Tasks: multi-modal messages are forwarded to the session, never parsed by the loop

The last spec artifact. Derived from design.md and testing-plan.md. Gate-less (issue-281): it advances on shape.

Execution DAG

Task list

[x] 1. Write cli/the_loop/channels/attachments.py with its unit tests, red first
- Attachment, the constants, kind_of, safe_name, vtt_to_text, github_attachment_urls, fetch_url (no auto-redirect, host check per hop, the cap), slack_attachments (with the transcript wait), github_attachments, render_section, record_lines.
- cli/tests/test_attachments.py with the T1 and T8 scenarios; it fails on import before the module exists — that is the red.
- Depends on: none
- Requirements: R1.2, R1.4–R1.6, R2.1–R2.4, R3.1, R3.2, R4.1, R6.2; abuse cases 1–5
- Test: T1 / T8 — pytest cli/tests/test_attachments.py
[x] 2. Declare the attachments directory in the state layout
- StateLayout.attachments_dir → <root>/local/attachments; a GENERATED_PATHS entry (portable=False, its why); the tree line and the classification row on docs/cli/state.md, with the no-retention sentence.
- Depends on: 1
- Requirements: R6.1, R6.3
- Test: T10 — pytest cli/tests/test_state_portability.py
[x] 3. Carry files on the normalized Slack message
- InboundReply.files: Tuple[Mapping, ...] = (); populated in handle_socket_event (both branches), fetch_replies, fetch_kickoffs, fetch_channel_messages.
- Depends on: 2
- Requirements: R1.1
- Test: T2 — the poll-read and socket scenarios (written in task 7; the readers are exercised by the existing channel suites meanwhile)
[x] 4. Fetch and render in the pipeline; name files in the snapshot and the kickoff
- SlackBotChannel.fetch_attachments; in process_reply the record text gains record_lines, the delivered text render_section, for the three relayed kinds; snapshot_thread appends 📎 per file; process_kickoff appends record_lines to the issue body.
- Depends on: 3
- Requirements: R1.2, R1.3, R1.7, R2.1, R2.2, R3.1–R3.4; abuse case 6
- Test: T2 — pytest cli/tests/test_attachments_integration.py -k "slack or voice or only or gate or snapshot or kickoff or stranger"
[x] 5. Append the section to the GitHub prompt
- Dispatcher._attachments_section(routed, work_item) after _render_prompt's render: bodies from the four containers, the configured host, the token from integrations.github.api.tokenEnv, reuse from disk.
- Depends on: 4
- Requirements: R4.1–R4.5
- Test: T2 — pytest cli/tests/test_attachments_integration.py -k github and T10 — pytest cli/tests/test_interaction_integration.py (a prompt without attachments unchanged)
[x] 6. Declare the scope and measure it
- files:read in slack-app-manifest.yaml with its reason; attachment_findings beside mention_findings, appended in probe_subscription.
- Depends on: 5
- Requirements: R5.1–R5.3
- Test: T1 — pytest cli/tests/test_attachments.py -k finding; T10 — the manifest suites
[x] 7. Write the integration scenarios
- cli/tests/test_attachments_integration.py, Gherkin docstrings, every T2 row of the plan; fakes at the edges (FakeSlackClient + files_info, a fake fetch, FakeTmux).
- Depends on: 6
- Requirements: all
- Test: T2 — pytest cli/tests/test_attachments_integration.py
[x] 8. Documentation and the decision record
- docs/guide/slack.md: the manifest copy gains files:read; a new Images and voice notes section; a line under Limits. docs/capabilities/channels.md and docs/capabilities/webhook-triggers.md: current-behaviour clauses and history rows. docs/decisions/decision-135.md + its index row. evidence/documentation.md.
- Depends on: 7
- Requirements: the loop's rule — docs change in the same PR
- Test: T12 — pytest cli/tests/test_docs_parity.py and make lint
[x] 9. Verify and record
- make check; fill testing-plan.md § Verification results; write evidence/verification.md, self-review.md, security-review.md, pull-requests.md; the T11 procedure written down and marked not executed here.
- Depends on: 8
- Requirements: all
- Test: All — make check

Dependency graph (DAG)

Linear, as drawn above: each task builds on the seam the previous one opened, and the integration scenarios (7) need every seam in place.

Checkpoints

After task 1: the module and its contract exist, offline-tested.
After task 4: Slack end to end, provable with the existing fake client.
After task 6: both channels and the scope; ready for the scenarios.
After task 9: evidence recorded; ready for review.

Out of scope for this DAG

No task calls a model on a file, adds a config key, changes the excerpt contract, or touches the Cursor adapter. If one appears to need it, stop and raise it on the ticket.

Review comments

Appended by the-loop's record-feedback hook when a human gate approves with comments.