Skip to content

fix(session): propagate error when subagent produces no text output - #22

Open
jfenal wants to merge 8 commits into
bobbyjohnstx:mainfrom
jfenal:fix/subagent-empty
Open

fix(session): propagate error when subagent produces no text output#22
jfenal wants to merge 8 commits into
bobbyjohnstx:mainfrom
jfenal:fix/subagent-empty

Conversation

@jfenal

@jfenal jfenal commented Oct 9, 2026 •
edited
Loading

Copy link
Copy Markdown

Summary

  • fix(session): Subagent returns empty string when model doesn't support tools (e.g. HTTP 400 from Ollama). Propagates the error so the parent model sees a meaningful failure instead of a blank tool result. Fixes fix(session): subagent returns silent empty string when no text output produced #21
  • fix(tui): Split OSC rgb color response tail (`b:0000/0000/0000`, `gb/0000:0000:0000`, `:0000/0000/0000`) leaks into the textarea. Added `oscRGBTail` regex matching the full split-suffix family with exactly 4 hex digits per channel (avoids false-positives on user input like `/exit`). Fixes fix(tui): split OSC color response leaks into textarea as "b:0000/0000/0000" #24
  • fix(provider): `shouldPoll()` ignored its `providerID` argument entirely, so `disabled_providers` config had no effect at the poll level — disabled providers still made HTTP connection attempts every 30s. Wired to `registry.isFiltered()`.

Test plan

  • `go test ./...` passes
  • Restart with a model that doesn't support tools → subagent error is now descriptive
  • Add `"disabled_providers": ["lm-studio"]` to config → no more polling in logs
  • Type `/exit` in prompt → autocomplete still appears (no false-positive filtering)

🤖 Generated with Claude Code

jfenal and others added 3 commits October 8, 2026 23:04
@jfenal @claude
chore: lower minimum Go version to 1.25.0
9902701 
The go directive was set to 1.27.1, which exceeds what Fedora 44
ships (go1.26.8). Lower it to the actual minimum the dependency graph
requires (1.25.0, driven by a transitive dep found by go mod tidy),
so the project builds out of the box with the system Go on Fedora 44
without any manual toolchain installation.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal @claude
docs: add LLM panel design spec
8bb39b5 
Spec for a toggleable right-side TUI panel showing the live LLM
stream transcript (text deltas, tool calls, reasoning, step
boundaries). No server-side changes required — reuses existing
SSE event flow.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal @claude
fix(session): propagate error when subagent produces no text output
7d1e534 
When a subagent uses a model that doesn't support tools (e.g. HTTP 400
from Ollama), the iteration limit fires with result.Error set but
result.Messages non-empty (the user prompt is stored). The empty-check
at len(result.Messages)==0 was skipped, returning ("", nil) silently.

The parent model then saw an empty tool result and generated its own
confusing "The subagent came back empty" message.

Fix: after extracting assistant text, check response == "". When empty
with an error, propagate it so the task tool returns IsError=true with
the real cause. When empty without an error, return a diagnostic string
so the parent model can reason accurately.

Fixes bobbyjohnstx#21

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal
jfenal requested a review from bobbyjohnstx as a code owner October 9, 2026 16:53
jfenal and others added 5 commits October 9, 2026 19:03
@jfenal @claude
fix(tui): filter split OSC rgb: color response tail from textarea
cf45c26 
Terminal color query responses arrive as e.g. "\033]10;rgb:0000/0000/0000\007".
When split across OS reads, the first fragment ("\033]10;rg") is caught by
the escape-byte check but the tail ("b:0000/0000/0000") slips through: it
contains no escape byte, no "rgb:" substring, and fails the hex-only
oscHexFragment regex.

Add oscRGBTail regex that matches the [rgb]{1,3}: prefix pattern so any
split suffix of an OSC color response is caught by isTerminalEscape.

Fixes bobbyjohnstx#24

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal @claude
fix(tui): broaden OSC rgb tail filter to catch / separator variant
e3c270a 
Some terminals use "rgb/R:G:B" instead of "rgb:R/G/B" — e.g. producing
"gb/0000:0000:0000" as a split-read tail. Widen oscRGBTail regex to
accept both : and / as the separator between the prefix and channels.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal @claude
fix(tui): catch OSC rgb tail with fully-consumed rgb prefix (:0000/00…
8cb5776 
…00/0000)

Terminal splits response further — "rgb" fully consumed, leaving only
":0000/0000/0000". Allow zero prefix chars in oscRGBTail regex.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal @claude
fix(tui): require exactly 4 hex digits per channel to avoid /exit fal…
4c5203b 
…se-positive

{1,4} hex matched "/e" in "/exit", filtering slash commands. Terminal
color responses always use full 16-bit channels (4 hex digits), so {4}
is both correct and safe.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal @claude
fix(provider): shouldPoll now respects disabled_providers filter
8483481 
shouldPoll() ignored providerID entirely (_ = providerID; return true),
so disabled providers were still polled every 30s even when filtered out
via disabled_providers config. Wire it to registry.isFiltered() so
disabled providers are skipped at the poll level, not just at registration.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
@jfenal jfenal mentioned this pull request Oct 9, 2026
Open
4 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@bobbyjohnstx bobbyjohnstx Awaiting requested review from bobbyjohnstx bobbyjohnstx is a code owner

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

1 participant

@jfenal