Coding agents overview

Compare maturity, setup, events, injection, capture, flush, permissions, and validation across five coding-agent hosts.

Validation depth differs by host. OpenCode has a real host-process prepare/capture E2E smoke, which is Supported at that boundary; its full host lifecycle is Not validated. Codex, Claude Code, and DSH real-host E2E are Not validated. Pi has package-load coverage, while its full recall host chain is Not validated.

All five hosts connect to one separately running PowerContext Server. They share a domain contract, not one Hook API or a unified HostAdapter SDK.

Maturity and installation

Host	Automatic path	Explicit path	Install/version status	Real-host validation
Codex	UserPromptSubmit	MCP + Skill	No formal minimum host version declared	Not validated
Claude Code	UserPromptSubmit	MCP + Skill	v1; Claude Code 2.1.133 is compatibility evidence, not a minimum	Not validated
DSH	agent/pre-step	Curated HTTP tools	Built lib/index.js required; formal minimum version Not validated	Not validated
OpenCode	Two-stage exact-turn cache and transform	Curated HTTP tools	Version-gated: >=1.18.21,<2	Prepare/capture smoke Supported; full lifecycle Not validated
Pi	before_agent_start + pending flusher	Native tools + /pc	Node >=22.19; Pi 0.82.x is compatibility evidence, not a minimum	Package load only; full chain Not validated

Not validated does not mean Unsupported. Not validated means the path may work but lacks the stated evidence. Unsupported means the integration does not provide the operation.

Setup installs host files or packages. It does not start the Server. doctor checks a bounded installation and activation boundary; it is not proof of a live prompt, recall, capture, permission UI, cross-session behavior, or real-host E2E.

When choosing a host, first ask whether you control a pre-prompt event and a transient injection channel, then examine capture and permission behavior. A “yes” in the matrix is not completion. Completion means a new session recalls from the same Scope, an ordinary task continues when the Server is down, and an explicit write reports that outage.

Minimal usage

powercontext setup <host> \
  --source oceanbase/powercontext \
  --ref master

powercontext server run
powercontext doctor <host>

Open a new host session and send one ordinary coding prompt. Use the explicit tool surface only when you intend to search or mutate Memory, create a Handoff, run Review, or export a Skill.

The shared sequence is useful for comparing failure behavior. A prepare failure should remove only the recalled contribution. A capture failure should not erase already validated transient context. Permission belongs to explicit operations and cannot be inferred from a successful automatic hook.

Common execution flow

host event
→ resolve Scope
→ prepare bounded untrusted context
→ optionally capture prompt Source / immediate flush
→ inject through host-specific channel
→ run model
→ Pi only: retry pending flush at lifecycle boundaries
→ explicit tools under host permission

That common shape contains five different event models:

Codex / Claude Code
UserPromptSubmit → prepare + prompt capture → transient context
explicit work → MCP + Skill

DSH
agent/pre-step → prepare + user-originated batch capture
→ await downstream decision → plugin message
explicit work → curated HTTP tools

OpenCode
chat.message → prepare + capture → cache by (sessionID,messageID)
experimental.chat.messages.transform → exact latest-turn match → inject
explicit mutation → context.ask

Pi
before_agent_start → prepare + prompt capture → immediate flush or pending position
→ augment systemPrompt
boundary events → retry pending flush within a bounded budget
explicit mutation → interactive confirmation; deny without UI

Capability matrix

Choose by lifecycle first, then use the matrix to confirm that the selected adapter has the exact behavior you need.

Capability	Codex	Claude Code	DSH	OpenCode	Pi
Automatic event	UserPromptSubmit	UserPromptSubmit	agent/pre-step	chat.message + message transform	before_agent_start
Injection	Hook stdout additional context	JSON additionalContext	Plugin message after downstream enter	(sessionID,messageID) exact-turn cache	Returned systemPrompt augmentation
Prompt capture	User prompt; character bound	User prompt; character bound	User-originated batch; secret-like filter	Eligible prompt; secret-like/byte filters	Eligible prompt; secret/byte filters
Final response capture	No	No	No	No	No
Explicit interface	MCP	MCP	Curated HTTP tools	Curated HTTP tools	Native tools + /pc
Handoff	MCP + Skill	MCP + Skill	Curated HTTP tools	Curated HTTP tools	Native tools
Permission	Host-gated	Host-gated	Pre-execute ask	context.ask	Interactive confirmation; deny without UI
Managed Skill export	Supported	Supported	Unsupported	Unsupported	Unsupported
Real-host E2E	Not validated	Not validated	Not validated	Prepare/capture smoke Supported; full lifecycle Not validated	Package load only; full chain Not validated

All five integrations bundle host guidance, but a bundled project-context Skill is not a managed Skill export. Managed export is a separate PowerContext operation with projection and drift contracts; current targets are Codex and Claude Code only.

