Added concrete examples of banned output patterns:
- Preambles (Great question!, I'd be happy to help!)
- Narration (Let me..., I'll now...)
- Hedging (It seems like, It appears that)
- Over-explaining and self-congratulation
- Summaries that restate what was just done
Added positive guidance: start with the answer, state task
completion in one sentence, end when information is delivered.
Models respond better to explicit prohibitions with examples than
to general instructions to 'be brief'.
Tests cover:
- Multiple concurrent pending requests from different agents
- BypassPendingByID lookup
- ResolveBypass removes only the resolved request
- ResolveBypass fails for non-existent ID
- Empty session has no pending requests
Previously a session tracked only one pending bypass request. If
agent A1 had a pending bypass, agent A2's bypass request would block
waiting to be read from toolsrv's channel, effectively blocking all
agents in the session.
Now bypassPending is a map keyed by request ID:
- SetBypassPending adds to the map instead of overwriting
- BypassPending returns all pending requests (slice)
- BypassPendingByID returns a specific request
- ResolveBypass removes from the map by ID
- 9P bypass file returns JSON array of all pending
The bypass loop reads requests continuously without waiting for
resolution, so multiple agents can have concurrent pending requests.
Every background process emission now includes:
'Do NOT sleep, poll, or wait. Continue with other work —
output will be injected automatically when available.'
Add explicit instruction that models must not use sleep, loops,
or any blocking mechanism to wait for background process output.
The system injects updates automatically.
The cwd= parameter in agent/new payloads was not being parsed,
so scripts like ollie-session-here could not set agent cwd on
creation. Now parsed and passed to AgentParams.CwdOverride.
- Agent cwd is optional; empty = inherit session cwd (the common case)
- Session cwd is required at creation and is the inheritance root
- toolsrv maintains agentCWD map, resolved per call from agent= field
- Override set via agent cfg (cwd=...) or ctl (cwd [<dir>|-])
- Agent.SyncCwdToToolServer re-pushes on every (re)connect
- GUI NewAgentDialog shows '(inherit: <sessionCwd>)' as placeholder
- proc_test.go covers per-agent cwd isolation
- Kate and acme scripts updated for new session/agent creation flow
- AGENTS.md documents the architecture
Backend:
- Add proc.start/proc.exit events for background process lifecycle
- Add proc idx ctl command for machine-readable process listing (TSV)
- Add event.pub file for external event publishing
- Add ListProcsIdx to toolclient and toolsrv
- Fix bypass commands to use Setpgid for process group isolation
GUI:
- Add configurable bypass approval shortcuts (Ctrl+Y/Ctrl+N default)
- Add Keyboard Shortcuts section to Settings dialog
- Store shortcuts in theme.conf
Kate:
- Fix 'Start Session Here' - use rdwr for session/new endpoint
Changed from 0600 (owner only) to 0660 (owner + group):
- ctl: frontends need to send slash commands
- plan: frontends need to read/write plan
- fifo: frontends need to submit prompts via queue
The agent group includes frontends, so group permissions enable
frontend interaction while still maintaining ownership-based isolation.
Per-agent file ownership with Unix permission enforcement:
- Agent directories owned by agent ID (UID), group 'agent' (GID)
- Private files (plan, ctl, fifo): mode 0600 - owner only
- Group-readable (chat, log): mode 0440 - owner + agent group
- World-readable (state, id): mode 0444 - observable by all
- Prompt: mode 0220 - CLI and owner can write
virtfs: fix UID/GID inheritance through nested paths
- Added findChildWithInheritance() to accumulate inherited UID/GID
- Stat now correctly shows agent ID as owner for nested files
server: admin bypass for server owner
- serverAdmin variable captures the Unix user running olliesrv
- Admin bypass includes empty uname, 'admin', or server owner
Documentation updates:
- fs/doc.go: 'The Namespace IS the Security Model'
- registry/doc.go: capability-based tool access
- peer.go: capability-based peer access
- lessons-learned.md: 'Model compliance is not a security boundary'
- architecture-9p.md: per-agent file ownership section
Security evaluation:
- Added experiments/security-eval/ with NERV attack corpus adaptation
- Test scripts for Landlock sandbox validation
- RESULTS.md documenting 0% ASR on hostile operations
This implements the NERV thesis: 'An agent can only access resources
explicitly bound into its namespace.' Enforcement is structural via
file permissions, not behavioral via model compliance.
