Fix F1 of plans/2026-07-11-nomos-agent-code-review.md, the last item.
buildTools called listToolsFull (a tools/list MCP round-trip) at the start of
EVERY chat turn, including every auto-continuation resume — the tool list is
static for the lifetime of one MCP connection, changing only when the api
process re-registers tools (a restart, which this client already detects and
reacts to via reconnectLocked). Re-fetching it every single turn was
avoidable network+parsing work on the hot path.
mcpClient now caches the parsed tool list after its first fetch, guarded by
its own mutex (kept separate from the request-serializing mu so a cache
check never contends with an in-flight doRequest call). reconnectLocked
clears the cache — an api restart may have changed what's registered, so a
stale cache would be wrong, not just slow. fleetSnapshot's get_health_summary
call is deliberately left uncached — it's meant to be "as of now."
Since each session gets its own client (the per-session pool from the
concurrency work), this caches per-task-conversation rather than globally: a
task's FIRST turn still pays the round-trip, every turn after reuses the
cached list — which is exactly the case that mattered (long-running,
heavily-autonomous tasks with many auto-continuation resumes).
Verified live via the api's request log: a brand-new session's first turn
made 3 MCP calls (initialize, tools/list, get_health_summary); a second turn
on the SAME session made exactly 1 (only get_health_summary) — tools/list
correctly skipped.
This completes the implementation order in
plans/2026-07-11-nomos-agent-code-review.md — every A/B/D/E/F finding from
the review (excluding C1, explicitly deferred per operator instruction) is
now fixed, tested, and verified live.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix B3 of plans/2026-07-11-nomos-agent-code-review.md. resumeSession's retry
loop (used by both auto-continuation and panel-answered questions) already
retried once on a transient LLM failure, but if BOTH attempts came back
empty/erroring, the code just logged and returned — the task was left at
whatever status it already had (typically 'executing' or 'awaiting_input')
with no outcome, no operator-visible signal beyond an inert error line
buried in the transcript, and no way to tell a genuinely stuck task apart
from one quietly still working.
On permanent failure, now calls store.completeTask(outcome='failure', a
summary built from the error) so the task board reflects reality instead of
showing a task that looks perpetually in-progress. Uses context.Background()
for that write, matching resumeSession's own persistence pattern, since the
context that led to the failure may itself be in a bad state. This doesn't
prevent the operator from continuing to work the task via a fresh chat
message afterward — it only replaces silent hanging with a real status.
A full live induction of a permanent LLM outage would require breaking the
model/API-key config for the whole nomos container — too invasive for this
fix's priority. Verified instead that the new branch stays correctly dormant
on the happy path: ran a real ask_operator → panel-answer → resume cycle
end-to-end and confirmed the task landed at status='executing' with no
outcome set, proving the failure-handling code doesn't false-positive on a
normal successful resume.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix A2 of plans/2026-07-11-nomos-agent-code-review.md. chatWith replayed a
session's ENTIRE message history into the LLM's context on EVERY turn, no
windowing, no token budget — confirmed against a documented production case
(a single turn with 70 tool calls, messages up to 106KB). Every subsequent
turn of a long-running or heavily-autonomous task re-sent that ever-growing
history in full — a real cost/latency/eventual-context-limit risk for
exactly the tasks this system runs longest (many auto-continuation cycles).
Design call (flagged in the review as needing one before implementation):
a fixed-size window for LLM replay specifically, not the UI's own transcript
view. Simplest option that still keeps roughly the current task's working
context; a token-aware trim or LLM-summarize-on-drop are documented as
stretch options if 30 proves insufficient in practice.
- store.go: new getRecentMessages(ctx, sessionID, limit) — last `limit`
messages in chronological order, plus whether older ones were omitted.
getMessages (used by the UI's GET /sessions/{id}) is untouched and stays
unbounded — the operator should still see a task's full history regardless
of length; only what gets sent to the model is bounded.
- agent.go: chatWith uses getRecentMessages(sessionID, historyWindowSize=30)
instead of the unbounded getMessages. When truncated, injects a system
note telling the model explicitly that older turns exist but aren't shown,
so it checks upsert_knowledge/search_knowledge rather than assuming
something wasn't done just because it isn't visible.
New cmd/nomos/store_test.go: real Postgres integration tests (mirroring
internal/db/integration_test.go's throwaway-database pattern, guarded by
OIKOS_TEST_DATABASE_URL). TestGetRecentMessages_Truncation is the direct
proof for this fix (35 messages → 30 returned, correctly ordered,
truncated=true; 5 messages → all 5, truncated=false) — both cases run
against a fully-migrated database, not mocked. Also added
TestProposePlan_AppendVsReplace, closing part of the review's test-coverage
finding (E) by permanently regression-testing the earlier append-vs-replace
plan fix (commit 5384499), which had only been verified manually until now.
Verified live: inflated a real session to 42 persisted messages via direct
SQL, then continued it with a real chat call — the turn proceeded normally
(multiple real tool-call iterations, no crash, no context-length error);
nomos stayed healthy throughout. A3's incremental persistence separately
confirmed to have caught the 7 real tool calls made before the client
connection was cut, cleanly closing out both fixes' interaction.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fixes D1-D3 of plans/2026-07-11-nomos-agent-code-review.md:
- D1: deleted isTaskTool — defined, never called (dispatch already checks
handleTaskTool's own `handled` return value).
