- /agent [filter] lists the run's agents (glyphs, elapsed, task) via the
native selector and switchSession()es to the agent's own session file
- run is restored in the agent session by deriving the run id from the
session file path (<stateDir>/runs/<runId>/sessions/...)
- main session file recorded in the run manifest so /back works even after
a reload while viewing an agent chat
- widget re-attach timer guard on session switch
The /agents-panel now fills the whole right column (40% width, 100% height,
top-right anchor). Under each running subagent it shows the agent's most
recent activity read from the tail of its session file: the last tool call
with its query (e.g. web_search: "...") or the last assistant text. Reads are
cached by file size so the 1.5s refresh stays cheap.
Adds a toggleable non-capturing overlay anchored top-right (36% width,
max 85% height, hidden under 100 columns) that shows the live agent tree:
statuses with icons and colors, elapsed time, auto-refreshed every 1.5s.
Because it is non-capturing, typing in the editor keeps working while the
panel is visible. The panel closes via the toggle or on supervisor shutdown.
The interactive /agents overlay with keyboard navigation and details is
unchanged.
Adds an overlay panel (ctx.ui.custom with overlay options) showing the full
agent tree: tree glyphs, colored statuses and icons, live elapsed time, and
keyboard navigation (up/down, enter toggles a detail pane with the agent's
task and result, esc closes). The panel refreshes itself every 1.5s while
open. Also fixes /research-pause and /research-resume to use the same
loop-state guards as research_control (previously they could desync the tree
from the loop). The live widget header now hints at /agents.
The main session and subagents share the same model backend; when that
backend serializes requests (cloud rate limits or a local server), subagent
streams queue the main chat. Adds two config knobs:
- agent_concurrency (default 3): a semaphore in AgentTree.start caps how many
subagent sessions stream simultaneously (slot is released on completion,
timeout, or setup error so failures cannot deadlock the queue).
- subagent_model (optional): routes spawned agents to a different model or
backend, e.g. ollama/gemma4:e4b, so subagents never contend with the main
session at all. Wired through the spawn tool, the main-session tools, and
the per-agent runtime tools.
Documents both in .hypothesis-machine.example.yaml and adds a concurrency-cap
test (43 tests passing, tsc clean).
Adds a widget above the editor (ctx.ui.setWidget) showing the research run
id, loop status, iteration, active agents with elapsed time, and recently
finished agents. Refreshed every 1.5s while work is running, hidden when
idle; timer is cleaned up on shutdown so reloads do not leak intervals.
- research_control pause/resume now validate the loop state before touching
the tree, so pausing a stopped loop can no longer leave the tree 'paused'
while the loop is not (spawn would fail until a restart).
- read_artifact confinement resolves symlinks (realpath) before checking the
project boundary, preventing symlink-based file leaks.
- Adds confinement tests (42 tests passing, tsc clean).
- web.ts: outbound downloads resolve DNS once, validate the address, and
connect to the validated IP directly, closing the DNS-rebinding TOCTOU
window; malformed SearXNG result URLs are skipped instead of failing the
whole search; web cache is bounded (evicts oldest past 500 entries).
- experiment.ts: reviewer must not be a descendant of the experiment author
(tool-level lineage check on top of the author-id check); stdout/stderr
buffers capped at 512 KB.
- Confirmed Pi's extension tool registry is per-extension-instance, so
reload re-registration is safe; no change needed there.
- Adds a lineage-independence test (39 tests passing, tsc clean).
- AgentTree.restart() revives a stopped run in the same session: resets the
root, archives old branches, and restores spawnability (stop no longer
cancels the root).
- ResearchLoop.start() restarts from stopped/completed and resets iteration
counters; setGoal is allowed in terminal states; supervisor 'start' revives
the tree when the root is not spawnable.
- Cancelled/failed/archived children no longer count toward the per-agent
child limit and archived agents are ignored by the duplicate-task guard.
- cancel() no longer overwrites an already recorded agent result.
- Add agent_timeout_seconds (default 1800) so hung agent sessions fail
instead of blocking wait() forever.
- FTS search now prefix-matches tokens (tolerates inflections) and safely
handles punctuation/FTS5 metacharacters instead of returning false
negatives or throwing.
- Add tests for restart semantics, child limits, result preservation,
agent timeouts, and FTS morphology/special characters (38/38 passing).
Emil