fix(qwen): document the 0.25.1 floor and reserve Qwen's launch placeholders (#1143)

Qwen Code through 0.25.0 treats every ${name} in a SubAgent prompt as a value
to fill, and throws "Missing context values for the following keys" when it
has none. Sixteen agents show ${name} in their code examples, so those
releases can't launch them (#1081). 0.25.1 (2026-10-11) leaves any name it
doesn't fill as written, so the agents keep their examples. No escape for
${...} is safe across bash, JavaScript and the other 14 tools anyway.

- Document "Qwen Code 0.25.1 or later" in integrations/qwen/README.md and
  in the README's Qwen section.
- Qwen does fill ${task_prompt}, ${hook_context} and
  ${external_inputs_override}. lint-agents.sh now rejects those three in
  agent bodies, so no agent text is replaced at launch. No current agent uses
  them, and the whole roster still lints clean.
- test-lint-qwen-placeholders.sh covers the three reserved names and four
  look-alikes that must pass. It fails against the previous linter, and CI
  runs it whenever the linter changes.

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Michael Sitarzewski
2026-10-15 00:33:20 -05:00
committed by GitHub
co-authored by Claude Opus 5.5
parent bd956db032
commit 652b2692f8
5 changed files with 54 additions and 0 deletions
+4
View File
@@ -108,3 +108,7 @@ jobs:
- name: Reject folded frontmatter fixtures
if: steps.rules.outputs.changed == 'true'
run: bash scripts/test-folded-frontmatter.sh
- name: Reject Qwen launch placeholders
if: steps.rules.outputs.changed == 'true'
run: bash scripts/test-lint-qwen-placeholders.sh
+2
View File
@@ -1003,6 +1003,8 @@ See [integrations/openclaw/README.md](integrations/openclaw/README.md) for detai
SubAgents are installed to `.qwen/agents/` in your project root (project-scoped).
> Requires Qwen Code 0.25.1 or later. Earlier releases can't launch agents whose code examples contain `${…}` placeholders ([#1081](https://github.com/msitarzewski/agency-agents/issues/1081)).
```bash
# Convert and install (run from your project root)
cd /your/project
+13
View File
@@ -5,6 +5,19 @@ Qwen Code uses project-scoped `.md` SubAgent files in `.qwen/agents/`.
The generated files come from `scripts/convert.sh --tool qwen`, which writes one
SubAgent Markdown file per agency agent into `integrations/qwen/agents/`.
## Requirements
Qwen Code **0.25.1 or later**. Earlier releases treat every `${name}` in a
SubAgent's prompt as a value to fill and refuse to launch the agent when they
have none ("Missing context values for the following keys: …"). Several agents
show `${name}` in their code examples, so on those releases they can't be
launched. From 0.25.1, Qwen leaves any `${name}` it doesn't fill as written
([#1081](https://github.com/msitarzewski/agency-agents/issues/1081)).
Qwen does fill three names: `${task_prompt}`, `${hook_context}` and
`${external_inputs_override}`. `scripts/lint-agents.sh` rejects them in agent
bodies, so no agent text is replaced at launch.
## Generate
From the repository root:
+10
View File
@@ -169,6 +169,16 @@ lint_file() {
warnings=$((warnings + 1))
fi
# 4b. Qwen Code substitutes ${task_prompt}, ${hook_context} and
# ${external_inputs_override} into a SubAgent's prompt at launch. From 0.25.1
# it leaves every other ${name} as written (older releases refuse any name
# they can't fill, #1081). One of these three in a body would be silently
# replaced, so the names are reserved.
if grep -qE '\$\{(task_prompt|hook_context|external_inputs_override)\}' <<<"$body"; then
echo "ERROR $file: uses a placeholder Qwen Code fills at launch (\${task_prompt}, \${hook_context} or \${external_inputs_override}) — rename it"
errors=$((errors + 1))
fi
local soul_headers=0
local agents_headers=0
local fence_marker="" fence_len=0 fence_indent=0 fence_line=0
+25
View File
@@ -0,0 +1,25 @@
#!/usr/bin/env bash
# Qwen Code fills three ${name} placeholders into a SubAgent's prompt at launch
# (#1081). The linter must reject those three in an agent body and keep
# accepting every other ${name}, which Qwen 0.25.1+ leaves as written.
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
scratch="$(mktemp -d)"
trap 'rm -rf "$scratch"' EXIT
write_agent() {
printf '%s\n' '---' 'name: Example Agent' 'description: Checks placeholder handling' 'color: blue' '---' \
'## Identity' '## Core Mission' '## Critical Rules' '```bash' "echo \"$1\"" '```' > "$scratch/agent.md"
}
for key in task_prompt hook_context external_inputs_override; do
write_agent "\${$key}"
if bash "$SCRIPT_DIR/lint-agents.sh" "$scratch/agent.md" > "$scratch/result.log" 2>&1; then
echo "FAIL: linter accepted the Qwen launch placeholder \${$key}" >&2; exit 1
fi
grep -q 'placeholder Qwen Code fills at launch' "$scratch/result.log"
done
for text in '${name}' '${BACKUP_DIR}' '${task_prompt_suffix}' '$task_prompt'; do
write_agent "$text"
bash "$SCRIPT_DIR/lint-agents.sh" "$scratch/agent.md" > "$scratch/result.log" 2>&1 \
|| { cat "$scratch/result.log" >&2; echo "FAIL: linter rejected ordinary text $text" >&2; exit 1; }
done
echo 'PASS: Qwen launch placeholders are rejected while other ${name} placeholders pass'