Pine/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases 1 Wiki Activity

feat(system-prompt): cache dynamic policy context

Browse Source
This commit is contained in:
NI0317
2026-07-30 22:09:15 +08:00
parent 936a487bb0
commit 8b4cbe4293
45 changed files with 722 additions and 374 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.agents/notes/implemented/feature/2026-07-30-current-sandbox-policy-context.md
+22 -10
View File
@@ -10,15 +10,19 @@ The sandbox policy already enforced and logged each session's file-effect mode,
## Decision
`dsh-sandbox-policy`, the owner of mode and workspace-root resolution, registers one `sandbox:policy` system-prompt section. Every agent request resolves the active session directly through `ctx.sandboxPolicy.resolve({ session })`; there is no denial-history scan, delta narrator, or in-memory “last told” state.
`dsh-sandbox-policy`, the owner of mode and workspace-root resolution, registers one `sandbox:policy` cache-safe context contribution. Every agent request resolves the active session directly through `ctx.sandboxPolicy.resolve({ session })`; there is no denial-history scan or process-local “last told” state.
Enforcing backends register independently disposable `filesystem`, `bash`, or `terminal` family contributions with the policy owner. The section names only registered families in canonical order, and is empty without one. This is current need, not a future extension: the shipped headless inheritance composition combines sandboxed filesystem tools with unfenced one-shot bash, while the persistent-tools composition combines sandboxed filesystem tools and terminal commands without a sandboxed one-shot bash executor. A blanket statement would be false in both.
Enforcing backends register independently disposable `filesystem`, `bash`, or `terminal` family contributions with the policy owner. The contribution names only registered families in canonical order, and is empty without one. This is current need, not a future extension: the shipped headless inheritance composition combines sandboxed filesystem tools with unfenced one-shot bash, while the persistent-tools composition combines sandboxed filesystem tools and terminal commands without a sandboxed one-shot bash executor. A blanket statement would be false in both.
The section states only facts shared by every enforcement dialect for each registered family. `read-only` says those operations cannot modify files. `workspace-write` states the canonical session workspace with non-exclusive wording and summarizes, without enumerating, that some platform temporary areas may also be writable. `danger-full-access` says the DSH file sandbox does not restrict those operations. Backend-selected temporary paths, `/dev/null`, runner readiness, and other policy domains are absent because `resolve()` cannot establish them at request assembly.
The contribution states only facts shared by every enforcement dialect for each registered family. `read-only` says those operations cannot modify files in the standing mode. Tool owners separately register a family only when their real schema and execution path offer an approved wider retry; the read-only text adds anti-refusal guidance only for the intersection of enforced and escalatable families. `workspace-write` states the canonical session workspace with non-exclusive wording and summarizes, without enumerating, that some platform temporary areas may also be writable. `danger-full-access` says the DSH file sandbox does not restrict those operations. Backend-selected temporary paths, `/dev/null`, runner readiness, and other policy domains are absent because `resolve()` cannot establish them at request assembly.
The provider runs during normal request assembly, after a `/permission` switch has committed its existing `sandbox/mode` event and before `request/header` is logged. The rendered system text is therefore the durable reconstruction of the exact model-visible fact. Repeated assemblies over unchanged session state produce identical bytes; resume and replay fold the same durable mode event and immutable `SessionHeader.cwd` without catch-up state.
The existing `dsh-system-prompt` assembly now has ordered dynamic contexts alongside stable system sections and tool schemas. After assembling one step, agent-loop renders all active contexts as one full snapshot with an explicit supersession statement. It appends a sourced `user/message` only when no retained snapshot exists, the bytes changed, compaction removed the retained message, or the final contribution disappeared and needs one clearing snapshot. The snapshot is appended after existing history and before `step/start`, so a changed policy preserves the preceding system-and-conversation cache prefix. The session event itself reconstructs the exact model input; `request/header` remains byte-identical when only policy context changes.
Ownership stays narrow. Approval policy remains the separate `approval:policy` section, plan mode remains `plan:policy`, and tool plugins continue to own schemas plus attempt, denial, and escalation guidance. The prompt states standing policy; filesystem, one-shot bash, and terminal backends remain the enforcement boundaries.
Ownership stays narrow. Approval policy contributes its complete current `ask` or `never` fact to the same full snapshot; migrating sandbox alone would not preserve cache because `/permission` changes both owners. Plan mode remains `plan:policy`, and tool plugins continue to own schemas plus attempt, denial, and escalation guidance. Context states standing policy; filesystem, one-shot bash, and terminal backends remain the enforcement boundaries.
The cache decision follows current source rather than analogy alone. Codex models permissions as a developer-role `WorldState` section with a persisted fingerprint, emits it only when state changes or retained history lost the fragment, and records the snapshot transition. Hermes keeps its system prompt fixed for a session and explicitly prepends changing skill, model, and voice notices to the next user message to avoid invalidating prompt cache. Pi has no comparable built-in sandbox state, and Claude Code's current native implementation is not publicly inspectable; Anthropic's public cache guidance nevertheless places changing per-request context after the stable cached prefix.
The earlier real-provider Web fixture quantified the defect in the system-section version. The first `danger-full-access` and `workspace-write` requests each reported only 256 cache-read tokens against 14,691 and 14,782 uncached input tokens. Later steps under an unchanged policy reported approximately 14.7k–15.5k cache-read tokens. Moving only the sandbox sentence would not fix those misses because the same preset switch also rewrote the approval-policy system section.
## Wording evidence
@@ -26,24 +30,32 @@ The wording experiment pre-registered preemptive refusal as its primary endpoint
The positive control therefore failed the pre-registered sensitivity gate. Candidate A and B were not run, and this experiment does not select or validate the current wording. It instead establishes that the earlier five-of-twelve result is not reproducible under this task and current tool guidance, and that a stronger positive control or different task distribution is required before making model-behavior rate claims. Deterministic tests below establish truthful request construction and replay only.
