@tallshort/pi-inclusive-cost

A Pi coding agent extension that shows inclusive cost for the current session, including supported subagents and delegated workflows, not just the main thread.

[!NOTE] This repository is the tallshort-maintained fork of serhii-chernenko/pi-inclusive-cost, which remains the upstream project.

What this fork adds

Compared with upstream, this fork adds:

• Attributed cost collectors for Pi Fabric, Pi Taskflow, and a broad set of named subagent extensions.
• Strict current-chat attribution with stable contribution identities, cumulative-cost deduplication, and no blanket scans of unrelated shared histories.
• Lifecycle-aware accounting for persisted foreground, background, and nested work, plus backend-specific resume, cache, and recompute semantics.
• A modular collector architecture with backend-specific source precedence, conservative fallbacks, and regression coverage.

Example statusline with illustrative values:

↑1.9M ↓176k R62M CH8.9% $34.609 (sub) 66.8%/272k (auto)             (openai-codex) gpt-5.6-sol • medium
$37.72 (incl)

The problem

Pi's built-in statusline sums cost from the current session's own message history only. If you use a subagent extension (e.g. one that spawns background workers, reviewers, researchers, etc.), those subagent sessions are written to their own nested session files and are invisible to the built-in cost total — so the number in your statusline can understate what a task actually cost by a wide margin.

What this does

The extension rebuilds a deduplicated total from persisted records attributable to the current chat.

Collectors

Collector	Attributed source
Standard Pi sessions	Root and linked child Session JSONL files containing usage.cost.total
@pi-archimedes/subagent	Attributed synchronous details.results[].usage.cost records
pi-subagents-lite	Foreground Agent tool-result cost and attributed background completion messages
@ian-pascoe/pi-minimal-subagents	Root Session registry turn-settled usage, keyed by stable agent and turn identity
pi-better-subagents	Attributed launch run IDs and their retained sandboxed per-run standard child Session usage
Explicit Herdr child Sessions	Attributed pi-herdr-agents launch paths and @henryqw/pi-subagent direct-session results
@bacnh85/pi-subagent	Attributed foreground tool-result usage and background completion messages
pi-subagents-j0k3r	Attributed subagent_* tool-result task usage, with an explicitly referenced standard session as fallback
@ferris1225/pi-subagents	Run IDs from attributed launch results and cumulative usage from persisted subagent_status results; explicit standard sessions are fallback only
Pi Dynamic Workflows	Attributed workflow run IDs and cumulative cost from the run's small persisted v2 head
Pi Fabric	Run IDs and cumulative usage from fabric_exec rendered output or structured agent audit details, with attributable .fabric exports as fallback
Pi Taskflow	Run IDs returned by taskflow, then usage-bearing trace events or an older snapshot fallback

Supported extensions

Extension	Collector	Notes
Standard full-session subagent modes	Standard Pi sessions	Layout-based compatibility for nested session.jsonl or parent-linked sibling sessions; not a blanket claim about any package named pi-subagents
@pi-archimedes/subagent	@pi-archimedes/subagent	Synchronous results expose stable child Session IDs and usage without persisting child Session files
pi-subagents-lite	pi-subagents-lite	Foreground calls persist cumulative cost; background completions are linked to attributed launch IDs
@ian-pascoe/pi-minimal-subagents	@ian-pascoe/pi-minimal-subagents	Registry V2 settled-turn usage is authoritative; legacy V1 records and child Session totals are not collected
pi-better-subagents	pi-better-subagents	Strictly attributed retained sandboxed child Sessions only; shared session trees and large output.log files are not scanned
@henryqw/pi-subagent	Explicit Herdr child Sessions	Direct mode only: attributed launch and completion records identify standard child Sessions; isolated mode has no collectible persisted usage
pi-herdr-agents	Explicit Herdr child Sessions	Attributed launch records identify standard child Sessions; fork-inherited parent entries are excluded
@gotgenes/pi-subagents	Standard Pi sessions	Independently maintained hard fork of @tintinweb/pi-subagents; supported through its tasks/*.jsonl layout
@nklisch/pi-subagents	Standard Pi sessions	Maintained MIT fork of @gotgenes/pi-subagents; retains the same tasks/*.jsonl persistence while adding ordered lifecycle interception
@xzzpig/pi-subagents	Standard Pi sessions	Fork of nicobailon/pi-subagents; persisted native children use the same nested-session mechanism
oira666_pi-subagent	Standard Pi sessions	Persisted RPC children use nested subagents/.../session.jsonl files under the parent session stem
@bacnh85/pi-subagent	@bacnh85/pi-subagent	In-memory foreground results are keyed by parent tool call; background completions are keyed by task ID
pi-subagents-j0k3r	pi-subagents-j0k3r	Parent tool results provide stable task identities and usage; explicit child sessions are fallback only
@ferris1225/pi-subagents	@ferris1225/pi-subagents	Launch results establish stable run IDs; persisted status results provide usage; detached completion text is not parsed
Pi Dynamic Workflows	Pi Dynamic Workflows	Exact parent Session identity and run ID locate a small cumulative run head; large event logs are not read
Pi Fabric	Pi Fabric	Requires the parent chat to retain the Fabric run ID in rendered output or structured audit details
Pi Taskflow	Pi Taskflow	Requires the parent chat to retain the Taskflow run ID

Support for a named subagent extension means its attributable persisted mode is covered; in-memory or explicitly non-persistent modes remain out of scope. In particular, @henryqw/pi-subagent isolated mode uses --no-session and does not persist its internally observed usage in the parent result, so only direct mode is collected. The generic first row is based only on the standard persisted layout.