- D2: recordTouched issued one SELECT per entity slug found in a tool call's
args; batched into one `WHERE slug = ANY($1)` query. Verified live: a turn
naming three separate entities recorded involves edges for all three via
the single batched lookup.
- D3: complete_task's outcome had a declared enum (success|failure|partial)
in its tool schema but nothing validated it — an out-of-enum value (model
typo or a weaker model not respecting the schema) silently persisted as-is,
with only "failure" special-cased (anything else became status='done'
regardless of what the value actually said). Now validated in
handleTaskTool: empty defaults to "success" (unchanged), a recognized value
passes through, anything else defaults to "partial" (safer than silently
treating an unrecognized value as success) with a warning logged. Verified
live: instructed the agent to call complete_task with outcome="unclear" —
persisted as outcome='partial', not the literal invalid string.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix A3 of plans/2026-07-11-nomos-agent-code-review.md. handleChat only ever
saved the assistant message ONCE, after a.chat(...) returned, using
ctx := r.Context() for that write — the same context that cancels the instant
the client disconnects (Stop button, tab close, network blip). A disconnect
mid-turn meant the final save ran with an already-cancelled context and its
error was never checked: the entire turn's tool-call history was silently
lost from the persisted transcript, even though real work (executions
launched, knowledge written) had already happened server-side.
Brought handleChat in line with resumeSession's existing pattern
(continue.go): insert a placeholder assistant row immediately, update the
SAME row after every tool call. The key fix is WHICH context the writes use —
a new pctx := context.Background() for every DB write in this handler
(session creation/touch, the user message, question auto-close, the
placeholder + incremental updates, the title update), while ctx/r.Context()
still gates the agent's own work (a.chat) and the SSE writes exactly as
before — a disconnect still correctly stops the agent from doing further
work, it just no longer also erases what it already did.
Verified live: sent a message requiring 6 tool calls (get_entity/
get_relations/get_blast_radius on two targets) and force-killed the client
connection mid-stream with curl -m 12 (confirmed via exit code 28). Before
this fix the persisted transcript would show 0 tool-call entries; after,
all 12 raw tool_use/tool_result entries (6 calls × 2) were present and
correctly attributed by tool name — proving both that progress survives an
abort and that the incremental writes aren't corrupting the data.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fixes B1 and B2 of plans/2026-07-11-nomos-agent-code-review.md together,
since the right granularity for B1 in the auto-continuation worker turned
out to require B2's restructuring anyway (see below).
B1: grep -rn "recover()" cmd/nomos/ internal/mcp/ internal/httpapi/ returned
nothing before this — every explicitly-spawned goroutine (continuation
worker, resumed chat turns, async execution dispatch, the SSE listener, two
duplicate sshExec implementations' output-collector goroutines) crashed the
whole process on an unhandled panic, not just that one goroutine. More
consequential post-concurrency: more simultaneous unattended background work
means more surface area for one bad input to end every running task.
New internal/safego package: Go(label, fn) launches fn in a goroutine with a
recover-and-log wrapper. Applied at every bare `go` spawn site across the
three packages. Two sites needed bespoke handling instead of the generic
helper because their callers block on a channel and a silent recover would
just make them hang until timeout: sshExec's output-collector goroutine (two
near-identical copies, internal/mcp/server.go and internal/httpapi/phase3.go)
and httpapi's ListenAndServe goroutine — both now recover AND send a
synthetic error result so the waiting select unblocks immediately instead of
waiting out the full timeout.
httpapi's sseListener got extra treatment: its per-notification handling was
extracted into handleNotification with its own recover, so a panic decoding
ONE malformed pg_notify payload can't kill the listener goroutine for every
connected SSE client — the outer goroutine spawn only needs to guard the
connection setup/reconnect code around it.
B2: cmd/nomos/continue.go's processContinuations used to run every pending
continuation SEQUENTIALLY in a plain for loop, in the SAME goroutine as the
ticker — meaning (a) task B's continuation waited for task A's full (up to
10-minute) resumed turn to finish first, undercutting this session's earlier
concurrency work on exactly the path autonomous tasks depend on most, and
(b) an unrecovered panic anywhere in that call chain didn't just crash the
process (B1) — even WITH B1's recovery wrapped only at the top-level worker
spawn, the panic would still unwind the ENTIRE ticker-loop goroutine,
silently ending auto-continuation for every task until nomos restarted.
Fixed by spawning each pending item via safego.Go individually: real
parallelism, and a bad item can now only ever take down its own goroutine.
Added internal/safego/safego_test.go: TestGo_RecoversPanic is the concrete
proof — a deliberate panic inside Go() that would otherwise crash the whole
test binary; reaching the assertion after it IS the evidence recovery works.
Verified live against the rebuilt containers: full chat turn round-tripped
correctly (hostname lookup, 2 iterations, normal completion) — no regression
from threading safego.Go through the tool-dispatch/continuation paths.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix A1 of plans/2026-07-11-nomos-agent-code-review.md. isAssent and
isTypedConfirmation used a space-padded word-boundary check for negation
words but a bare strings.Contains for assent/confirm words — confirmed live
via test probes: isAssent("...maybe yesterday's logs...") returned true
("yes" matched inside "yesterday"), and isTypedConfirmation("I haven't
confirmed anything yet") returned true ("confirm" matched inside "confirmed",
and "haven't" wasn't in negationWords — only "don't"/"do not" were).
isTypedConfirmation is the sole gate for DESTRUCTIVE actions, so the second
case meant a message merely stating something hadn't been confirmed could
read as an explicit confirmation.
