` on `gaia hub install` (matching `gaia agent install`), and `trusted=` into `installer.install`. It defaults to FALSE everywhere and is never inferred — an absent body, an empty body and an explicit false are all refusals. A missing opt-in is refused synchronously with 403 (not a generic 500, and not an opaque background failure), naming the exact remedy, so a TUI or UI can render "Trust & Install" and retry. The id checks run first, so an unknown or reserved id still answers 404/400 rather than prompting for trust on an agent that cannot exist.
github-actions
Bot
added
documentation
Contributor
|
Verdict: Approve with suggestions. This moves agent install/uninstall/catalog into the daemon so the TUI, Two non-blocking things to consider before merge:
Nice work — the fail-loudly discipline and the deliberate "keep it deleted" guard tests are exactly the right instincts. Real-world evidence
🔍 Technical details🟢 Minor — 🟢 Minor — real-world evidence for the CLI surface (see visible section above) Strengths
No prompt-injection content detected in the diff. |
`gaia tui install nope` (a typo), `install bash` (subprocess-only) and `install chat` (a real agent) all returned the same sentence, and it led with `builtin_specs()` — a Python symbol the user cannot act on — while burying the one useful clause, the list of installable agents, mid-way. The refusal now leads with the user's next move and demotes the internal symbol to a labelled developer note at the end: no installable agent 'nope' — installable: email. See the full catalog with `gaia hub list`. (Developer note: ...) It also stops reporting a real agent as a typo. Three cheap local probes — the install sentinels, the cached hub index, and this environment's agent entry-point names (metadata only, nothing imported) — distinguish "installed but not supervised", "published on the hub but not runnable here", and "a GAIA agent, just not published as a sidecar". Classification is advisory: each probe is guarded so a failure logs and falls back to the generic wording rather than masking the refusal. Uninstall named what failed and where to look but never what to do; it now points at `gaia hub list --installed`.
The uninstall guard refused correctly and named the pids, then handed the
user two commands that both fail:
kill 88154, 88156 -> kill: illegal pid: 88154, (the ", ".join)
gaia kill -> refuses; its surface is --port/--lemonade, so
no invocation targets an agent by install path
It now prints a single space-separated `kill -9 <pids>` (taskkill /F on
Windows) and says why plain kill is not enough: these processes ignore
SIGTERM, which is how they outlived the daemon that spawned them in the
first place. `gaia kill` is gone from the message.
Guarded by tests that RUN the advice rather than read it: the generated
command is executed against real throwaway processes and their death is
asserted, and a companion test pins that the old comma form leaves the
first pid alive (its exit code is not the tell — bash complains and still
exits 0, zsh reports `illegal pid` and exits 3). Every `gaia ...` string
this package emits is also parsed against the real argparse surface, which
caught two docstrings advertising `gaia hub install` with no agent id.
kovtcharov-amd
previously approved these changes
Jul 27, 2026
kovtcharov-amd
left a comment
kovtcharov-amd
left a comment
Collaborator
There was a problem hiding this comment.
Approving — daemon-side install/uninstall/catalog is careful, security-conscious work: the trust gate is default-deny and enforced twice, agent-id path traversal is validated at the boundary (traversal ids return 404 with the victim dir surviving, tested), directory mutation is guarded by a registry pid-check plus a live-process scan (tested against a real child process), a SHA mismatch leaves nothing installed, and the slot always releases. Fail-loudly throughout, with cold-state/concurrent-install/token-hygiene tests.
One nit: HubUnavailableError maps to 503 on the catalog route but 502 on install — same failure, two codes; 503 reads more accurately for both.
kovtcharov-amd
previously approved these changes
Jul 27, 2026
kovtcharov-amd
enabled auto-merge
July 27, 2026 08:57
# Conflicts: # docs/plans/email-agent-tui-port.md
kovtcharov-amd
approved these changes
Jul 27, 2026
pull Bot
pushed a commit
to bhardwajRahul/gaia
that referenced
this pull request
Aug 13, 2026
# GAIA v0.23.0 Release Notes GAIA v0.23.0 makes the agents easier to get, safer to run, and easier to extend. You can now browse, install, and run agents straight from the terminal with `gaia hub`, and add new capabilities to an agent as signed, auditable skills. Under the surface it's a security release: the local API and MCP bridge no longer expose themselves to the network by default, the confirmation prompt that pauses an agent before it sends mail, writes a file, or runs a command now works from the terminal and over the local API and MCP — not just inside the graphical app — and an agent can no longer quietly reach into your `~/.gaia` config or slip crafted SQL into the database agent. Connecting a Microsoft account is now an explicit Personal-or-Work/School choice with a zero-setup sign-in. **Why upgrade:** - **Get agents from the terminal** — `gaia hub` browses, installs (behind a trust prompt for unverified agents), runs, and removes agents without leaving the shell. - **Every agent asks before it acts** — the confirmation gate for sending mail, writing files, and running commands now works in the terminal, over the local API, and across MCP tools, not only in the Agent UI. - **Safer by default** — the MCP bridge binds to localhost, the local API refuses credentialed cross-origin requests from arbitrary sites, MCP servers launch without a shell, and an agent can't write into `~/.gaia` or reach the database with crafted SQL. - **Build and share skills safely** — `gaia skill` makes skills first-class: create, import, sign with trust tiers, and audit them before sharing. They're opt-in — you add the ones you want. - **Connect a Microsoft account without a secret** — Personal and Work/School are now separate connectors with device-code sign-in and no client secret required. <Note> **The email agent is beta and CLI-first this release.