Nw Speculative Dispatch Agent Skill

Speculative Dispatch Methodology

Speculative dispatch is a technique where the orchestrator generates N candidate

implementations of the same TDD step in parallel, scores each by measured

properties, and picks the best — while logging every candidate, including

discarded ones, for human audit.

1. When to Use Speculative Dispatch

**Use speculative dispatch when the decision is ambiguous and at least 3 feasible

candidate strategies exist.**

| Signal | Use speculative dispatch? |

|--------|--------------------------|

| Step has ≥3 plausible implementation strategies | YES |

| Decision is clear from prior context | NO — implement directly |

| Only 1-2 strategies, clear winner | NO — implement directly |

| Performance matters and strategies have measurable trade-offs | YES |

| Domain is novel and "correct" structure is unknown | YES |

Speculative dispatch is orthogonal to TDD stages — it applies at any phase

(RED unit-test authoring, GREEN, COMMIT under the 3-phase canon per ADR-025;

or RED_UNIT, GREEN, COMMIT under the legacy 5-phase contract) where ambiguity

is genuine and 3 candidates are feasible within the time budget.

2. The Three-Candidate Rule

Generate exactly 3 candidates per speculative step by default:

| Role | Strategy |

|------|----------|

| minimal-change | Inline, no new abstractions. Lowest complexity, fewest lines. |

| refactor-heavy | Extracts helpers, adds guards, defensive patterns. Higher complexity. |

| pattern-extraction | Generalises to a factory/template. Highest lines. Justified if reuse follows. |

Three candidates is the minimum viable set for meaningful scoring. More than 5

candidates adds noise without proportional signal. Each candidate must be

behaviourally correct before scoring — incorrect candidates are eliminated before

scoring, not by scoring.

3. Auditability Mandate

ALL candidates must be logged — winner AND losers.

Discarded candidates have audit value. A human reviewer validating the

orchestrator's pick must be able to inspect every alternative considered, not

just the selected one. This is not optional.

What the audit log captures per candidate

candidate_id:     unique name within the step (e.g. "minimal-change")
step_id:          identifies the TDD step these candidates competed on
timestamp_iso:    ISO-8601 when the trace was written
files_modified:   tuple of relative paths modified by this candidate
tests_added:      tuple of test file paths added
tests_pass:       True / False — did the full test suite pass?
rationale:        human-readable explanation of this candidate's approach

Why losers matter

A loser trace explains:

• What the orchestrator tried and why it was discarded.
• Whether the discard was due to test failure (hard signal) or lower score (preference).
• Whether a future reviewer would make the same call.

Without loser traces, the audit log is a winner's narrative — it cannot

be challenged or validated.

4. Scoring Model

Candidates are ranked by a composite 5-tuple score. Tuple ordering implements

priority naturally: Python's built-in > comparison on tuples is sufficient.

score(metrics) -> (tests_pass: int, -complexity_delta, -lines_added, 0, -runtime)

Priority order (element 0 dominates):

| Priority | Metric | Direction |

|----------|--------|-----------|

| 1 | tests_pass | True (1) > False (0). Hard gate — a failing candidate never beats a passing one. |

| 2 | complexity_delta | Lower is better. Negated so higher score = simpler. |

| 3 | lines_added | Fewer is better. Negated. Tiebreaker when complexity is equal. |

| 4 | reserved | 0 — placeholder for future metrics (coverage delta, type-error count). |

| 5 | test_runtime_seconds | Faster is better. Negated. Final tiebreaker. |

Correctness gate: a candidate with tests_pass=False is ALWAYS dominated by

any candidate with tests_pass=True, regardless of all other metrics. This

prevents the orchestrator from ever selecting a broken candidate on the grounds

that it is "simpler".

5. pick_best Rationale Requirements

The pick_best function must return a rationale string that:

• Names the winner candidate_id.
• Names every loser candidate_id.
• States the winner's key metrics (tests_pass, complexity_delta, lines_added).
• States why each loser was discarded (test failure, or specific metric

comparison).

