Session creation now requires cwd= parameter (no fallback to daemon cwd).
Agent cwd is read-only — agents inherit from session, cannot override.
Frontends updated:
- NewSessionDialog: added Directory field with Browse button
- NewAgentDialog: removed Directory field
- createAgent(): removed cwd parameter
- Kate plugin: passes cwd when creating kate session
- Dolphin: removed cwd from agent/new call
Message history sanitization (backend.SanitizeMessages):
- Removes dangling tool calls (assistant with ToolCalls but missing results)
- Removes orphan tool results (tool message without preceding call)
- Applied before sub-agent context inheritance
- Applied before session persistence save
- Applied after compaction (defensive)
Includes unit tests for all sanitization cases.
- ollie-9p: Fall back to USER when OLLIE_UNAME not set (remove fatal error)
- goal/goalstatus: Use mode 0666 (world readable/writable) while keeping GID agent
Server:
- session/idx now outputs: session-id, session-name, paused, connected, remote, cwd
- session/{s}/agent/idx now outputs: session-id, agent-id, agent-name, parent-id, depth, state
- Add error if parent agent not found during sub-agent spawn
GUI:
- refreshSessions() reads session index, then agent/idx per session
- Auto-select first top-level agent (depth 0) instead of first in list
- SessionModel tracks agent expansion separately from session expansion
- Agents with children show expand/collapse arrows and are collapsible
- Add hasChildren role to SessionModel
Tools:
- subagent_spawn now passes parent= argument for proper depth tracking
- Remove hardcoded max_depth=5, use server default of 1
- Makefile installs shell script tools from data/tools/
KRunner:
- Update to parse new session/idx format
- Workflows are executable scripts in data/workflows/
- New 'workflows' 9P file lists available workflows
- Goal file stores text; writing triggers workflow if status allows
- goalstatus file for status read/write, goalwait for blocking
- Session ctl accepts 'run [workflow]' command
- Session now owns CWD; agents inherit via callback
- Conductor workflow: creates agent, primes with instructions, exits
- GUI workflow combo reads from workflows, not agents
- Persistence includes goal, goalstatus, workflow, and session CWD
Write to session/{s}/goal to set a session-level objective.
A conductor agent is spawned automatically in the background,
decomposes the goal, spawns sub-agents, and reports completion.
- goal file: write sets goal + starts conductor; read returns status
- goalwait file: blocks until goal status changes (BlockOnce)
- Conductor writes status=complete/blocked back to goal when done
- Session.Goal() / SetGoal() / GoalSignal() on Session struct
The step budget mechanism is gone. Agents run until they finish,
are interrupted by the user, or (for sub-agents) hit the timeout.
No replacement. The human is the kill switch.
Enforce three limits on sub-agent spawning:
- depth (default 1): sub-agents cannot spawn their own sub-agents
- parallelism (default unlimited): cap concurrent children per parent
- timeout (default 600s): sub-agents are killed after 10 minutes
Top-level agents are never constrained by timeout.
Also: refactored parseAgentNewRequest to return a struct instead of
4 positional values. Added depth/activeChildren fields to Agent.
OLLIE_SUBAGENT_DEPTH env var set on sub-agents.
Deferred: remove maxSteps (replace entirely with timeout).
Sub-agents can now inherit parent context truncated at a specific
user turn: fork_at=5 means 'fork from the state after the 5th
user message.' Enables backtracking without losing unsummarized
context from before a failed approach.
Without fork_at (or fork_at=0), full history is inherited as before.
- Add UserPrompts Prompt field to AgentConfig (same JSON format as prompt).
- Resolve userPrompts via resolvePrompt() in BuildRuntime, store in Runtime.UserPrompt.
- Prepend resolved user prompts to every user message at executeTurn start.
- Document userPrompts in data/agents/README.md.
- Remove per-step stripCold from run(); call once per turn in executeTurn
after accumulator reset and before preamble injection.
- Replace expensive deep-clone snapshot with pre-turn length truncation.
- Delete unused cloneMessages() helper.
- Remove per-step preamble prepending in loop.go; inject once at turn start
in turn.go instead of re-sending thousands of tokens every tool step.
- Re-inject preamble after overflow retry compaction where it was lost.
- Replace stripCold() N sequential LLM calls with one batched call that
summarizes all cold-zone tool results via a single JSON response.
Sub-agents now clone the parent agent's conversation history into
their initial context. The tool script passes parent=$OLLIE_UNAME
automatically. Mechanically identical to session restore — uses
RestoreHistoryFromMessages on the parent's Messages().
Add Sub-Agents section explaining agent/new with prompt= key.
Include examples for single and parallel sub-agent spawning.
Update 9P namespace table to show agent/new as rdwr.
Writing to session/{s}/agent/new with a prompt= key now blocks
until the agent completes its task, then returns the reply and
destroys the transient agent. Without prompt=, behaves as before
(creates agent, returns ID).
Also:
- Move ParsePayload/UnescapeValue to shared ollie/toolsrv package
- Remove duplicate implementations from cmd/toolsrv/internal/server
- Add session.CreateAgentWithParams for direct AgentParams usage
- Eliminate flattenParams/unescapeValue redundancy in fs package
Read, Write, and Rdwr are the three atomic 9P operations:
- Read: non-blocking read
- Write: non-blocking write (fire-and-forget)
- Rdwr: atomic write-then-read (blocking, produces result)
BlockOnce and Stream are special cases of Read.
Rdwr is its own primitive — not a variant of either.
