Mitigate ARG_MAX/E2BIG by rewriting inline prompt expansion to stdin in AWF agent command by Copilot · Pull Request #1999 · github/gh-aw-firewall

Copilot · 2026-04-15T21:00:55Z

Large gh-aw workflows can inline 100KB+ prompts into argv via --prompt "$(cat ...)", which combines with container env size and can exceed Linux ARG_MAX, failing process start with E2BIG. This change adds an AWF-side mitigation by keeping prompt payload on disk and passing it via stdin instead of argv.

Command rewrite for inline prompt expansions
- Detects gh-aw style inline prompt patterns in agentCommand:
  - --prompt "$(cat /path/to/prompt.txt)"
  - --prompt $(cat /path/to/prompt.txt)
- Rewrites execution form to:
  - cat '/path/to/prompt.txt' | ... --prompt -
- Preserves normal literal prompts (no rewrite).
ARG_MAX diagnostics improvement
- Expands existing --env-all warning from env-only size to estimated combined argv + env size.
- Surfaces clearer risk signal before runtime execve failure.
Focused coverage
- Adds unit tests for:
  - quoted inline $(cat ...) rewrite
  - single-quoted prompt-file paths (including spaces)
  - unquoted inline $(cat ...) rewrite
  - non-inline literal --prompt no-op behavior

# before
node ... --prompt "$(cat /tmp/gh-aw/aw-prompts/prompt.txt)" --allow-all-tools

# after
cat '/tmp/gh-aw/aw-prompts/prompt.txt' | node ... --prompt - --allow-all-tools

Agent-Logs-Url: https://github.com/github/gh-aw-firewall/sessions/4052214b-06e7-4406-aeb0-e3958af8a77e

Copilot

Pull request overview

Mitigates Linux ARG_MAX / E2BIG failures caused by large inline prompt expansions (e.g. --prompt "$(cat ...)") by rewriting the agent command to feed the prompt via stdin, and improves preflight diagnostics by warning on estimated combined argv + env size.

Changes:

Adds a command rewriter in generateDockerCompose() to detect --prompt $(cat ...) / --prompt "$(cat ...)" and convert it to cat 'file' | ... --prompt -.
Enhances the existing --env-all warning to estimate combined argv + env bytes (instead of env-only).
Adds unit tests validating the rewrite behavior for quoted/unquoted forms and a no-op for literal prompts.

Show a summary per file

File	Description
`src/docker-manager.ts`	Introduces inline-prompt rewrite + enhanced `argv+env` sizing warning; switches compose command to use rewritten agent command.
`src/docker-manager.test.ts`	Adds test coverage for prompt rewrite variants and ensures literal prompts are not rewritten.

Copilot's findings

Tip

Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Files reviewed: 2/2 changed files
Comments generated: 2

Copilot · 2026-04-16T00:14:01Z

+  const rewrittenCommand = `${agentCommand.slice(0, inlinePrompt.start)}--prompt -${agentCommand.slice(inlinePrompt.end)}`;
+  if (rewrittenCommand === agentCommand) return agentCommand;
+
+  logger.warn(`Rewriting inline prompt expansion to stdin pipe to avoid ARG_MAX/E2BIG: ${promptPath}`);
+  return `cat ${shellQuote(promptPath)} | ${rewrittenCommand}`;


The rewrite currently prepends cat ... | to the entire agentCommand. This changes shell semantics when agentCommand is a compound command (e.g. contains &&, ;, an existing pipe, or redirects): stdin gets consumed by the left-most command and the pipeline exit status/precedence can differ from the original, potentially breaking execution.

Also, shellQuote(promptPath) forces single-quoting, which will prevent any $VAR expansion that may have worked in the original $(cat $VAR/...) or $(cat "$VAR") form.

Consider either (a) restricting the rewrite to “simple” commands (no shell operators / variable expansions in the prompt path) and otherwise leaving the command unchanged, or (b) rewriting in-place to use --prompt - with an input redirection that applies only to the copilot invocation, while preserving the original path quoting/expansion semantics.

