Research synthesis — #2007 (3.2.0 training bugs Robert witnessed)

Date: 2026-06-16. Squad: debugger-debbie (repro-on-HEAD) · planner-priti (ticket map + sequencing) · architect-alphonso (systemic) · python-pedro (command-drift + guard). All opus, profile-loaded. Inputs: debbie-repro-triage.md, priti-ticket-mapping.md, alphonso-systemic.md, pedro-command-drift.md.

Framing note: this synthesis names focus areas / milestone goals / follow-on missions, not patch version numbers — versioning is a PO/release decision superimposed later.

Headline (four-way convergence)

#2007's 16 field-observed bugs collapse into the exact two classes the epic named, and both are non-adoption of authorities that already exist — the same disease this whole effort targets:

  1. Command-contract drift — prompts/skills/docs reference CLI surfaces that don't exist or are internal-only.
  2. Read-path / error-fidelity non-adoption — commands bypass the typed mission-context resolver, or flatten its typed error into a generic one.

Critically (debbie + alphonso, code-verified): the single typed resolver already exists and is correct (resolve_mission_read_path / resolve_action_context → ExecutionContext, mission_runtime/resolution.py:682). It is #2007's prescribed "single typed authority", not a thing to build. Adoption demonstrably works (the already-fixed bugs prove it); the live bugs are call-sites validating against the wrong base, flattening the typed error, or skipping auto-select. Finish + adopt, not build.

Empirical triage (debbie, on current HEAD — not 3.2.0-as-shipped)

Status Bugs Note
REPRODUCES (still real) #1, #4, #5, #8, #12, #13, #14, #15, #16 (9) the actionable set
PARTIALLY-FIXED #2 (charter status side-effects) residual
ALREADY-FIXED → CORRECTED: NOT FIXED (operator override) #6, #7, #9, #10, #11 (5) The static "fixed" verdict is REJECTED. These reproduced in a real run; the #6 fix (#1944/#1965) is verifiably in v3.2.0 (ancestor of HEAD) yet #6 still fired ⇒ the fix is present but doesn't cover the real path. Carried OPEN; re-investigated for the missed path (incomplete-coverage / unwired / alternate-trigger). Close none.
NEEDS-LIVE-REPRO #3 (typed-state UX, not a hard defect)

Caveat: #7's check is fixed, but the auto-commit-not-firing I saw this session (setup-plan leaving plan.md uncommitted) is a separate trigger — worth a follow-up, not part of #7.

Top 3 still-real P0s: #15 next collapses ActionContextError(→STATUS_READ_PATH_NOT_FOUND) into MISSION_NOT_FOUND (runtime_bridge.py:3128-3134) — highest blast radius, next is the primary agent entrypoint and it lies about the failure; #8 decision open rejects valid coord-aware handles (decision.py:103-109 second escape-check authority); #4 setup-plan hard-requires --mission (agent/mission.py:1248-1250) — the one I hit this session.

Does #2007 change the current opener? — NO (priti + alphonso, unanimous)

  • The naming routing rider (current opener) is severable, in-flight, and its safety was panel-confirmed. #2007 neither blocks it nor flips the safety-vs-impact values call. It overlaps the rider only at the #1888 surface — and even there, bug #10 is a different finalize defect than the rider's phantom-path fix (and #10 is already fixed). Keep the opener.
  • #2007 does supply the missing real-world severity evidence for the read-side adoption grain — the panel's "safe entry inside the impact plan." But its loudest class (command-drift) is orthogonal to the write-side topology keystone and pulls toward the cheap-safe end. So #2007 hardens the case to do read-side adoption + the snippet guard as the next focus — it is not the data that says "lead with the write-side topology redesign." That stays its own later focus (alphonso: keep it out of #2007's core).

The two coherent focus areas #2007 resolves into

Focus A — Command-contract-drift guard (DevEx-enabler; bounded; low-risk; sibling of the rider's ratchet)

  • Repoint the 15 drifted SOURCE references (pedro): 11× doctrine list/show in spec-kitty-charter-doctrine/SKILL.md, 1× in spec-kitty-mission-system/SKILL.md, and 3 behavioral in software-dev/plan/prompt.md (context resolve missing required --action; setup-plan no-flag-first — the drift that bit me this session). (#9 import & #13 worktree repair are not in SOURCE.)
  • Add the command-snippet CI guard: a ~120–180 LOC generalization of the existing FR-018 test_docs_cli_reference_parity.py, reusing scripts.docs._typer_walker.walk(), in the existing docs-contract gate (no new CI job). Three finding codes: unregistered-path, unknown-flag, internal-as-public. Catches snippet drift, not behavioral drift.
  • Plus the simple literal repoints for the reproducing surface bugs (#1, #5, #13→doctor workspaces --fix, #16 contract decision).