** It runs locally and never sends, forwards, or deletes without your confirmation — that safety gate is verified. This cycle was mostly robustness and correctness: sturdier Outlook and calendar handling, honest reporting when a scan is truncated, and a long list of fixes (see Bug Fixes). It's still early — a full inbox triage can currently time out on larger mailboxes, and autonomy is experimental and not yet wired up in the packaged sidecar. Treat its output as a draft to review, and please report what you run into. </Note> ## Breaking Changes ### `GAIA_MICROSOFT_TENANT` is gone The Microsoft connector was split into two explicit connectors — Personal and Work/School — each with its own hard-coded tenant, so the `GAIA_MICROSOFT_TENANT` environment variable no longer does anything and has been removed (PR [amd#2729](amd#2729)). If you set it to work around the old single-connector tenant guessing, drop it and pick the connector that matches your account instead (see *Microsoft accounts* below). ## What's New ### Install and run agents from the terminal — `gaia hub` Getting an agent used to mean the graphical app or a manual pip install. Now the hub is in your shell: `gaia hub list` shows the catalog, `gaia hub install <agent> --trust` installs one (the `--trust` is required for an unverified agent — it will not install silently), and `gaia hub uninstall <agent>` removes it. The install → run → uninstall round-trip works end to end against the live catalog, with the trust prompt actually enforced (PRs [amd#2484](amd#2484), [amd#2530](amd#2530), [amd#2708](amd#2708)). Try it: `gaia hub list`. ### Every agent asks before it acts — beyond the Agent UI The confirmation prompt that pauses an agent before a consequential action — sending or deleting mail, writing a file, running a shell command — used to work only inside the Agent UI; from a terminal, the local API, or an MCP tool call those actions could run unprompted. This release closes those paths: the gate now fires from a plain terminal, through the `gaia api` server, and across MCP tool calls, classifying a tool as read-only or mutating and failing closed when unsure. The agent stops and asks before the action, declining leaves nothing changed, and setting `GAIA_AUTO_APPROVE_TOOLS=1` in your environment is the explicit way to opt out (PRs [amd#2475](amd#2475), [amd#2544](amd#2544), [amd#2846](amd#2846), [amd#2854](amd#2854)). ### Safer by default — a security-focused release Several local exposures are closed this release. The MCP bridge binds to `127.0.0.1` by default instead of every interface, so it isn't reachable from other machines on your network unless you pass a bind-all host, and it can now require an `--auth-token` that is actually enforced rather than ignored. The local API server no longer echoes an arbitrary origin back with credentials allowed — a cross-origin request from a site that isn't allow-listed is refused. MCP servers are launched without going through a shell, so a server name can't smuggle shell metacharacters; an agent can no longer write into your `~/.gaia` configuration; and SQL supplied by the model is blocked from reaching the database agent's statements rather than being executed (PRs [amd#2246](amd#2246), [amd#2238](amd#2238), [amd#2344](amd#2344), [amd#2844](amd#2844), [amd#2847](amd#2847), [amd#2860](amd#2860)). ### Build and share skills — `gaia skill` A skill gives an agent a new capability from a folder with a manifest — no new Python, no forking the agent. `gaia skill create <name>` scaffolds one, `gaia skill import` adds a skill so an agent can discover it, and `gaia skill list` / `info` show what's installed and the permissions each one declares. Sharing is guarded: skills carry signatures with trust tiers — an unsigned or untrusted skill is capped at the lowest tier, and tampering is caught by checksum — a pre-publish audit rejects a skill that attempts prompt injection or `shell=True`/`eval`, and `gaia skill migrate` converts skills authored in other formats. Skills are opt-in: no agent loads them automatically yet, so you add the ones you want (PRs [amd#2669](amd#2669), [amd#2692](amd#2692), [amd#2702](amd#2702), [amd#2693](amd#2693)). Try it: `gaia skill list`. ### Microsoft accounts: Personal and Work/School, no secret required Connecting a Microsoft account is now two clear choices instead of one connector guessing your tenant. `gaia connectors list` shows Microsoft as two distinct connectors — Personal and Work/School — each with a zero-setup device-code sign-in (a code and a URL to visit) and no client secret required for a public app registration (PRs [amd#2718](amd#2718), [amd#2364](amd#2364)). ### Lemonade Server 11.5.0 This release runs against Lemonade Server 11.5.0 — the version installed by `gaia init` and pinned across CI and the installer. ## Bug Fixes A selection of the user-visible fixes this release — the full list is in the changelog below. - **Triage now paginates large inboxes and reports truncation honestly** (PR [amd#2646](amd#2646)) — no more silently dropping mail past a hidden limit. - **Thread messages come back sorted and numbered** (PR [amd#2570](amd#2570)) — "reply to 3" hits the message shown at position 3, not raw backend order. - **The inbox pre-scan stops reporting a guess as a verdict** (PR [amd#2587](amd#2587)) — an uncertain classification is surfaced as uncertain. - **An email conversation survives its turns** (PR [amd#2837](amd#2837)) — a follow-up question keeps the session context instead of starting over. - **A bare reconnect no longer guts a mailbox** (PR [amd#2733](amd#2733)) — reconnecting an account keeps its existing grants instead of wiping them. - **Restore from Trash anytime** (PR [amd#2542](amd#2542)) — undo an archive/trash without a narrow time window; the dead permanent-delete path was removed. - **A low-priority sender no longer forces a promotional label** (PR [amd#2774](amd#2774)) — sender priority stops overriding the actual content classification. - **The agent survives an OpenMP double-init** (PR [amd#2508](amd#2508)) — a mid-conversation native-library clash no longer kills the run. - **Large tool results are truncated to