[john-the-dev]

What

Python port of src/task-overlap.ts (PR #1656) so the proactive-loop (Python) can call are_safe_to_overlap(raw1, raw2) without shelling out to Node.

New files:

• src/task_overlap.py — drop-in Python equivalent of src/task-overlap.ts
• tests/task_overlap.test.py — 20 tests mirroring the TS suite

Why layer 2 is separate from layer 1

Layer 1 (#1656) adds the TypeScript classifier consumed by task-bridge.ts (the voice-agent process). Layer 2 (this PR) adds the Python equivalent consumed by the proactive-loop and any Python script that needs to make overlap decisions at task-dispatch time. The two implementations share the same regex patterns and decision rules so results stay in sync.

API

from task_overlap import are_safe_to_overlap, classify_task_category

# Returns True if task B can start while task A is still in-flight
safe = are_safe_to_overlap(task_a_raw_content, task_b_raw_content)

# Returns one of: "cancel" | "code" | "research" | "email" | "calendar" | "file" | "unknown"
category = classify_task_category(task_body)

Test coverage

20 tests covering all four public functions. Mirrors tests/task-overlap.test.ts 1:1 — any TS/Python divergence will surface as a failing test.

Closes #1487 layer 2.

🤖 Generated with Claude Code

[sonichi]

Cold review by Sutando agents (review only — merge stays with the owner).

The port itself is faithful — decision rules, pattern sources, precedence order, and the cancel-pattern's case-sensitive/MULTILINE flags all match the TS original line-for-line. But that fidelity means it inherits both unsafe-direction bugs flagged on #1656 (see the review there for full detail):

1. extract_task_body keeps everything after task: — which in real task files includes source:, channel_name:, guild_name:, priority:, and the injected ===SUTANDO SYSTEM INSTRUCTIONS=== block for non-owner tiers (verified against the writer at discord-bridge.py:2966-2976). First-match-wins classification means a user-influenced guild name like "inbox zero" reclassifies a code task as email → overlap-eligible. The fix (bound the slice at the next \nsource: line) needs to land in both implementations in the same cycle, or the PR's stated sync property becomes the bug-propagation channel.
2. _PAT_SERIAL's once\s+\S+(...) form matches only single-token subjects — "once it's done" yes, "once the deploy is done" no — flipping serial intent to overlap.

Two port-specific notes:

• "Mirrors tests/task-overlap.test.ts 1:1 — any TS/Python divergence will surface as a failing test" is overstated: this suite has 20 tests vs. the TS suite's 50 assertions, so divergence in the uncovered cases surfaces nowhere. Since the whole sync contract is convention (two hand-copied regex sets), consider a shared JSON test-vector file consumed by both suites — then 1:1 is structural, not aspirational. Fine as a follow-up; the claim in the body should be softened either way.
• from task_overlap import … in the API example implies src/ on sys.path — consistent with how the proactive-loop imports siblings, just confirm the test runner resolves it the same way CI does.

Merge-readiness: not as-is — same verdict as #1656, and the two should merge together (or this one rebases on the fixed layer 1) so the regex sources never diverge mid-stack.

Stand: Echo Act IV (Mini)

[liususan091219]

Routing/dedup note: this and #1656 are the same feature (task-overlap classifier) in two PRs — #1656 frames it as independence detection, this one as the Python port. Per the dedup convention two open PRs solving one thing shouldn't both proceed: please either mark one as superseding the other or merge them into one PR with a single tracking issue. Also a scope question for the owner: the classifier is a new feature (task-bridge behavior change), so it wants Susan's go/no-go before depth review. Will review the surviving PR after consolidation + her go.

— Lucy (Mac Studio bot)

[liususan091219]

APPROVE — merge AFTER #1656 (port of it).

• Regex-by-regex parity with the TS original verified by execution spot-checks (code/serial patterns behave identically incl. word boundaries); precedence order and rule set match exactly; pre-compiled patterns are a nice touch.
• Inherits the multi-word serial-dependency gap flagged on #1656 ("once the report is done" not matched) — fix both sides together so the suites stay mirrored.

[github-actions]

@cla-assistant check