refactor: retire the legacy timeboxing agent, and the 34 modules only it reached

[hugolytics]

Problem

TimeboxingFlowAgent — 8,936 of its own lines, plus the 34 modules only it
reached — has been unreachable from any live entry point since
_timebox_backend() started answering harness unless
FF_TIMEBOX_BACKEND=legacy, a variable set nowhere outside
tests/conftest.py. It was still registered on the AutoGen runtime
unconditionally, still instantiated in the running bot with nothing routed
to it, and four live send_message call sites still addressed it directly
— each one a bare Exception("Recipient not found") waiting for the day the
class was deleted without those sends being closed first. This PR closes
those four sends, ports the one behaviour the harness genuinely lacked (a
refine-loop no-progress cap legacy had and the harness's uncapped
NeedsAnotherTurn did not), then deletes the class, the constraint-review
surface only it wrote to, and every module reachable only through it —
across thirteen commits, each task-reviewed, with the deletion landing only after the sends were closed and the cap was ported.

Measured numbers

• src/: 74 files changed, 442 insertions(+), 21,830 deletions(-) — 56 modules deleted
  (git diff --stat 31e88b5..HEAD -- src/)
• tests/: 100 files changed, 902 insertions(+), 15,295 deletions(-) — 77 test files deleted (67 in the retirement commit)
  (git diff --stat 31e88b5..HEAD -- tests/)
• Fast suite: 3,136 tests before → 2,630 after (one pre-existing
  environmental failure —
  test_the_deployed_profile_matches_the_repository — on both sides)
• Wall-clock: ~28s, against the 31.6s baseline (well inside the 10%
  budget)
• scripts/dev/tests/orphans.py: 0 newly orphaned modules
• scripts/dev/tests/dangling.py 42d9eb2..HEAD: 54 deleted modules, 0
  surviving src/ importers
• -m slow still collects 89 tests (untouched by this cut)
• The AST honesty guard landed in Task 8 with a measured, per-file
  allowlist: 34 files, 149 offences (24 __new__, 125 private writes) —
  this is the starting point Project 2 (test-suite composability) shrinks
  from, not the "few dozen" the original spec estimated

Also in the fix wave after the whole-branch review

Three constraint-review button ids from legacy thread roots had no handler once constraint_review.py went — the exact dead-button failure the retired-cards handler exists to end. Ten ids are retired now, not seven. docs/architecture/agents.md no longer calls the deleted class the primary planner.

Two spec deviations from docs/superpowers/specs/2026-09-08-test-suite-composability-design.md

1. mcp_clients.py scope. The spec said the whole file goes. Measured:
   only McpCalendarClient and CalendarDaySnapshot leave it —
   ConstraintMemoryClient stays, because it is the default
   tasks_defaults_memory_backend and has tests of its own that have
   nothing to do with the legacy agent.
2. settings.timeboxing_memory_backend. The spec said this setting and
   its validator go with agent.py. Measured: kept — it is also read by
   runtime.py's graphiti startup checks and by tasks' defaults memory;
   only agent.py's own branch on it was deleted.

(A third correction, not a deviation in behaviour: Project 2's "Injection —
none" section listed ConstraintMemoryClient as deleted in project 1 —
corrected to say it survives, and that its workbench= seam — measured
0.5s → 17.9s without it — is the one project 2 actually adds. The spec's
"allowlist shrinks to TBPlan" line now states the measured 34-file/
149-offence starting point it shrinks from.)

