savvy-progress

savvy-progress adds /agents-info, a panel that lists the session’s subagents (running, finished and planned) with model, effort, context fill, estimated cost and elapsed time. It also draws a progress bar above the prompt for savvy-flow, a skill from the same repo that plans a task, hands the pieces to tiered worker subagents and reviews what comes back. The panel works with any subagent; the bar needs savvy-flow.

What it does

With any subagent

/agents-info opens the panel, and running it again closes it. Every subagent the session starts gets a row with:

• The description the model gave it, its type, and its model with the effort level.
• Context fill as a percentage and in tokens, estimated cost, and elapsed time, which ticks every second while it runs.
• A mark for running, done or error (a run that ends without an answer), and round 2 when a task with the same description runs again.
• A thin bar under the row. Subagents that do not report steps, which is every subagent outside savvy-flow, show their context fill there in grey.

The header gives the total cost and tokens of all runs and the time from the first start to the last finish. Finished runs fold away under their own heading, and Collapse shrinks the panel to one line of icons and the totals.

With savvy-flow

• A bar above the prompt: the flow’s title, the phase (Plan, Design, Tasks, Review, Done), accepted tasks out of planned (Tasks 3/7), a percentage, a ×N button with the crew size that opens the panel, and a ✕ that clears the bar. Tasks handed to workers but not yet accepted fill a fainter layer.
• Planned tasks at the bottom of the panel, each with its tier, the model that tier runs on and the tasks it waits for. A task leaves the list when a subagent starts with the same description.
• Step progress: each worker reports its own plan of three to eight steps, so its bar shows 4/6 and the step in progress instead of context.
• Costumes: Fable workers are astronauts, heavy ones detectives with a magnifier, careful ones engineers with a wrench, medium ones chefs tossing an omelette, light ones racers with a flag. An Explore subagent is a pirate; any other subagent is a plain crab.
• The panel opens by itself once per flow.

The bar shows up only when something reports through the mod’s progress tool, as savvy-flow does, or when a savvy-* worker starts. Without savvy-flow there is no bar, no planned list and no step progress.

In the Claude Desktop Code tab the code draws each row and the bar as SVG, with pixel crabs that walk while their agent runs; prefers-reduced-motion stops the animation. In the terminal the same rows are text: block-character bars and a colored mark per run, no crabs.

Install

Needs Claude Code 2.1.287 or later, plus the savvy-flow plugin from the same marketplace for the progress bar, planned tasks and step progress.

claude plugin marketplace add JohnnyVizz/claude-kit
claude plugin install savvy-progress@claude-kit

For the bar, add claude plugin install savvy-flow@claude-kit. Restart the session, then run /agents-info. Installed as a plugin, the skill is namespaced: /savvy-flow:savvy-flow <task>.

If the mod does not show up after a marketplace install, the author’s fallback is to load it from a clone:

git clone https://github.com/johnnyvizz/claude-kit.git ~/claude-kit
claude --plugin-dir ~/claude-kit/plugins/savvy-progress

The panel speaks English or Russian. The language option defaults to auto, which follows Claude Code’s language setting, then the system locale, and falls back to English; set en or ru in /config.

How it works

The mod registers two tools the model can call:

• mcp__savvy-progress__progress: the orchestrator reports the plan (title, number of tasks, each task’s tier and what it waits for), the phase, the tasks accepted after review, and the end of the flow.
• mcp__savvy-progress__step: a worker reports steps done and planned, with a short note. Calls from the main session are ignored.

Their descriptions sit in the tool list the model sees, so they cost a few tokens per request, and savvy-flow is what tells the model to call them. The rest comes from engine events and costs no tokens: agent.spawn gives each new subagent’s description, type and model; turn.step gives the token usage of every request a subagent makes; turn.complete says whether it ended with an answer. On tool.call for the Agent tool it reads only the subagent type, to count savvy-* workers.

It reads Claude Code’s settings and the locale variables to pick the language. It keeps the flow, the last 200 runs and the panel toggles in the engine’s plugin state, makes no network calls and starts no processes.

Limitations

• Totals cover subagents only. The main session’s own tokens and cost are not in the panel.
• Cost is an estimate: token counts times a per-model price table in the module (PRICES in hooks/register.tsx), not a bill. Models the table does not recognize are priced as Opus 5.5.
• The context percentage assumes a 1M-token window for every model except Haiku, which gets 200k.
• The events and methods the mod uses can change between Claude Code versions.