feat(btw): let /btw see main-session tool names and outcomes (amend visible-text-only scope)

[Yeachan-Heo]

Migrated from #6035 (originally opened by @yazzang-homelab).

Description

During a tool-heavy main turn, /btw receives almost nothing about what the main session is doing. AgentSession.#projectBtwVisibleText (packages/coding-agent/src/session/agent-session.ts, dev d23a4f955) keeps only text blocks of user/assistant messages, so assistant turns that consist only of tool calls disappear entirely. In an observed session the side chat saw only the original request, two resume inputs, and the side question, and correctly answered that no tool results, logs, or file contents were visible.

This is the current, intended contract, not a defect: the #2825 privacy dispositions (#2825 (comment), item 4) state that the frozen scope "projects only visible user/assistant text; custom/developer/tool/image/thinking/provider payload state is dropped", following the owner review on #2516 that asked for "an explicit, tested content-flow boundary". Changing what the side chat sees is therefore a product decision, which is why this is an issue and not a PR.

Use Case

The most common /btw question while the main agent is working is "what is it doing / why did that happen?". With tool-only turns dropped, that question gets the least context exactly when it is asked most.

Area

Session management

Proposed Solution

A worked implementation exists for reference only (not opened as a PR): https://github.com/yazzang-homelab/gajae-code/tree/feat/btw-tool-activity (head 3faeca9cb, 2 commits on dev 5c6197ecd, 6 files, +282/−4).

• Each assistant turn that called tools gets one [main tool activity] read ok, bash error x2 line: sanitized tool name ([A-Za-z0-9_.:-], ≤64 chars) plus ok / error (from toolResult.isError) / pending (only the newest turn with stopReason: "toolUse") / unknown (no result recorded, e.g. provider-executed or orphaned calls).
• Results are paired per assistant turn by toolCallId; arguments, intent, result content, details, and images are not inputs to the projection.
• Adjacent same-role turns in the projected scope are merged, because projecting tool-only turns otherwise produces consecutive assistant messages that strict-alternation providers (Bedrock Converse) reject.
• Visible text lines starting with the marker are rewritten to (quoted main tool activity) so user/assistant text cannot forge activity lines.
• btw-user.md states that the scope is captured at panel open and that ok is the tool's error flag, not task success.

Tests: sentinel matrix in test/agent-session-btw.test.ts (arguments, intent, command, result, details, error text, in-flight command all absent from the provider-bound payload; role sequence; per-turn id reuse; unknown vs pending; marker forgery) and formatter/prompt-parity cases in test/btw-contract.test.ts. The three new session tests fail on dev and pass on the branch; the 32 btw-related test files pass (645/0) and bun --cwd=packages/coding-agent run check passes. No live Bedrock/Anthropic request was made.

Known limits of that sketch:

• Tool names are still a free-form channel set by MCP servers / ACP agents / the model; a registered-tool allowlist was not added because the frozen scope has no registry snapshot.
• The scope is frozen at panel open, so follow-ups do not see later tool activity (unchanged behavior).
• Control inputs such as resume still enter the scope as ordinary user text; that is a separate question and not addressed.
• scope.messages has no byte bound (unchanged); the sketch adds ≤74 B per call plus ≤22 B per turn.

Alternatives Considered

Option	What changes	Cost
Keep the visible-text-only contract	Nothing	Mid-task /btw keeps seeing only user messages in tool-heavy sessions
Tool names + outcomes (sketch above)	Amends dispositions item 4 to admit tool name and error flag	New metadata channel (tool names); needs owner sign-off on the boundary
Names + outcomes with a registered-tool allowlist	Unregistered names collapse to a fixed token	Requires snapshotting the tool registry into the frozen scope; larger diff
Outcome counts only (e.g. "3 tool calls, 1 error")	No names cross the boundary	Much less useful for "what is it doing?"
Let the main agent summarize before /btw	Documentation only	Relies on the model writing visible text; no code guarantee

—
[repo owner's gaebal-gajae (clawdbot) 🦞]