run-agent.sh: prompt passed as a single argv exceeds MAX_ARG_STRLEN on long threads #123

New Issue

2026-08-30T12:57:05+02:00

ffaerber commented

2026-08-30 12:57:05 +02:00

Symptom

Every agent invocation on a long issue thread fails immediately, on every model, with:

opencode attempt 1/3 for @junior (ollama-cloud/kimi-k2.7-code:cloud, timeout 1200s)
rc=126
--- events (0 lines) ---
--- stderr (trace) ---
/tmp/agents-scripts/run-agent.sh: line 202: /usr/bin/timeout: Argument list too long

All three attempts fail identically after failing over to a second provider, which rules out the model, the provider and the timeout. rc=126 with zero event lines means the process was never executed.

Seen on ffaerber/homelab run 1919 (issue #226).

Cause

run-agent.sh:201:

timeout -k 30 "$AGENT_TIMEOUT" \
  opencode run --model "$MODEL" --auto --format json "$PROMPT" ${FILES:-}

"$PROMPT" is passed as a single argv element. Linux caps an individual argument at MAX_ARG_STRLEN = 32 pages = 131,072 bytes, independent of the much larger total ARG_MAX. Exceed it and execve fails with E2BIG; bash attributes the error to the binary it was about to run, which is why it reads as a timeout problem rather than a prompt problem.

Measurement on the thread that broke


newest 100 comments (what the script keeps)	127,716 B — 124.7 KiB
`MAX_ARG_STRLEN`	131,072 B — 128.0 KiB
headroom before the prompt template, issue body and instructions are added	~3.3 KiB

The thread had been sitting just under the cliff for a while. Adding one ordinary 3.3 KiB comment tipped it over, and it would have failed on the next comment from anyone.

Why the current guard does not help

thread comments fetched: 100 (newest 100 kept) caps by count. The limit being hit is in bytes. One hundred short comments is fine; one hundred long ones is not — and this thread's largest single comments are 24 KB and 19 KB, so the count cap can pass while the byte total is triple the limit.

Suggested fix

Stop putting the prompt on the command line:

printf '%s' "$PROMPT" > /tmp/agent_prompt.txt
timeout -k 30 "$AGENT_TIMEOUT" \
  opencode run --model "$MODEL" --auto --format json --prompt-file /tmp/agent_prompt.txt ${FILES:-}

or pipe it on stdin if opencode run accepts -. Either removes the ceiling entirely.

Additionally, budget the thread in bytes rather than comments — e.g. keep newest comments until a 60 KiB budget is spent, then insert a truncation marker. That keeps the prompt bounded no matter how the thread grows, and leaves room for the template.

Impact

Any issue whose thread grows past ~120 KiB becomes permanently unworkable for every agent, and the failure gives no hint of the real cause. ffaerber/homelab#226 is in that state now.

## Symptom Every agent invocation on a long issue thread fails immediately, on every model, with: ``` opencode attempt 1/3 for @junior (ollama-cloud/kimi-k2.7-code:cloud, timeout 1200s) rc=126 --- events (0 lines) --- --- stderr (trace) --- /tmp/agents-scripts/run-agent.sh: line 202: /usr/bin/timeout: Argument list too long ``` All three attempts fail identically after failing over to a second provider, which rules out the model, the provider and the timeout. `rc=126` with zero event lines means the process was never executed. Seen on `ffaerber/homelab` run 1919 (issue #226). ## Cause `run-agent.sh:201`: ```bash timeout -k 30 "$AGENT_TIMEOUT" \ opencode run --model "$MODEL" --auto --format json "$PROMPT" ${FILES:-} ``` `"$PROMPT"` is passed as a **single argv element**. Linux caps an individual argument at `MAX_ARG_STRLEN` = 32 pages = **131,072 bytes**, independent of the much larger total `ARG_MAX`. Exceed it and `execve` fails with `E2BIG`; bash attributes the error to the binary it was about to run, which is why it reads as a `timeout` problem rather than a prompt problem. ## Measurement on the thread that broke | | | |---|---| | newest 100 comments (what the script keeps) | 127,716 B — **124.7 KiB** | | `MAX_ARG_STRLEN` | 131,072 B — 128.0 KiB | | headroom before the prompt template, issue body and instructions are added | ~3.3 KiB | The thread had been sitting just under the cliff for a while. Adding one ordinary 3.3 KiB comment tipped it over, and it would have failed on the next comment from anyone. ## Why the current guard does not help `thread comments fetched: 100 (newest 100 kept)` caps by **count**. The limit being hit is in **bytes**. One hundred short comments is fine; one hundred long ones is not — and this thread's largest single comments are 24 KB and 19 KB, so the count cap can pass while the byte total is triple the limit. ## Suggested fix Stop putting the prompt on the command line: ```bash printf '%s' "$PROMPT" > /tmp/agent_prompt.txt timeout -k 30 "$AGENT_TIMEOUT" \ opencode run --model "$MODEL" --auto --format json --prompt-file /tmp/agent_prompt.txt ${FILES:-} ``` or pipe it on stdin if `opencode run` accepts `-`. Either removes the ceiling entirely. Additionally, budget the thread in **bytes** rather than comments — e.g. keep newest comments until a 60 KiB budget is spent, then insert a truncation marker. That keeps the prompt bounded no matter how the thread grows, and leaves room for the template. ## Impact Any issue whose thread grows past ~120 KiB becomes permanently unworkable for every agent, and the failure gives no hint of the real cause. `ffaerber/homelab#226` is in that state now.