The statusline shows $X.XX (incl) whenever attributable delegated work exists. /cost keeps the compact three-line summary: Direct (this thread), Inclusive (+ subagents & delegated work), and Subagent/delegated contributions folded in. Refresh happens on session start, at the end of every turn, and when the agent run ends.

See docs/DESIGN.md for persistence layouts, source precedence, resume/cache/recompute behavior, contribution keys, and deduplication invariants.

It intentionally does not perform its own token-based cost estimation—it sums monetary cost fields already persisted by supported runtimes, such as Session usage.cost.total or Dynamic Workflows tokenUsage.cost. These values are themselves client-side estimates computed from reported token counts and local model pricing, not authoritative provider bills. When a runtime marks token usage as estimated but has no monetary cost, that unknown cost is omitted rather than treated as free.

Install

pi install npm:@tallshort/pi-inclusive-cost

or directly from source:

pi install git:github.com/tallshort/pi-inclusive-cost

Restart Pi, or run /reload.

Commands

• /cost — direct cost, inclusive cost from subagents and other delegated work, and the number of subagent/delegated contributions folded into the total.

Not currently collected

Extension	Reason
@esso0428/pi-subagents	Defaults to in-memory children and custom .output transcripts without attributable priced usage
@narumitw/pi-subagents	Starts children with --no-session and exposes neither stable jobId + usage.cost data nor attributable child Sessions; upstream support is required first
wj-pi-subagents	Starts children with --no-session; child replies contain text but no usage/cost

@tintinweb/pi-subagents is intentionally not given a package-specific collector: its optional reportUsage mode reports child cost through the parent Pi tool result, so collecting the same child usage again could double-count it.

The unscoped pi-subagents name in the preserved upstream limitation below did not include a package identity or version range. Treat it as a description of the observed artifacts-only layout, not as a claim about every similarly named npm package or fork.

Known limitation: newer artifacts-only subagent storage isn't covered yet

pi-subagents has changed how it persists at least one subagent mode (parallel/fork batches) between versions. Older versions always wrote a full session file for every subagent invocation, nested or as a flat sibling — the two mechanisms above. A newer mode instead writes only to <project-cwd>/.pi-subagents/artifacts/<runId>_<agent>_<index>_meta.json (plus a _transcript.jsonl with a different record schema), with no corresponding session file anywhere — nested or sibling. That cost is currently invisible to this extension.

This wasn't skipped by accident: there's no reliable way to attribute a given artifacts file back to one specific chat. The only correlation lead found (a session_info entry naming the subagent, written into the parent's own session file) turns out to depend on the pi-intercom package being installed — not a safe thing to build a core feature on. A worktree-wide sum without that correlation would also reintroduce exactly the cross-chat over-counting problem this extension was built to avoid in the first place (multiple unrelated chats in the same project folder getting lumped together). Until there's a first-party, version-stable way to scope this correctly, it's left uncovered rather than shipped fragile or wrong.

Collector-specific limitations for Fabric and Taskflow are documented in docs/DESIGN.md.

Don't expect this to match your provider's billing dashboard exactly

usage.cost.total is computed locally from token counts and a local price table — it is not your provider's authoritative bill. Even with accurate, current pricing, a locally-summed total can diverge from what your provider's own usage dashboard reports, for reasons outside this extension's control: reporting lag on the provider's side, per-request billing multipliers instead of per-token metering, or negotiated/enterprise rates that differ from the public price list. If you need a number that reconciles with billing, treat this as a fast local signal, not a substitute for your provider's own dashboard.

For a full project-wide audit across every chat (not just the current one), use a dedicated usage-dashboard tool — this extension is meant for a quick, live, per-chat signal in the statusline.

Releases

The first scoped release must create the public npm package before Trusted Publishing can be attached. From the clean release commit:

npm login
npm publish --access public
npm trust github '@tallshort/pi-inclusive-cost' \
  --repo tallshort/pi-inclusive-cost \
  --file publish.yml \
  --allow-publish \
  --yes
git tag v0.2.0
git push origin main v0.2.0

For later releases, bump version in package.json, commit, then push the matching tag. The tag workflow type-checks, tests, publishes through npm Trusted Publishing, and creates a GitHub Release with generated notes. It skips npm publication when that exact package version already exists, so the first-release tag only creates the matching GitHub Release. No npm token is stored in the repository; npm trust requires npm CLI ≥11.15.0.

License

MIT