[dsyme]

Summary

When the initially selected tasks and their local fallbacks find no actionable work, Repo Assist should try useful generic repository work even if those tasks were not randomly selected.

• Emit generic_fallback_tasks in the real task-selection JSON: testing (9), performance (8), take the repository forward (10), then docs/QA (7), engineering (4), and code improvements (5), excluding initially selected tasks.
• Emit no generic fallback plan for explicit command mode or scheduled maintenance-only mode.
• Require the agent to inspect a concrete relevant source/test/docs area or existing check for each assessed category, not infer there is nothing useful from issue titles or test filenames.
• Track already assessed tasks across local and generic fallbacks; visit each category at most once to prevent 1↔2 / 5↔9 cycles. Stop the sweep at the first verified worthwhile work item and apply normal validation, memory and reporting rules.
• Preserve honest noops after the bounded sweep; do not manufacture low-value changes or bypass maintainer decisions, dependencies policy, command-mode exclusivity, maintenance-only mode or PR limits.
• Update both workflow and user documentation so the three selected tasks are explicitly a starting allocation, not a restriction to those tasks.
• Add deterministic planner tests to the existing CI test discovery.

This branch independently targets main, not the separate fetch-compatibility branch. It incorporates the newly merged #407 to resolve their overlap and explicitly preserves maintenance-only mode; the net PR diff contains only the fallback workflow, documentation and tests.

Validation

• python3 -m unittest discover -s tests -p 'test_repo_assist_*.py' -v: 23 tests passed (11 generic-fallback tests plus 12 existing maintenance/capacity tests), exercising the actual inline task-selection Python, including all 120 possible initial three-task combinations, empty backlogs, explicit/blank commands, scheduled capacity/manual exemptions and run-seed reproducibility.
• gh aw compile workflows/repo-assist.md: passed with gh-aw v0.93.3; freshness reports Yes.
• Shellcheck passed for the preparation script; generated command-mode wiring and read-only GitHub agent permissions were verified.
• git diff --check: passed.
• The generic planner is tested deterministically; live model adherence to the sweep/visited-task instructions has not been exercised. No manual workflow dispatch.

Sample workflows/*.lock.yml files are intentionally ignored here, so the regenerated sample lock is validated locally rather than force-added. Current compiler scaffolding warnings remain visible. Until #408 merges, hosted compilation still encounters the two unsupported native fetch declarations on main. That PR has now confirmed all 5 hosted workflows compile, then exposed 14 additional unchanged reusable samples with unsupported native fetch entries; those are outside this fallback-only change.

Compatibility / safety

The current-compiler compatibility changes (removing native Copilot web-fetch and normalizing request_review to request-review) are already on main through #407. This PR does not broaden network allowlists or write privileges. The existing memory validator is unchanged and no migration is added.

A read-only safety review found no confirmed vulnerability in the changed scope. The new command value is passed as data through an environment variable and checked only for nonblank content; it is never evaluated as code. The test harness uses disposable files and mocked inputs, with no API/network calls or credential use. Runtime telemetry values and actual agent adherence remain unverified; no live execution is claimed.

[dsyme]

Hosted validation after integrating current main: all 23 tests passed (11 fallback tests and 12 maintenance/capacity tests), including all 120 initial selections and the new scheduled-cap/manual-exemption cases. Run: https://github.com/githubnext/agentics/actions/runs/40116214786. CI then fails only at the two inherited hosted-workflow native fetch declarations fixed independently in #408. The merge conflict with the now-merged maintenance-only change is resolved; this PR remains a three-file fallback-only diff against main. Local focused compilation/freshness also passed. No workflow was manually dispatched.