Testing plan: a read-only tmux observer blocks every comment the poller forwards

Derived from bugfix.md and design.md, beforetasks.md — each task's _Test:_ names a row of the matrix below. Authored at the test-planning node and completed at the verification node.

This file is executable content. It names commands an agent will run, so review it like code. Credentials appear by reference only; this plan needs none.

What has to be proved, and the one thing that is hard about it

The bug reproduces only on tmux ≥ 3.7, and this repository's CI (and the machine this work item was executed on) runs tmux 3.4. So the plan is split deliberately:

• T1 proves the mechanism on a live tmux, with a genuine read-only client attached — that an unbracketed paste-buffer of a carriage return submits, and that it is accepted while send-keys would be resolved against the observer. This is real, and it runs wherever tmux exists.
• T2 pins the contract that makes the mechanism reachable — the exact argv sequence. This is what protects the fix on machines where the failing tmux is not installed, and it is the reason R4.2 asks for the whole list rather than "no send-keys".
• T11 records the version evidence — that send-keys acquired the guard after 3.6 — so the claim in bugfix.md is checkable rather than asserted.

What cannot be proved here is stated once, in Verification results, rather than implied: no tmux ≥ 3.7 is available in this environment, so the end-to-end refusal is demonstrated from tmux's own source and version history, not from a failing command.

Test matrix

#	Type	Applies?	Scope / what it proves	Where it runs
T1	Integration (live tmux)	yes	An unbracketed CR paste submits a bracketed-pasted prompt, with and without a read-only client attached; a real #{client_readonly} client is present for the second half	manual procedure, output committed as evidence
T2	Unit	yes	TmuxRunner.deliver issues exactly load-buffer, paste-buffer -p, load-buffer, paste-buffer — no send-keys, both buffers -d, both tempfiles unlinked	uv run --project cli python -m pytest -q cli/tests/test_tmux_runner.py
T3	Unit	yes	giveup_notice builds a body carrying the marker, the attempts and the recovery, and takes no parameter that could carry a comment body	uv run --project cli python -m pytest -q cli/tests/test_poller.py
T4	Integration (scenario)	yes	Gherkin-documented: a poller that exhausts the budget posts one notice; the ledger is written first; a failing post changes nothing	uv run --project cli python -m pytest -q cli/tests/test_poller_integration.py
T5	Integration (scenario)	yes	The webhook→tmux pipeline's recorded argv no longer contains send-keys	uv run --project cli python -m pytest -q cli/tests/test_tmux_runner_integration.py
T6	Security / abuse case	yes	No payload-derived text reaches an argv or the notice body; the notice carries the self-comment marker so it cannot resume the session	uv run --project cli python -m pytest -q cli/tests/test_poller.py -k notice
T7	Contract (OpenAPI / GraphQL SDL)	n/a — no API surface changes. deliver is internal; the poller exposes no route.
T8	End-to-end	n/a — an end-to-end run needs a live harness TUI, a GitHub repository and the daemon; T1 covers the tmux half against a real server and T4 covers the poller half against a real ledger.
T9	UI / visual	n/a — no UI is touched.
T10	Snapshot	n/a — no snapshot artifacts in this repository.
T11	Manual exploratory	yes	tmux version archaeology: send-keys carries no read-only guard at 3.4/3.5a/3.6 and does at master; a read-only client is genuinely attached during T1	manual procedure, output committed as evidence
T12	Performance / load	n/a — one extra tmux round trip per delivered event, on a path that already writes a tempfile and probes liveness. Not measurable against the harness's own latency.
T13	Accessibility	n/a — no user interface.
T14	Migration / upgrade	n/a — no persisted shape changes. The poll ledger's gaveUp record is written exactly as before; the notice adds no field.
T15	Regression suite	yes	The whole Python suite, to prove nothing else depended on send-keys or on the poller's give-up branch being silent	uv run --project cli python -m pytest -q cli

Scenarios & requirement trace

Row	Requirement(s)	Scenario / case
T1	R1.1, R1.3, R1.4	A prompt pasted bracketed is submitted by an unbracketed CR paste, while tmux list-clients reports client_readonly=1
T2	R1.2, R1.5, R3.3, R3.4, R4.1, R4.2	The argv sequence; -d on both pastes; both tempfiles gone; unchanged session_missing behaviour
T3	R2.1, R2.2, R2.3, R2.6	giveup_notice content and signature
T4	R2.1, R2.4, R2.5, R4.3	Scenario: A comment abandoned after the retry budget is reported on the ticket · Scenario: A give-up is recorded even when the ticket cannot be told
T5	R1.2, R3.1, R3.2	Scenario: An event is delivered into a live tmux session without send-keys
T6	R2.6, R2.2	Negative: the notice for a comment whose body is <script>/marker-forgery text contains none of it, and carries <!-- the-loop:agent-comment -->
T11	R1.1 (root cause)	Version archaeology + a real read-only client
T15	R3.1, R3.2	No regression anywhere else in the CLI

Verification environment

• Repositories: this repository only.
• Services / containers: none. T1 and T11 need a running tmux on the machine (tmux 3.4 here) and a pty for the read-only attach, obtained with script -q -c "tmux attach-session -r -t <session>" /dev/null.
• Fixtures & data: none. The poller tests use the existing in-repo fakes (an injected runner in place of gh, a WorkItemStore under tmp_path).
• Credentials: none. No test invokes a real gh; post_issue_comment takes an injectable runner and every test supplies one.
• Bring-up: uv sync · Tear-down: tmux kill-server after T1/T11.
• If bring-up fails: record it under Verification results, leave the dependent activities unticked, and escalate.

Evidence plan

Row	Evidence	Path under evidence/
T2, T3, T4, T5, T6, T15	test summaries (counts, duration) and the red run that preceded the fix	unit-and-integration.md, red.md
T1, T11	the live tmux transcript: session set-up, the read-only client's #{client_readonly}, the two pastes, the pane's own output, and the per-version cmd-send-keys.c counts	manual.md

Nothing captured here contains a token, a hostname or personal data: the transcripts are local tmux session names and pytest counts.

Verification activities

• [x] T1 — live tmux: bracketed paste + CR paste submits, with a read-only client attached
• [x] T2 — uv run --project cli python -m pytest -q cli/tests/test_tmux_runner.py
• [x] T3 — uv run --project cli python -m pytest -q cli/tests/test_poller.py
• [x] T4 — uv run --project cli python -m pytest -q cli/tests/test_poller_integration.py
• [x] T5 — uv run --project cli python -m pytest -q cli/tests/test_tmux_runner_integration.py
• [x] T6 — uv run --project cli python -m pytest -q cli/tests/test_poller.py -k notice
• [x] T11 — tmux version archaeology, recorded with the commands that produced it
• [x] T15 — uv run --project cli python -m pytest -q cli
• [x] Red run — every new assertion, failing against unfixed code, committed before the fix

Verification results

Not yet executed.

Activity	Command / procedure	Outcome	Evidence

Review comments

Appended by the-loop's record-feedback hook when a human gate approves with comments (issue-109). Append-only and attributed.