pine-of-glass

Small observability and context management tools for Pi in the terminal.

1. contextimate breaks down what is filling your context window: sysprompt, AGENTS.md, Skill frontmatter, Tool schemas, and session material. Toggle with ctrl+o on start and /reload.
2. traceline collapses tool calls to one trace line each, so you can see the full arc of what Pi did (path taken, context read, bloated tool results). Toggle with ctrl+t (expands/collapses both thinking blocks and tools).
3. cachemire shows you what's happening to your cache and warns you about events
4. meantime (work-in-progress, feature flagged) understand latency from inference speed (TTFT, TPS) and potentially diagnose other things slowing down your agent like orcehstration or CI

See each extension's own README.md for details and docs/ for deeper reference.

Installation

Requires Pi 0.86.0 or later.

• From GitHub: pi install git:github.com/tmustier/pine-of-glass
• From npm (installs contextimate, traceline, and experimental cachemire and meantime): pi install npm:pine-of-glass. Meantime stays inert until its config sets "enabled": true.

Themes

All four extensions use Pi's selected theme. Running status uses its accent colour; warning, success and error states use their corresponding roles. The system theme lets Pi match these colours to your terminal. Existing notices, /cache and /pace ledgers, and live clocks recolour when you change themes. Historical facts and column alignment stay unchanged.

These real Pi 1.1 captures show the same ledgers with supplied dark and light terminal palettes:

Compatibility

The locked test runtime is Pi 0.87.1. Theme switching and startup have also been tested on Pi 1.1, but full Pi 1.1 compatibility remains under review. In particular, Traceline's Drill viewer may lose non-PNG previews on kitty-protocol terminals after Pi's image-conversion changes. See the release notes for validation and known limits.

Screenshots

Traceline

ctrl+t collapses each tool call and reasoning run to a single line. In fullscreen mode, click a compact call or reasoning preview to expand it independently, or use ctrl+t to reset the whole view.

Contextimate

Each new session and /reload now lists what's in your context window before you type a word. Toggle level of detail with ctrl+o.

Summary mode:

Compact mode, one aligned line per skill and tool:

Expanded mode, with per-section sources and a schema field tree for every active tool (excerpt):

Cachemire

You see cache-related warnings inline and above the editor only when the cache needs attention. /cache gives you a full session review by turn.

Meantime

Feature-flagged off by default. Once its config sets "enabled": true, a tempo line above the editor decomposes the current wait (waiting / thinking / writing / tools), slow calls earn a notice with a named cause, and /pace shows the per-call tempo ledger:

  call    ttft   think   write   tools   total     out  tok/s
     1    1.9s     42s    8.0s     31s   1m22s    2.1k     48
     2    1.7s    3.0s     12s    4.2s     17s    0.8k     41
     3     14s    6.1s    5.0s     22s     26s    0.7k     35  slow start (median 1.9s)
  totals: 3 calls · waiting 18s · thinking 51s · writing 25s · tools 57s · harness 4.6s
  timed 1h42m   ██▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒  active 10m (10%) · idle 1h32m

Plans

1. The current goal is to provide an interface that makes tool and agent behaviour more legible for humans using pi in their terminal interactively.
2. Once that's solid, we can help both humans and agents to more easily analyse previous sessions and traces, including when pi is running in RPC mode, remotely, or there is a large number of agent sessions we need to read to get insights from.

Note: A nice side benefit of 1. is that tools like traceline can help agents running interactive pi subagents in tmux to monitor them without risking context bloat from tool outputs.

Status: There are several things left to do in 1., including

• maturing cachemire (implemented, live-tested, conformant with the family design language, and now shipped as experimental while the UX settles; see issue #6),
• making contextimate more useful beyond upfront context (for example, are my MCP servers and skills efficient when loaded?)
• and other refinements tracked in issues.