revert: remove hooks-dedupe module, batching guidance block, git-ops idempotency section (evidence refuted)

[Brian Krabach (bkrabach)]

Summary

A payload-level investigation found that three recent changes were built on a measurement artifact: analyses counted the same events 2-4x across duplicate snapshot copies of session files — proven by identical tool_call_id + nanosecond timestamps across the "duplicates." Deduped, the phenomena each change was built to address vanish. This PR reverts all three on that evidence. Not merging — opened for review.

---

Revert 1 — remove hooks-dedupe module entirely

Added as hooks-tool-dedupe in #323, renamed to hooks-dedupe in #326.

• Deleted modules/hooks-dedupe/ (whole directory).
• Removed its hooks: block entry (and the now-empty hooks: key) from behaviors/agents.yaml.
• Repo-wide grep for hooks-dedupe / hooks_tool_dedupe / tool_dedupe / DedupeHook confirms no dangling functional references remain.

Rationale: the module coalesced "same-batch duplicate reads" that do not exist — deduped count is 0 across 314 sessions / 10,320 reads; measured real value 19k tokens ($0.003) across 5 runs vs. the claimed 16.4% — a ~4,800x overstatement.

Kept: tests/test_hook_module_classification.py (added by #326). It guards a real amplifier-core loader fragility — _guess_from_naming()'s first-match keyword scan can misclassify a hooks-* module as type=tool — which is independent of whether hooks-dedupe itself exists. Removed only the two hooks-dedupe-specific pieces: the hardcoded assert "hooks-dedupe" in HOOK_MODULE_IDS, and the test_hooks_dedupe_passes_real_validation end-to-end test (plus its now-unused ModuleValidationError import). The general, dynamically-parametrized test_hook_module_classifies_as_hook test is untouched and still exercises every remaining hooks-* module.

Revert 2 — remove PR #320's batching guidance block

In modules/tool-delegate/amplifier_module_tool_delegate/__init__.py, removed the "BATCH YOUR DELEGATIONS..." block (through the "...it was independent" bullet) from the delegate tool's rendered description, and restored the original line it replaced: - Launch multiple agents concurrently when tasks are independent, at the top of "Agent usage notes:".

• PR feat(delegate): plan-then-dispatch guidance — batch independent delegations in one turn #320's two context-file edits (context/agents/multi-agent-patterns.md, context/agents/delegation-instructions.md) are untouched — out of scope for this revert.
• PR chore(context): remove legacy scaffolding shown to cause over-delegation (-27% tokens, quality preserved) #327's own deletions inside this same string (the CRITICAL/ALWAYS/NEVER preamble and the "DEFAULT TO DELEGATION" line) are untouched — those deletions stay.

Rationale: measured inert in a clean same-commit 5v5 A/B (treatment waves median 4 vs. control median 3 — the earlier apparent win was a version confound), while costing ~175 tokens on every request.

Revert 3 — remove PR #317's "## Idempotency Discipline" section

Removed the section from agents/git-ops.md (between "Git Safety Protocol" and "Common Git Commands").

Rationale: its evidence ("git-ops re-ran identical clone x8 / ls-remote x7 under context truncation") is refuted — deduped session data shows 8 clones of 8 different repos, once each, with per-repo ls-remote calls. The thrash never happened.

---

What is deliberately KEPT

• tests/test_hook_module_classification.py's general loader-classification regression test (test_hook_module_classifies_as_hook), and its historical docstring/assert-message narrative of the original hooks-tool-dedupe naming bug.
• PR feat(delegate): plan-then-dispatch guidance — batch independent delegations in one turn #320's two context-file sections (multi-agent-patterns.md, delegation-instructions.md).
• All of PR chore(context): remove legacy scaffolding shown to cause over-delegation (-27% tokens, quality preserved) #327's deletions (the "6d6 deletion set"), including inside the same tool-delegate string this PR also edits.

---

Verification

• Tests: uv run pytest tests/ -q → 1640 passed, 1 skipped (down from 1642 by exactly the two hooks-dedupe-specific instances removed above: the deleted async test, and one fewer parametrize instance of test_hook_module_classifies_as_hook now that hooks-dedupe no longer exists on disk).
• Lint/types: python_check clean on all touched files — only pre-existing, unrelated ruff warnings remain elsewhere in tool-delegate/__init__.py (none introduced by this change).
• Dangling references: repo-wide grep confirms modules/hooks-dedupe/ is gone with no dangling functional references. The 3 remaining textual mentions of hooks-tool-dedupe/hooks-dedupe are historical narrative inside the kept regression test's module docstring and assert message, explaining why that general test exists.

🤖 Generated with Amplifier

[Brian Krabach (bkrabach)]

Evidence-hygiene revert — documenting comment

A payload-level investigation into the claims underlying the three items this PR removes (hooks-dedupe module, the batching guidance block, and the git-ops idempotency section) found that the underlying measurements were artifacts, not real phenomena:

• Duplicate session event files on disk: session event files existed 2-4x on disk as snapshot copies. The original analyses counted the same events multiple times as a result — confirmed by identical tool_call_id + nanosecond timestamps across the "duplicate" rows.
• Deduped re-analysis: once duplicate snapshot copies are removed, the phenomena the original PRs were built on disappear — 0 same-batch duplicate reads across 314 sessions.
• git-ops "clone x8": re-examined at the payload level, this was 8 different repos, each cloned once — legitimate, non-duplicative work, not context-truncation thrash.
• Batching guidance block: measured inert in a clean, same-commit 5v5 A/B test — no measurable effect on delegation/batching behavior.

Given all three premises are refuted by the underlying data, this PR reverts the corresponding additions:

• removes the hooks-dedupe module
• removes the batching guidance block
• removes the git-ops idempotency section

Correcting comments are being posted on the originating PRs (#317, #320, #323) linking back to this revert and the evidence above.

---

Merge process note: this PR is self-authored; self-approval is not possible on this repo, so per the established maintainer admin-merge pattern (documented precedent: ~25 PRs this week), this is being merged via explicit user direction with admin override of the review-required gate. CI is green (6-matrix + license/cla, all passing).