- Replaced the ad-hoc space-padding/prefix-check negation logic with proper
tokenization (wordTokenRe) + containsPhrase, matching WHOLE tokens/phrases
only — never a mid-word substring. Handles curly apostrophes too (a
pre-existing gap: the old straight-quote-only check would have missed
"don't" typed with a smart quote).
- Added contracted negatives (haven't, hasn't, isn't, wasn't, aren't, can't,
cannot, won't, wouldn't, shouldn't, didn't, doesn't) to negationWords.
Deliberately did NOT add a bare "not" — too broad, would false-negative
ordinary assent like "go ahead, this is not risky".
- Added regression tests for both confirmed cases plus a couple of adjacent
ones (eyesight/isn't, can't confirm) so a future change can't silently
reintroduce either bug.
All existing assent/confirmation tests pass unchanged — this is a pure
robustness fix, not a behavior change for any previously-correct case.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix 3 of plans/2026-07-11-concurrent-task-execution.md, the throughput one.
nomos held exactly one *mcpClient for the whole process, shared by every
/chat goroutine. Its mutex was held for the full duration of each tool
round-trip, and `run` executes its SSH command SYNCHRONOUSLY inside that
round-trip (capped at up to 10 minutes) — so while Task A was mid-`run`,
every other task's tool calls, even a trivial get_entity, queued behind that
single lock. Tasks could think (LLM calls) in parallel but never act in
parallel.
The MCP server has no per-connection state to protect (newServer returns one
shared *mcp.Server instance whose handlers close only over the DB
connection pool, already safe for concurrent use) — the mutex existed purely
because the client reused one stateful transport session. So the fix doesn't
touch the server at all:
- New mcpClientPool (cmd/nomos/main.go): one *mcpClient per session id,
created lazily (a real MCP initialize handshake) on first use and cached;
session-less traffic (the ephemeral no-DB-store path, the structured
/query endpoint) gets its own fixed, reused key instead of a fresh
connection per request. Idle clients (20 min past last use — long enough
to outlive a single slow `run`) are evicted on a 5-minute sweep ticker.
- agent.go: `client *mcpClient` → `clients *mcpClientPool`; every call site
(buildTools, fleetSnapshot, the tool-dispatch loop) now resolves its own
session's client via clients.get(sessionID) instead of reaching for one
shared field. A task's own tool calls stay sequential (already true — the
agent loop calls tools one at a time within a turn) but no longer block
anyone else's.
- main.go: handleQuery takes the pool instead of a client (keyed "query", a
fixed non-session slot); shutdown calls pool.closeAll().
Verified live against the deployed stack: fired a slow-but-ungated command
(`ping -c 15 127.0.0.1`, read-only per policy's allowlist, no approval
needed) as Task A, then — 2s into A's run — a trivial hostname lookup as
Task B, both through the real /chat endpoint. Task A's ping genuinely ran
~14.3s (confirmed via its own execution record and the agent's reported
output). Task B returned in 6s total, well before A finished — proving it
was never queued behind A's connection. Before this fix, B would have been
forced to wait out A's entire ~14.3s hold on the single shared client.
This completes plans/2026-07-11-concurrent-task-execution.md's required
scope — only the explicitly optional/deferred Fix 4 (a concurrency/cost cap,
pending real usage data) remains.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix 1 of plans/2026-07-11-concurrent-task-execution.md — the safety-critical
one. The assent window (and destructive window) were keyed purely by agent
id ("assent_window.agent:<uuid>"). With one agent:nomos entity serving every
concurrent task, this meant approving Task A's plan opened a window that ANY
concurrently-running task's config-mutation/destructive actions could also
ride, auto-executing without their own approval.
- store.go / agent.go: assentWindowActive/openAssentWindow and
destructiveWindowActive/openDestructiveWindow/destructiveWindowKey all gain
a sessionID parameter; keys become
"assent_window.agent:<id>.session:<sessionID>" and
"destructive_window.agent:<id>.target:<slug>.session:<sessionID>". Missing
session id fails closed (no window) rather than falling back to the old
agent-wide key.
- continue.go: the auto-continuation worker's window check moved from once-
per-batch to once-per-pending-item, scoped to that item's own session —
it was previously checking ONE agent-wide window for a batch that can span
multiple tasks.
- agent.go tool-dispatch: injects `_session_id` into a COPY of the wire args
sent to the MCP server (never into the args used for the emitted/logged/
persisted tool call, and never part of any tool's declared InputSchema —
invisible to the model) so the gating checks on the OTHER side of the
process boundary know which task is asking.
- internal/mcp/server.go: assentWindowActive/destructiveWindowActive/
classifyAndGate gain the same sessionID parameter, read from
args["_session_id"] at the three call sites (request_execution's
apt_upgrade/pct_create branches, and the shared classifyAndGate used by
restart/pct_exec/systemctl/run).
Verified against the live stack with the exact scenario from the plan: opened
an assent window for session A only, then called `run` with an identical
config-mutation command for session A (window open) and session B (same
agent, no window). A auto-ran (execution status completed); B correctly
queued for approval (pending_approval) instead of bleeding through — proven
at both the MCP response text and the executions table.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Closes the gap that made the knowledge loop optional/implicit: every
non-trivial task now has an EXPLICIT first plan step (research) and last
plan step (write back), not just background behavior the model might skip.
New MCP tools (the agent had no way to do these before — only REST endpoints
existed, unexposed to it):
- update_entity_attributes(slug, attributes): shallow-merge new/changed facts
into an entity (an IP, a version, a discovered port) so a future task
doesn't have to rediscover them from scratch. No approval required — this
updates the knowledge graph, not live infra.