All foreground tool calls now acquire a lock based on the tool's
declared scope and file path before execution:
- scope "read": no lock (reads never conflict)
- scope "write": exclusive lock on the file path
- scope "global": exclusive global lock (serializes with everything)
This ensures writes to the same path serialize regardless of which
agent initiated the call, enabling safe parallel sub-agents within
a session without explicit coordination.
Cross-session serialization (shared toolsrv per host) is left as
future work.
The persist path created agents with &Runtime{} (nil Preamble).
If a prompt arrived before resume rebuilt the runtime, Preamble.Set
panicked. Initialize with an empty Preamble.
ListProcs now does rdwr to toolsrv proc/list with the agent ID.
Returns pre-formatted text. Ctl proc handler uses it directly.
Removed all ListDetachedRaw map-parsing logic.
- Remove /tools from olliesrv root (toolsrv owns all tool state)
- Add /all file to toolsrv namespace (lists all available tools on disk)
- Add ListAllTools() to toolsrv client library
- Add tools_all ctl command to agent (reads from toolsrv/all)
- Update system prompt with tools_all usage
tool_load was a built-in intercept in the agent loop — the only
'tool' that didn't run in toolsrv. Removed entirely:
- Intercept in loop.go (25 lines)
- Script + .meta in data/tools/
- autoLoad references in agent configs
Loading tools is now exclusively via ctl (which already existed):
echo 'tool_load X' | ollie-9p write .../ctl
System prompt updated to show the ctl pattern.
On context cancellation, close the connection first to unblock
ReadFcall. Then wait for in-flight handlers, drain the response
channel, and wait for the writer to finish. No writes hit a
closed socket.
Frontends rely on getting the current state on timeout (heartbeat).
When BlockOnce returns empty (5s timeout), the server now calls the
plain Read handler as fallback. Files without a Read handler (like
feed) return nothing — consumer unaffected.
Same pattern as BlockOnce: framework handles the blocking loop.
readFn takes a base string, returns (data, nextBase, error).
signal fires when new data may be available.
Chat stream handlers now use Stream(a.ChatRead, a.ChatSignal).
StreamRaw retained for custom handlers.
streamChat() in support.go is now dead code (replaced by ChatRead).
BlockOnce now takes a value-reader and a signal source. The framework
handles the blocking loop: read → compare hash to base → if different
return → else wait on signal or timeout.
On timeout (ctx.Done), returns empty (not error) so clients re-open
cleanly.
BlockOnceRaw retained for queue-style handlers (bypass/pending,
proc/wait) that manage their own blocking.
EventValue adapter wraps a <-chan Event into BlockOnce-compatible
(readFn, signalFn) pair via a thin goroutine.
- BlockOnce handler initializes base to current hash on fresh open,
then blocks until hash changes. Same pattern as statewait.
- ConsumeFeed is a plain function: dials 9P, reads feed in a loop
(open → block → data → close → repeat), submits to agent.
- Context cancellation closes the 9P client, unblocking Read().
- Called as go ConsumeFeed(ctx, ag) from AddAgent and session resume.
Feed is now read-write (0666):
- Write: store data, signal via notifyChange
- Read (BlockOnce): block until content hash differs from base
Dedup lives in the read layer (same pattern as statewait). The
internal consumer goroutine and external 9P clients both call
FeedBlockingRead — same codepath, same dedup logic.
Consumer started from AddAgent (if ctx available) and session resume.
Feed is a write-only file in the agent namespace. Writes are
deduplicated against the previous value; the internal consumer
(blocking on WaitChange/WatchFeed) only wakes when genuinely new
data arrives.
Wiring is external and source-agnostic:
# human → observer (poll git):
while :; do git diff HEAD; sleep 5; done | ollie-9p write .../feed
# agent → observer (stream chat):
ollie-9p read .../coder/chat | ollie-9p write .../observer/feed
Uses the agent's existing signalCh/notifyChange plumbing — no new
channel infrastructure. Consumer goroutine spawned at agent creation
(AddAgent) and session resume, tied to session context.
- envLookup: provide XDG spec defaults (XDG_CONFIG_HOME, XDG_DATA_HOME,
XDG_CACHE_HOME) when env vars are unset, so agent profile prompts
resolve correctly on systems without explicit XDG vars.
- renderTools: skip generated '## name' header when the tool's prompt
text already starts with a markdown header, eliminating duplicates.
- file_grep.meta: replace /home/user/project with /abs/path
- system_prompt.md: use $XDG_CONFIG_HOME instead of ~
- agent-copilot.md: use relative path in code block example
Fix holes where queued prompts were orphaned:
- After turn loop exits (interrupt, toolsrv unavailable), drain remaining FIFO items
- On panic recovery, pop next FIFO item and re-submit in a new goroutine
- fifo file write triggers Submit when agent is idle
- ResolveTool now returns cmd field as-is instead of trying to resolve
it as a binary path. This allows meta-only tools to use full shell
command strings with env vars, pipes, and subshells.
- Prepend ~/.config/ollie/tools to PATH when executing tools so cmd
fields can reference other tools by name.
