Streaming partials were appended to log.raw per chunk, each carrying the
full cumulative content, so one response left dozens of partial lines
persisted. GUI clients parsing the snapshot re-rendered the growing block
once per partial (O(N^2)) and replayed all historical partials on every
reconnect, causing intermittent rendering loops.
Partials are no longer persisted: AppendBlock writes only finalized
blocks to log.raw; SetPartial broadcasts the in-flight block without
storing it. New chat.raw StreamRaw file is the authoritative live JSONL
source (finalized history replay, then live deltas); log.raw is a
finalized-only one-shot snapshot. GUI streams chat.raw, never polls
log.raw.
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'.
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 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- All non-historical references in docs and code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.
Add per-model pricing to /models output. Format:
backend<tab>model[<tab>in<tab>out<tab>cache_read<tab>cache_write]
Pricing sources:
- Anthropic: hardcoded from official pricing, includes cache rates
- OpenRouter: parsed from API response pricing field
- Other backends: static lookup table fallback, marked with (e)
Changes:
- backend: Add ModelPricing/ModelInfo types, ModelLister interface,
static price table (Claude, GPT, Gemini, DeepSeek), LookupStaticPricing()
- openai: Parse pricing from API, implement ModelsInfo()
- anthropic: Implement ModelsInfo() with hardcoded cache rates
- fs/cache: Use ModelsInfo when available, fall back to static lookup,
format prices per 1M tokens with (e) suffix for estimates
- agent/cost: Use shared LookupStaticPricing instead of duplicate table
- Add quirks package for stupid model behavior workarounds
- ShellInvokesNativeTool blocks shell(cmd="tool_name") patterns
- Add client_9p tool: native wrapper for ollie-9p operations
- Block ollie-9p in shell — use client_9p instead
- Update all prompts to use client_9p, not shell+ollie-9p
- Clarify 9P namespace is complete (tools are NOT in 9P)
- Registry.All() lists all available tools for validation
Add explicit guidance to Autonomous Operation section:
- New unacceptable behaviors: giving up before checking skill_list or trying to load tools
- New checklist: must check skills, try loading tools, attempt even uncertain options before reporting failure
Added explicit unacceptable behaviors:
- Mentioning a tool without calling it
- Acknowledging a tool exists but not using it
Added bold statement: 'Knowing a tool exists is not the same as using it.
Your response should contain tool calls, not descriptions of tools you
could call.'
Agents were using shell to invoke tools (cat, grep, ollie-9p) instead of
calling the actual tools. Added unmistakable WRONG/RIGHT code blocks
showing the correct pattern. Shell is ONLY for git, make, npm, etc.
Rewrote system prompt to enforce immediate tool use:
- New 'Autonomous Operation' section: act first, report results
- Explicit list of unacceptable behaviors (narrating intentions, asking
permission for routine ops, producing text when tools should be called)
- Tools section: 'Use them without hesitation', concrete examples
- Skills section: 'if the task needs it, load it' — no asking
- Stronger sub-agent prefix: 'Do NOT respond with a plan. Call tools.'
Timeout fix:
- Add Timeout field to ToolInfo (protocol) and MetaFile (metadata)
- proc.go respects tool-declared timeout before falling back to 30s default
- subagent_spawn.meta declares timeout=0 (no timeout) so the tool is
never killed prematurely while waiting for the sub-agent to finish
- Tool schema declares timeout with 'do not set' guidance to prevent
the LLM from adding a short timeout
Premature response fix:
- Inject behavioral prefix into sub-agent prompt: complete all work
before responding, report results not intentions
- Sub-agent's final text is returned to parent; this instruction ensures
it contains accomplished work, not a plan
Add peer/ directory to each agent's 9P namespace. Agents communicate
by writing to peer/{name}, which delivers to the target's prompt handler.
Only declared peers can be messaged — the directory is the ACL.
Implementation:
- Agent struct: peers map + AddPeer/RemovePeer/Peers methods
- fs/spec.go: peer/ Each node (write-only entries), peeradd/peerdel/peers ctl commands
- Bidirectional: peeradd A on B also adds B on A
- Peers constrained to same session
- Persisted with session state (PersistedAgent.Peers field)
- peeradd/peerdel trigger immediate session save
Docs updated: system_prompt.md, AGENTS.md, README.md, architecture-9p.md,
architecture-core.md, architecture.md, usage.md.
- Add MEMO_TOOLS=1 env var to memo script; when set, all printed
instructions reference native tool names instead of memo paths
- Set MEMO_TOOLS=1 in all memory tool .meta wrappers
- Add memory_nap tool for compressions
- Add memory_zoom tool for tree navigation
- Add part/T pagination args to memory_wake
- Update system prompt to use memory_zoom tool call
- Fix inject ctl: submit as user message when agent is idle
- Add OptMem memory section to system_prompt.md with tool-based API
- Fix KRunner plugin icon: resolve via QStandardPaths instead of theme name
- Fix desktop file icon: use absolute path to bypass stale system icon
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
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).
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.
- 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.
- 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
- 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
- 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
- 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
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
Ollie Agent
Levi Neely