[posthog]

Problem

• Inbox research runs and scouts that name a repository the team's GitHub app cannot read fail after three identical clone attempts, with a generic "git exited with code 128" message.
• The run has no usable token for that repository. Git then asks for a username, and every retry reuses the same credential, so it fails the same way.
• Each attempt captures a RepositoryCloneError. Access denials make up most of the RepositoryCloneError issue since fix(tasks): stop a run planning when it has no repository checkout #117823 merged. That PR's .git guard works as intended. It now fails these runs that were already broken.
• The rest of the issue is a different problem: sandbox DNS failures and concurrent writes to the pack files.

Origin

• Scout: signals-scout-inbox-validation
• First signal: 2026-10-20
• Inbox report: open
• Task started by: auto-start, after the report was rated P2 and ready to fix

Changes

• When GitHub denies access, the run fails at once with a message the person can act on: "Could not clone owner/repo: GitHub denied access. Check that the PostHog GitHub app is installed on this repository's account and can read this repository."
  • A new RepositoryAccessError is non-retryable and a NonReportableError, the same contract as CredentialUnavailableError. It does not reach error tracking.
  • The match uses git's output: "could not read Username", "Authentication failed for", or "remote: Repository not found".
• Other clone failures, such as "Could not resolve host", keep the transient RepositoryCloneError and its three retries.
• The resume fallback for a missing remote branch runs first, as before.

How did you test this code?

• Added test_clone_failure_stops_retries_only_when_github_denies_access to test_provision_sandbox.py, with three cases: no credentials, repository not found, and DNS failure.
• Test rationale: the test fails if an access denial goes back to retrying or to error tracking, or if a network failure stops retrying. The nearest test, test_clone_failure_records_failed_latency_and_captures_command_result, covers only the generic failure path.
• Ran test_provision_sandbox.py locally. test_desktop_bootstrap_relaunch_replaces_a_running_build fails, and it also fails on master in this sandbox, as fix(tasks): stop a run planning when it has no repository checkout #117823 noted.
• Not run: repo-wide mypy, the workflow tests, and any manual run in a live sandbox.

Release status

• No feature flag controls this change

Docs update

None.

🤖 Agent context

Autonomy: Fully autonomous

Agent: PostHog Desktop (Claude Code), claude-opus-5-5

• Skills invoked: /writing-tests, /writing-pr-descriptions.
• Error-tracking events and worker logs showed git's stderr for each clone failure after fix(tasks): stop a run planning when it has no repository checkout #117823. Most were access denials from read-only and report runs, where the token mint returned no token.
• Considered and rejected: skipping the clone in the workflow when the token is empty. That path still fails at the .git guard, and it needs a workflow patch.
• Open question for the owners: the read-only token contract says that a missing token must not fail the run. Since fix(tasks): stop a run planning when it has no repository checkout #117823, a read-only run that pins an inaccessible private repository fails. This PR makes that failure fast and clear, but does not decide the contract.
• fix(tasks): attach the github installation that reaches the task repository #119056 is related. It picks an installation that can reach the repository at task creation for the write path. This PR covers the clone-time failure that remains.

---

Created with PostHog Desktop from this inbox report.

🤖 Generated with Claude Code

[trunk-io]

😎 Merged successfully - details.

[parameterai]

Risk: No findings

This change stops retrying sandbox repository clones when git's output shows GitHub denied access: a new fatal, non-reportable RepositoryAccessError replaces the transient RepositoryCloneError for those failures, with a user-actionable message. The risk surface is minimal — the new error context deliberately omits stdout/stderr and exposes only a boolean for token presence, so no credentials can leak through the new path.

_{Sentinel reviewed b68b80f · Review settings}

[github-actions]

🤖 CI report

⚠️ Trunk lane — backend Python lane

This PR is assigned to the backend Python lane. It runs backend Python tests and may merge in parallel with PRs in other lanes.

✅ Duplication (Python) — clean

New Python code duplication introduced by this branch. Fails at 70+ tokens in app code, or 150+ tokens when both copies live in test files. Advisory while the gate proves itself: extract a shared helper instead of copying.

✅ Duplication (TypeScript) — clean

New TypeScript code duplication introduced by this branch. Fails at 70+ tokens in app code, or 150+ tokens when both copies live in test files. Advisory while the gate proves itself: extract a shared helper instead of copying.

[stamphog]

Approved.

The access-denial handling has current-head independent assurance from PostHog Review (Standard), and no unresolved concerns were supplied or found. The verified Inbox provenance exempts this machine author from the bot-author refusal rule.

• PostHog Review (Standard) reviewed the current head: no findings.

Gate mechanics and policy version

Gate		Result
prerequisites	✓	all clear
deny-list	✓	no deny categories matched
size	✓	39L, 2F substantive, 99L/3F incl. docs/generated/snapshots — within ceiling
tier	✓	T1-agent / T1b-small (99L, 3F, single-area, fix)
stamphog 2.6.0		.stamphog/policy.yml @ b68b80f · reviewed head b68b80f

[trunk-io]

   

View Full Report ↗︎ ⋅ Docs

[deployment-status-posthog]

Deploy status

Environment	Status	Deployed At	Workflow
dev	✅ Deployed	2026-10-21 13:35 UTC	Run
prod-us	✅ Deployed	2026-10-21 13:54 UTC	Run
prod-eu	✅ Deployed	2026-10-21 13:56 UTC	Run