- create_relationship(source, target, type): record a discovered edge
(depends-on, hosts, provides, ...). Idempotent, FK-validated against the
ontology's relationship_types, no approval required.
SOUL.md: restructured the task loop so step 1 is explicitly "gather
knowledge, not just status" (get_entity_knowledge, search_knowledge,
get_relations, get_blast_radius, http_get) and the last step before
complete_task is explicitly "write back" (update_entity_attributes,
create_relationship, upsert_knowledge) — both called out as real plan
entries the operator should see in propose_plan, not silent side-work. This
is what prevents the graph drifting from reality and is the concrete
mechanism behind "tasks compound."
propose_plan's tool description reinforces the same first-step/last-step
convention at the call site.
Verified against the live stack: both tools registered and callable via MCP;
update_entity_attributes merged an attribute correctly; create_relationship
rejected an invalid type (FK violation, clear error) and succeeded with a
valid type+direction, confirmed idempotent (2 calls, 1 row).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause: proposePlan unconditionally deleted and replaced the whole
session_plan_steps list on every call. The model isn't strictly held to
"call propose_plan once with the full list" — nothing stopped it (and
production evidence + live testing showed it happening) from calling
propose_plan once per step as it worked. Each such call wiped every
already-completed step, so the operator only ever saw the model's latest
single step ("1/1") instead of the real, growing plan.
Fix, two layers:
- store.go: proposePlan now only does a destructive replace when no step
has left 'pending' yet (a genuine pre-execution revision). Once any step
has started, a new call APPENDS after the current max seq instead of
wiping — so the panel accumulates the full history regardless of how the
model chooses to call the tool. plan.proposed now carries `appended` so
the frontend knows whether to replace or append.
- workspace.ts: plan.proposed handler respects `appended` (update vs set).
- tasks.go / SOUL.md: strengthened the propose_plan description and task-
loop guidance to call it ONCE with the complete step list end-to-end,
using update_plan_step (not re-calling propose_plan) to advance — fixing
the root behavioral cause, with the store-side append as a safety net
that holds even if the model still calls it incrementally.
Verified: forced the exact incremental-call pattern (propose_plan with 1
step, mark it running, propose_plan again with 1 more step) — the second
call appended at seq 2 instead of erasing seq 1, and its plan.proposed
event carried appended=true.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replaces the chat right rail's ad-hoc Digest+Graph stack with a single
TaskContextPanel that renders the task's live working state, driven by the
always-on events stream (not the per-turn chat SSE) so it keeps updating
during server-side auto-continuation/resume:
- GoalHeader: goal + status pill (planning/executing/awaiting_input/done/
failed), sourced from the sessions list.
- PlanProgress: ordered steps with live status icons + progress bar, hydrated
via new GET /sessions/{id}/plan; clicking a step with a target opens its
EntitySheet (no fake "jump to transcript" — bits-ui Collapsible content
isn't force-mounted, so a DOM-scroll jump would silently no-op for
collapsed tool groups).
- OperatorQuestion: the pinned structured question card (prompt/why/entity
chips/option buttons/free-text), hydrated via new GET /sessions/{id}/
questions; answering POSTs to the existing answer endpoint.
- SessionGraph upgraded to a live entity panel: entity.touched pulses the
node (animated ring) and shows "Now touching <slug>"; health.changed shows
a transient diff badge for touched entities.
- SessionDigest gains a success/failure/partial outcome banner and now also
refetches when the task's status changes, not just on session switch.
Two bugs found and fixed while wiring this up:
- workspace.ts's status-refresh trigger only covered goal.set/task.status;
question.raised/answered didn't refresh the sessions list, so GoalHeader's
pill went stale after answering via the panel (resumeSession runs entirely
server-side — no client 'done' event to piggyback a refresh on). Now every
status-affecting event triggers the (debounced) refetch.
- Forgot to rebuild the nomos container after adding the /plan and
/questions endpoints, so they silently fell through to the old default GET
handler — caught via a live curl diff against the running container,
not a code read.
Verified end-to-end against the live stack: goal/plan/question all update
without a reload as the agent works; answering a question via the panel
resumes the agent and the header pill correctly flips to Executing;
entity.touched pulses the live graph.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The last backend piece: when the agent hits a decision only the operator can
make, it surfaces a structured question instead of guessing or stalling.
- ask_operator(prompt, why?, options?, context_entities?): nomos-local tool
that records a session_questions row, moves the task to awaiting_input, emits
question.raised, and ENDS the turn (the agent loop returns after it, so the
agent can't barrel past its own question). The prompt becomes the assistant's
visible message so the question also shows inline in the transcript.
- Two resume paths, both close the question + emit question.answered + return
the task to executing:
- Panel: POST /sessions/{id}/questions/{qid}/answer → resumes the agent in the
background with the answer injected (reusing the continuation machinery,
refactored continueSession → resumeSession). Returns 202; the reply lands via
message polling.
- Chat reply: the next chat message on a task with an open question IS the
answer — auto-closed in handleChat; the turn itself is the resume.
Verified end-to-end: forcing a decision paused the task at awaiting_input with
the structured question (prompt/why/options/entities); a panel answer resumed
the agent (it acknowledged host:strong and continued); a plain chat reply
auto-closed a second question. Cleanup + tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Gives a task a legible, live-advancing plan via three more nomos-local tools:
- set_goal(goal): records the task goal, status → planning, emits goal.set.