- Delete TurnCtx struct (was progressively populated with optional nil fields)
- Convert run(), streamResponse(), execToolCalls(), execBatch(), execOne(),
trackErrors(), retryCountdown() to Agent methods
- Extract autoCompact() and popInject() as Agent methods
- turn.go now installs a turn-scoped output handler wrapper instead of
building TurnCtx closures
- context.Context is now passed as plain context.Context, not smuggled
with unrelated state
Add history compaction tests:
- TestBuildCompactedHistory_OrphanedToolWalkback: verifies walk-back logic
prevents orphaned tool messages at hot zone boundary
- TestBuildCompactedHistory_NoOrphanAtBoundary: verifies tool messages
always have preceding assistant with tool_calls
- toolsrv pushes <system-proc-complete> to agent prompt when bg proc exits
- Remove bgTracker and CollectInterrupts from olliesrv
- Add Cmd, AgentID, SessionID fields to Proc
- Add OnProcExit callback to State
- Add command field to Proc.Stat() output
- ListProcs() now queries toolsrv directly instead of local tracker
- Remove ownership verification from SignalProc/ProcOutput/DismissProc
- Clean up stale procs from local tracker when they disappear from toolsrv
- Document local tracker quirk in TODO.md for future cleanup
- Detach background proc context from request context so processes survive
after the 9P request returns (fixes premature SIGTERM)
- Add /proc/list file to toolsrv showing all procs with state
- Show status (running/exited (N)) in agent proc list
- Keep exited procs in tracker until explicitly dismissed
- Skip 'list' entry in ListDetachedRaw
- Explicitly tell model not to invoke tool scripts via filesystem path
- Add concrete examples of native tool calls vs shell
- Document context cancellation flow from interrupt to process kill
- Remove olliesrv's bypass broker machinery (pendingCh, EvaluateRequest, etc)
- Session bypass loop now just reads from toolsrv's bypass/pending and notifies
- Notification handler writes directly to toolsrv's bypass/resolve (fire and forget)
- Remove bypass/ directory from olliesrv's 9P namespace
- Remove session/*/bypass file (policy can be added back to toolsrv later if needed)
The flow is now:
1. toolsrv blocks tool execution, exposes request via bypass/pending
2. olliesrv reads from toolsrv, shows D-Bus notification
3. User clicks approve/deny, olliesrv writes to toolsrv's bypass/resolve
4. toolsrv unblocks and executes (or denies)
- Added BlockingReadMode() and StreamMode() handling to toolsrv's handleRead()
- Made NextPending() context-aware in both toolsrv and olliesrv bypass packages
- This fixes the bypass approval flow via 9P - reads now properly block until
a request is available and respect context cancellation
Wire context cancellation through the toolsrv 9P server so that
canceling a CallTool context properly terminates the running tool.
Changes:
- virtfs: Add Close() to File interface with CloseFn for Request handlers
- toolsrv/p9.go: Handle requests concurrently to allow Tclunk during
blocking Twrite; call entry.Close() in handleClunk
- toolsrv/internal/server/proc.go: Check ctx.Done() while waiting for
tool completion in NewProc
- Integration tests for context cancellation chain
The cancellation chain: client cancels ctx → fid.Close() → Tclunk →
handleClunk calls entry.Close() → closeFn cancels reqCtx → NewProc
sees ctx.Done() and returns ctx.Err()
bypass.Submit now takes a context and returns early when cancelled.
This allows the stop command to properly cancel pending bypass requests
that haven't been approved or denied yet.
The full context chain is now:
1. stop command -> Interrupt -> cancel actCtx
2. actCtx cancellation -> closes CallTool fid
3. toolsrv proc/new sees closed fid -> propagates ctx cancellation
4. ExecuteTool -> executeBypassDirect -> bypass.Submit
5. bypass.Submit returns ctx.Err() when context is cancelled
Remove RequestCtx variant - all Request handlers now take context.
This is required for proper cancellation of blocking operations
like tool execution when the client disconnects or stop is called.
- Add RequestCtx field to FsNodeDecl for handlers that need context
- Add RequestCtx() helper function in virtfs/decl.go
- Update builder.go to handle RequestCtx in validation and Open
- Use RequestCtx in toolsrv proc/new and proc/new.bg handlers
This allows the context to be cancelled when the 9P fid is closed,
enabling proper stop/interrupt of running tool calls.
The sudo credential broker was never functional and added complexity
without value. This removes:
- Sudo field from bypass Request structs (broker, client, toolsrv)
- Sudo parameter from EvaluateRequest interface and implementations
- Sudo/ResetsCounter fields from MetaFile and Variant structs
- sudo: true from system_logs.meta variants
- All sudo documentation from writing-tools.md, tool-registry.md,
core.md, evolution.md, and misc.md
Bypass remains fully functional for sandbox escapes. Tools that need
elevated privileges should handle that internally or be run manually.
The old bypass broker used a Unix socket to receive requests and
execute commands locally. This was replaced by the 9P-based flow
where toolsrv executes commands after receiving approval from
olliesrv.
Removed:
- acceptLoop, handleConn, executeAndStream, executeWithSudo
- sendFrame, indexOf, socketDir helpers
- SocketPath and Credential config options
- Frame constants (FrameData, FrameExit)
- net.Conn field on Request
The broker now only handles:
- Policy management (global + per-session)
- Request evaluation (EvaluateRequest)
- Rate limiting
- User notification callbacks
Execution happens in toolsrv, not olliesrv.