valid JSON** (PR [amd#2645](amd#2645)) — an oversized result no longer produces unparseable output. - **A missing model surfaces as a real 404** (PR [amd#2245](amd#2245)) — the builder names the missing model instead of a generic placeholder. - **GPU is detected on all platforms and default_device is honoured** (PR [amd#2244](amd#2244)). - **A sidecar that is alive but has stopped serving is now detected** (PR [amd#2707](amd#2707)) — a wedged agent process is caught instead of hanging. - **The model-slot lease is held across inference, not just the load** (PR [amd#2394](amd#2394)) — a second agent can't evict the model mid-generation. - **A browser that never launched is surfaced** (PR [amd#2507](amd#2507)) — a failed OAuth browser open reports an actionable error instead of hanging. - **Stop actually aborts in-flight streaming** (PR [amd#2166](amd#2166)) — the Agent UI Stop button ends generation immediately. ## Full Changelog **331 commits** since v0.22.0: <details> <summary>Expand full changelog (331 commits)</summary> - `96ce0d1e` — feat(email): recognize the third mailbox connector — Gmail, Outlook personal, Microsoft work (amd#2896) - `88ecc18e` — fix(email): scoped 'anything suspicious?' query no longer dumps the full triage report (amd#2910) - `1c75ccee` — fix(skills): reject version pins GAIA cannot read instead of matching them (amd#2928) - `29a5c364` — docs(plans): assess readiness for the generic gaia-agent (amd#2926) - `e3b28653` — fix(tui): stop cancel-then-resend from racing the daemon's session lock (amd#2912) - `b38f5e60` — fix(tui): anchor TTFT on first inference token, use real token counts (amd#2911) - `779d7b00` — fix(email): reply/draft/send actions no longer report failure after they already succeeded (amd#2908) - `f8d4610c` — ci(eval): queue eval runs instead of cancelling the pending one (amd#2921) - `611b1928` — feat(hub,skills): pre-publish security audit gate for marketplace skills (amd#2702) - `c83047b9` — feat(skills): gaia skill migrate — OpenClaw/Hermes skills to GAIA format (amd#2693) - `ca17067a` — test(skills,hub): cover the signing and lane checks amd#2668/amd#2692 shipped untested (amd#2907) - `a951567a` — fix(eval): drop the judge temperature pin the model now rejects (amd#2905) - `236f8a58` — feat(skills): gaia skill publish/install with signature-backed security tiers (amd#2692) - `d26b15da` — ci(eval): trigger the Gemma eval on the PR diff, not the push (amd#2897) - `be15258a` — fix(cpp): normalize CRLF so Windows-authored skills parse identically (amd#2906) - `84415b06` — feat(cpp): text extraction and chunking (amd#2822) - `6352f264` — fix(database): make the SQL read-only authorizer disarm on Python 3.10 (amd#2904) - `4709a728` — feat(hub,skills): publish and serve skills as a first-class hub catalog lane (amd#2668) - `245cb72e` — fix(api): refuse approval-gated tools instead of auto-approving them (amd#2854) - `e803103a` — fix(mcp): enforce --auth-token on the MCP bridge instead of ignoring it (amd#2844) - `a6dc1fa3` — fix(ci): run the Gemma-4-E4B agent eval in PowerShell on its Windows pool (amd#2773) - `75b61f4f` — feat(cpp): SKILL.md format parser and validation (amd#2824) - `b89c6414` — feat(cpp): gaia::HttpClient — a general HTTP client abstraction (amd#2809) - `5e6b1d43` — fix(security): launch MCP servers without a shell, protect ~/.gaia from agent writes (amd#2847) - `768be2e2` — feat(skills): ship a ten-skill starter pack with a guide and honesty guards (amd#2697) - `340bf2c2` — feat(cpp): interactive TUI — event loop, streaming render, modals (amd#2825) - `acaad161` — feat(cpp): native OpenAI tool calling and conversational response mode (amd#2821) - `23ae5d63` — feat(cpp): SQLite integration and gaia::Database (amd#2816) - `4f2a6686` — feat(cpp): gaia::VectorIndex — flat vector index with persistence (amd#2807) - `b1a7e8df` — fix(email): ship with Agent Skills off until the eval gate covers them (amd#2848) - `34c26577` — ci(webui): run the Agent UI Vitest job on Node 22 so the suite executes (amd#2898) - `20036474` — docs: escape MDX-breaking literals so Mintlify validation passes (amd#2903) - `07b0858e` — ci(tui): run on stacked PRs, add -race, lint, and a per-OS test matrix (amd#2696) - `e19d46be` — feat(cpp): harden the coding toolbelt — stale-write rejection, ignore-aware search, persistent shell (amd#2823) - `7621b223` — ci(claude): move every Claude workflow to Opus 5 and run the audits nightly (amd#2859) - `d086faa4` — fix(claude): correct the agent/skill prompts and make plain language the default (amd#2862) - `990d19ff` — fix(cpp): gate MCP tools behind user confirmation in the C++ SDK (amd#2851) - `878d473f` — fix(mcp): gate MCP write tools behind the user confirmation prompt (amd#2846) - `c2c68064` — fix(security): stop LLM-supplied SQL reaching DatabaseAgent statements (CWE-89) (amd#2860) - `e936d3d9` — docs(release): rebuild v0.23.0 notes to hardware-verified features only - `be03c7c8` — fix(init): stop reporting success gaia init did not deliver (amd#2889) - `75804a15` — docs(dev): require --extra-index-url for every uv pip install on Linux (amd#2878) - `4a24d663` — ci(email): run the triage eval on PRs that touch the email agent (amd#2849) - `c3d9ec11` — fix(code): route orchestrated tool calls through the confirmation gate (amd#2853) - `c59e5d8a` — fix(hub): stop the terminal hub publishing against a core that cannot serve it (amd#2712) - `04f18294` — chore(deps-dev): bump the agent-ui-dependencies group in /src/gaia/apps/webui with 6 updates (amd#2750) - `ce4fcbf3` — fix(web): bracket pinned IPv6 addresses in URLs (amd#2739) - `e7e93287` — fix(installer): install the terminal hub on macOS and Windows (amd#2708) - `3cda9f77` — feat(cpp): MCP server registry — resolve server ids from mcp.json (amd#2820) - `c09764da` — fix(ci): validate the STX CMake cache instead of trusting bin/cmake.exe (amd#2818) - `616014d7` — docs(plans): scope C++ framework parity for domain-specific agents (amd#2806) - `f6433039` — ci(security): audit allowlist soundness + add PSIRT/CVSS triage skill (amd#2752) - `5c9f2b60` — chore(deps): bump the github-actions group with 2 updates (amd#2751) - `c103f6cd` — feat(discovery): real macOS and Linux branches for day-zero scanners (amd#1956) (amd#2747) - `e7e362e1` — chore(deps-dev): update mcp requirement from \<2.0,>=1.1.0 to >=1.1.0,\<3.0 in the python-dependencies group (amd#2749) - `ca9e5461` — chore(deps): bump the root-npm-dependencies group with 2 updates (amd#2748) - `f62ca240` — fix(website): unbreak the Railway deploy, red since July 31 (amd#2855) - `88d130ec` — feat(email): render the triage list from the scan, not from the model (amd#2858) - `bf9eb183` — ci(windows): pin ffmpeg version to drop the gyan.dev dependency (amd#2852) - `c531ca69` — fix(email): a conversation now survives its turns — session_id on /query (amd#2837) - `bcde95ee` — fix(email-agent): stop meeting/invite answers from inventing what tools never said (amd#2833) - `162274a5` — fix(tui): 'triage my inbox' draws its card under the question again (amd#2845) - `bbf69fd6` — feat(email,skills): bundled skills + account-keyed skill-set selection (amd#2695) - `abed9f1c` — fix(security): close find -exec / write side-doors in shell command whitelist (CWE-184) (amd#2740) - `0d1258a1` — fix(email-agent): get_thread renders a table card instead of relying on model prose (amd#2788) - `300163dd` — fix(email-agent): search_messages defaults to metadata-only, fixing context overflow on counting questions (amd#2782) - `cc331b78` — fix(email): low-priority-sender match no longer forces PROMOTIONAL (amd#2774) - `68736511` — fix(email): meeting proposal in a confidently-classified message no longer vanishes from needs_you (amd#2779) - `e135b8ef` — fix(ci): post PR reviews from the workflow instead of hoping the model does (amd#2719) - `efd3812b` — feat(email/tui): resolve "reply to 1" to the message the card actually shows (amd#2761) - `ed70db73` — fix(email-agent): search_messages states an exact, stable message count (amd#2760) - `42fdfdcd` — fix(email/tui): one triage card that tells you what to do, not what was classified (amd#2757) - `3803aa06` — fix(email): stop leaking classifier internals into triage card rationale (amd#2754) - `2c11ac0b` — fix(tui): sanitize agent error text (amd#2753) - `9bf0042a` — docs(tui): rewrite the terminal hub README for newcomers (amd#2717) - `1cdd9746` — docs(skills): implementation spec for Agent Skills v2 adaptive skills (amd#2685) - `50a5b862` — fix(website): pin the deploy toolchain so CI and production build alike (amd#2691) - `500cdac6` — fix(mcp): cap the mcp dependency below 2.0 (amd#2694) - `3686e069` — docs(guides): document the terminal hub, and how to actually get it (amd#2698) - `99508ec5` — docs(plans): preserve the TUI packaging design and binary-distribution plan (amd#2699) - `21c02f69` — fix(hub): mark the email agent verified so first install is not refused (amd#2703) - `8d07ee3d` — docs(website): correct every deployment instruction in the README (amd#2689) - `dff8f1b2` — refactor(memory): rename synthesis Skill dataclass to DistilledProcedure (amd#2684) - `d2832fd5` — # feat(email): opt-in on-device SLM classifiers for phishing and triage category (amd#2568) - `a02f2f0c` — fix(connectors,email): a bare reconnect no longer guts a mailbox (amd#2733) - `bc0d4632` — fix(website): reach the hero terminal, marquee, and code blocks by keyboard (amd#2706) - `82635362` — fix(website-router): track apex Worker, fail loudly on origin errors (amd#2688) - `5d269094` — feat(skills): SKILL.md loader, validator, discovery + gaia skill CLI core (amd#2669) - `0ae019c8` — fix(daemon): detect a sidecar that is alive but has stopped serving (amd#2707) - `4551ec5b` — fix(website): stop offering Intel Macs a DMG they cannot run (amd#2701) - `bc8a7e29` — docs(spec): Gatekeeper blocks the browser download, not curl | sh (amd#2732) - `0d170869` — fix(website): make the hub agent rows readable in both themes (amd#2711) - `d5ae430d` — fix(tui): a consequential readiness check now holds the screen (amd#2731) - `d6bd047f` — fix(email): stop Gmail 429-ing every scan — chunk at 25 and retry the rate limit (amd#2727) - `93216bc0` — chore(connectors): remove GAIA_MICROSOFT_TENANT — dead since the connector split (amd#2729) - `e3a6958a` — fix(tui): show a failed tool's own error instead of an "Invalid card" box (amd#2726) - `6efc520b` — fix(email-agent): surface connector errors from autonomy runs instead of a bare HTTP 500 (amd#2640) - `f0543406` — feat(connectors): split Microsoft into Personal + Work/School connectors (amd#2718) - `b8a97cbb` — fix(connectors): stop requiring a client secret from secretless public PKCE clients (amd#2630) - `70e4eb12` — fix(website): stop a published agent rendering twice on the hub page (amd#2690) - `282e30e0` — fix(tui): name the binary the user actually invoked, and fix the setup hint (amd#2700) - `d95866ff` — fix(installer): repair the Lemonade download URLs and add macOS support (amd#2704) - `3bc0b612` — fix(hub): stop components publishing against a core that cannot serve them (amd#2705) - `c068166d` — Email Triage draft/proposal SDK (amd#2551) - `cb90bab3` — fix(email): priority senders never force urgent; informational tail auditable (amd#2658) - `ab41b115` — fix(email): stop the assistant from narrating what the turn's tools don't support (amd#2659) - `7a93bfbe` — fix(email-agent): propagate the autonomy kill switch to the scheduler (amd#2657) - `4de65cff` — fix(email-agent): require the conflict tool for conflict verdicts (amd#2656) - `84404a6e` — perf(email): metadata-first scan + read-mail pre-scan coverage (amd#2661) - `733822e2` — fix(tui): restore the attention card on direct chat --agent launches (amd#2655) - `d4d5a7cc` — fix(mcp): cap the mcp dependency below 2.0 - `1d867eed` — fix(email): strip infrastructure banners from bodies before the prompt (amd#2650) - `22d8b2d4` — fix(email-agent): kill a running autonomy cycle and keep partial reports (amd#2652) - `a4f4b39b` — fix(email): paginate the triage scan and report truncation honestly (amd#2646) - `7adc4882` — fix(tui): dedup attention rows and stop the card interrupting turns (amd#2648) - `24e8d5ab` — fix(email): thread summaries keep the newest message's open asks (amd#2644) - `c3ef3796` — fix(agents): truncate large tool results to valid JSON (amd#2645) - `d633b387` — docs(connectors): document which Google scopes to declare in the Console (amd#2612) - `4ad367a9` — fix(connectors): derive --grant-agent scopes from the agent's own declaration (amd#2610) - `f7d5d7d9` — chore(deps): bump the root-npm-dependencies group across 1 directory with 3 updates (amd#2503) - `c84d9a79` — fix(tui): make the terminal hub readable on a light terminal background (amd#2611) - `bcffbc8d` — feat(email): recover the attention view and waiting-on-you detector onto main (amd#2604) - `b63a9f0d` — feat(email): guided Outlook mailbox setup — walk, verify, and answer questions in chat (amd#2598) - `e0d5014c` — fix(daemon): dev-mode start-agent refuses a checkout mismatch instead of silently serving a stale build (amd#2592) - `29dc0f9f` — feat(email): find meeting proposals during the inbox scan (amd#2589) - `22b0fa04` — fix(email): pre-scan stops reporting a guess as a verdict (amd#2587) - `c211e870` — fix(daemon): start the daemon clock so scheduled work can fire (amd#2586) - `358fd6e1` — fix(email-agent): add preference removal tools and a truthful read-back (amd#2520) (amd#2541) - `d01b9d43` — fix(email): don't cancel a retrying agent for a recoverable tool error (amd#2572) - `7fdcd6b2` — fix(email): surface degraded memory state and diagnose the real cause (amd#2577) - `1c42d842` — fix(email): autonomy /run refuses while off; add gaia email autonomy CLI (amd#2578) - `25b6cacd` — fix(email): draft_reply/draft_forward compose the body, don't ask for it (amd#2576) - `a050c64d` — fix(email): briefing carries a structured breakdown, not one sentence (amd#2575) - `8da002f0` — fix(email): resolve relative snooze/schedule times agent-side (amd#2574) - `ee4af04b` — fix(email): get_thread returns messages sorted and numbered, not raw backend order (amd#2570) - `9269a306` — fix(email): give list_inbox/search_messages a combined envelope budget (amd#2546) - `c4495a0f` — fix(agents): dispatch Python-call-style embedded tool syntax (amd#2573) - `f6103a03` — feat(email-agent): broaden autonomy candidates, add undo surface and per-message decisions (amd#2545) - `d975853c` — fix(email): normalize calendar time bounds to RFC 3339 before Google (amd#2579) - `fdf665f0` — feat(website): redesign landing, Agent Hub, and agent detail pages (amd#2566) - `000ab88e` — fix(ui): reject null bytes in upload-path; unrot 8 stale UI/journey tests (amd#2565) - `0d6ca966` — feat(tui): tool-confirmation modal for destructive/external actions (amd#2544) - `4131420a` — fix(agent): recover from context overflow on NPU/FastFlowLM (amd#2543) - `35062526` — fix(email): restore from Trash anytime, drop dead permanent_delete (amd#2542) - `2a1767e0` — chore(release): bump the hub component manifests to 0.23.0 - `b5d5bf54` — feat(hub): publish the terminal hub and Agent UI as R2 hub packages (amd#2530) - `014cbcbc` — fix(email): catch the contract guards up to schema 2.6, and de-race the heartbeat test (amd#2549) - `9f70be13` — fix(ui): serve the Lemonade start hint instead of hardcoding a dead command (amd#2510) - `26127e80` — feat(release): publish the terminal hub binary and install it (amd#2522) - `59374508` — fix(tui): unbreak build_tui on main — allowlist the bare-host remedy (amd#2548) - `1b189215` — fix(daemon): unbreak Unit Tests on main — remedy docstring names an unparseable command (amd#2534) - `9bdd7eb3` — docs(release): name the Lemonade version v0.23.0 actually ships (amd#2509) - `21d40c33` — fix(llm): detect Lemonade on macOS; stop printing commands that don't work (amd#2497) - `e272fbae` — fix(tui): keep a valid hub selection when switching tabs (amd#2482) - `1d0454b6` — fix(tui): prove the mailbox is usable before the gate clears a launch (amd#2494) - `ee54b779` — feat(email): agent-led mailbox onboarding — the agent sets up its own access (amd#2496) - `cc73e244` — fix(tui): stop the hub offering agents it cannot run or launch (amd#2492) - `5b378ebe` — fix(tests): make unit suite hermetic by blocking real network connections (amd#2500) - `238fe9ac` — feat(tui): install, run and uninstall agents from the TUI (amd#2484) - `ce730876` — feat(tui): draw tool_result render cards, starting with the inbox pre-scan (amd#2485) - `7fdd43ea` — chore(deps-dev): bump electron from 43.1.1 to 43.2.0 in /src/gaia/apps/jira/webui in the jira-app-dependencies group (amd#2501) - `c940fd0a` — chore(deps-dev): bump electron from 43.1.1 to 43.2.0 in /src/gaia/apps/example/webui in the example-app-dependencies group (amd#2502) - `7f569605` — feat(daemon): install, uninstall and catalog hub agents from the daemon (amd#2477) - `61ac101c` — docs(plans): design the TUI user journey around the email agent (amd#2480) - `fd66ba95` — fix(email): fail loudly on a dead worker and reconcile the package docs (amd#2479) - `1340b3b7` — fix(agents): actually ask before running confirmation-gated tools (amd#2475) - `5f283e10` — feat(tui): control API + MCP server for driving the live TUI (amd#2478) - `12016d5d` — feat(tui): stream agents over the daemon HTTP/SSE relay (amd#2476) - `554ef27c` — chore(deps): bump electron from 43.1.1 to 43.2.0 in /hub/agents/emr/python/gaia_agent_emr/dashboard/electron in the emr-dashboard-dependencies group (amd#2504) - `4822e2ed` — chore(deps-dev): bump the agent-ui-dependencies group in /src/gaia/apps/webui with 6 updates (amd#2505) - `3fc0d2e1` — chore(deps): bump the github-actions group with 4 updates (amd#2506) - `47c2d1b2` — fix(connectors): surface a browser that never launched (amd#2507) - `1763606f` — fix(agents): stop the OpenMP double-init from killing the agent mid-conversation (amd#2508) - `56215070` — fix(hub): carry requirements.min_lemonade_version through the manifest parser (amd#2493) - `02cf9984` — fix(mcp): keep console logs off stdout in stdio transports (amd#2473) - `d6c02c2f` — chore(deps): bump Lemonade Server to v11.5.0 (amd#2424) - `2898f1b5` — fix(agents): don't dedup errored mutation retries (amd#2464 batch dead-end) (amd#2465) - `867bc677` — fix(email): recall last archive batch so undo reaches across turns (amd#2458) - `9073ec20` — fix(email): strip LLM quoting from ARCHIVE_MESSAGE_BATCH ids (amd#2457) - `529d11b1` — ci(review): allow fork-PR checkout under pull_request_target (checkout@v7) (amd#2461) - `524d3282` — fix(email-agent): make undo window configurable for chat-speed bulk ops (amd#2449) - `61dc3a1f` — fix(email-agent): surface actionable Lemonade-down copy in gaia email -q (amd#2453) - `d78115a9` — fix(memory): guard against self-supersede hiding recalled preferences (amd#2452) - `6bf96e56` — fix(email-agent): isolate per-provider failures in read fan-out (amd#2451) - `acd20400` — fix(agents): reject unexpected tool kwargs with a structured error (amd#2450) - `f99ccaea` — fix(daemon): drop --reload from dev-mode email sidecar spawn (macOS) (amd#2442) - `c2337178` — fix(email-agent): verify archive left inbox + fix same-day search miss (amd#2438) - `3f6af0ff` — fix(email): don't misclassify timeouts as Lemonade-down (amd#2139 follow-up) (amd#2454) - `9386e1b1` — fix(email-agent): resolve draft/reply target from sender or topic (amd#2403) (amd#2437) - `229e61da` — fix(email-agent): never auto-archive IMPORTANT / security-sender mail (amd#2426) (amd#2435) - `e6ac57f9` — fix(email-agent): persist preferences to state.db so they survive without the embedder (amd#2427) (amd#2434) - `4301c897` — fix(agent-email): actionable copy for Lemonade-down /query errors (amd#2432) - `1bb2cccd` — fix(daemon/email-agent): dev-mode sidecar 'Empty module name' on macOS (bad PYTHON_KEYRING_BACKEND) (amd#2443) - `7a44bfd5` — fix(email): bulk-archive undo survives the whole run via a per-turn batch handle (amd#2439) - `5101b853` — fix(daemon): re-forward OAuth tokens on expiry so sidecars self-recover (amd#2436) - `55010368` — fix(email-agent): applying an existing label fails with 'Invalid label' (T14) (amd#2433) - `ce83feb9` — fix(email-agent/ui): de-jargon the send confirmation surface (amd#2407) - `4bb196fb` — fix(connectors): name consumers-tenant migration on personal-account app rejection (amd#2391) - `6b29a4d6` — fix(email-agent): construct with zero connectors instead of 502 (amd#2423) - `f3af7bff` — fix(agent-ui): treat ctx_size=0 as unknown, not a too-small window (amd#2402) - `f22b0f4a` — fix(ci): run evidence stage via direct claude CLI, not the GitHub-coupled action (amd#2430) - `60334202` — fix(email-agent): add live mailbox connection-status tool (amd#2405) - `4c370b23` — fix(ui): surface actionable sidecar HTTP errors instead of generic crash card (amd#2422) - `7fc559af` — test(ci): evidence lane exercises UI-backed routes + spot-regresses adjacent ops (amd#2421) - `9be31036` — security(ci): harden the evidence stage against env dumps / credential flows (amd#2417) - `0f709733` — fix(daemon): validate custody rag/query 'k' to a bounded positive int (amd#2390) - `ca50bf8a` — ci(review): broaden evidence gate + require a verdict-linked evidence section (amd#2415) - `59c736df` — ci(review): fold gaia-testing evidence into the PR review comment (Phase 1: CLI/API/MCP) (amd#2414) - `1a233d94` — test(connectors): fix main red — amd#2408 install test vs amd#2410 trust gate (amd#2412) - `78de5bfd` — fix(electron): resume periodic update checks after a no-feed start (amd#2389) - `a20c04b4` — fix(agent-ui): left-align installed-agent hub cards on home screen (amd#2398) - `236ba233` — fix(connectors): register hub-installed sidecar agents so email grant works on fresh install (amd#2411) - `b343f396` — fix(security): harden install trust gate and analyze_data_file sandbox (amd#2410) - `765cb544` — fix(daemon): hold the model-slot lease across inference, not just the load (amd#2394) - `27be002e` — fix(onboarding): neutral NPU wording in first-run Hardware check (amd#2399) - `c86a46e5` — docs(release): note email is Linux + API/CLI only this release (Windows amd#1648) - `e658cb15` — docs(testing): bind real-world evidence contract to the changed surface (amd#2376) - `feb00155` — docs(email): correct earn-trust claim — positive-outcome accrual not yet wired (amd#2392) - `7aee219a` — fix(email): commit autonomy dedup INSERT so it survives headless teardown (amd#2393) - `6b6a8dab` — fix(daemon): map NotGrantedError to 403 in forward_all route (amd#2395) - `5b2e6be8` — docs(release): keep only verified user-facing features in What's New - `e11740e7` — docs(release): fix Agent UI heading — drop 'update' (auto-update was trimmed) - `a011e33d` — docs(release): trim v0.23.0 notes to features verified working - `ce5ddd84` — docs(release): correct v0.23.0 notes to match verified behavior - `c62a3e27` — Release v0.23.0 - `5f624706` — fix(hub): importable CLI wheel agents + chat distribution via gaia init (amd#2373) - `3ad39e60` — fix(sidecar): actionable user-mode binary error (amd#2347) (amd#2357) - `bfc0f5b8` — fix(lemonade): recognize grouped amd_gpu/nvidia_gpu device keys in validation (amd#2368) - `0553c8b0` — feat(email): full autonomy — earn-trust engine, learning loop, scheduled driver (amd#2363) - `0284300b` — feat(connectors): support work/school Outlook + zero-setup device-code sign-in (amd#2364) - `1aaba5cc` — feat(lint): require security suppressions to be reviewed in an allowlist (amd#2343) - `19b1a4c5` — test(daemon): de-flake sidecar stop test on the pid-liveness check (amd#2349) - `174e9d0c` — refactor(chat): extract ProfileSpec, honest manifest, lazy RAG — one class → separable profiles (amd#2323) (amd#2362) - `57f970b0` — chore(audit): drop the security dimension from the weekly audit (amd#2348) - `23181040` — feat(security): proactive Claude security-audit workflow + CVSS/SARIF tooling (amd#2346) - `d605caec` — fix(security): enforce --allowed-paths sandbox on file read tools (amd#2344) - `1c8a91c9` — fix(security): remove pre-existing bandit HIGH findings and enable the HIGH gate (amd#2350) - `4b5c16b7` — ci(labeler): add tui/daemon/sidecar auto-label rules (amd#2356) - `a905b057` — chore(deps): bump the github-actions group with 2 updates (amd#2341) - `b3f793af` — fix(routing): default unknown language to TypeScript, not a process kill (amd#2337) - `48286dbd` — fix(init): stop Rich eating bracketed tokens in gaia init output (amd#2340) - `638a7643` — docs(skills): add porting-agent-to-hub — the legacy-agent port flow (amd#2338) - `d2c00b55` — fix(hub): harden agent-archive extraction against path traversal (amd#2342) - `a4417656` — fix(ci): repair the startup-failing GAIA CLI aggregate workflow (amd#2307) - `0835a250` — fix(ci): make the weekly eval and runner heartbeat monitor actually run (amd#2306) - `977c158f` — fix(cli): add --layout hub to gaia agent init for the agent-first hub tree (amd#2295) - `558a73e6` — chore(deps): bump the github-actions group with 3 updates (amd#2294) - `f8bff000` — chore(deps-dev): bump electron from 43.1.0 to 43.1.1 in the root-npm-dependencies group (amd#2292) - `fb179e32` — ci(eval): gate the Gemma-4-E4B consolidation on agent evals (amd#2283) - `6b2aa2db` — chore(deps-dev): bump electron from 43.1.0 to 43.1.1 in /src/gaia/apps/jira/webui in the jira-app-dependencies group (amd#2291) - `5837bdac` — chore(deps-dev): bump electron from 43.1.0 to 43.1.1 in /src/gaia/apps/example/webui in the example-app-dependencies group (amd#2290) - `3ce4f790` — test(agents): stub live Lemonade probe in context-overflow tests (amd#2288) - `1886a92f` — chore(deps): bump electron from 43.1.0 to 43.1.1 in /hub/agents/emr/python/gaia_agent_emr/dashboard/electron in the emr-dashboard-dependencies group (amd#2289) - `0a308449` — feat(agents): consolidate every agent onto Gemma-4-E4B at one context size (amd#2284) - `4392e06a` — chore(deps-dev): bump the agent-ui-dependencies group in /src/gaia/apps/webui with 3 updates (amd#2293) - `d9b11ec7` — refactor(hub): agent-first layout — hub/agents/\<id>/\<lang> (amd#2060) - `5f15b333` — feat(daemon): broker-wire the remaining direct model-load surfaces (amd#2286) - `3affe149` — fix(ui): detect GPU from real Lemonade payload shapes (amd#2285) - `4f888b78` — fix(hub): report file-based custom agents' real health, not always 'error' (amd#2277) - `36c29873` — fix(cli): detect GPU on all platforms and honour default_device (amd#2244) - `6c6f8a34` — test(audit): coverage for schedule CLI, perf-vis, VLM extraction, PDF gen/export; fix silent table-row loss (amd#2259) - `2425afe8` — fix(builder): surface model-not-found (404) instead of a generic placeholder (amd#2245) - `3a0eebe7` — ci(hub): wire nine hub package test suites into CI; fix .cjs docs-link guard gap (amd#2258) - `3e82ca61` — docs(connectors): correct the client_id_hash claim in the OAuth runbook (amd#2264) - `9f7dca84` — fix(ci): fix Lemonade startup and bash-on-PATH in the doc walkthrough (amd#2281) - `53f4b529` — fix(daemon): owner-only DACL for the Windows launch secret (amd#2250) (amd#2282) - `ee8825fa` — fix(security): require allowed_dir in compute_file_hash (amd#2280) - `46fb7687` — spec(factory): define the dogfooding loop — Claudia as live validation runtime (amd#2234) - `1a679f02` — feat(ui): always-available OAuth client field in Settings (amd#2104 interim) (amd#2265) - `45c796b1` — test(jira): unit tests for JiraAgent HTTP boundary and config discovery (amd#1991) (amd#2263) - `bed702eb` — feat(ci): execution-based weekly doc walkthrough (amd#2278) - `f8f309e6` — fix(ui): stop silently accepting unimplemented agent_mode 'autonomous' (amd#2257) - `3b0938ed` — fix(ci): weekly audit cross-links the prior parent instead of auto-closing it (amd#2254) - `78306480` — test(audit): risk-bearing coverage — jira, trust-gate, flag-precedence, amd#1655 boundaries (amd#2253) - `1aa39ca2` — fix(observability): real system-metrics polling + real rollbackAction (amd#2251) - `384b3ca8` — fix(security): bind MCP bridge to loopback, not all interfaces (amd#2246) - `037f4370` — fix(cli): implement/gate stubbed api-status, eval flags, schedule --skill (amd#2247) - `2b6ed53e` — fix(security): rate-limit exposed routes, harden JS ReDoS/XSS/cleartext-logging (amd#2237) - `6c6b3baf` — fix(packaging): stop the amd-gaia[agents] extra from downgrading the core wheel (amd#2262) - `8744a51a` — fix(security): validate user-influenced file paths (py/path-injection) (amd#2252) - `c921fa7b` — feat(email): quality + robustness batch (amd#2110/amd#2113/amd#2114/amd#2115/amd#2116) (amd#2192) - `b9391ec2` — fix(security): parameterize SQL, redact sensitive logs, harden ReDoS regexes (amd#2239) - `09e0bac4` — fix(security): stop leaking stack traces at API boundaries + least-privilege workflow permissions (amd#2236) - `3df5a3db` — fix(security): tighten API CORS — no wildcard origin with credentials (amd#2238) - `a3db04ef` — feat(connectors): OAuth forward-out to sidecars (V2-14) (amd#2203) - `e0886cbf` — refactor(daemon): reconcile the clocks into one daemon-owned scheduler (V2-15) (amd#2199) - `15459153` — feat(daemon): /host/v1 custody API v1 with per-agent scoping (V2-12) (amd#2197) - `e226d584` — feat(webui): group session sidebar by agent + session-state polish (amd#2193) - `6e48cb22` — feat(daemon): host-owned model-slot broker serializes loads (V2-11) (amd#2194) - `c42bb354` — fix(website): correct hub install command and clarify agent availability (amd#2207) - `2f18de85` — fix(tests): align email CLI dispatch test with amd#2191 thin-client contract (amd#2209) - `a2ad4eff` — fix(tests): repair stale gaia.ui.email_sidecar.manager import after amd#2144 (amd#2208) - `2b89643c` — fix(webui): repair broken main — AgentHubView imported deleted AgentHubGrid (amd#2206) - `68773d46` — test(eval): sidecar eval harness + distributed-seams suite (V2-19) (amd#2202) - `d7cffc72` — feat(agent-ui): first-run onboarding wizard — hardware pre-flight, in-app model download, connect-on-install (amd#2204) - `5bc7d325` — feat(electron): in-app install, R2 auto-update feed, gaia:// deep links (amd#2196) - `eb6b34bf` — feat(daemon): one-time versioned migration of ~/.gaia state (V2-13) (amd#2200) - `73b8f98c` — refactor(api): remove the last in-process email mount; relay via daemon (amd#2176) (amd#2205) - `e4ea0e33` — feat(connectors): grant the mailbox to the email agent in the same connect flow (amd#2195) - `2bf405db` — feat(webui): in-app Hub page with catalog lanes + install trust gate (amd#2201) - `459efba9` — feat(api): relay /v1/\<agent>/query through the daemon (V2-17) (amd#2198) - `ada5b95a` — feat(cli): gaia email attaches to the daemon — thin client (V2-8) (amd#2191) - `a49d257d` — fix(webui): reachable Agent Hub + installed-agent discovery and per-session picker (amd#2190) - `7c98b51a` — docs(connectors): rewrite the Google client-ID walkthrough for the current console (amd#2189) - `a94a8126` — feat(daemon): deliver sidecar launch secret via 0600 file, not bare env (amd#2149) (amd#2186) - `5601e2a5` — feat(daemon): streaming SSE