- propose_plan(steps[]): persists ordered steps (clean replace for v1 — a
revision starts a new list), status → executing, emits plan.proposed with
the persisted steps (id+seq) so the panel can address them.
- update_plan_step(seq, status, execution_id?): advances a step, stamping
started_at/finished_at, emits plan.step.started/finished. Anchors the event
to the step's target entity when it has one.
Belt-and-suspenders: when an execution linked to a step reaches a terminal
state, the api auto-closes the step (closePlanStepForExecution in
emitExecutionEvent) and emits plan.step.finished — so the board stays honest
even if the agent forgets to close a step it started.
Verified end-to-end: a goal-driven task fired goal.set → plan.proposed →
2× step.started/finished → task.status on the SSE stream; both steps persisted
done with start/finish timestamps; status progressed planning→executing→done.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the compounding knowledge loop the task model is built around:
- complete_task(outcome, summary): a nomos-LOCAL, session-scoped tool (the
shared MCP server has no session id). Introduces the local-tool mechanism —
buildTools appends task tools, the agent loop routes them to handleTaskTool
instead of the MCP client. Sets the task's terminal status/outcome/summary,
mirrors it onto the task entity, and emits task.status.
- Knowledge → task linkage: after a successful upsert_knowledge in a task,
nomos links the note to the task entity (documents) and emits
knowledge.recorded, so the task's outcome view shows what it learned. The
note's about-link to the involved entity (written by upsert_knowledge) is the
retrieval path future tasks use.
- SOUL: every chat is a task loop — retrieve prior knowledge FIRST
(get_entity_knowledge on the target), plan, execute, record learnings, then
complete_task. Scales down for trivial read-only tasks.
- deleteSession now cleans up the task entity, its relationships, and its
task-scoped events (was orphaning them); the knowledge doc itself and its
about-links survive, as knowledge should outlive the task.
Verified end-to-end: a task recorded a note and completed; task.status +
knowledge.recorded hit the SSE stream; status=done/outcome=success persisted;
the note linked to both lxc:caddy (retrieval) and the task; a future
get_entity_knowledge(lxc:caddy) surfaces it; delete cleaned edges+events (0/0/0)
while the knowledge survived.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
As the agent runs a task, record which entities each tool call references:
write an idempotent task —involves→ entity relationship and publish one
entity.touched event per entity (correlation_id = session, data {slug,tool}).
Extracted from tool ARGS only — never results — so a bulk fleet query can't
drag every entity into the task graph; bulk/no-slug tools stay silent.
Emitted from the nomos agent loop rather than the shared MCP wrapper, which
has no session id. The involves edges make a task's graph neighborhood its
involved-entity set (queryable via get_relations) — the substrate for the
knowledge loop; the events are the live pulse the context panel consumes in
phase 6.
Verified end-to-end on the local stack: a chat referencing lxc:caddy/lxc:gitea
produced entity.touched on the browser SSE stream with slug+tool+correlation,
and exactly one involves edge per entity despite repeated touches.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Migration 018 adds goal/status/outcome/summary/entity_id to agent_sessions
and creates session_plan_steps + session_questions. Registers a 'task'
entity type and an 'involves' (task→entity) relationship in the ontology
so each session anchors its knowledge and involved-entity edges on the
existing relationships graph.
nomos createSession now mints a task:<session-id> entity (type task) and
links it via agent_sessions.entity_id — best-effort so chat never blocks on
it. listSessions/GET /sessions surface the new task fields.
No behaviour change yet; this is the data foundation for the task board and
live context panel. Verified end-to-end against the local stack: migration
applied, ontology ingested (60 types/47 rels), a new session mints a linked
task entity and the API returns status/goal/entity_id.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
From the last (successful) TypeType deploy session, two gaps the operator hit:
1. Knowledge write-back — the missing half of the loop.
The agent could read the knowledge base (search_knowledge/get_entity_knowledge)
but had no way to WRITE it, so everything it learned (the Dragonfly memlock
rlimit gotcha, the NAT-hairpin DNS issue, etc.) lived only in an ephemeral
chat message and was lost — the system could never actually "get better."
This is the `upsert_knowledge` MCP tool the 2026-07-08 gaps plan called for.
- internal/mcp/server.go: upsert_knowledge(title, content, about?, tags?,
kind?) writes a document/investigation/runbook entity + knowledge_entities
row (search column is generated), upserts by slug so re-titling updates in
place, and optionally links it to the entity it's about so
get_entity_knowledge surfaces it there.
- SOUL.md: capture non-obvious findings/deploys/gotchas as part of finishing
work, not only when asked "what did we learn".
2. "I had to ask for status multiple times."
The clearest cause: a long working turn (64 tool calls) that exhausted the
iteration cap ended with a bare "max iterations reached without final
answer" — a dead end that forced the operator to ask what happened.
- cmd/nomos/agent.go: on exhaustion, make one final no-tools LLM call
(finalSummary) asking for a status report — what was accomplished, current
state, what remains — so the turn always ends with a real outcome.
- maxIterations 25 -> 40 (the decomposed per-step pct_create flow legitimately
needs more steps).
- SOUL.md: always end a turn with a clear outcome; never end silently or on a
bare tool call — the operator can't see the tools working and reads silence
as "nothing happened".
