elophanto/EloPhanto v2026.05.19

v2026.05.19 — Watch the agent think. Fast Codex by default.

on GitHub

Headline: streaming chain-of-thought in the main chat

Watching the agent work used to mean staring at a blank screen while Codex ground for 30-180 seconds before responding. Same problem ChatGPT and Claude solved with streaming reasoning. EloPhanto now does it too — for both user-driven chat AND autonomous mind cycles.

Codex's Responses API streams reasoning-summary chunks; we wire those through a new AGENT_THOUGHT gateway event into the dashboard's main chat as dim italic 💭 lines:

─ You
  go check alphascala.com

  💭 Loading the homepage and scanning the navigation
  · browser_navigate — load homepage
  💭 Extracting menu labels and category structure
  · browser_extract — pull nav items
  💭 Cross-referencing the broker section with prior knowledge
  · knowledge_search — find existing AlphaScala notes

─ AlphaScala  codex/gpt-5.5
  Here's what's on the site…

Autonomous mind cycles get the same treatment — when no one is typing:

◆ mind cycle #14 — waking to think
  💭 Reviewing today's identity focus and recent dream history
  · goal_dream — generate candidates for the creation lens
  💭 Picking the candidate that fits the affect snapshot
  · goal_create — commit chosen candidate
◆ mind: created a new goal from required dream phase

Only Codex exposes streaming reasoning summaries today; OpenRouter / Z.ai / Kimi calls still surface tool calls + planner thoughts via step_progress lines but not internal monologue.

Live activity in the main chat (not just the bottom ticker)

The chat was previously "user message → final response" and nothing in between. Every interesting beat (tool calls, checkpoints, mind cycles) sank into the 5-line events ticker at the bottom which is fine for debugging but unwatchable as a stream. Now layered onto the chat:

• step_progress → dim · tool_name — first line of planner thought line per tool fire. Scheduled tasks tagged [mind] so autonomous activity is visually distinct.
• goal_checkpoint_complete → ✓ checkpoint N: title line in OK-green. Gives the chat a tangible progress beat between long stretches of dim tool calls.
• mind_wakeup → ◆ mind cycle #N — waking to think frames the upcoming autonomous burst instead of mystery activity.
• mind_action → ◆ mind: <summary> closes the cycle's narrative beat.

All lines are dim+indented so the agent's final responses still pop visually.

Fast Codex by default

config.demo.yaml shipped with reasoning_effort: medium for planning and coding. On Codex/gpt-5.5 that produces 90-180s latency on tool-heavy planning prompts (versus 5-15s with low). Fresh installs now default to low across all four task types (planning, coding, analysis, simple).

Existing installs need to either re-run elophanto init or edit config.yaml directly — the existing medium could be intentional, so no auto-migration.

README clarity

Added a one-time callout right after the diagnostics block explaining that elophanto lives inside .venv/ (not on global PATH) and how to invoke it: either source .venv/bin/activate once or prefix any elophanto <cmd> with ./start.sh. New operators no longer hit zsh: command not found: elophanto on their second shell.

Commits

adba49c Stream Codex chain-of-thought into the main chat (covers autonomous mode too)
3424edf config.demo: reasoning_effort low across all task types (was medium for planning/coding)
8752641 Dashboard: stream live activity into the main chat (not just the ticker)
7b5a7eb README: explain that elophanto lives inside .venv (one-time callout)