Models that don't support native function calling emit tool calls as
plain text. Instead of treating this as ToolUnsupportedError, parse the
text and execute the tool calls normally.
Handles both direct JSON and shorthand (key=value) formats with relaxed
JSON support (unquoted keys).
The ! prefix allowed arbitrary command execution without sandbox
protection. Since agents share the same Submit path as users, this
was a sandbox bypass vector. Remove it entirely; users who need
shell access can use execute_code (sandboxed) or run commands
directly in their terminal.
- Add CWD field to PersistedSession for per-directory filtering
- /sessions now shows only sessions matching current cwd (-a for all)
- /save <name> creates a named checkpoint
- /resume <id> restores a saved session (messages + task state)
- Autosave (already existed) now persists cwd with each save
- Add pubsub.Bus to agent; all events published on "event" topic
- Remove EventHandler param from Submit (and internal methods)
- Queue() self-drains via goroutine when agent is idle
- /i submits directly when idle, injects when running
Session identity is now a numeric UID (starting at 10000) assigned at
creation and never changed. Session names can be renamed freely without
affecting the 9P identity.
- AgentCoreConfig.Uname carries the principal
- pushSessionEnv sets OLLIE_UNAME on the execute server
- startMount gated by WithMount() option (not env var)
- Close() uses context timeout to avoid hanging on fusermount
The execute server now starts ollie-9p-mount when OLLIE_SESSION_ID is
set, giving each agent session its own FUSE proxy that attaches to
olliesrv with Uname=session_id. Tools see the full namespace through
$OLLIE but all 9P operations carry the session's identity.
Close() tears down the mount on session destroy.
resolveExecPrompt spawns x/prime which uses envsubst to expand
${OLLIE_SESSION_ID} in the system prompt template. The variable was
never in the process environment at resolution time — it was only
set on the execute server after the prompt was already rendered.
Add env []string parameter to resolvePrompt/resolveExecPrompt and
BuildAgentEnv so callers can inject session-scoped vars into the
subprocess environment. Generate session ID before BuildAgentEnv in
the agent-switch path so it is available at prompt resolution time.
- Wire PlanFile from $OLLIE/s/{id}/plan into agentConfig so plan
re-injection is no longer dead code
- Add missing failureWindow/maxFailures/blockDuration constants
- Replace hand-rolled contains() with strings.Contains
- Fix rm -rf .. regex to match end-of-string and whitespace
- Add TestPlanReinjection covering periodic plan re-surfacing
- Add loop_maxsteps_helpers_test.go with mockResponse/newState/sequentialStream
- AgentCoreConfig.MaxSteps: when non-zero, overrides the value loaded
from the agent JSON (cfg.Env.maxSteps). Zero means "use the agent
JSON value, or unlimited if absent". Callers (9p frontend, httpgw,
CLI) can now inject a per-session limit at construction time without
touching any agent config file.
- /maxsteps [n] command: shows the current limit (0 = unlimited) or
sets it live during a session. Follows the same pattern as /model.
Added to /help listing.
Adds a monotonically increasing tool-call counter (tcct) to the agent.
- atomic.Int64 toolCallCount on the agent struct
- ToolCallCount() int64 on the Core interface
- IncrToolCallCount func() int64 on agentConfig, wired in executeTurn
- Incremented in execOne after the pre-tool hook returns (blocked calls
don't count) and before cfg.Exec is called (timeout/error still counts)
The counter never resets. Consumers use modulo arithmetic:
tcct % 10 == 0 -> fires every 10 actual executions
The primary consumer is ollie-9p, which exposes tcct as a read-only
9P file that pre_tool / post_tool hooks can inspect via cat.
After interrupt, tool results with cancelled status are removed but
the assistant message kept its ToolCalls referencing them, causing
'Improperly formed request' errors.
Now also strips cancelled tool call IDs from assistant messages.
Only removes the assistant message entirely if it has zero remaining
tool calls AND no text content.
- Add configurable retryBaseDelay and streamDropBaseDelay for tests
- Fix hook kill: don't wait for done channel after Kill() since child
processes may keep stdout/stderr pipes open
- Update tests to use short delays
Tests now run in ~0.4s instead of ~59s.
- On interrupt or error, keep user message, assistant response, and
completed tool results (success or error)
- Only remove cancelled/interrupted tool results
- Use deep copy for message snapshots to prevent aliasing bugs
- Snapshot now only used for context overflow retry
Identical read-safe tool calls in the same batch now execute only
once. Duplicates get the result copied with the correct ToolCallID.
Fixes a race where concurrent goroutines both missed the cache
before either could store, causing double execution.
On user interrupt (stop/CTRL+C), the session history now rolls back
to the pre-turn snapshot instead of keeping partial/cancelled tool
results in the model's context. This prevents raw errors like
'context canceled' from poisoning subsequent turns.
Tool results interrupted by cancellation now report 'tool execution
interrupted by user' instead of the raw Go error chain.
Levi Neely
Levi Neely