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Operator: "I'd like to be able to see in the chat what the agent is actually
running, right now I just wait while nothing happens." Two compounding gaps:
1. The auto-continuation worker (cmd/nomos/continue.go) had zero live push —
its result only appeared on a manual page reload, so approving a plan and
watching the chat looked completely dead even while the agent was actively
working.
2. Even with polling, continueSession only persisted ONE message at the very
end of a continuation — a continuation that runs several tool calls before
concluding would still show total silence for however long that took.
Fixed both:
- web/src/lib/stores/chat.ts: polls the current session's messages every 3s
between turns (never while a live stream owns the message list) and merges
in anything new. Started after a live turn ends and when a session loads;
stopped on new-chat/session-switch.
- cmd/nomos/store.go: insertMessageReturningID/updateMessage — lets a message
be created as a placeholder and updated in place.
- cmd/nomos/continue.go: continueSession now inserts a placeholder the
instant it starts (renders as the existing "thinking" dots — immediate
feedback that something is happening) and updates that SAME row after
EVERY tool call, not just at the end. A poll within ~3s of any tool call
landing shows it — individual `run` commands appear as the agent issues
them, not just the final rolled-up summary.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Closes the two remaining open points from the auto-continuation work.
1. Atomic pct_create (observability, the bigger of the two):
pct_create used to bundle create + apt install + post_install script into
one black-box multi-minute SSH call — the agent got back a single opaque
success/fail with no way to see (or fix) which step actually broke.
Removed the whole post-create provisioning block (and the now-dead
provisionScript/sanitizePkgs helpers + their tests). pct_create is now
create + start + register ONLY — fast, and its result is fed back to the
agent via auto-continuation almost immediately. The agent installs
packages and runs setup as its OWN sequence of `run` calls against the new
lxc:<hostname>, observing each command's real output and able to diagnose
and retry exactly the step that failed — the same recovery loop already
proven for the general case, now applied to installs too, instead of
requiring a separate black-box mechanism.
- services/post_install removed from the pct_create params struct and
from the MCP tool schema/SOUL.md docs.
- SOUL.md: explains the new flow, moves the Docker CLI gotcha and DNS
troubleshooting guidance to be steps the agent runs itself.
2. Scoped destructive window (targeted autonomy for recovery):
Verified live in the previous session that a destructive recovery (a
failed destroy needing stop-then-destroy on the same container) required
TWO separate typed confirmations for what was clearly one recovery
action. Added a narrow, TARGET-scoped 15-minute grant
(destructive_window.agent:<id>.target:<slug> in autonomy_settings,
shared key format across cmd/nomos and internal/mcp) that opens only
after an EXPLICIT typed confirmation (never loose assent) or an explicit
button-approval of a destructive step, and only ever covers further
destructive commands against that SAME target. A different target always
needs its own fresh confirmation — this narrows risk instead of loosening
it globally, unlike broadening the general assent window to cover
destructive actions would have.
- cmd/nomos/store.go: openDestructiveWindow/destructiveWindowActive/
executionTarget.
- cmd/nomos/agent.go: opens the window when a typed confirmation grants a
destructive chat-assent execution.
- internal/mcp/server.go: `run` tool checks the window before gating a
destructive command; auto-runs if active.
- internal/httpapi/phase3.go: DecideApproval opens the same window when a
destructive execution is approved via the button/API, for parity with
the chat-assent path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Verified live: the new auto-continuation worker (previous commit) worked end
to end for the happy path (provision -> auto-verify -> report success, zero
operator ticks). But testing a failure-recovery case (a destroy that failed
because the container was still running) surfaced a real bug: continueSession's
emit closure only captured "text" events, so when chatWith ended the turn on
an "error" event (LLM returned an empty/refusal response, internal retry also
empty), the worker persisted a completely blank, uninformative "auto" message
— no sign anything had gone wrong, undermining observability of the very
mechanism just built.
- Capture "error" events and, if the turn produced no text/tool_calls at all,
persist an explanatory placeholder instead of blank.
- Add one outer retry of the whole chatWith call when the first attempt
produces nothing — the principle behind this whole feature ("don't give up
on the first error") should apply to the continuation mechanism itself, not
just the homelab commands it's continuing.
Also verified live: recovery-from-failure works via the normal chat path once
prompted, and cleaned up the test container.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The root cause behind "the agent stops at the first error and doesn't recover":
provisioning executions run ASYNCHRONOUSLY (pct_create fires the SSH work in a
goroutine and returns "running" immediately), so the agent's turn ENDS before
the result exists. The agent literally isn't running when the step fails — it
can't react to a failure it never observes. The only thing that fed results
back was the operator typing "continue" after every async step: the human was
the event loop. (In the flagged 18-message session the operator typed
continue/proceed/?? eight times while the agent correctly diagnosed each failure
but couldn't advance a step on its own.)
This makes the system the event loop instead:
- migrations/017: nomos_plan_executions links each gated execution to the chat
session that started it.
- cmd/nomos: after a tool result, any "execution <uuid>" it started is linked
to the session. A background worker (continue.go) polls for those executions
reaching a terminal state and — while the agent has an open assent window (an
approved plan is in flight) — re-invokes the agent with the result
("execution X completed/failed: <result>"), so it proceeds to the next step
or diagnoses+fixes the failure, with no operator tick. Guarded against loops
(mark-continued before running) and bounded by the 30-min window.
- chatWith(): chat() variant that injects the finished-execution note after
replayed history without persisting a fake user turn.
- DecideApproval: approving a step by ANY route (button or chat-assent) now
opens the assent window, so auto-continuation works regardless of how the
operator approved — previously only typing "go ahead" opened it.