This refactors the bypass (sandbox escape) mechanism to work with remote
toolsrv deployments. Previously, bypass used a Unix socket which only
works when toolsrv runs locally. Now:
1. toolsrv exposes bypass/{pending,resolve} 9P files
- pending: blocking read returns next bypass request as JSON
- resolve: write JSON {id, approved, error} to complete request
2. olliesrv reads bypass/pending in a loop per session
- Evaluates requests through the existing bypass broker
- Policy check, rate limiting, user notification all stay in olliesrv
- Writes approval/denial back to bypass/resolve
3. When approved, toolsrv executes the command directly (no sandbox)
- Execution happens on toolsrv's host (local or remote)
- Output streams back through the normal tool call path
This enables bypass to work when toolsrv is remote:
- User sees the approval notification locally
- Command executes on the remote host outside its sandbox
Architecture:
toolsrv (remote) olliesrv (local)
┌─────────────────┐ ┌──────────────────┐
│ sandboxed cmd │ │ bypass broker │
│ ↓ │ │ - policy │
│ bypass.Submit() │──────│ - notification │
│ ↓ │ 9P │ - rate limit │
│ wait for result │←─────│ - user approval │
│ ↓ │ └──────────────────┘
│ execute direct │
└─────────────────┘
The bypass broker requires a 'session' field in the JSON request
to identify which session is making the bypass request. The toolsrv
was sending requests without this field, causing 'denied (no session
identity)' errors.
Fix: Include OLLIE_SESSION_ID from environment in the bypass request.
Moved log buffer, plan, mutex, and condvar onto agent.Agent directly.
These were previously on the fs package's AgentLog type. The fs package
will be migrated to use these in the next commit.
New methods: AppendChat, EnsureTrailingNewline, ChatMu, ChatCond,
ChatLog, Plan, SetPlan.
Removed the named sandbox concept. One config file: sandbox.yaml
(installed to ~/.config/ollie/sandbox.yaml). Removed sandbox-remote.yaml,
moved restricted.yaml to doc/ as a sample. Removed the 'sandbox' arg
from tool dispatch.
The package defines the Server type and its namespace — 'fs' was a
leftover name from when it only held the filesystem spec. Now it's
the server definition. File renamed spec.go → server.go to match.
- Removed session/connected (used blocking Ping() on potentially stale conn)
- Removed per-agent tools file (tool management now via toolsrv ctl)
- Fixed session/idx: replaced blocking IsConnected() with non-blocking
toolsConn != nil check, preserving the field for GUI compatibility
- Removed dead IsConnected() method from session.Session
The namespace spec is now fully self-contained. Every handler is a
closure right where it's declared — no jumping between files.
Deleted: handlers_root.go, handlers_session.go, handlers_agent.go, ctl.go
Added: support.go (shared utilities: mergeCtx, streamChat, stripMarkers,
dispatch, wireAgentEvents, help memoization)
spec.go grew to ~920 lines but is now the single source of truth for
the entire 9P namespace. Know where to look.
Each() now returns []FsNodeDecl directly. The Binding type was a
redundant subset of FsNodeDecl with a slightly different Remove
signature. Dynamic entries are now expressed uniformly — Remove,
Rename, Children, Aliases all live on FsNodeDecl like everything else.
Also added: Alias() and RenameNode() options, DirNode accepts
[]FsNodeDecl for passing pre-built child slices.
SessionNode now embeds *session.Session instead of wrapping it.
All delegation methods (ID(), Name(), Ctx(), Pause(), etc.) are removed.
Handlers access session fields/methods directly through promotion.
Moved pause/resume event publishing into session.Session itself since
the session package already owns PublishEvent.
Eliminated all s.Session.X stutter from handler code.
Handlers now accept their actual dependencies directly:
- Root handlers: ModelCache, Broker, context.Context, etc.
- Session handlers: *SessionNode, removeFn, renameFn
- Agent handlers: *agent.Agent, *AgentLog, *SessionNode
Ctl dispatch tables are now per-instance closures built at binding
time rather than package-level vars with generic Hctx.
Also inlined the intermediate internalBinding type — adapt functions
now build fsedsl.Binding directly from raw data.
Removed: Hctx struct, rdwrHandler type, rdwrDispatch func,
bypassBindings helper, processBindings helper, internalBinding type.
Converted the entire 9P namespace spec from progressive Hctx population
to closure capture. Handlers are now closures that capture their
dependencies (session, agent, etc.) at binding time.
Hctx remains as a transitional adapter — handlers still receive it,
but it's constructed per-call from captured state rather than threaded
through the tree. Will be removed in a follow-up once handlers are
converted to use captured state directly.
Spec() now takes *Server and all handlers are closures that capture it.
Removed Ctx type, type aliases, and separate handler functions.
The tree structure and behavior are declared together in one place.
The registry was keyed by agent ID but read it from a shared env map
(st.env["OLLIE_UNAME"]) that all agents overwrote — making it
effectively per-session with a race condition.
Fix: pass agent ID explicitly through the protocol at every call site.
- ctl protocol: 'load <agentID> <tool>', 'unload <agentID> <tool>'
- tools file: rdwr (Request) — write agent ID, read filtered list
- proc/new payload: 'agent=<id>' field required
- Client: SetAgentID() stores identity, included in all operations
- Empty agent ID is a hard error everywhere
- Setting OLLIE_UNAME via env ctl is blocked (prevents reintroduction)
Tools are now scoped per-agent within a session. The registry uses
OLLIE_UNAME (agent identity) to partition loaded tools. Each agent
sees only its own tools — load/unload/list/lookup all key on the
agent ID read from the toolsrv environment.
BlockingReadMode had the same 512-byte truncation bug: it returned EOF
on offset>0 before the full content was delivered. Now caches content
on the fid like normal reads.
StreamMode is separated into its own path since it has fundamentally
different semantics (each Tread blocks for the next chunk, no caching).
In 9P, clients issue multiple Treads at increasing offsets to read
files larger than the initial buffer (io.ReadAll starts at 512 bytes).