reverse-proxy for agent routes (amd#2150) (amd#2188) - `1465bc73` — feat(agents): proactive lifecycle hooks with approval-gated proposals (amd#1484) (amd#2187) - `2c4ff395` — test(chat): mirror the amd-gaia floor guard from amd#2169; harden version parsing (amd#2184) - `2271278c` — fix(llm): make swallowed model-load failures loud in _ensure_model_loaded (amd#2185) - `2e0294f9` — feat(hub): add multi-component type discriminator to the manifest (amd#1716) (amd#2183) - `b78b71d1` — test(chat): guard the amd-gaia dependency floor against amd#2112 regression (amd#2174) - `33caec54` — refactor(ui): extract _best_effort_cancel helper for the relay cancel paths (amd#2173) - `ba58e31a` — feat(email): fast dev-iteration loop for the email agent SDK (amd#2083) - `f79c4177` — fix(webui): pin explicitly-set session titles — stop auto-retitle churn (amd#2165) (amd#2171) - `ae741690` — fix(email): error on an explicitly-targeted unconnected mailbox instead of substituting (amd#2164) (amd#2172) - `4dafa091` — fix(email): normalize date operators in search_messages (amd#2161) (amd#2170) - `dd2fa328` — fix(email): raise amd-gaia floor to match get_embedding_model_for_device (amd#2112) (amd#2169) - `011b6e3c` — fix(email): default calendar list to a forward window when range args are absent (amd#2162) (amd#2168) - `ed7d2969` — fix(ui): propagate cancel to the email sidecar on relay timeout/crash (amd#2158) (amd#2167) - `d8daf66d` — fix(agent): abort in-flight streaming generation on Agent-UI Stop (amd#2166) - `c907f6c4` — feat(memory): adaptive, review-gated onboarding conversation (amd#1955) (amd#2143) - `def5e979` — fix(ui): stop blaming Lemonade for cancelled/empty chat turns (amd#2141) - `553cb541` — fix(email): consolidate Agent UI pre-scan across every connected mailbox (amd#2129) - `5446af2c` — fix(agent-email): drop $orderby from Outlook get_thread to avoid Graph InefficientFilter (amd#2140) - `8ba66f4d` — fix(email): REST triage honors the LLM's is_spam verdict (amd#2125) - `aa44c2c8` — refactor(daemon): daemon-supervised agent sidecars (V2-6) (amd#2144) - `ac3cc4cd` — fix(eval): fail the briefing eval loudly on a zero-case/zero-judged run (amd#2123) - `c9566b19` — chore(deps): bump Lemonade Server to v11.0.0 (amd#2130) - `47c1142a` — feat(ui): route email chat through the sidecar /v1/email/query loop (V2-10) (amd#2136) - `b340427d` — docs(skill): correct the release skill against what v0.22.0 actually did (amd#2133) - `9f1ece0f` — feat(webui): render→component map + generic render primitives (V2-9) (amd#2131) </details> Full Changelog: [v0.22.0...v0.23.0](amd/gaia@v0.22.0...v0.23.0) --- ## How these notes were verified Every **What's New** entry was exercised on real hardware, not inferred from green CI. Re-verified end to end after merging current `main` (334 commits since v0.22.0), because the delta touched the MCP bridge, hub, skills, and connectors — the exact surfaces the notes assert. <details> <summary>Verification results, what broadened, and what was excluded</summary> **Re-verified against the current pin and kept:** `gaia hub` install/run/uninstall (trust gate enforced); the confirmation gate now covering terminal + `gaia api` + MCP tool calls (fail-closed classifier proven model-free — amd#2846/amd#2854); localhost binding + stricter CORS + enforced MCP `--auth-token` + shell-less MCP launch + `~/.gaia` write guard + CWE-89 SQL block (amd#2844/amd#2847/amd#2860, live probes); `gaia skill` create/import/list/info plus signed-tier + pre-publish-audit + migrate (amd#2692/amd#2702/amd#2693); the Microsoft connector split (two connectors, secretless PKCE — amd#2896's third connector is Gmail, so the wording is not stale). **Broadened this cycle:** the confirmation-gate claim (blocking gap amd#2846 merged) and the security section (four verified hardening items). The skills section is stated **opt-in** (amd#2848: no shipped agent loads skills by default). **Removed:** the onboarding / day-zero discovery headline — the scanner works on macOS, but the first-run onboarding flow it feeds is not ready to ship, so it is not claimed (the underlying commit amd#2747 remains in the changelog). **Excluded (not claimable):** a skills-by-default claim (false — amd#2848); the C++ SDK (real and CI-green but ships no user-facing artifact and predates this release); the prebuilt terminal-hub binary ("not yet distributed"). Claims that failed live verification remain filed as open bugs, not shipped: amd#2883, amd#2884, amd#2885, amd#2893, amd#2894. Email ships as a beta note whose only asserted behavior — never sending or deleting without confirmation — is verified. </details> ## Release checklist - [x] `util/validate_release_notes.py` passes - [x] `docs && npx mintlify validate` passes - [x] `src/gaia/version.py` → `0.23.0`; `LEMONADE_VERSION` → `11.5.0` - [x] webui `package.json` / `package-lock.json` → `0.23.0` - [x] Navbar label → `v0.23.0 · Lemonade 11.5.0` - [x] All 334 commits in range represented in the changelog - [x] Branch merged current with `main`; `setup.py` keeps `mcp<2.0` - [ ] Review from @kovtcharov-amd addressed --------- Co-authored-by: Kalin Ovtcharov <kalin@extropolis.ai> Co-authored-by: k <k@e>
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.
Installing an agent required the web UI — no CLI path, no daemon API — so a headless machine had no way to provision a sidecar. Adds catalog, install, install-status and uninstall to the daemon with
gaia hub list|install|uninstallon top, so the TUI, the CLI and the web UI share one implementation and one integrity check.Installing a non-verified agent runs third-party code, so it needs an explicit trust opt-in: an omitted, empty, false or loosely-truthy body is refused with a 403 naming
--trust, before anything is downloaded. Install is driven from the hub manifest's server-computed SHA-256, so the placeholder-SHA lock is off the install path entirely.Test plan
pytest tests/unit/test_daemon_hub_routes.py— 54 passgaia hub install email→ refused, names--trust; nothing written to~/.gaia/agents/emailgaia hub install email --trust→ installs;gaia hub list --installedshows itgaia hub uninstall emailwith the sidecar running → stops it first, or aborts loudly if the pid survives