🚀 1

junior commented

2026-08-30 12:57:35 +02:00

Assessment

Confirmed the failure point: .gitea/workflows/scripts/run-agent.sh:203 passes "$PROMPT" as a single argv element, and the issue thread can exceed Linux's per-argument ceiling MAX_ARG_STRLEN (~128 KiB).

opencode run --help shows no --prompt-file option; the positional message argument is an array, so the simple fix is to pass the prompt on stdin instead of the command line (or break it into multiple positional args, but stdin is cleaner).

opencode run with no positional message should read from stdin, as is conventional for CLI message tools. If it does not, the fallback is to split the prompt into several positional chunks — but that would still risk MAX_ARG_STRLEN on any single chunk > 128 KiB.

Recommended approach

Two related changes in run-agent.sh:

Immediate fix — avoid passing $PROMPT as an argv element:
```
printf '%s' "$PROMPT" | timeout -k 30 "$AGENT_TIMEOUT" \
  opencode run --model "$MODEL" --auto --format json ${FILES:-} \
  >/tmp/events.jsonl 2>/tmp/agent_err.log || rc=$?
```
This removes the per-argument ceiling entirely. Need to verify opencode run actually reads stdin when no message positionals are given; if it does not, we can pipe via a here-document or use a temp file plus multiple positional fragments.
Byte budget on the thread — add a helper before /tmp/thread.md is consumed that trims the kept comments to ~60 KiB (or some safe budget), inserting a … (older comments truncated) marker. This bounds the prompt even when the thread grows.

Effort estimate

Small: one shell script, ~10–20 lines changed.
Risk: low, but needs verification that opencode run accepts stdin without the positional message.

Who should build

This is a single-file shell change with a config/guard tweak — within the @junior lane. I can implement it once @pm gives the go-ahead. If opencode run stdin behavior turns out to be non-standard, I may need a quick check from @senior or @lead on the safest fallback.

No images or complex code involved, so this is text-only and safe for me.

details

🔧 9 tool calls · in 63585 · out 1333 tokens · subscription · model ollama-cloud/kimi-k2.7-code:cloud

🔧 read: {"filePath":"/workspace/gitea/agents/run-agent.sh","limit":250}
🔧 grep: --prompt-file|prompt-file
🔧 glob: 🔧 **read**:
🔧 glob: 🔧 **glob**:
🔧 grep: opencode run
🔧 read: .gitea/workflows/scripts/run-agent.sh
🔧 bash: opencode run --help 2>&1 || true

_{tokens — input 63585 (fresh 63585 · cache 0w / 0r) · output 1333 · reasoning 0 · subscription}

## Assessment Confirmed the failure point: `.gitea/workflows/scripts/run-agent.sh:203` passes `"$PROMPT"` as a single `argv` element, and the issue thread can exceed Linux's per-argument ceiling `MAX_ARG_STRLEN` (~128 KiB). `opencode run --help` shows no `--prompt-file` option; the positional `message` argument is an array, so the simple fix is to **pass the prompt on stdin instead of the command line** (or break it into multiple positional args, but stdin is cleaner). `opencode run` with no positional message should read from stdin, as is conventional for CLI message tools. If it does not, the fallback is to split the prompt into several positional chunks — but that would still risk `MAX_ARG_STRLEN` on any single chunk > 128 KiB. ## Recommended approach Two related changes in `run-agent.sh`: 1. **Immediate fix** — avoid passing `$PROMPT` as an argv element: ```bash printf '%s' "$PROMPT" | timeout -k 30 "$AGENT_TIMEOUT" \ opencode run --model "$MODEL" --auto --format json ${FILES:-} \ >/tmp/events.jsonl 2>/tmp/agent_err.log || rc=$? ``` This removes the per-argument ceiling entirely. Need to verify `opencode run` actually reads stdin when no message positionals are given; if it does not, we can pipe via a here-document or use a temp file plus multiple positional fragments. 2. **Byte budget on the thread** — add a helper before `/tmp/thread.md` is consumed that trims the kept comments to ~60 KiB (or some safe budget), inserting a `… (older comments truncated)` marker. This bounds the prompt even when the thread grows. ## Effort estimate - Small: one shell script, ~10–20 lines changed. - Risk: low, but needs verification that `opencode run` accepts stdin without the positional message. ## Who should build This is a single-file shell change with a config/guard tweak — within the `@junior` lane. I can implement it once @pm gives the go-ahead. If `opencode run` stdin behavior turns out to be non-standard, I may need a quick check from @senior or @lead on the safest fallback. No images or complex code involved, so this is text-only and safe for me. <details> <summary>details</summary> 🔧 9 tool calls · in 63585 · out 1333 tokens · subscription · model ollama-cloud/kimi-k2.7-code:cloud 🔧 **read**: `{"filePath":"/workspace/gitea/agents/run-agent.sh","limit":250}` 🔧 **grep**: `--prompt-file|prompt-file` 🔧 **glob**: `` 🔧 **read**: `` 🔧 **glob**: `` 🔧 **glob**: `` 🔧 **grep**: `opencode run` 🔧 **read**: `.gitea/workflows/scripts/run-agent.sh` 🔧 **bash**: `opencode run --help 2>&1 || true` <sub>tokens — input 63585 (fresh 63585 · cache 0w / 0r) · output 1333 · reasoning 0 · **subscription**</sub> </details>

Sign in to join this conversation.

2 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: gitea/agents#123