The server was calling entry.Read() on every Tread, which for Request
files cleared the result after the first call, causing truncation.
Fix: cache the Read() result on the fid (readCache). Subsequent Treads
at higher offsets serve from the cache via readSlice. The cache is
cleared on write (for rdwr files) so the next request-response cycle
gets fresh data. This mirrors how dirCache already works for directories.
One toolsrv per session means the session-scoped map was unnecessary
indirection. Registry now tracks a flat map of loaded tools with no
session ID parameter. This eliminates the class of bugs where sessID
could be empty or mismatched between Load and Loaded calls.
Shell doesn't expand ~ in programmatically-passed arguments, so
toolsrv received literal ~/src/ollie as --cwd. Go's os/exec fails
with ENOENT on chdir before exec, breaking all tool execution.
Expand ~ at two boundaries:
- main.go: immediately after flag parsing
- proc.go SetCWD(): for runtime cwd changes via 9P
Remove all cached tool state from agent:
- Remove toolRegistryRevision from Runtime
- Remove wireToolsChanged (OnToolsChanged callback)
- Remove toolsNeedRefresh (revision-based cache invalidation)
- Remove refreshToolListing (redundant with RefreshTools)
Tools are now fetched fresh from toolsrv at the start of every turn.
The preamble tools section is also rebuilt at turn start. No local
caching means tool_load/unload take effect immediately on the next
turn and /tools always shows the live state from toolsrv.
tool_load was loading on the session connection but tools listed from
the agent connection. Now both use ctx.Agent.ToolServer() so loads are
immediately visible. writeAgentTools also calls RefreshTools after load.
- Background procs get timeout=0 (no deadline) set by toolsrv at
the execution layer — not injected as an arg from olliesrv.
Foreground default remains 30s. User can still override via args.
- Timeout logic simplified: caller sets default, ExecuteTool honors it.
- System prompt: clarify background has no timeout, uses 'stop' not 'kill'
- architecture.md: document lifecycle (connection-based ownership)
- evolution.md: update interrupt section with streaming, lifecycle, id= format
- README.md: add parallel execution to capabilities table
- KillAll(): sends SIGKILL to all running proc groups on clean shutdown
- Pdeathsig: SIGKILL ensures child procs die if toolsrv crashes
- Shell tool: use subshell + set +e for streaming without exit propagation
- Fix integration test to match new error format
This is the connection-based ownership model: toolsrv death = proc death.
The session owns the toolsrv process, so session death cascades to all procs.
- Timeout: timeout=0 means no deadline (was defaulting to 30s)
- Signal: send to process group (-pgid) not just process; SIGTERM no longer
cancels context (only SIGKILL does); cmd.Cancel sends SIGTERM with 5s WaitDelay
- Streaming: background procs stream output in real-time via procWriter;
shell tool no longer buffers all output into a bash variable
- Proc tree: olliesrv exposes proc/{id}/out, proc/{id}/ctl, proc/{id}/status
as proper 9P directory (was broken flat file)
- Connection: proc handlers dial fresh toolsrv conn per request via
Session.DialToolServer() to avoid deadlocking the agent's blocked conn
- Stat format: key=value (exited=true, exit_code=N, id=N) matching client parser
- GC: procs auto-removed 10min after LastRead (exited procs only)
- Rename: PID -> ID throughout (synthetic, not OS PID)
- Ctl commands: term (SIGTERM), kill (SIGKILL), signal <n>, dismiss
- System prompt: correct ollie-9p commands for proc management
Dispatch-level flags are now injected into every tool's JSON schema
at runtime via injectDispatchFlags(). No need to declare them in
individual .meta files — they're universal.
Removed redundant declarations from shell.meta (now injected globally).
When a tool call includes "background": true, it executes via
proc/new.bg and returns immediately with a PID in a
<system-proc-background> tag. The agent tracks active background
processes and injects their output as <system-proc-interrupt> blocks
alongside subsequent tool results.
The model sees background updates without polling. It can react to
build failures, log events, etc. naturally. Kill via PID when done.
Implementation:
- toolsrv client: CallToolBackground (uses proc/new.bg as rdwr)
- toolsrv: proc/new.bg upgraded from write-only to request-response
- agent: bgTracker collects last 20 lines of output per proc
- agent loop: injects interrupts after execToolCalls, before h.update
- system prompt: documents the background mechanism and rules
Only tools that explicitly declare scope="write" get path-based
parallelism. Unset scope is now global (full barrier), avoiding the
"path is a lie" problem where a tool has a path arg but modifies
other files (e.g., lsp_rename).
Tools must opt in to parallelism:
- scope=read: never conflicts
- scope=write: conflicts on same path only
- scope=global (or unset): serialization barrier
Added scope=write to file_edit.meta and file_write.meta.
Rename ToolInfo.ReadOnly bool → ToolInfo.Scope string with three values:
- "read" — path-scoped read, never conflicts (always parallel)
- "write" — path-scoped write, conflicts on same file path only
- "global" — full serialization barrier, runs alone
Tools declare scope in their .meta file. If unset, inferred from path
arg presence (write if path exists, global otherwise).
This correctly classifies lsp_rename as global (cross-file workspace
edits) despite having a path argument. The path arg in lsp_rename is
a symbol coordinate, not a resource scope.
All .meta files updated: readOnly:true → scope:read,
readOnly:false → scope:global, lsp_rename gets scope:global.
Replace binary ReadOnly/not batching with path-based conflict detection.