Copilot · 2026-04-16T00:14:01Z

+    it('should rewrite inline --prompt $(cat ...) to stdin pipe to avoid ARG_MAX expansion', () => {
+      const configWithInlinePrompt = {
+        ...mockConfig,
+        agentCommand: 'node /tmp/gh-aw/actions/copilot_driver.cjs /usr/local/bin/copilot --prompt "$(cat /tmp/gh-aw/aw-prompts/prompt.txt)" --allow-all-tools',
+      };
+      const result = generateDockerCompose(configWithInlinePrompt, mockNetworkConfig);
+      const agent = result.services.agent;
+
+      expect(agent.command).toEqual([
+        '/bin/bash',
+        '-c',
+        'cat \'/tmp/gh-aw/aw-prompts/prompt.txt\' | node /tmp/gh-aw/actions/copilot_driver.cjs /usr/local/bin/copilot --prompt - --allow-all-tools',
+      ]);
+    });
+
+    it('should rewrite inline --prompt $(cat ...) when cat path is single-quoted', () => {
+      const configWithQuotedPath = {
+        ...mockConfig,
+        agentCommand: 'node /tmp/gh-aw/actions/copilot_driver.cjs /usr/local/bin/copilot --prompt "$(cat \'/tmp/gh-aw/aw-prompts/prompt with spaces.txt\')" --allow-all-tools',
+      };
+      const result = generateDockerCompose(configWithQuotedPath, mockNetworkConfig);
+      const agent = result.services.agent;
+
+      expect(agent.command).toEqual([
+        '/bin/bash',
+        '-c',
+        'cat \'/tmp/gh-aw/aw-prompts/prompt with spaces.txt\' | node /tmp/gh-aw/actions/copilot_driver.cjs /usr/local/bin/copilot --prompt - --allow-all-tools',
+      ]);
+    });
+
+    it('should rewrite unquoted --prompt $(cat ...) form', () => {
+      const configWithUnquotedInlinePrompt = {
+        ...mockConfig,
+        agentCommand: 'node /tmp/gh-aw/actions/copilot_driver.cjs /usr/local/bin/copilot --prompt $(cat /tmp/gh-aw/aw-prompts/prompt.txt) --allow-all-tools',
+      };
+      const result = generateDockerCompose(configWithUnquotedInlinePrompt, mockNetworkConfig);
+      const agent = result.services.agent;
+
+      expect(agent.command).toEqual([
+        '/bin/bash',
+        '-c',
+        'cat \'/tmp/gh-aw/aw-prompts/prompt.txt\' | node /tmp/gh-aw/actions/copilot_driver.cjs /usr/local/bin/copilot --prompt - --allow-all-tools',
+      ]);
+    });
+
+    it('should not rewrite regular literal --prompt values', () => {
+      const configWithLiteralPrompt = {
+        ...mockConfig,
+        agentCommand: 'copilot --prompt "hello world"',
+      };
+      const result = generateDockerCompose(configWithLiteralPrompt, mockNetworkConfig);
+      const agent = result.services.agent;
+
+      expect(agent.command).toEqual(['/bin/bash', '-c', 'copilot --prompt "hello world"']);
+    });


Tests cover literal paths (including single-quoted paths with spaces), but there’s no coverage for prompt file paths that rely on shell expansion (e.g. $RUNNER_TEMP/prompt.txt or "$GH_AW_PROMPT"). Given the rewrite logic currently normalizes then single-quotes the extracted path, adding a test for an env-var-based path would help catch regressions/semantic changes once the quoting/rewriting strategy is adjusted.

Replace 'cat file | command --prompt -' with 'command --prompt - < file' to preserve shell semantics for compound commands (pipes, &&, ;). Preserve original path quoting so env-var paths like $RUNNER_TEMP expand correctly at runtime instead of being force single-quoted. Add test coverage for env-var prompt paths and compound commands. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

github-actions · 2026-04-16T00:33:39Z

✅ Coverage Check Passed

Overall Coverage

Metric	Base	PR	Delta
Lines	85.35%	85.56%	📈 +0.21%
Statements	85.24%	85.41%	📈 +0.17%
Functions	87.96%	88.10%	📈 +0.14%
Branches	77.95%	78.01%	📈 +0.06%

📁 Per-file Coverage Changes (1 files)

File	Lines (Before → After)	Statements (Before → After)
`src/docker-manager.ts`	86.8% → 87.5% (+0.68%)	86.4% → 86.9% (+0.53%)

Coverage comparison generated by scripts/ci/compare-coverage.ts

github-actions · 2026-04-16T00:35:28Z

Smoke test results (run 24485544682)

