Testing plan: a forwarded event carries the instruction, not GitHub's metadata

Derived from the approved requirements.md and design.md, before tasks.md. Authored at test-planning, completed at verification.

Test matrix

#	Type	Applies?	Scope / what it proves	Where it runs
T1	Unit — comment surfaces	yes	R1: issue_comment, pull_request_review_comment, pull_request_review carry body + address + author and nothing else; anchor before body	uv run pytest cli/tests/test_excerpt.py
T2	Unit — every other event	yes	R2: lifecycle, label, CI and status events keep what makes them actionable; an unknown event distils rather than falling back to raw	uv run pytest cli/tests/test_excerpt.py
T3	Unit — caps and JSON validity	yes	R3: a 10 KB body truncates the field only; the excerpt parses; URL and anchor survive	uv run pytest cli/tests/test_excerpt.py
T4	Integration (scenario) — ingress parity	yes	R4.1: the poller's synthesised event and the equivalent webhook event render the same fields, through one function	uv run pytest cli/tests/test_excerpt_integration.py
T5	Integration (scenario) — the gates are untouched	yes	R5.1: authorization, self-comment detection, control parsing and reaction targeting still decide correctly on an event whose excerpt omits their inputs	uv run pytest cli/tests/test_excerpt_integration.py
T6	Security / abuse case	yes	Abuse cases 1–4 of requirements.md: forged JSON in a body stays inside the string; a crowding body is bounded with its URL intact; an unlisted hostile field never reaches the prompt; a malformed container yields {} rather than raising	uv run pytest cli/tests/test_excerpt.py -k abuse
T7	Regression — existing dispatcher/interaction suites	yes	R5.2: the $payload_excerpt placeholder contract, its position above/below the directive, and every existing delivery test still hold	uv run pytest
T8	Measurement — before/after cost	yes	The non-functional numbers claimed in requirements.md § Introduction and design.md are measured, not asserted	uv run python docs/specs/issue-243/evidence/measure_prompt.py
T9	Contract (OpenAPI / GraphQL SDL)	n/a — the change touches no API surface. docs/api-specs/openapi describes the control plane; the excerpt is prompt text inside the dispatcher
T10	End-to-end	n/a — an end-to-end run needs a real gh, real credentials and a live tmux harness; none exist in this environment. T4/T5 cover the seam between ingress and prompt, which is what changed
T11	UI / visual	n/a — the-loop has no product UI (design.uiArtifacts.format: html, unused here)
T12	Snapshot	n/a — the excerpt's exact bytes are asserted field-by-field in T1–T3, which is a stricter and more readable check than a golden file
T13	Performance / load	n/a — the function is a dict walk over ≤ 8 fields; the change strictly reduces work and output size
T14	Accessibility	n/a — no human-facing UI
T15	Migration / upgrade	n/a — no state, no config key, no on-disk format changes. An operator's custom template keeps working unchanged (R5.2), which T7 pins
T16	Manual exploratory	n/a — no interactive surface changed; the reviewer reads the measured before/after in T8 instead

Scenarios & requirement trace

Row	Requirement(s)	Scenario / case
T1	R1.1, R1.2	A conversation comment renders body, URL and author, and no issue, sender or API URL
T1	R1.3	An inline review comment renders path and line before body
T1	R1.4, R1.5	A review renders state, body, html_url, author — the author as a login string
T2	R2.1, R2.2	A labeled issue event renders the entity's four fields plus the label's name
T2	R2.3, R2.4	A failed check_run renders name/status/conclusion/URL and its output summary
T2	R2.5	A status event renders its root-level fields
T2	R2.6, R2.7	An event with no rule distils the containers it has; a payload with none renders {}
T3	R3.1, R3.2, R3.3	Scenario: a comment carrying a 10 KB log is truncated to its body alone
T4	R4.1, R4.2	Scenario: the poller and the webhook render the same comment identically
T5	R5.1	Scenario: the gates read the payload, not the excerpt
T6	Abuse cases 1–4	Forged JSON inside a body; crowding body with surviving URL; unlisted hostile field absent; malformed container tolerated
T7	R5.2	The existing test_interaction.py / test_routing.py assertions on the placeholder and delivered prompts
T8	Non-functional cost	Measured baseline and post-change prompt sizes for the same event

Verification environment

• Repositories: this repository only.
• Services / containers: none. No gh, no tmux, no network: the tests exercise pure functions and the dispatcher's fake-tmux seam the existing suite already uses.
• Fixtures & data: one realistic GitHub issue_comment payload (hand-built in the test module and in evidence/measure_prompt.py, shaped after GitHub's documented webhook payload — two user objects, reactions, labels, the full issue).
• Credentials: none. No test reads an environment variable or a secret.
• Bring-up: uv sync · Tear-down: none.
• If bring-up fails: record it under Verification results, leave the dependent activities unticked, and escalate.

