[hkwuks]

Summary

An Agent working inside TaskFlow tends not to look outside the repository. This change asks it where the answer comes from at the three points where that answer can change, records the lookups it makes, and refuses a claim to have looked that nothing backs.

v1 of this PR (still in the branch history, as 96044df) shipped only the field — a ## External Facts section approve validates. That recorded a judgement without touching the tendency to write no reflexively, and the v1 PRD said so in its own risk section. The user's correction on 2026-11-01 is what this PR now answers. v1's documents are archived under TaskFlowDocs/2026-10-31-external-facts/old/v1/.

TaskFlow traceability

• Task: TaskFlowDocs/2026-10-31-external-facts/ (v2)
• Scope: hooks/external-evidence (new), hooks/hooks.json, hooks/task, hooks/smoke-test, hooks/README.md, skills/taskflow/SKILL.md, skills/taskflow/references/{artifacts,runtime}.md, CHANGELOG.md, CONTRIBUTING.md (archive ordering), and the archive of this task itself
• Base branch: main
• Target repository: hkwuks/TaskFlow

Why a field was not enough

Every source this workflow points an Agent at is inside the repository:

where	what it says
SKILL.md:54	inspect lists seven sources — all internal
summarize-state	opens the session with internal conclusions
prompt-reminder	every turn, about task state only
capability-evidence	counts Skill/Agent/mcp__* and nothing else, so a web lookup is invisible

Nothing in that list prompts the question, so a no written by default is indistinguishable from one reached by thinking.

What the three pause points do

Before Phase 1 converges, at the Phase 3 spec decision — defining an interface or a compatibility boundary is exactly when "only our own code" stops being true — and once more before approval, against Steps that did not exist when the first answer was given.

The inspect enumeration is deliberately not extended. Its entries are "open the repo and look"; a judgement question among them would dress a decision up as an inspection. A smoke assertion counts the entries so a ninth cannot be added quietly.

The isolation, which is the design

taskflow/evidence   ← Skill / Agent|Task / mcp__*   → releases a stage
taskflow/external   ← WebSearch / WebFetch          → releases nothing

capability-evidence is untouched, so a built-in tool still cannot satisfy the gate (R12, 2026-09-27). The two predicates live in two files with two consumers: one case with a shared fall-through would put the meanings back together. The suite proves the separation rather than asserting it — a task holding two real lookups and no invocation must still be refused, and refused for the missing stage.

approve reads external for the opposite purpose: a yes needs every - Source: backed by a recorded lookup. achieved/2026-09-08-runtime-hooks is the case in point — it names Anthropic's hooks reference and Codex's hooks guide in prose, with no URL and no date, in a repository with no way to tell whether either was opened.

A no must now point at where the work is settled (a backticked path or path:line); "this is an internal change" is refused, because it is the sentence an unexamined no produces.

Verification

• The Hooks workflow reports pass for this branch — linked below once it runs
• bash hooks/smoke-test — ALL SMOKE PASSED, ~20 new assertions
• bash hooks/release-check . — STATUS: pass
• git diff --check — clean
• Four mutations, each reddening its own named assertion and restored byte-identically:

mutation	assertion that fired
drop the lookup-backing check	FAIL the unbacked source was not named
never fail on an unbacked source	FAIL the unbacked source was not named
drop the no pointer requirement	FAIL external facts message for no reason names nothing
add a ninth inspect entry	FAIL the inspect enumeration changed size

• Results recorded in the TaskFlow Plan — ## Verification / Review

The archive is in this pull request

hooks/task complete has been run on 2026-10-31-external-facts, and the move to achieved/ plus the Todo rewrite are the third commit here. Archiving after the merge would have left them on no branch — landable only by pushing to main outside a pull request, or by a second pull request carrying bookkeeping alone, both of which CONTRIBUTING.md forbids.

That document already covered the archive by name; the practice had drifted from it because hooks/task complete prints a teardown line that waits for the merge, and that condition is about git worktree remove, not about when to archive. CONTRIBUTING.md and SKILL.md now say which is which, and a smoke assertion requires both to keep saying it.

Two defects the fixture caught

• a|b recorded as search|a|b — one record with a forged field boundary. The assertion compares the whole line, not the count. The first fix then ate the separator by running tr over the joined string, which the same assertion caught.
• the isolation fixture shared a capability store the earlier cases had pushed out of step, so approve refused for the wrong reason. It has its own worktree now.

Review boundaries

• No secrets or opaque remote payloads included
• No unrelated task or user files changed
• Remote/base assumptions stated
• Known limitation, stated in the Plan rather than papered over: a model that answers no with a plausible pointer still gets through. Enforcing otherwise would demand external evidence from tasks that genuinely need none, which would make most of this repository's work lie. The field records a decision and the record makes it checkable; it does not make the decision honest.
• PostToolUse for WebSearch/WebFetch is documented by Claude Code's hooks reference (read 2026-11-01) but not yet observed in a running session — runtime.md records it as documented, not measured, and says what moves it to measured.