worker: prompt-too-long hint is code-flow-specific; misleading for design agents #95

New issue

Closed

opened 2026-04-19 17:42:16 +00:00 by claude-desktop · 0 comments

claude-desktop commented

2026-04-19 17:42:16 +00:00

Collaborator

User story

As the operator debugging a failed run, I want the prompt-too-long hint in the service log to point at the real cause for the agent that failed, not a one-size-fits-all rule. Right now it sends designers down a rabbit-hole that doesn't apply to them.

What happens today

src/worker.ts:217–228 catches "prompt is too long" / "context_length_exceeded" and emits a fixed hint:

prompt-too-long failure on <repo>#<issue> — skills must use Edit/Write + git CLI
for code changes, not mcp__forgejo__update_file / create_file / delete_file (see #79)

That hint is code-flow-specific. It's the right rule for dev / boss / reviewer (where overflow was first observed: dev feeding an entire 400-line file as mcp__forgejo__update_file content — #79).

When designer overflows — as observed today on task d3883982 (#70) — the root cause is Penpot MCP accumulation (100+ create_frame / create_text round-trips), not a misuse of any forgejo file tool. The hint points at an irrelevant skill rule and covers the actual cause (raise context window / compact history).

Acceptance criteria

Code

src/worker.ts splits the hint by agent role:
- Code-flow agents (dev, boss, reviewer): current wording; refs #79.
- Design agents (designer, design-reviewer): new wording pointing at #94 ("designer needs 1M context") and the Penpot-accumulation cause — use claude-opus-4-7[1m] and/or split the task across multiple dispatches.
The stored result.error field (shown in the dashboard) mirrors the log hint, so the operator sees the right message in the task detail view too.
Uses the same agent classification mechanism already used in src/webhook-routing.ts (the REVIEWER_AGENTS / DESIGN_AGENTS sets, or equivalent).

Tests

worker.test.ts: a design-agent config + a prompt-too-long error yields the Penpot-flavored hint, not the mcp__forgejo__*_file one.
Code-agent + prompt-too-long yields today's hint (regression guard).
Unknown agent name falls through to a generic hint ("raise context window or split the task").

Out of scope

Fixing the underlying overflow (tracked in #79 for code agents, #94 for designer).
Rewriting #79's skill-level rule — still correct for code-flow.
A structured-error schema. The hint is a string today; keep it a string.

References

Failure that surfaced this: designer task d3883982-7b23-4083-ac18-38b42b98e633 on #70.
src/worker.ts:217 — the isPromptTooLong branch.
#79 — rule for code-flow agents (mcp__forgejo__*_file).
#94 — designer needs 1M context window.
src/webhook-routing.ts — existing agent classification (REVIEWER_AGENTS, design-reviewer-author routing).

Dependencies

Blocked by: nothing.
Blocks: operators interpreting designer overflow failures correctly.
Branch off: main.

## User story As the **operator debugging a failed run**, I want the `prompt-too-long` hint in the service log to point at the real cause for the agent that failed, not a one-size-fits-all rule. Right now it sends designers down a rabbit-hole that doesn't apply to them. ## What happens today `src/worker.ts:217–228` catches `"prompt is too long"` / `"context_length_exceeded"` and emits a fixed hint: ``` prompt-too-long failure on <repo>#<issue> — skills must use Edit/Write + git CLI for code changes, not mcp__forgejo__update_file / create_file / delete_file (see #79) ``` That hint is **code-flow-specific**. It's the right rule for `dev` / `boss` / `reviewer` (where overflow was first observed: dev feeding an entire 400-line file as `mcp__forgejo__update_file` content — #79). When `designer` overflows — as observed today on task `d3883982` (#70) — the root cause is Penpot MCP accumulation (100+ `create_frame` / `create_text` round-trips), not a misuse of any forgejo file tool. The hint points at an irrelevant skill rule and covers the actual cause (raise context window / compact history). ## Acceptance criteria ### Code - [ ] `src/worker.ts` splits the hint by agent role: - Code-flow agents (`dev`, `boss`, `reviewer`): current wording; refs #79. - Design agents (`designer`, `design-reviewer`): new wording pointing at #94 ("designer needs 1M context") and the Penpot-accumulation cause — use `claude-opus-4-7[1m]` and/or split the task across multiple dispatches. - [ ] The stored `result.error` field (shown in the dashboard) mirrors the log hint, so the operator sees the right message in the task detail view too. - [ ] Uses the same agent classification mechanism already used in `src/webhook-routing.ts` (the `REVIEWER_AGENTS` / `DESIGN_AGENTS` sets, or equivalent). ### Tests - [ ] `worker.test.ts`: a design-agent config + a prompt-too-long error yields the Penpot-flavored hint, not the `mcp__forgejo__*_file` one. - [ ] Code-agent + prompt-too-long yields today's hint (regression guard). - [ ] Unknown agent name falls through to a generic hint ("raise context window or split the task"). ## Out of scope - Fixing the *underlying* overflow (tracked in #79 for code agents, #94 for designer). - Rewriting #79's skill-level rule — still correct for code-flow. - A structured-error schema. The hint is a string today; keep it a string. ## References - Failure that surfaced this: designer task `d3883982-7b23-4083-ac18-38b42b98e633` on #70. - `src/worker.ts:217` — the `isPromptTooLong` branch. - #79 — rule for code-flow agents (`mcp__forgejo__*_file`). - #94 — designer needs 1M context window. - `src/webhook-routing.ts` — existing agent classification (`REVIEWER_AGENTS`, design-reviewer-author routing). ## Dependencies - **Blocked by:** nothing. - **Blocks:** operators interpreting designer overflow failures correctly. - **Branch off:** `main`.

claude-desktop added the

area:agents

type:user-story

labels

2026-04-19 17:42:21 +00:00