Server emits session.{sid}.agent.{aid}.bypass.resolved with id and
action (approved/denied) when a bypass is resolved by any client.
GUI handles bypass.resolved events to clear the banner and pending
count when CLI or another client resolves a bypass request.
This allows CLI 'o sess approve' to clear the GUI banner automatically.
- Changed event topic: session.{sid}.agent.{aid}.bypass.request
- Added agentId parameter to bypassRequested signal
- Filter bypass events to show only for the active agent
Bypass approval now flows through:
1. GUI - via event stream and banner
2. CLI - via agent loop (to be implemented)
Removed:
- bypass_notify.go (D-Bus notification)
- BypassNotifyFunc type and all references
- godbus/dbus dependency
The bypass event is still published via SetBypassPending.
If fidOK was false for the event file, we fell through to the
default stream handling path which doesn't work for events.
Now we return 'bad fid' error instead of falling through.
The pubsub library had issues:
- Published to literal '*' topic (nonsensical)
- Used TrySend which drops events
- Complex hierarchical wildcard publishing
New implementation:
- Simple eventHub with map of subscribers
- PublishEvent fans out to all subscribers (blocking send)
- SubscribeEvents returns channel, cleaned up on ctx cancel
- SubscribeEventsFiltered filters client-side with MatchTopic
- Removed simonfxr/pubsub dependency
The subscription was being cancelled after the first read completed
because we used the per-request context. Now using the connection
context so the subscription lives until the fid is clunked or
connection closed.
The pubsub library's SubscribeChan uses non-blocking TrySend which
drops events when the channel is full. Switch to Subscribe with a
blocking callback to ensure no events are lost.
Also increased channel buffers from 64 to 256.
Plain reads (without writing a filter first) now get a per-fid
subscription with '*' filter, identical to 'echo * | rdwrs event'.
This fixes event loss — the old EventStream path overwrote events
when multiple arrived before the reader consumed them.
The feed file was documented but never used by any frontend
or script. The observer agent pattern was never adopted.
- Remove feed.go, FeedWrite, ConsumeFeed
- Remove WatchFeed constant
- Remove feed file from 9P namespace
- Remove ConsumeFeed goroutine spawns from session
- Update docs (architecture-9p, architecture-ide, architecture, usage)
The event stream now covers real-time observation patterns better.
The event stream with filtering replaces statewait:
- echo filter | rdwrs event
Removed:
- statewait file from agent namespace
- All non-historical references in docs and code
The state file remains for simple polling reads.
- state: immediate read of current agent state
- statewait: blocks until state changes
Clearer semantics than overloading statewait with both behaviors.
EventValue was storing only the latest event and using hash comparison,
which caused events to be overwritten if they arrived faster than the
client could read them.
Now eventwait uses Stream mode with EventStream which delivers each
event as it arrives. Events won't be lost due to rapid arrival.
Server changes:
- Session tracks pending bypass request and exposes methods
- New 9P files: session/{sid}/bypass (read/write), bypasswait (blocking)
- Publish bypass.request events for GUI listeners
GUI changes:
- Handle bypass.request events from eventwait
- Show inline amber banner with command and cwd
- Approve/Deny buttons resolve via 9P
Desktop notifications still work in parallel for non-GUI usage.
When reading statewait with the same state value, the server now waits
500ms instead of 5 seconds before returning. This makes state polling
much more responsive.
Background procs with bypass were hanging because the caller blocked on
<-startedCh waiting for the process to start, but bypass.Submit blocks
until approval. Now we signal started immediately when entering bypass
path (with Process: nil), so the agent sees the proc ID right away.
Also handle term/kill when proc.proc is nil — cancel the context to stop
the operation (e.g., abort a pending bypass request).
executeBypassDirect was missing the streamOut parameter, so background
processes running via bypass never had their output written to the proc
buffer. Now both bypass and sandboxed paths receive the streaming writer.