A rationale that omits any candidate is an audit violation. Reviewer agents

check that all candidate_ids appear in the rationale string.

6. Storage Layout

<root>/
  .nwave/
    speculative/
      <step_id>/
        traces.jsonl      # one JSONL line per candidate, in write order

• One file per step. All candidates for a step share the file.
• Append-only. Never overwrite. Each write_trace call appends one line.
• read_traces(step_id, root=root) returns all candidates for the step.
• Human-readable with jq or any JSONL viewer.

Inspection example

cat .nwave/speculative/ws-prepended-with/traces.jsonl | python -m json.tool

7. Stage Cascade Fit

Speculative dispatch is stage-agnostic. It applies at any TDD phase where a

genuine implementation choice exists:

| Stage (3-phase canon / legacy 5-phase) | Application |

|----------------------------------------|------------|

| RED (unit-test authoring) / RED_UNIT | Competing test decompositions (example-based vs property, flat vs parametrised). |

| GREEN | Competing implementations of a non-trivial function. |

| COMMIT | Competing refactor strategies (extract method vs extract module vs inline). |

Do not apply speculative dispatch to mechanical steps (adding an import,

renaming a variable, fixing a typo). Reserve it for decisions with measurable

trade-offs.

8. Integration with nwave_ai.speculative

Two modules provide the walking-skeleton implementation:

from nwave_ai.speculative.audit import CandidateTrace, write_trace, read_traces
from nwave_ai.speculative.score import CandidateMetrics, score, pick_best

Workflow

# 1. Build and run candidates (external to this module)
#    Each candidate modifies files, runs tests, records outcome.

# 2. Write a trace for each candidate — ALL of them, pass or fail
trace = CandidateTrace(
    candidate_id="minimal-change",
    step_id="step-42",
    timestamp_iso=datetime.utcnow().isoformat() + "Z",
    files_modified=("src/foo.py",),
    tests_added=("tests/test_foo.py",),
    tests_pass=True,
    rationale="Single-expression — no new abstractions.",
)
write_trace(trace, root=Path("."))

# 3. Build metrics for each candidate
metrics_map = {
    "minimal-change": CandidateMetrics(
        tests_pass=True, complexity_delta=1, lines_added=3, test_runtime_seconds=0.5
    ),
    ...
}

# 4. Recover all traces and pick best
traces = read_traces("step-42", root=Path("."))
winner, rationale = pick_best(traces, metrics_map)

# 5. Apply winner's changes; discard losers (but keep audit log)
print(f"Selected: {winner.candidate_id}")
print(f"Rationale: {rationale}")

9. Anti-Patterns

| Anti-pattern | Problem |

|--------------|---------|

| Logging only the winner | Destroys audit trail. Reviewers cannot validate the pick. |

| Scoring before correctness gate | A failing candidate may win on complexity. Never allowed. |

| Fewer than 3 candidates | Two candidates is a coin flip, not speculative dispatch. |

| More than 5 candidates | Noise overwhelms signal; budget exceeded. |

| Selecting by single metric | Composite score required; single-metric selection misses trade-offs. |

| Discarding traces after pick | Audit log is permanent. Delete only on explicit retention-policy trigger. |

10. Empirical Baseline (walking skeleton)

Walking skeleton verified 2026-05-05 (tests/speculative/test_walking_skeleton.py):

Three candidates implement prepended_with(value, prefix) -> bool:

| candidate_id | tests_pass | complexity_delta | lines_added | selected |

|---|---|---|---|---|

| minimal-change | True | 1 | 2 | YES |

| pattern-extraction | True | 4 | 10 | no |

| refactor-heavy | True | 5 | 12 | no |

Rationale produced:

Selected: minimal-change (tests_pass=True, complexity_delta=1, lines_added=2).
Discarded candidates:
  - refactor-heavy: tests_passed, complexity_delta=5, lines_added=12.
  - pattern-extraction: tests_passed, complexity_delta=4, lines_added=10.

All three candidate_ids appear in the rationale. Auditability mandate satisfied.