- SOUL: the agent is told it will be auto-re-invoked when async steps finish —
don't poll get_execution_status, don't wait for "continue"; end the turn and
keep going step by step until the goal is verified or a genuine blocker.
This is the root fix, not another per-command patch: you can't enumerate every
failure of an unbounded action space, but you can give the agent a loop that
observes each result and adapts — because "do anything" always includes "the
first attempt failed."
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three fixes for the session where the agent proposed a plan, waited
for 'proceed', then re-queued instead of being auto-approved:
1. Chat-assent fallback: when the operator says 'proceed' but the
preceding turn had NO pending approvals (agent proposed plan in text
without calling request_execution), inject a system note telling the
agent to execute the plan now. Opens the assent window so subsequent
config_mutation commands auto-run.
2. SOUL.md: instruct agent to ALWAYS call request_execution/run when
proposing a plan, not wait for 'proceed' first. This ensures a
pending approval exists for chat-assent to grant.
3. SOUL.md: stronger Docker instructions — Debian 13's docker.io package
installs the daemon but NOT the docker CLI binary. Must use
get.docker.com in post_install. Added 'handling errors' section:
diagnose, try alternatives, continue — don't stop after one failure.
Agent stopped after every approval step, forcing operator to type
'continue' 7× per deploy session. Root causes and fixes:
1. Compound read-only commands (e.g. 'systemctl status; journalctl')
defaulted to config_mutation — now splits on ;/&&/||/| and classifies
as read_only if all segments are inspection verbs. Added grep, wc,
sort, uniq, cut, tr, dpkg -l, apt list, docker stats to allowlist.
2. curl|sh was classified destructive, forcing typed confirmation for
legitimate installs (get.docker.com). Demoted to config_mutation —
loose assent grants it, no typed phrase needed.
3. SOUL.md said 'STOP after queuing' — replaced with 'continue working
on non-blocked steps'. Added assent window section instructing agent
to carry out the full plan after approval.
4. Assent window: when operator approves a plan via chat assent, a
30-minute window opens where config_mutation commands auto-run
without re-approval. Agent writes expiry to autonomy_settings; MCP
run tool checks it before gating. Destructive never auto-runs.
5. System note after approval now says 'CONTINUE executing the full
plan — do not stop and wait for continue.'
Root cause of "chat gave me no further feedback — had to go to Ops": two
compounding bugs, found by reading the actual production session transcript.
1. InlineApproval.svelte — all of last session's live-status/self-heal work —
was never imported or rendered anywhere. Chat.svelte had its own separate,
much dumber approval bar (no status tracking, no destructive handling, just
silently disappears after clicking) that WAS the one users actually saw.
Deleted the dead bar and its state; InlineApproval now renders per-message.
2. chat.ts's extractApprovals hardcoded `tool.name === 'request_execution'`,
so any approval raised by the newer `run` tool was invisible — no card, no
feedback, nothing to self-heal, forcing the operator to the Ops page with
zero acknowledgement in the conversation. This was the actual proximate
cause of last night's destroy-135 session. Fixed to match on response
shape, not tool name, so it doesn't silently break again for the next new
gated tool.
3. Nomos was telling operators "type something like 'I confirm destroy 135'"
for destructive actions (SOUL.md) but no backend path ever consumed that
phrase — chat-assent explicitly (and correctly) excludes destructive from
loose assent, but I never built the alternative. Added
isTypedConfirmation() (cmd/nomos/assent.go): stricter than loose assent,
requires an explicit "confirm" statement, only applies to destructive-
flagged pending approvals.
4. InlineApproval's completed-state hardcoded "Provisioned successfully" —
wrong/confusing for a destroy or arbitrary `run` command. Now says
"Completed on <target>" and shows the actual command output, verified live
against the real destroy-135 execution.
Verified live in a real browser against the production API/DB (dev server
proxying to :8090): the historical stuck session now retroactively renders
both executions as resolved with correct wording and real output.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Implements the first slice of plans/2026-07-10-general-gated-execution.md:
Nomos gets one general execution tool instead of only a fixed action enum,
gated by an automatic risk classifier, and approval can be granted by the
operator just replying in chat instead of clicking a button.
- internal/policy/command.go: ClassifyCommand(cmd, declaredRisk) — rule-based
read-only allowlist + destructive denylist, default-escalate to
config_mutation for anything else. Classification can only ESCALATE the
caller's declared risk, never de-escalate it (destructive always wins even
if declared read_only). Compound commands (&&, ;, |, $()) never qualify for
the read-only fast path. Full test corpus.
- internal/mcp/server.go: new `run` MCP tool — target (host:/lxc:), command,
purpose, optional declared_risk. Read-only commands execute immediately;
everything else queues an approval exactly like pct_create today, executed
via httpapi's existing executeApprovedAction. Also fixes a real latent bug:
pct_exec resolved an LXC's host attribute without the "host:" prefix, so it
could never find the Proxmox host — new resolveExecTarget/resolveRunTarget
helpers (mcp + httpapi) fix this for both the new `run` action and existing
actions that route through the same execution path.
- internal/httpapi/phase3.go: "run" case in executeApprovedAction; fixes two
bugs found while wiring this up — (1) DecideApproval hardcoded risk_class to
'config_mutation' on every approve, silently corrupting the audit ledger for
every other risk class; (2) denying/revoking an approval never updated the
linked execution's status, so it stayed 'pending_approval' forever instead
of reflecting the decision.