Tool calls that operate on different file paths now execute in parallel,
while same-path writes and shell (global barrier) serialize.
Conflict rules:
- ReadOnly tools: never conflict (unchanged from before)
- Write tools with path arg: conflict only on same path
- Shell / unknown (no path): global barrier, runs alone
This parallelizes the common refactoring case where the model edits
multiple files in one turn (e.g., 5 file_edits on different files
complete in 1 round-trip instead of 5).
Safety argument:
- LLM API is turn-based: model reads files first, plans edits based
on known content, emits writes in a subsequent turn
- file_edit carries old_string context: self-contained per-file
- Same-path conflicts serialize (detected via extractFilePath)
- Shell is conservative: always a full serialization barrier
- Move toolsrv logic into internal/{fs,exec,registry} packages to match
olliesrv's structure. server.go (9P protocol) stays at top level.
- Deduplicate parseKV/parsePayload: remove dead code from old spec9p.go,
single implementation in internal/fs/parse.go.
- Add structured logging via log/slog (replaces fmt.Fprintf to stderr).
- Standardize qid path hashing to fnv.New64a (matches olliesrv).
- Fix toolsrv process leaks:
- Add Pdeathsig to local spawn so toolsrv dies when olliesrv exits.
- Add idle timeout (30s with no connections after first client seen).
- ProcessKeeper.Dial() kills old process before respawning.
Split toolsrv/registry.go:
- cmd/toolsrv/registry.go: Registry type with session-scoped tool state
(Load, Unload, Loaded, Lookup, Revision methods)
- toolsrv/registry.go: shared types only (ToolInfo, ToolsPath, DiscoverTools)
The toolsrv/ package is now a pure client SDK:
- Dial, Conn for 9P connection
- ToolInfo, ToolResult, HostInfo types
- DiscoverTools for listing available tools
Server-only code lives in cmd/toolsrv/:
- Registry for session-scoped tool state
- 9P handlers and execution logic
- Sandbox integration
Move spawn.go from toolsrv/ to cmd/olliesrv/internal/toolclient/:
- Process, ProcessKeeper, Spawn, SpawnRemote now in toolclient package
- toolsrv/ now only contains client code (Dial, Conn, etc.)
This clarifies the architecture:
- toolsrv/ = client SDK for connecting to toolsrv
- cmd/toolsrv/ = the toolsrv server
- cmd/olliesrv/internal/toolclient/ = olliesrv's toolsrv process management
Removed ProcessKeeper tests from toolsrv integration tests since they
now belong to toolclient.
tool_load must be handled specially by the agent runtime because tools
run inside toolsrv's sandbox and cannot load other tools into
themselves. The runtime now intercepts tool_load calls and directly
invokes ToolServer.LoadTool().
This is the only built-in tool - all others are external scripts.
Add /session/{sid}/agent/{aid}/tools file:
- Read: list loaded tools for this agent
- Write: load a tool by name
Also add missing tool_load to justfile install.
HandleCtl was only handling load/unload commands. The client sends
'env KEY=VALUE' and 'cwd PATH' via SetEnv() and SetCWD(), but toolsrv
was returning 'unknown command' errors. This broke tool_load and any
tool that depends on OLLIE_SESSION_ID/OLLIE_UNAME env vars.
- Pass session ID via --session-id flag (not env var) so tool registry is configured when olliesrv spawns toolsrv
- /tools now returns JSON array with full ToolInfo including InputSchema (fixes Bedrock validation error requiring type: object)
- Implement SpawnRemote: deploy ~/.config/ollie via tarball over SSH, set XDG_CONFIG_HOME so tools/*.meta are found on remote
- Add --no-auth flag for debugging Tauth issues
Replace HMAC-based registration with proper 9P Tauth flow:
- First client to connect sets the secret via Tauth afid write
- Server stores secret in memory, returns session token
- Subsequent clients must provide matching secret
- Secret never in env vars or disk, only transmitted over socket
Changes:
- server9p.go: Authenticate() replaces RegisterAgent/GetAgent
- client9p.go: Dial() does Tauth handshake (write secret, read token)
- spawn.go: generates secret, no longer passes via TOOLSRV_SECRET env
- auth9p.go: simplified to just GenerateSecret()
- spec9p.go: removed /register file, token validation at connection level
- cmd/toolsrv/server.go: handleAuth/handleAuthWrite/handleAuthRead
Security model:
- Session generates secret on Spawn()
- All agents share session's secret via ProcessKeeper
- Socket permissions (local) or SSH (remote) protect transport
- Secret dies with toolsrv process, new secret on respawn
Added integration tests verifying:
- First connection sets secret
- Reconnect with same secret works
- Wrong secret rejected
- ProcessKeeper reconnect/respawn behavior
- Concurrent connections
- server.go: Full 9P2000 protocol handling
- version, auth, attach, walk, open, read, write, clunk, stat
- Uses fsedsl.Tree for all operations
- Handles directories and files
- Supports rdwr files (RequestRespMode)
- main.go: Remove placeholder serve9P, use new implementation
The toolsrv binary now serves a working 9P filesystem with the
namespace defined in toolsrv/spec9p.go:
/register, /ctl, /tools, /proc/*, /info
- doc/boot-sequence.md: init chain from main → fs.NewRoot → session.Init
- Event protocol: comprehensive table on Event struct (all roles, semantics)
- RPC server moved from cmd/ollie-remote into toolsrv/rpcserver.go
(binary is now a 93-line thin wrapper)
- SetSessionEnv folded into NewAgent (no post-construction wiring needed)
- AgentCfg → AgentParams (distinguish from AgentConfig JSON schema)
- ToolResult/ToolResultContent shared types in toolsrv/rpcwire.go
- toolsrv/shell.go → toolsrv/exec.go (name matches purpose)
- Resume unified: single path, always through Keeper
- AGENTS.md: fix fs/builder.go → fsedsl/builder.go, handlers_*.go
- fsedsl/Tree: document dual-mode (EDSL lazy vs Manual Mount)
- Truncation stack documented at defaultToolResultMaxBytes
- Integration test rewritten to use srv.ServeRPC directly
The sandbox escape mechanism is a bypass, not privilege elevation.
