The method

Agentic coding rollouts often fail not because the solver makes no progress, but because it stops at the wrong point: it repairs the most visible symptom and leaves a secondary requirement unresolved, submits after narrow validation, or keeps editing after reaching a correct solution and undoes it. A privileged advisor intervenes at exactly those moments. It never edits the repository or acts on the task itself.

Review gating

The advisor does not review every turn. With \(n_t\) interventions used so far, a budget of \(B\) interventions and a review interval \(k\), the gate at solver turn \(t\) is

\[
g_t = \mathbf{1}\left[n_t \lt B\right] \cdot \mathbf{1}\left[\mathrm{Finish}(\tilde{y}_t) \lor (t \bmod k = 0)\right]
\]

The advisor reviews the draft \(\tilde{y}_t\) when interventions remain and either the draft calls the finish tool or the turn is a multiple of \(k\). When periodic_review_interval is None, the periodic term is dropped and only proposed finishes are reviewed.

Decision

When the gate is open, the advisor produces a decision

\[
d_t = A_\phi\left(x, z, h_t, \tilde{y}_t, p_{t-1}, H_{\lt t}\right)
\]

from the public task \(x\), the privileged material \(z\), the bounded public history \(h_t\), the draft, its own previous plan, and the cues already delivered. The decision is a strictly validated typed object; see the decision schema. The advisor judges functional correctness, not similarity to the reference: a correct implementation that looks nothing like the reference is correct.

Selective intervention

The committed response \(y_t\) is the draft \(\tilde{y}_t\) when the gate is closed or the action is CONTINUE. Otherwise it is a revision

\[
\hat{y}_t \sim \pi_\theta\left(\cdot \mid x, h_t, I(d_t)\right)
\]

On an intervention the solver regenerates its turn with the public history plus a solver-safe cue \(I(d_t)\). The finish tool is withheld from the revision unless the action is CONFIRM_AND_SUBMIT, so the solver must take one real action before it can finish.

Statefulness

The advisor is stateful across reviews of one attempt. It carries its own plan \(p_t\), a short to-do list for the task and where the solver stands against it, so its sense of progress is continuous. It also sees the cues \(H_{\lt t}\) it has already delivered, so a repeat intervention can change the angle instead of restating an argument the solver has already worked through.

Notation and code

Symbol	Meaning	In the code
\(x\)	public task specification	TaskMaterial.instruction
\(z\)	privileged material: reference solution, hidden tests, private notes	TaskMaterial.files
\(h_t\)	public solver history at turn \(t\)	the messages passed to Advisor.review
\(\tilde{y}_t\)	the solver's draft response	the draft passed to Advisor.review
\(t\)	solver turn	Advisor.solver_calls
\(B\)	per-attempt intervention budget	AdvisorConfig.max_interventions
\(n_t\)	interventions used so far	Advisor.interventions
\(k\)	periodic review interval	AdvisorConfig.periodic_review_interval
\(g_t\)	review gate	the first lines of Advisor.review
\(d_t\)	the advisor's decision	AdvisorDecision
\(p_t\)	the advisor's private plan	AdvisorDecision.plan, carried as Advisor.plan
\(H_{\lt t}\)	cues already delivered	Advisor.delivered_cues
\(I(d_t)\)	the solver-facing cue	CompiledCue, from SelfCheckCue or ReasoningPrefillCue
\(\hat{y}_t\)	the solver's revision	what regenerate(Revision) returns
\(y_t\)	the committed response	what Advisor.review returns

One review

sequenceDiagram
    participant Agent as Agent loop
    participant Rev as Advisor.review
    participant Solver as Solver model
    participant Adv as Advisor model
    Agent->>Solver: public history and tools
    Solver-->>Agent: draft
    Agent->>Rev: review(history, draft, regenerate)
    alt gate closed: no finish, not a k-th turn, or budget spent
        Rev-->>Agent: the draft
    else gate open
        Rev->>Adv: task, bounded history, draft, plan, delivered cues, privileged material
        Adv-->>Rev: typed decision
        alt CONTINUE
            Rev-->>Agent: the draft
        else intervene
            Rev->>Solver: history plus cue (finish tool withheld unless CONFIRM_AND_SUBMIT)
            Solver-->>Rev: revision
            Rev-->>Agent: the revision
        end
    end