Session context
hh inspect --context answers four questions about a recorded session:
which model ran, through which provider, for how long, and at what token
cost.
hh inspect last --context # the panel
hh inspect a1b2c3 --context --json # the same data as one JSON object
a1b2c3 · context
name auth refactor
model Claude Opus 4.8 (claude-opus-4-8)
provider Anthropic
context limit 1,000,000 tokens
context used 51,200 tokens · 5.1% of limit · latest turn, self-reported
capture transcript · claude-code
tokens self-reported, accumulated
input 2,400
output 680
cache read 96,000
cache write 4,000
reasoning —
total 103,080 derived
estimated ≈ 140 tokens · size-based (bytes ÷ 4), not self-reported
messages 1 user · 3 agent · 1 tool call
created 2023-11-14 22:13:20 UTC
last activity 2023-11-14 22:13:23 UTC
duration 3s wall · 3s active
est. cost $0.10 at list price
models claude-opus-4-8 (steps 5–6)
The panel uses one accent colour and aligned columns. It honours NO_COLOR
and prints plain text when stdout is not a terminal.
Where each value comes from
| Row | Source |
|---|---|
| name | hh run --name, or hh rename afterwards. |
| model | The model field the agent's own transcript reports on each response. If the model changes mid-session, this row shows the last one and the models row lists every model with its step range. |
| provider, context limit | A table compiled into hh. Halfhand makes no network calls, so nothing is looked up at run time. |
| context used | The latest response's prompt size as the transcript reported it (input + cache read + cache write), and its share of the context limit. |
| tokens | The transcript's own usage objects, summed over the session as events are recorded. |
| total | Derived when you read it — never stored, so it cannot drift from its parts. Input + output + cache read + cache write. Reasoning tokens are a breakdown of output and are not added again. |
| estimated | Recorded payload bytes ÷ 4. See below. |
| messages | Counts of user_message, agent_message and tool_call events. |
| created, last activity | Session start, and the timestamp of its last recorded event. |
| duration | Wall is start to end. Active is the first to the last agent event (any event except terminal output and lifecycle markers), so idle time before the first prompt and after the last answer is excluded. |
| est. cost | Every response priced with the model that wrote it, at the table's list price. |
| capture | transcript when an adapter tailed the agent's transcript, terminal for PTY and file-system capture only, mcp for MCP proxy traffic, effects for file changes only (hh watch); then the adapter that ran. |
Two token families, never blended
Halfhand reports two kinds of token number. They answer different questions and are never added together or put in the same field.
- Self-reported counts are what the agent's transcript says it used.
They are exact for that agent's accounting. Only sessions recorded with a
structured adapter have them (today: Claude Code and Claude Desktop).
In JSON they live under
usage, which carries"source": "self_reported". - Estimated counts come from the size of what Halfhand recorded (bytes
÷ 4). They are rough, they exist for every session, and they are always
labelled as an estimate. In JSON they live under
estimated, with"source": "size_based"and"method": "bytes_div_4".
A session with no self-reported usage shows dashes in the token rows and still shows its estimate. Halfhand never substitutes one family for the other.
The dash convention
A value Halfhand does not know prints as — in the terminal and null in
JSON. It never prints as 0. Zero asserts a known value; a dash admits
ignorance. In particular:
- A model that is not in the compiled-in table shows its id with "(not in the model table)", and provider, context limit, usage percentage and cost are dashes.
- A count the transcript never reported (Claude does not report reasoning tokens separately) is a dash, not 0.
- If any response in a session came from a model the table does not know, the cost is a dash. A partial sum would understate the cost, so the total is unknown rather than wrong.
- A session recorded before Halfhand tracked usage (migration 0007) keeps its model, but its token counts are dashes. The usage it stored back then covers only the last response, not the whole session.
The model table
The table lists the Claude models whose provider, context window and list
prices are known when hh is built. Prices are first-party API list prices
per token, and cache writes use the 5-minute rate. They are not your invoice:
batch, partner-platform and negotiated pricing differ. A model the table
lacks is not an error. It gets dashes until a future hh release adds it.
JSON
hh inspect --context --json prints one object:
| Field | Type | Description |
|---|---|---|
schema |
integer | 2. |
session |
object | { id, short_id, adapter_status, adapter_degrade_reason }, as on event objects. |
name |
string | null | The session name. |
model |
object | { id, display, provider, context_limit }. Each is null when unknown. |
context_used |
object | { source: "self_reported_latest_turn", tokens, percent_of_limit }. percent_of_limit is a number with one decimal place, or null. |
usage |
object | As on the session object: self-reported counts and the derived total_tokens. |
estimated |
object | { source: "size_based", method: "bytes_div_4", tokens }. |
messages |
object | { user, agent, tool_calls }. |
capture_tier, adapter_id |
string | null | As on the session object. |
created_at, created |
integer, string | Session start: unix ms, and RFC 3339 UTC. |
last_activity_at, last_activity |
integer | null, string | null | Last recorded event: unix ms, and RFC 3339 UTC. |
duration_ms, duration |
integer | null, string | null | Wall-clock duration: ms, and ISO 8601 (PT1H49M40S). |
active_duration_ms, active_duration |
integer | null, string | null | First to last agent event: ms, and ISO 8601. |
estimated_cost_usd |
string | null | Decimal US dollars with six places ("0.102000"). A string, so no precision is lost or invented. null when unknown. |
cost_basis |
string | "list_price". |
models |
array | [{ model_id, first_step, last_step }], in order. |
The same name, model, usage and duration fields are on every hh list --json
session object. See the JSON reference.