Focus B — Read-path / error-fidelity adoption (the field-proven read-side of the impact surface; the panel's #1832-shape "safe entry")

  • Cheapest highest-leverage cut (alphonso): the typed-error pass-through — preserve ActionContextError.code + checked paths end-to-end, no reclassification — closes #12/#14/#15 with no resolver change.
  • Then the call-site adoptions: #4 (exact-one auto-select), #8 (resolve identity before path validation).
  • #2 charter status side-effect-free + one normalized hash (separable).
  • This is the centre of mass (priti: 7 of 16 bugs) and one surface at six call-sites, not six bugs.

Out of focus (its own later work): the write-side coordination/topology redesign (#1716) and the ExecutionContext builder-hardening (still mutable; branch_name ≠ branch_ref.target_branch lives inside the SSOT — resolution.py:793-801). #2007 does not pull these forward.

Tracker reality (priti)

  • Net-new: 14/16. Existing: 2 — #1890 (bug 13 coord-worktree repair, OPEN) and #1891 (bug 16 agent action implement --json, OPEN). Both unmilestoned.
  • 5 bugs already fixed on HEAD → close-with-evidence (claim-exempt per the standing rule).
  • The epic #2007 and its components are unmilestoned — the milestone inversion persists (the real P0/launch work isn't on the active milestone).
  • Launch-relevant: #6 (submodule root — already fixed) and #14 (read-path miss in the implement loop — still real, the read-side of launch-blocker #1716).

Recommendation (focus/sequencing — PO assigns versions later)

  1. Keep the naming routing rider as the opener (unchanged; don't balloon it).
  2. Fold Focus A (command-drift guard) in as a bounded sibling — it's an architectural-consistency guard exactly like the rider's ratchet, low-risk, and fixes the SOURCE drift that mis-steers agents (incl. the plan-prompt drift I hit). Or run it as a tiny parallel mission if we want the rider pure.
  3. Make Focus B (read-path/error-fidelity adoption) the next named milestone slice — field-proven, the panel's read-side lead, "finish + adopt" not build; lead it with the typed-error pass-through cut.
  4. Keep the write-side topology redesign (#1716) as a distinct later focus.
  5. Tracker hygiene now (CORRECTED — close NOTHING): the operator's standing override (2026-06-16) is that no #2007 bug is closed — a bug that surfaced in a real run is not "fixed" because the code looks fixed; assume we missed something. The "already-fixed 5" are re-investigated for the missed path and carried OPEN. Hygiene that IS safe: record #1890/#1891 as the two existing residuals; milestone the live #2007 work onto the active cycle (fix the inversion); open the net-new sub-issues along the two-class / six-call-site decomposition. No closures, no claims, until a live repro proves a fix.

Epistemic correction (operator, 2026-06-16) — supersedes the "already-fixed" rows above. Static code-reading cannot retire a bug witnessed in a live session. Verified: #6's fix is in v3.2.0 yet #6 reproduced ⇒ present-but-incomplete/unwired/alternate-path. Treat all 16 as live until a real run proves otherwise. See debbie-missed-path-reinvestigation.md.

Missed-path re-investigation outcome (debbie, 2026-06-16) — the "already-fixed 5" are OPEN

Version facts: HEAD's 22 commits past v3.2.0 are docs/planning only → HEAD behaves == v3.2.0 here. Robert ran v3.2.0rc45; the #7/#11 fix was already in rc45 → stale-binary ruled out; fixes present and still failed. Pinned root causes:

Bug Disposition What we missed (root cause)
#6 submodule root OPEN, reproduces on HEAD Second authority — fix patched locate_project_root, but the live assert_initialized guard calls resolve_canonical_root (core/paths.py:284-288), which on a submodule .git file walks UP into the parent repo. Upgrading would not fix it.
#7 spec_committed:false OPEN, reproduces on HEAD is_committed (missions/_substantive.py) checks coord-ref + coord-worktree HEAD only — no primary-branch leg; coord-priority feature_dir feeds it a coord path. Secondary: _commit_to_branch (mission.py:1178-1195) silently swallows commit failures → commit_created: None + untracked artifact (the auto-commit-not-firing observed this session).
#9 stale import guard OPEN Guard for stale specify_cli import snippets is absent (tool_surface lints CLI paths, not Python imports). Bad import not in SOURCE, but absence ≠ fixed → extends IC-06's guard.
#10 finalize exit 1 OPEN, needs live repro glob classifier is correct, but the validate_all overlap gate (mission.py:3333) never consults create_intent, and --validate-only create_intent is built from a possibly-stale frontmatter snapshot. Needs Robert-env --json.
#11 finalize coord surface OPEN, needs live repro Coord-read not reproduced on HEAD; surviving trigger = fail-closed require_exists=True pre-read (mission.py:2752 → _read_path_resolver.py:316-367) aborts on a materialized-but-empty coord worktree BEFORE the primary read. Needs Robert's exact mid8/coord topology.

Structural root: the "single resolver" consolidation is not behavior-equivalent across input classes — two root resolvers (#6), coord-only committedness (#7), fail-closed pre-read gating primary (#11). This is the concrete mandate for the read-path/error-fidelity adoption focus: make the resolvers behavior-equivalent, not just nominally "single." #6 and #7 are fixable now with pinned coordinates; #10/#11 need a live repro in Robert's monorepo/coord environment. Close none.

Edit this page