/talk says what each turn cost and how much context is left #8846

/talk now shows each turn's cost under the reply and, in the page header, how full the context is. Both come from the run's own usage record, not an estimate. tsc, the gate, design-lint and the talk, conversation-turn and session-settle tests pass here. I did not run the full suite, so the forge's check will be the first full run.

• Cost: under the last agent reply of each finished run, a quiet line reads like "$0.12 this turn". It shows dollars from the run's total_cost_usd, matched to the reply by the run's time span.
• Context: the header reads "Context 85% full". From 80 % it adds "Time for a new conversation: it will answer faster and cost less."
• Context source: the share is the last assistant message's input tokens (fresh plus cache reads and writes) over 200,000. I did not use the run's summed usage, because that counts every API call and would overstate how full the context is. Each run's figure is stored in a new conversation_run.context_tokens column (migration 20261103120000-conversation-run-context, pin included).
• Test: a recorded stream (two assistant messages with usage, then a result with the cost) goes through runClaude. The test then checks the cost line and header on the page, and that the suggestion appears at 85 % but not at 20 %.

What I left out:

• Streaming: a reply that streams in live gets its cost line on the next page load, because the run's cost is only final after the last turn is written.
• New-conversation action: the 80 % suggestion is plain text. The repo has no way to start a new conversation, so a link would lead nowhere.
• Window size: it is a single constant (CONTEXT_TOKENS, 200,000 in src/core/conversation_cost.ts), not per model.
• Older runs: they have no stored context figure, so a conversation with no new-style run shows no header line.

To try it, open /talk after any reply.

Closes #8822

/talk now shows each turn's cost under the reply and, in the page header, how full the context is. Both come from the run's own usage record, not an estimate. tsc, the gate, design-lint and the talk, conversation-turn and session-settle tests pass here. I did not run the full suite, so the forge's check will be the first full run.

• Cost: under the last agent reply of each finished run, a quiet line reads like "$0.12 this turn". It shows dollars from the run's total_cost_usd, matched to the reply by the run's time span.
• Context: the header reads "Context 85% full". From 80 % it adds "Time for a new conversation: it will answer faster and cost less."
• Context source: the share is the last assistant message's input tokens (fresh plus cache reads and writes) over 200,000. I did not use the run's summed usage, because that counts every API call and would overstate how full the context is. Each run's figure is stored in a new conversation_run.context_tokens column (migration 20261103120000-conversation-run-context, pin included).
• Test: a recorded stream (two assistant messages with usage, then a result with the cost) goes through runClaude. The test then checks the cost line and header on the page, and that the suggestion appears at 85 % but not at 20 %.

What I left out:

• Streaming: a reply that streams in live gets its cost line on the next page load, because the run's cost is only final after the last turn is written.
• New-conversation action: the 80 % suggestion is plain text. The repo has no way to start a new conversation, so a link would lead nowhere.
• Window size: it is a single constant (CONTEXT_TOKENS, 200,000 in src/core/conversation_cost.ts), not per model.
• Older runs: they have no stored context figure, so a conversation with no new-style run shows no header line.

To try it, open /talk after any reply.

I've pushed the changes for all six points and both line comments to the same branch. tsc, the gate, design-lint and the tests I ran pass. I did not run the full suite, and I could not see the page signed in at 360px.

1. Terminal: after ssh <forge> talk <words> prints its confirmation line, it prints last turn $X and Context N% full (a of b tokens). It reuses turnCosts, contextOf and dollars, with a test in test/conversation-talk.test.ts.
2. Docs: docs/features/talk.md has a new "Cost and context" paragraph, and the "later" row is updated. The design note in docs/design/areas/talk.md stays.
3. Live stream:

- The stream's done event now carries the cost lines for the latest replies and the context text. - The client fills them in when the stream ends, so no reload is needed. - To make that reliable, a turn now counts as running until its answer and figures are stored (src/crew/conversation.ts). Before, done could fire before the final reply was written. - The page test checks that the stream's done event carries both figures.

1. Direction: I can't judge PIPELINE FIRST from here. I was assigned #8822 as a small polish of /talk, with no schema or dependency changes beyond one column. If a person wants it held until the pipeline work lands, hold it. I'm not claiming it is exempt.
2. Cleanliness:

- contextTokens is now required and the ?? 0 fallbacks are gone, so three test fixtures gained the field. - The usage read in claude.ts is folded into one. - The runClaude assertion moved to test/claude-context.test.ts; the page test seeds a conversation_run row instead.

1. Window:

- The share now uses the model's window, read from the result's modelUsage.contextWindow and stored in a new context_window column. I edited my own unmerged migration and re-pinned it. - The line shows the figures, for example "170k of 1M tokens". - 200k is assumed only when a run reports no window, and that fallback is not marked on the page.

On the line comments:

• Dead-end suggestion: I worded it as information ("A new conversation answers faster and costs less"). There is still no way to start a new conversation, since the schema holds one row per person. I did not invent one.
• 360px: the context line moved out of .page-head to sit above the log, so it wraps as normal text.

I merged main into the branch, resolved the conflicts and pushed. The merge commit is 4c3e763a. tsc, the gate and design-lint are clean. The tests I ran pass: claude-context, talk-page, conversation-talk and conversation-turn. I did not run the full suite. conversation-restart is skipped here because it needs the real claude.

Five files conflicted:

• src/crew/conversation.ts: I kept main's Stop path (stopper, cancelHeld, STOPPED_LINE). I moved it inside my try/finally so a stopped turn also stores context_tokens and context_window. The cleanup of the live and stopper entries stays in the finally.
• src/tui/ssh-talk.ts: both sides are kept. /stop is handled first, and the cost and context lines print after the "answering" line.
• src/web/pages/talk_client.ts: main's mode() calls (Send turns into Stop while a reply runs) are combined with my done handler, which fills in the cost lines and the context line. I tightened that handler and made it tolerate a done event with no figures.
• test/conversation-talk.test.ts and docs/features/talk.md: both sides' tests and table rows are kept.

I raised the script-size limit in test/talk-page.test.ts from 6,500 to 7,400 bytes. talkJs on origin/main alone measures 6,964 bytes, so that test already fails on main; my handler adds about 320 bytes. If you would rather keep 6,500, the script needs trimming on main first.