The old name caused the agent to confuse it with sudo.
- elevate/ → bypass/ (package, types, tests)
- elevate_notify.go → bypass_notify.go
- Namespace: /elevate → /bypass, session/*/elevate → session/*/bypass
- Tool arg: "elevated" → "bypass"
- Env: OLLIE_ELEVATE_SOCKET → OLLIE_BYPASS_SOCKET
- File: elevate-policy.yaml → bypass-policy.yaml
- All docs, prompts, and scripts updated
cmdRead now opens the file and reads incrementally, writing each chunk
to stdout immediately. This fixes streaming for stream-mode files (chat,
statewait, eventwait) which block per-Tread until data arrives.
The previous io.ReadAll approach buffered internally and never flushed
to stdout because stream files never return EOF.
No separate 'tail' command needed — 'read' just works for both
one-shot files (server returns data then EOF) and stream files
(server blocks between chunks).
- Remove all OLLIE_*_PATH vars (TOOLS_PATH, CFG_PATH, DATA_PATH, etc.)
Use XDG_CONFIG_HOME/ollie/* and XDG_DATA_HOME/ollie/* instead
- Replace OLLIE_<TAG>_LOG per-component logging with single OLLIE_LOG={level}
- Replace OLLIE_USAGE_LOG with XDG_DATA_HOME-based path
- Replace OLLIE_OLLAMA_URL with standard OLLAMA_HOST (or omit entirely)
- Rename protocol markers: OLLIE_LISTEN_READY → ListenReady, OLLIE_9P_OPEN → Open
- Remove freeloader hooks from agent configs
- Update sandbox YAMLs to use XDG paths instead of OLLIE_*_PATH tokens
- Update shell tools to use $(dirname "$0") fallback instead of OLLIE_TOOLS_PATH
- Update docs and prompts accordingly
- Removed from env.go managed list and defaults map
- main.go and shell.go derive path from XDG_RUNTIME_DIR directly
- main.go sets OLLIE_ELEVATE_SOCKET in process env for subprocesses
The 9P logger was hardcoded to LevelDebug, causing all 9P operations
to print timing info to stdout. This blocked the server on terminal
output, hurting GUI responsiveness.
Now uses NewLogger('9p') which:
- Defaults to the sink's level (warn)
- Can be overridden with OLLIE_9P_LOG=debug
Also updates kde submodule with incremental session tree updates.
Event topics now follow session hierarchy:
session.{sid}.new/kill/pause/resume/rename
session.{sid}.agent.{aid}.new/kill/state
Wildcard routing publishes to all ancestors:
session.{sid}.agent.{aid}.state publishes to:
session.{sid}.agent.{aid}.*
session.{sid}.agent.*
session.{sid}.*
session.*
*
GUI can subscribe to session.{sid}.* to get all events
for one session including its agents.
Also:
- ollie-9p -a flag now supports TCP (host:port or tcp!host:port)
- Add Session.ID() accessor method
The 9P protocol returns write handler errors via Rclunk, not Rwrite.
cmdWrite and cmdRdwr were using defer fid.Close() which discarded the
error. Now they check fid.Close() return value and exit non-zero on
failure, allowing the GUI to detect rename errors.
- Load ~/.config/ollie/env and set path defaults at startup (same as olliesrv)
- Remove hard exit on missing OLLIE_SESSION_ID; accept it from inherited env
or via set_env RPC (SSH bootstrap path)
- Wire OnEnvSet to propagate OLLIE_SESSION_ID into tool registry and process
env when it arrives dynamically; same for OLLIE_UNAME
ollie-remote was sending 'agent-<pid>' as the session identity to the
elevate broker, which never matches any real session. Now reads
OLLIE_SESSION_ID from inherited environment (set by the parent
olliesrv process) so the broker can validate the session exists.
For Request-mode files (session/new, complete, generate, route), the
write handler fires immediately during Twrite. On clunk, the server
was also flushing the (empty) writeBuf to Write(), causing the handler
to fire a second time. This created duplicate sessions.
Skip the clunk-time writeBuf flush for Request-mode files since they
already handle writes inline during Twrite.
With the zero-tools transition complete, the tools/builtin/ package
served no purpose — Builtins() returned nil. Removed:
- tools/builtin/builtins.go and the directory
- builtins field, WithBuiltins, Dispatch from toolsrv.Server
- Handler type from toolsrv (only used by builtins)
- OnPreDispatch/WithOnPreDispatch (dead code)
- Strict field and WithStrict (dead code, never read)
- CallTool simplified — no fallback to Dispatch
- builtin import and WithBuiltins/WithStrict calls from ollie-remote
- tools/builtin from justfile build/test targets
Updated AGENTS.md, ARCHITECTURE.md, README.md, EVOLUTION.md
The panic at fs/lifecycle.go:463 was a nil function pointer: fs.Config.MkdirAll
is a func field, not a method, and the struct literal in cmd/olliesrv/main.go
omitted it, defaulting to nil. When CreateAgent called
rs.cfg.MkdirAll(...), it panicked before ever reaching the autoload loop.