Evidence plan

Row	Evidence	Path under evidence/
T1–T3, T6	red run (tests failing against the undistilled tree)	red.md
T1–T7	unit + integration run output, counts, duration	unit-and-integration.md
T7	full-suite run, plus ruff, ruff format --check, pyright, markdownlint	lint-and-typecheck.md
T8	the measurement script and its before/after output	measure_prompt.py, baseline.md, after.md

Verification activities

• [x] T1 — uv run pytest cli/tests/test_excerpt.py
• [x] T2 — uv run pytest cli/tests/test_excerpt.py
• [x] T3 — uv run pytest cli/tests/test_excerpt.py
• [x] T4 — uv run pytest cli/tests/test_excerpt_integration.py
• [x] T5 — uv run pytest cli/tests/test_excerpt_integration.py
• [x] T6 — uv run pytest cli/tests/test_excerpt.py -k abuse
• [x] T7 — uv run pytest
• [x] T8 — uv run python docs/specs/issue-243/evidence/measure_prompt.py

Verification results

Activity	Command / procedure	Outcome	Evidence
T1–T3, T6 (red)	uv run pytest cli/tests/test_excerpt.py -q, against the pre-change distiller	fail — 21 failed, 6 passed (the six pass in both trees; the file says which and why)	red.md
T4, T5 (red)	uv run pytest cli/tests/test_excerpt_integration.py -q, with 08b7bd6:excerpt.py restored under the wired dispatcher	fail — 4 failed	unit-and-integration.md
T1, T2, T3, T6	uv run pytest cli/tests/test_excerpt.py -q	pass — 28 passed	unit-and-integration.md
T6	uv run pytest cli/tests/test_excerpt.py -k abuse -q	pass — 8 passed, 20 deselected	unit-and-integration.md
T4, T5	uv run pytest cli/tests/test_excerpt_integration.py -q	pass — 4 passed	unit-and-integration.md
T7	uv run pytest cli -q	pass — 2156 passed, 1 skipped	unit-and-integration.md
T7	ruff check, ruff format --check, pyright cli, markdownlint-cli2 "**/*.md", validate_config.py	pass — clean	lint-and-typecheck.md
T8	uv run python docs/specs/issue-243/evidence/measure_prompt.py	measured — excerpt 4,014 → 203 chars (−94.9%), prompt 6,676 → 2,865 (−57.1%), and the excerpt parses	baseline.md, after.md

Not executed: none. Every activity in the checklist ran.

Corrected after execution: two numbers the specs carried before the code existed. The distilled excerpt is 203 characters, not the ~238 design.md estimated (the estimate counted a field the design later dropped — see below), and the integration suite is 4 tests, not the 5 the plan sketched: ingress parity for a conversation comment and for the review surfaces fit one test each rather than three.

Changed during implementation: the issue and pull_request containers do not carry an author. The design said every container would; the first run of test_a_labeled_issue_carries_the_entity_and_the_label_that_is_the_event showed what that means in practice — GitHub's issue.user is whoever opened the item, not who acted, so carrying it as author invites a session to reply to the wrong person. Those events carry a top-level actor (from router.event_actor) and nothing else; design.md records the rule.

Residual risk

One assumption is not testable here: that the fields the allow-list keeps are the ones a session actually needs. It is bounded rather than eliminated — the excerpt is context, not an input to any decision (R5.1), and every carried object keeps its html_url, so a session that needs a field the excerpt no longer shows can fetch the object. The failure mode is a session asking one extra question, never a wrong action; the loud alternative (the previous behaviour) was a truncated, unparseable excerpt.

Review comments

None yet.