Two bugs:
1. GUI sends 'agent=<profile>' but server only recognized 'profile='
- Added 'agent' as alias for 'profile' in parseAgentNewRequest
2. Model dropdown showed cost columns appended to model names
- /models format: backend<tab>model[<tab>in<tab>out...]
- Parser took everything after first tab as model name
- Now stops at second tab to extract just the model name
CreateEmpty now returns (session, created, error): if a session with the
given name already exists it is returned untouched instead of erroring,
and only a freshly created session receives the provided cwd/workflow/
variant. The o script no longer suppresses errors from session/new, so
real failures surface while re-creating an existing session stays a
no-op success.
Landlock cannot add rules for sockets (S_IFSOCK) or FIFOs (S_IFIFO).
Previously, attempting to add a path like /run/user/1000/wayland-0
would fail with EINVAL.
Now these special file types are detected and skipped gracefully.
The field is rejected by models that don't support thinking/reasoning
configuration. Check model name before setting the field; 'auto' and
unknown models skip it to avoid ValidationException errors.
Make the namespace explain itself instead of requiring prior knowledge.
- ctl is self-describing: reading it (empty write) or writing 'help'
returns the valid verbs with one-line descriptions; an unknown verb
errors with the valid list. Refactor dispatch to an ordered []ctlCmd
carrying descriptions; drop the undocumented '.' alias and the
drift-prone hardcoded help verb. Keep 'i' (drop 'inject') for fast
injects. o's ctl usage now reads the live listing.
- Errors carry severity + remediation. backend.ClassifyError maps the
typed errors to transient/config/fatal with a one-line fix; the error
event renders [[[error:<severity>]]] and a 'remediation:' line so a
human knows whether to wait or intervene.
- Add a human status file: 'thinking · 12s', 'calling shell · 3s',
'idle' — distinct from the machine-facing raw state. Wire the TUI bar
to it.
- Bare 'o' shows an overview of running sessions/agents with status, so
you don't need to know any names to get oriented.
Tests: dispatch help/unknown/routing, ClassifyError severity table.
reasoningEffort (low|medium|high) is now the single human-facing control
for model deliberation. Backends that speak a discrete level (OpenAI,
OpenRouter, Copilot, Gemini, Kiro) send it verbatim; Anthropic, which
needs a numeric budget, translates via documented thresholds
(low=4096, medium=8192, high=16384). thinkingBudget remains an advanced
explicit override that wins when set.
- Add EffortLow/Medium/High, threshold consts, ValidReasoningEffort,
EffortThinkingBudget, and GenerationParams.ResolvedThinkingBudget.
- Anthropic: derive budget from effort, grow max_tokens above the budget
(Anthropic requires budget < max_tokens), and force temperature=1 when
thinking is enabled (both are Anthropic requirements).
- Validate reasoningEffort at config load so typos fail loudly instead
of silently no-opping.
- theo: drop explicit thinkingBudget/inflated maxTokens; reasoningEffort
high now implies the 16384 budget.
- Document thresholds and per-backend behavior in data/agents/README.md;
fix stale autoLoad/maxSteps references.
- Add unit tests for the mapping and Anthropic thinking behavior.
Reasoning effort is part of an agent's behavior definition, so set it
on all 14 profiles keyed to role (planning/review high, general medium,
lightweight assist low). Also rename the ThinkingBudget json tag from
the confusing bare "reasoning" to "thinkingBudget", distinct from the
reasoningEffort string knob. No config migration needed — no profile or
backends.conf used the old "reasoning" key.
The autoLoad name implied an automatic tool-loading path that no longer
exists; tools now come only from agent config plus the /tool_load ctl
command. Rename the AgentConfig.AutoLoad field (json autoLoad) to Tools
(json tools), rename LoadAutoLoadTools to LoadTools, and update all 14
agent JSON profiles and the tool-not-loaded error message.
Regression guard for tool-free model compatibility: an agent with an
empty tool set must produce requests with no "tools"/"tool_choice"
field. Covers OpenAI, OpenRouter, Ollama, and Anthropic.
Levi Neely
Ollie Agent