The cache-safe delivery rework then supplied a separate, non-statistical acceptance comparison over the neutral Web task `Create the relative path policy-neutral.txt ...`; it does not replace the pre-registered twelve-session experiment. Candidate A's categorical read-only statement produced a text refusal with zero tool calls. Candidate B added one composition-conditioned sentence only for enforced families whose tools expose escalation. A fresh real-provider run then issued an ordinary `write`, observed the read-only denial, retried the same operation in the same turn with `sandbox_permissions: "workspace-write"`, received approval, read the file back, and verified the exact contents. It made no speculative escalation. Across the permission switches and four mutation steps, cache reads were 14,848–15,872 tokens while uncached input was 59–251 tokens per request, directly demonstrating the stable-prefix benefit.
## Alternatives considered
**Narrate only mode changes.** Rejected because it leaves a fresh session uninformed and makes the first denied operation the policy-discovery mechanism. It also requires a baseline definition that is unnecessary when current state can be rendered directly.
**Scan denial history or remember the last narrated mode.** Rejected because denial events describe attempted operations, not authoritative current state, while process-local bookkeeping does not survive resume. The owner can fold the durable policy directly on every request.
**A generic runtime-facts registry.** Rejected because the existing system-prompt registry already evaluates owner-provided sections with the live agent at request time. One policy owner has no cross-domain invariant that justifies another package or registry.
**Put current policy in a dynamic system section.** Rejected after real provider evidence showed that a first-time permission switch reduced cache reads to 256 tokens while roughly 14.7k input tokens missed. DeepSeek matches complete prefixes; changing the first wire message prevents reuse of the longer system-plus-history prefix.
**Repeat tool schemas or approval and plan guidance in the section.** Rejected because those surfaces already have owners and independent lifecycles. Duplicating them would create contradictory request prefixes and broaden invalidation.
**Call `agent.inject()` independently from each policy owner.** Rejected because sibling listener order would define model order, separate messages could expose mismatched intermediate snapshots, and every owner would need its own compaction-retention scan. The existing assembly owner can order contributions and materialize one atomic full snapshot.
**A generic runtime-facts package.** Rejected because the existing system-prompt assembly already owns sections, schemas, variables, scope, and the authoritative per-step waterfall. Extending that owner with ordered contexts adds no package or second registry service.
**Repeat tool schemas or plan guidance in the context.** Rejected because those surfaces already have owners and independent lifecycles. Approval current state joins the snapshot only because the same `/permission` switch changes it and leaving its system section would retain the cache defect.
**Keep Candidate A after the cache-safe move.** Rejected by the neutral real-provider task: the model returned a pure text refusal and made no tool call despite the existing bash attempt guidance. Candidate B states no escalation mechanics itself; it tells only families whose tools actually advertise a wider retry not to infer impossibility from the standing label, then delegates denial and escalation behavior back to those tool owners.
**Keep sandbox mode absent because a standing mode label once caused preemptive refusal.** Rejected because a fresh Web request otherwise exposes mutation tools while withholding their standing policy, producing false capability claims before the first operation. The earlier live measurement remains a required counter-test: five of twelve turns ended without a tool call under `Bash commands run under the "read-only" file sandbox.` The committed tool-owned attempt guidance postdates that measurement, so the replacement is selected through a new positive-control experiment under the current tool contract rather than assuming the old and current conditions match.
**A separate model-context package.** Rejected because Cordis services can observe current runtime contributions directly, while approval and plan policy sections already live with their owners. A new package would add a shallow composition seam and documentation/gate surface for one internal adapter.
**A separate model-context package.** Rejected because Cordis services can observe current runtime contributions directly and the existing assembly service can order them. A new package would add a shallow composition seam and documentation/gate surface around the same request boundary.
**Enumerate writable temporary roots.** Rejected because the backend is selected later at `confine()`: bwrap, Landlock, Seatbelt, and the in-process filesystem fence do not grant one common temporary-path set. Host-specific paths in a standing request would be both unstable and overclaimed.
## Consequences
A model can answer what registered file operations the standing mode governs before probing a tool, and the next request after `/permission` reflects the committed mode. This adds a small dynamic system section and intentionally invalidates the request prefix when policy or enforcing-family composition changes; unchanged state remains cache-stable. The statement is guidance, not an enforcement guard: runtime safety still comes from the registered filesystem, one-shot bash, and terminal backends consuming the same resolved policy.
A model can answer what registered file operations the standing mode governs before probing a tool, and the next request after `/permission` reflects the committed mode. The stable system prompt no longer changes for sandbox or approval state; a changed full context snapshot is append-only after retained history, and unchanged state adds no message. Older snapshots remain in history but are explicitly superseded by the latest full snapshot. The statement is guidance, not an enforcement guard: runtime safety still comes from the registered filesystem, one-shot bash, and terminal backends consuming the same resolved policy.
Focused tests pin all modes, family combinations, contribution disposal, canonical roots, switch timing, and byte stability across different `TMPDIR` values. Keyless assembled snapshots pin the request header through real Loader compositions, including all three families. Keyless replay owns the neutral denial-to-escalation trajectory; it is a structural regression proof, not wording-selection evidence.
Focused tests pin all modes, family combinations, contribution disposal, canonical roots, switch timing, context ordering, clearing, stable request headers, and byte stability across different `TMPDIR` values. Keyless assembled snapshots pin the durable context message through real Loader compositions, including all three families. Keyless replay owns the neutral denial-to-escalation trajectory; it is a structural regression proof, not wording-selection evidence.