✅ GitHub MCP: "feat: add smoke test for Copilot CLI offline BYOK mode" / "fix(api-proxy): fix Gemini API_KEY_INVALID with credential isolation"
✅ Playwright: github.com title contains "GitHub"
✅ File write: /tmp/gh-aw/agent/smoke-test-claude-24485544682.txt created
✅ Bash verify: file read back successfully

Overall: PASS

💥 [THE END] — Illustrated by Smoke Claude

github-actions · 2026-04-16T00:36:59Z

Smoke test results:

feat: add smoke test for Copilot CLI offline BYOK mode ✅ (GitHub MCP review)
fix(api-proxy): fix Gemini API_KEY_INVALID with credential isolation ✅ (GitHub MCP review)
safeinputs-gh PR query ❌ (tool unavailable in this run)
Playwright github.com title contains "GitHub" ✅
Tavily web search ❌ (Tavily MCP unavailable)
File write/readback in /tmp/gh-aw/agent ✅
Discussion interaction ✅
npm ci && npm run build ✅
Overall status: FAIL

🔮 The oracle has spoken through Smoke Codex

lpcox · 2026-04-16T03:49:19Z

abandoning since copilot cli does not support --prompt -

Initial plan

8c2a75a

Copilot AI assigned Copilot and lpcox Apr 15, 2026

Copilot started work on behalf of lpcox April 15, 2026 21:01 View session

Copilot AI linked an issue Apr 15, 2026 that may be closed by this pull request

E2BIG: large prompts inlined into argv via --prompt "$(cat ...)" exceed ARG_MAX #1998

Open

Copilot AI added 4 commits April 15, 2026 21:07

fix: rewrite inline prompt cat expansion to avoid E2BIG argv failures

b8c973a

Agent-Logs-Url: https://github.com/github/gh-aw-firewall/sessions/4052214b-06e7-4406-aeb0-e3958af8a77e

docs: clarify prompt rewrite ARG_MAX mitigation internals

7d6d23c

Agent-Logs-Url: https://github.com/github/gh-aw-firewall/sessions/4052214b-06e7-4406-aeb0-e3958af8a77e

fix: harden inline prompt rewrite parsing for quoted paths

52d5be9

Agent-Logs-Url: https://github.com/github/gh-aw-firewall/sessions/4052214b-06e7-4406-aeb0-e3958af8a77e

fix: avoid regex redos in inline prompt rewrite logic

6b0f9c6

Agent-Logs-Url: https://github.com/github/gh-aw-firewall/sessions/4052214b-06e7-4406-aeb0-e3958af8a77e

Copilot AI changed the title ~~[WIP] Fix E2BIG error due to large prompts in argv~~ Mitigate ARG_MAX/E2BIG by rewriting inline prompt expansion to stdin in AWF agent command Apr 15, 2026

Copilot finished work on behalf of lpcox April 15, 2026 21:13

Copilot AI requested a review from lpcox April 15, 2026 21:13

lpcox marked this pull request as ready for review April 16, 2026 00:09

lpcox requested a review from Mossaka as a code owner April 16, 2026 00:09

Copilot AI review requested due to automatic review settings April 16, 2026 00:09

Copilot started reviewing on behalf of lpcox April 16, 2026 00:10 View session

Copilot AI reviewed Apr 16, 2026

View reviewed changes

github-actions Bot added the smoke-claude label Apr 16, 2026

github-actions Bot mentioned this pull request Apr 16, 2026

[aw] No-Op Runs #1993

Closed

lpcox closed this Apr 16, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Mitigate ARG_MAX/E2BIG by rewriting inline prompt expansion to stdin in AWF agent command#1999

Mitigate ARG_MAX/E2BIG by rewriting inline prompt expansion to stdin in AWF agent command#1999
Copilot wants to merge 6 commits intomainfrom
copilot/fix-arg-max-exceeded-error

Copilot AI commented Apr 15, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 16, 2026

Uh oh!

Copilot AI Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

lpcox commented Apr 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

Copilot AI commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Copilot's findings

Uh oh!

Copilot AI Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Apr 16, 2026

✅ Coverage Check Passed

Overall Coverage

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

lpcox commented Apr 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Copilot AI commented Apr 15, 2026 •

edited

Loading