Scope, injection, capture, and flush

Scope selects a data partition. It is not authentication. Derive it from a stable, authorized workspace or project identity, and test that two Scopes cannot reuse each other’s transient context.

PreparedContext is bounded, temporary, and untrusted. Put it only in the current-turn injection channel. Do not copy it into a transcript, checkpoint, or durable Source. A stale turn key must never inject context into a later turn.

Capture and Memory are separate stages:

prompt capture
→ Source position
→ optional flush / scheduler
→ extraction candidate
→ Memory commit or no-op

Codex and Claude Code capture one user prompt. DSH can use all batch text for recall but captures user-originated text only. DSH, OpenCode, and Pi skip secret-like capture; that filter does not prevent recall transmission. Pi and OpenCode enforce UTF-8 byte bounds, while Codex and Claude Code use a character-count capture bound. None captures the final assistant response automatically.

Flush only advances processing toward a captured Source position. A timeout or lagging cursor cannot prove that Memory exists. Codex, Claude Code, DSH, and OpenCode support optional immediate flush. Pi also records pending positions and retries them on lifecycle boundaries. Only Pi has that deferred boundary-flush model.

Automatic and explicit paths

Automatic recall, capture, and optional flush are auxiliary. Expected Server, validation, size, deadline, or capture failures normally fail open so the host task can continue. Claude Code, DSH, OpenCode, and Pi guard recall and capture independently. Codex keeps expected failures independent, but an unexpected recall exception can skip capture for that invocation; do not claim stronger isolation than the code provides.

Explicit durable operations are different. Memory writes, Handoff commit, Review decisions, and Skill export must expose failure and remain behind host permission. Pi’s no-UI denial is verified. Codex and Claude Code no-UI durable-mutation denial is Not validated; do not copy Pi’s contract onto them.

Negative experiments

Run the checks that apply to the selected host:

• Stop the Server and confirm an ordinary prompt still reaches the model.
• Return 401/403 and keep it distinct from an empty successful result.
• Return an unknown PreparedContext schema and confirm nothing is injected.
• Exceed the response-size or deadline budget and confirm the turn remains bounded.
• Redirect an authenticated request and claim refusal only where that host implementation enforces it.
• Make recall fail while capture succeeds, then reverse the failures; include an unexpected Codex recall exception.
• Reuse, reorder, or delete a turn/session key and confirm stale context is rejected or a race cannot inject the wrong turn.
• Deny an explicit mutation, including a headless/no-UI case where the host defines one.
• Keep the flush cursor behind the Source position and confirm bounded retries stop without claiming Memory exists.
• Inspect diagnostics and confirm they contain no prompt, Scope, PreparedContext, response body, or token.

Currently unsupported or not validated

Item	Status
Final assistant-response or transcript capture in these automatic paths	Unsupported
Automatic Task Outcome recording	Unsupported
Managed Skill export for DSH, OpenCode, or Pi	Unsupported
Codex real host-process E2E and formal minimum host version	Not validated
Claude Code real host-process E2E, formal minimum version, and no-UI mutation denial	Not validated
DSH real-host E2E and formal minimum host version	Not validated
OpenCode out-of-order prepare race behavior	Not validated; exact-turn matching prevents wrong-turn injection but may miss a newer injection
Pi full recall host chain, formal minimum Pi version, and all lifecycle-flush boundaries	Not validated
Managed exported Skill loading/execution in Codex or Claude Code	Not validated
One unified HostAdapter SDK	Unsupported as a current repository abstraction

Choose a host

Use the smallest host surface that completes the task. Broader integration adds more behavior to verify and more content that may cross the Server boundary.

Need	Read next
Short UserPromptSubmit, MCP, and project Skill	Codex or Claude Code
Pre-step middleware and curated HTTP tools	DSH
Two-stage exact-turn cache and context.ask	OpenCode
before_agent_start, pending flusher, and no-UI denial	Pi

Completion checklist

Treat the integration as complete only when these results are observable:

• The Server runs separately.
• Setup and doctor are not presented as E2E evidence.
• The host version or compatibility claim matches the maturity table.
• One ordinary prompt proves the intended event and transient injection path.
• Scope comes from stable authorized identity, with a cross-Scope negative test.
• Capture content, secret policy, and character/byte bounds are documented and observed.
• Flush and Memory commit are verified as separate stages.
• Automatic failures fail open only to the extent current code and tests prove.
• Permission and no-UI behavior are tested only for the selected host.
• A bundled Skill is not reported as managed Skill export.
• Unsupported and Not validated remain distinct in deployment notes.

If the real host has not launched and processed a prompt, report the completed contract, service-chain, or doctor tier. Do not call it real-host E2E.

Continue with Hooks and plugins before selecting a host-specific setup.