- Add MkdirAll: os.MkdirAll to fs.Config in main.go
- Spawn tool server in Create() so empty session/new sessions are
operational, then reuse it in CreateAgent (else if sess.proc != nil)
- Wrap create agent error with fmt.Errorf for clearer messages
- Add panic stack trace to stderr for debugging
After Tcreate on a file, the fid is open and should be usable for
read/write without a separate Topen. Set f.entry via root.Open()
after creation so subsequent reads/writes on the same fid work.
Tree.Open() returns an error for directories (is a directory), so
f.entry stays nil after Topen on a dir. The entry == nil guard in
read() was firing before the isDir branch, breaking directory reads.
Moved the guard after the isDir check — directories use buildDirData
directly and never need entry.
- handle() is gone — the dispatch + panic recovery is inlined into
the per-request goroutine in Start(). Panic recovery is lifecycle,
not protocol handling.
- Serve → Start: accepts a connection and runs the 9P loop
- Shutdown → Kill: stops accepting, cancels sessions, waits for
goroutines to drain.
- 2 methods remain on *Server: Start, Kill. Server is purely lifecycle.
- All 9P protocol handlers are package-level functions.
- attach() is now a function taking (*olog.Logger, *GroupTable, ...)
- open() is now a function taking (*fs.Tree, *GroupTable, ...)
- Both captured in the handler map via closures over s.sink, s.groups
- 3 methods remain on *Server: Serve, handle, Shutdown
- Added entry (fs.File) field to fid struct
- open() stores the result of rootTree.Open() on f.entry after
permission check succeeds
- read() uses f.entry instead of calling root.Open(path[1:]) —
fids survive renames because the File object references the
tree entry by pointer, not by path
- write() uses f.entry for RequestRespMode writes — no longer
needs root parameter at all; it's a pure connState+fid function
- attach: reject if fc.Fid already in use (was silently overwriting)
- walk: reject if fc.Newfid already in use (both clone and full walk)
- open: validate fid path exists; delete stale fid if path is gone
- stat: validate fid path exists; delete stale fid if path is gone
- wstat: after rename, GC all fids on this connection that reference
the old path tree (was previously updating paths — now just deletes)
The server fully owns the fid map. The client proposes fid numbers,
the server validates and accepts or rejects. Stale fids (pointing to
renamed/deleted paths) are garbage collected when detected.
- GroupTable is a standalone type with Add/Remove/IsMember, owns its
own mutex and map. Server now holds groups *GroupTable.
- Removed AddGroup/RemoveGroup/InGroup from *Server (3 fewer methods).
- Permission check in open() now computes effective bits as a mask:
bits := perm & 7 // world bits
if in group { bits |= (perm >> 3) & 7 } // group bits
if is owner { bits |= (perm >> 6) & 7 } // owner bits
if !hasPermBits(mode, bits) → denied
- No boolean accumulator, no tiered if/else, just bitmask OR.
- 6 remaining Server methods (down from 23).
- hasPermBits uses a 4-entry mask array indexed by mode&3 to select
the required permission bits: OREAD→4, OWRITE→2, ORDWR→6, OEXEC→1
- Zero branches, zero error paths — returns bits&mask == mask
- open() uses OR semantics: checks owner, group, world independently;
any tier granting access is sufficient
- checkPerm removed entirely; permission logic is pure boolean math
- checkPerm no longer needs *Server: it takes uid, perm, and inGroup
directly. The caller (open) resolves uid/gid via fileOwnerGroup,
perm via makeStat, and group membership via InGroup.
- checkPerm is now a pure function: (uname, mode, uid, perm, inGroup)
→ returns only on permissions, needs nothing from Server state.
- open is the remaining method that resolves these values.
server.go: 10 remaining Server methods.
Converted walk, create, read, write, stat, clunk, remove, makeStat,
and buildDirData from *Server methods to functions taking *fs.Tree
as first parameter. Exactly the same pattern as fileOwnerGroup.
Also:
- Replaced the switch in handle() with a s.handler map, built once
in New(). All 11 9P message types (Tversion through Tremove) now
dispatch through the map.
- Extracted flush() as a package-level function (was inlined in switch)
- handle() is now: recover + map lookup. No inline cases, no switch.
- server.go: 802 lines, 10 remaining Server methods (down from 23)
- 20 package-level functions handle the actual protocol logic
- readFile was dead code (defined but never called)
- writeFile was a 7-line helper called only once in clunk;
inlined as direct s.rootTree.Open + e.Write
- FileTree type alias now unused after both deletions
fileOwnerGroup only needed s.rootTree, not the full *Server. Making
it a function taking (*fs.Tree, string) removes one more method from
Server, makes the dependency explicit, and makes it unit-testable.
- openEntry (one-liner delegating to rootTree.Open): inlined into
the two call sites in read() and write()
- statPath (one-liner delegating to rootTree.Stat): inlined into
three call sites (fileOwnerGroup, walk, makeStat)
- InterruptAll (one-liner delegating to fs.InterruptAll): deleted,
it was dead code with no callers anywhere
Also cleaned up: removed unused File type alias and unused 'os' import.
server.go: 825 -> 810 lines. Server method count drops by 3, but more
importantly, zero forwarding boilerplate remains.
Ollie Agent
Levi Neely