Coverage: three real regressions, named rather than explained away

Per-line coverage diff against PR #396 (covdiff.py, root-argument bug
fixed and re-verified in Task 7) found several surviving files that
lost covered lines a now-deleted test used to reach. Two of the original
five candidates are dead code, not regressions — their only consumer,
CalendarEventWorkerAgent (agents/schedular/agent.py:~901), is never
registered in runtime.py (git grep -n CalendarEventWorkerAgent returns
only its own definition), so nothing live reaches these lines at all. Most
of durable_constraint_store.py's lost lines are the same shape — callers
nothing calls. What's left is three real, live-code-now-untested lines:

• agents/timeboxing/durable_constraint_store.py:506 — the
  if not rows and search_names re-query, an ordinary-day path (not an
  edge case): the rest of the file's lost lines (326-329, 334-336, 357,
  539, 561) are get_store_info and callable(...) fallbacks the live
  client never takes — dead code, not a regression
• debug/diag.py:46-48 — the except Exception arm of with_timeout,
  imported by four live agents; the surviving test only covers the
  TimeoutError arm
• contracts.py:51 — the one ordinary-day loss: an all-day event arrives
  as {"date": ...} rather than {"dateTime": ...}, and haunt/reconcile.py
  imports contracts for exactly this parse

durable_constraint_store.py:506 and contracts.py:51 are ticketed —
tickets/recover_two_lines_the_retirement_left_untested.md — rather than
fixed in this pass; debug/diag.py:46-48 is an accepted known gap.

Dead-code cleanup candidates the retirement exposed (not fixed this round)

Named but not removed — a source change, out of scope for a docs/test-
retirement PR:

• llm/factory.py:151-158 — the calendar_submitter and timebox_patcher
  branches; their only users (submitter.py, patching.py) are deleted
• tmbx/journal/store.py:31-34 — journal_sessionmaker; its only caller
  was the deleted agent.py:339
• agents/schedular/models/calendar.py:185, 246, 351-352, 367-371, 392-402, 408-411 and agents/schedular/models/core.py:143 — _parse_event_type,
  _parse_time, the _check_time_combo validator, the colorId field:
  CalendarEventWorkerAgent, their only consumer, is never registered in
  runtime.py
• agents/timeboxing/durable_constraint_store.py:325-329, 334-336, 357 —
  get_store_info (zero callers) and the callable(...) fallbacks the
  live client never takes

(Three more class-(a) findings from the same coverage diff —
adapters/calendar/models.py, tool_result_models.py,
constraint_review.py — were already deleted in this branch, Tasks 6–7.)

Follow-up tickets filed

• tickets/harness_bounded_brief_says_what_it_withheld.md — new. The
  harness puts every applicable constraint row into the brief today (nothing
  withheld), but the day the brief is bounded, nothing makes the truncation
  speak — legacy rediscovered that rule three times (3dea6ae, e0c1f30,
  Journal: ACCEPTED is the fallback, so an unattended auto-commit becomes a fake approval label #177).
• tickets/test_suite_docs_after_legacy_retirement.md — new, filed and
  dispatched to a sonnet subagent per CLAUDE.md; its commit is folded
  into this branch. Brought tests/README.md's "What's out" section, the
  two superseded root calendar docs, the timeboxing agent's README.md +
  AGENTS.md, core/README.md, and slack_bot/README.md back in line
  with what the retirement actually deleted.
• tickets/skeleton_pre_generation.md — status line updated (not new): the
  confirm/undo wiring it describes was retired with the legacy agent on
  this branch; Undo lives on ff_harness_undo / handlers.act_harness_undo
  now.
• tickets/recover_two_lines_the_retirement_left_untested.md — new. The
  two real coverage regressions the corrected classification leaves
  (durable_constraint_store.py:506, contracts.py:51), each with the
  one test that would cover it. Not blocking.

The live-drive checklist — for Hugo, not for an agent

This is the one gate the suite cannot stand in for: commit dda88f4
records every unit test passing while the live bot went to legacy. The
how-to is in the memory note "Driving a live Slack timeboxing session".

• /timebox in the timeboxing channel opens a session root with the
  working card as first reply
• A plain "plan tomorrow" in a normal channel: the receptionist hands
  off, and a session root appears in the timeboxing channel (or where I
  typed, if none is configured) — no :warning: Exception: Recipient not found
• A second message in the open DM session continues it in the same
  thread
• Press Undo on a committed candidate: the calendar write reverses and
  the card says so
• Press a button on an old legacy stage card in history: it rewrites to
  the retired-flow sentence

🤖 Generated with Claude Code