fix: make tool progress readable and preserve project context #988

Closed
opened 2026-10-27 13:38:34 +00:00 by sam · 0 comments
Owner

Problem

Telegram tool progress is currently displayed as a stream of raw inline entries, for example:

  • tool names followed by JSON schema placeholders such as [string]
  • raw JSON arguments and oversized output
  • repeated output without readable grouping
  • full error dumps and turn-budget failures

Agents also search broad locations such as /home/sam instead of staying focused on the active project/worktree, read directories as files, and emit malformed shell commands. Useful project context is not being retained cleanly across the turn.

Scope

  • Render tool progress as compact, readable streaming fenced code blocks (Telegram-safe).
  • Remove schema placeholders and redundant raw invocation JSON from user-visible output.
  • Truncate or offload oversized results without consuming the entire turn context.
  • Preserve concise success/failure status and actionable errors.
  • Default agent discovery and file operations to the actual project/worktree context where task context provides one.
  • Improve context retention across tool calls.
  • Add validation or guardrails for malformed shell commands where practical.

Acceptance criteria

  • A normal tool call is shown as a readable streaming status block, not raw JSON.
  • Tool failures remain visible but are concise and actionable.
  • Large results are bounded/offloaded and do not exhaust the turn budget.
  • Project-scoped tasks do not begin by searching unrelated home-directory data.
  • Regression tests cover rendering, failure output, truncation/offload, and project-root selection.
  • Telegram output is verified against a live deployed instance.

Imported from beads archie-core-1787522397855-2-792cc4d6 · GitHub #604 · type task · priority P2 · created 2026-08-23 · closed 2026-08-23

## Problem Telegram tool progress is currently displayed as a stream of raw inline entries, for example: - tool names followed by JSON schema placeholders such as `[string]` - raw JSON arguments and oversized output - repeated output without readable grouping - full error dumps and turn-budget failures Agents also search broad locations such as `/home/sam` instead of staying focused on the active project/worktree, read directories as files, and emit malformed shell commands. Useful project context is not being retained cleanly across the turn. ## Scope - Render tool progress as compact, readable streaming fenced code blocks (Telegram-safe). - Remove schema placeholders and redundant raw invocation JSON from user-visible output. - Truncate or offload oversized results without consuming the entire turn context. - Preserve concise success/failure status and actionable errors. - Default agent discovery and file operations to the actual project/worktree context where task context provides one. - Improve context retention across tool calls. - Add validation or guardrails for malformed shell commands where practical. ## Acceptance criteria - A normal tool call is shown as a readable streaming status block, not raw JSON. - Tool failures remain visible but are concise and actionable. - Large results are bounded/offloaded and do not exhaust the turn budget. - Project-scoped tasks do not begin by searching unrelated home-directory data. - Regression tests cover rendering, failure output, truncation/offload, and project-root selection. - Telegram output is verified against a live deployed instance. --- _Imported from beads `archie-core-1787522397855-2-792cc4d6` · GitHub #604 · type task · priority P2 · created 2026-08-23 · closed 2026-08-23_
sam added the mediumin-progresstask labels 2026-10-27 13:38:34 +00:00
sam self-assigned this 2026-10-27 13:38:35 +00:00
sam closed this issue 2026-10-27 13:38:35 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: sam/archie#988