fence_twin_acks — AgentTeamsModule

Reads the operator's acknowledgements of kept pre-fencing twins. A twin that fence_twins cannot prove is template text is kept, and fleet reports it as REVIEW on every run. When the operator decides to keep such a twin, a row in this ledger stops that REVIEW, and only that.

Source: agentteams/fence_twin_acks.py — integrity-pinned.

The ledger

<output dir>/references/fence-twin-acks.csv, committed with the team:

file,heading,digest,reason
agents/technical-validator.agent.md,## Accuracy Rules,3f1c0e9a2b7d4c51,CH-08..10 are project rules the template never carried; kept until moved to notes.

Column	Value
file	The path relative to the output dir, as the merge notice prints it.
heading	The twin's heading line.
digest	fence_twins.ack_digest(file, heading, twin): 16 lowercase hex characters. The kept-twin notice prints the whole row, with a placeholder for the reason.
reason	Why the twin stays. At least 40 characters after whitespace is collapsed, the same bar as redteam.registry.MIN_REASON_CHARS.

A matching row changes the merge notice from twin kept (not provably template) to twin kept (acknowledged), with the reason. fleet does not count that notice as REVIEW. It lists it in the target's twins_acknowledged, so the acknowledgement stays visible in the fleet report. The first applied fleet run that shows a given acknowledgement still reports REVIEW (below).

• It never deletes. The row is read only after the drop decision, and only to pick the notice for a twin already being kept. A twin that is provably template text is dropped whether or not a row names it.
• It binds to the exact twin. An edit to the twin's text, or a change of file or heading, changes the digest. The row then stops applying, and the kept notice returns, saying that the acknowledgement on record no longer matches. Whitespace changes do not count as edits.
• Invalid rows are skipped. A wrong header, a malformed digest, an absolute path, a line that is not a heading, or a reason under 40 characters prints a warning and leaves the twin under REVIEW.
• Unused rows are reported. A row naming a file that the run does not render prints a warning.
• Trust model. The digest is unkeyed, and the ledger is team content, so anyone who can write the team's files can add a row. That is no wider than what such a writer can already do: unfenced text that is not a twin is never reviewed, and deleting the twin also clears its REVIEW. The ledger adds a committed, reasoned record, and fleet keeps listing the twin.
• Changes are flagged for operator review. Because no row is tied to the operator, fleet reports an acknowledged twin as REVIEW until an earlier applied fleet run has shown that exact acknowledgement (see fleet: fence-twin acknowledgements are reviewed once). A row added or changed by anyone, committed or not, is flagged once and cannot clear REVIEW in the run that first applies it.
• A twin whose span cannot be bounded (for example, a heading that appears twice unfenced) cannot be acknowledged; resolve it by hand.

Name	Purpose
LEDGER_REL	references/fence-twin-acks.csv, relative to the output dir.
COLUMNS	The required header: file, heading, digest, reason.
MIN_REASON_CHARS	40.
DIGEST_HEX	16.
TwinAck(file, heading, digest, reason)	One validated row.

load(output_dir)

Reads the ledger under output_dir. Returns {file: (TwinAck, ...)}, which is empty when there is no output dir or no ledger. cli.render_pipeline.twin_sources calls it once per run and attaches each file's rows to its TwinSource.

parse(text, *, source='references/fence-twin-acks.csv')

Validates ledger text and groups the rows by file. Each skipped row prints a warning to stderr.

ledger_row(file, heading, digest)

Returns the CSV row to record for a twin, with a placeholder where the reason goes.