- cmd/nomos/assent.go: deterministic (not LLM-judged) chat-assent detection.
Scoped to the immediately-preceding assistant turn's pending approvals only
— an old "yes" can't retroactively approve something new. Destructive-risk
actions are excluded from loose assent. Approves via the same HTTP decision
endpoint the UI button calls, so both paths share one audit trail.
- web/.../InlineApproval.svelte: self-healing poll — a pending approval card
now picks up being decided via ANY path (chat assent, Ops page, Matrix), not
just its own button. Previously the banner stayed stuck showing
Approve/Deny even after the action had already run elsewhere.
- nomos/SOUL.md: `run` is now the general capability ("no fixed menu, only a
risk gate"); documents chat-assent behavior and the destructive exception.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of "asks permission but never acts": the approved pct_create
execution failed to parse because the LLM emitted `"privileged":0` /
`"nesting":1` (numbers) into strict `bool` fields, so the container was
never created. Compounded by a hardcoded template name (debian-13.0-1)
that no longer exists on the host, and no way for the agent to read the web.
- flexBool: accept 0/1, "true", bool for privileged/nesting (the exact prod failure)
- pct_create template pre-flight: list host cache, validate/auto-pick newest debian
- pct_create services[] + post_install: one approval provisions a working service
- new http_get MCP tool (sanitized, size-capped, SSRF-guarded) — agent can read repos/sites
- request_execution description: target=host, full JSON schema + example
- SOUL.md: agent CAN fetch the web; prefer one-step provisioning
- default model deepseek-v4-flash -> v4-pro; maxIterations 15 -> 25
- unit tests for flexBool, template resolve, pkg sanitize, HTML sanitize + SSRF block
Verified live on host:strong with a throwaway VMID 999: template auto-resolved,
container created + booted, services installed, post_install ran, then destroyed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Empty/refusal responses retried once, then surfaced as errors instead of silent blanks
- Chinese refusal boilerplate detected via denylist + non-ASCII heuristic
- Bulk-tool preference added to SOUL.md (list_lxcs over per-entity get_lxc_state)
- Tool results truncated to 4KB on persist; get_state_snapshot filters null-state entities
- Session delete (DELETE /sessions/{id} + confirm-on-second-click UI)
- Session titles auto-generated from assistant answer instead of raw user message
Problem: every host/service/lxc/etc. entity_status row was permanently
stuck at 'unknown' since creation. Verified against the live DB:
metric_samples had 17,559 rows, 100% attached to type='check' probe
entities and 0% to any real monitored entity; only 25 check entities
ever had real health written. check_defs.entity_id (the probe's own
bookkeeping entity) and check_defs.target_id (the host/service actually
being observed) were both real fields, but the scheduler wrote
UpsertEntityStatus/InsertMetricSample/emitSchedulerEvent keyed by
entity_id instead of target_id — so every check ran and every result
was real, it just landed on the wrong row. This is the mechanism behind
observed drift: the agent's dashboard/health tools reported the
internal probes' state, never the actual fleet.
Change:
- scheduler.go: runCheck/resolveSignal now resolve targetID from
cd.TargetID (falling back to the check's own id if unset) and write
status/metrics/events there. Signals stay keyed by the check entity,
unchanged, matching their existing resolution logic.
- Added a staleness sweep to housekeeping(): an entity whose last
observation is older than 3x its fastest enabled check's interval
(floor 5m) is marked 'stale' and emits health.stale, so a stalled
scheduler or disabled check_def can no longer look like current data
forever.
- migrations/016: deletes the now-orphaned check-entity entity_status
rows so dashboard/fleet-health rollups stop double-counting probes as
monitored entities. Historical metric_samples on check entities are
left as-is (time-series data, not safe to reattribute).
- openapi.yaml + regenerated gen code: Entity gains health/last_check_at;
'stale' added to the health enum everywhere it's used.
- dashboard.go / GetFleetHealth / nomos's get_health_summary MCP tool:
exclude type='check' entities from rollups.
- nomos/agent.go: replay prior turns' tool_use/tool_result pairs into
the conversation instead of dropping them (previously only final text
was replayed, forcing the agent to re-derive fleet state every turn),
and inject a compact live fleet-health snapshot into the system prompt
each turn so it starts oriented instead of spending an iteration on
discovery.
Risk: config_mutation (schema-adjacent — new migration, no destructive
DDL, additive DELETE only on orphaned rows). No behavior change until
oikos-api/oikos-scheduler/nomos are rebuilt and redeployed.
Verification: go build/vet clean across the repo. Ran this worktree's
own API binary against the live dev Postgres on an alternate port
(read-only from the live containers' perspective) and confirmed
/api/v1/entities now returns health/last_check_at, and the dashboard
health rollup dropped from double-counting to an honest 168 unmonitored
entities (matches reality pre-deploy — the live scheduler hasn't run
the fixed code yet). Confirmed check_defs.target_id correctly maps
multiple checks to host:hubris via direct psql query.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
After an api (MCP server) restart, nomos held a dead session id and every
tool call failed with "unexpected end of JSON input" until nomos was manually
restarted — which happens on every deploy. The MCP client now detects a
rejected session (4xx or empty body) and transparently re-initializes and
retries once. Also raise the SSE scanner buffer to 4MB so large tool results
don't exceed the 64KB default token limit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
http.FileServer canonicalizes /index.html -> "./", which for /ui/ produced a
301 redirect loop and made the control room unreachable. Serve embedded files
directly with http.ServeContent instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
