fix(session): fire system.transform before messages.transform by ualtinok · Pull Request #19961 · anomalyco/opencode · GitHub

Issue for this PR

Closes #19960

Type of change

Bug fix
New feature
Refactor / code improvement
Documentation

What does this PR do?

Reorders the plugin hook execution so experimental.chat.system.transform fires before experimental.chat.messages.transform. Previously, messages.transform ran first in prompt.ts while system.transform ran later inside llm.ts — plugins couldn't inspect the final system prompt when mutating messages.

The system prompt assembly (agent prompt + env + skills + instructions + user.system) and the system.transform hook + header caching logic were extracted from llm.ts into shared helpers (LLM.buildSystem, LLM.transformSystem). These are now called in prompt.ts before messages.transform, and llm.ts receives the already-finalized system prompt.

Also adds sessionID and model to the messages.transform input (previously empty {}), so plugins can correlate transformations across hooks.

agent.ts's independent system.transform call (for agent generation) is unchanged — different code path.

How did you verify your code works?

Typecheck passes clean
Full test suite: 1625 pass, 0 fail, 8 skip (identical to dev baseline)
Verified LLM.buildSystem produces the same system prompt shape as the old inline assembly in llm.ts
Verified LLM.transformSystem preserves the header caching rejoin behavior

Checklist

I have tested my changes locally
I have not included unrelated changes in this PR

/review

This looks like a lot more changes than simply changing the orders just at a glance

Suggestion for the human to consider: this summary failure path persists msg.error, but it does not set msg.finish. Since failed compaction summaries are later returned as the last assistant message, a subsequent loop can see an errored summary with no finish and attempt the compaction path again instead of treating the turn as complete. Consider setting msg.finish = "error" before updateMessage, matching the existing compact-result error path below.

Suggestion for the human to consider: this new pre-stream failure path records msg.error, but it leaves msg.finish unset. The prompt loop only exits early when the latest assistant has a non-tool finish, so a later loop() call on the same session can treat this errored assistant as unfinished and retry the same user turn. This probably should mark the assistant as finished, consistent with other error paths that use finish = "error".

This looks like a lot more changes than simply changing the orders just at a glance

I will explain the reasoning when I get back home.

Hey @rekram1-node, I made some changes after your first review. Let me first explain why it looks more changes than needed:

The original plan was just reordering the 2 hooks but as LLM.stream contract has direct callers, we needed to preserve backwards compatibility to make the least changes. That was wrong call, we now removed the backward compatibility and all callers including compaction route uses the same contract.

I've been testing this PR on my local since the implementation and there's no problems observed. If you need longer explanation of the surface, I can give more detailed answer. I believe this PR is ready to ship.

Can we have this reviewed? thanks

…malyco#19961/anomalyco#31867 bundle merge

moved system assembly to caller

Keep the volatile current date out of the cached system prefix for DeepSeek models (both the V2 runner and V1 path) and instead append it to the trailing user turn, add SystemContext.omit, and rework DeepSeek reasoning handling so reasoning_content is only replayed across assistant tool turns (and caching hints are skipped) — maximizing DeepSeek prompt-prefix cache reuse. Ported from anomalyco#31867 (adapted for the Effect layer; the prepare() date test passes an explicit system prompt since anomalyco#19961 moved system assembly to callers). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add an experimental subagent-interrupt capability (behind OPENCODE_EXPERIMENTAL_SUBAGENT_INTERRUPT): an Interrupt service with pending/terminal state, task_steer/task_cancel/task_abort tools (gated by a new `interrupt` permission), a POST /session/:id/interrupt endpoint, and a turn-boundary consume in the prompt loop that injects an escaped steer/cancel frame plus a visible transcript marker. Cancel records a terminal reason and force-breaks after a grace window; abort writes a marker, records the terminal, and cancels the BackgroundJob immediately. The TUI gains an esc-with-reason flow, footer button, and distinct interrupt markers. Ported from anomalyco#32425 (adapted for the Effect layer; merged with the prior task.ts/registry.ts/prompt.ts changes from anomalyco#29447/anomalyco#32122/anomalyco#12520/anomalyco#32192/anomalyco#19961). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Keep the volatile current date out of the cached system prefix for DeepSeek models (both the V2 runner and V1 path) and instead append it to the trailing user turn, add SystemContext.omit, and rework DeepSeek reasoning handling so reasoning_content is only replayed across assistant tool turns (and caching hints are skipped) — maximizing DeepSeek prompt-prefix cache reuse. Ported from anomalyco#31867 (adapted for the Effect layer; the prepare() date test passes an explicit system prompt since anomalyco#19961 moved system assembly to callers). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add an experimental subagent-interrupt capability (behind OPENCODE_EXPERIMENTAL_SUBAGENT_INTERRUPT): an Interrupt service with pending/terminal state, task_steer/task_cancel/task_abort tools (gated by a new `interrupt` permission), a POST /session/:id/interrupt endpoint, and a turn-boundary consume in the prompt loop that injects an escaped steer/cancel frame plus a visible transcript marker. Cancel records a terminal reason and force-breaks after a grace window; abort writes a marker, records the terminal, and cancels the BackgroundJob immediately. The TUI gains an esc-with-reason flow, footer button, and distinct interrupt markers. Ported from anomalyco#32425 (adapted for the Effect layer; merged with the prior task.ts/registry.ts/prompt.ts changes from anomalyco#29447/anomalyco#32122/anomalyco#12520/anomalyco#32192/anomalyco#19961). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Keep the volatile current date out of the cached system prefix for DeepSeek models (both the V2 runner and V1 path) and instead append it to the trailing user turn, add SystemContext.omit, and rework DeepSeek reasoning handling so reasoning_content is only replayed across assistant tool turns (and caching hints are skipped) — maximizing DeepSeek prompt-prefix cache reuse. Ported from anomalyco#31867 (adapted for the Effect layer; the prepare() date test passes an explicit system prompt since anomalyco#19961 moved system assembly to callers). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add an experimental subagent-interrupt capability (behind OPENCODE_EXPERIMENTAL_SUBAGENT_INTERRUPT): an Interrupt service with pending/terminal state, task_steer/task_cancel/task_abort tools (gated by a new `interrupt` permission), a POST /session/:id/interrupt endpoint, and a turn-boundary consume in the prompt loop that injects an escaped steer/cancel frame plus a visible transcript marker. Cancel records a terminal reason and force-breaks after a grace window; abort writes a marker, records the terminal, and cancels the BackgroundJob immediately. The TUI gains an esc-with-reason flow, footer button, and distinct interrupt markers. Ported from anomalyco#32425 (adapted for the Effect layer; merged with the prior task.ts/registry.ts/prompt.ts changes from anomalyco#29447/anomalyco#32122/anomalyco#12520/anomalyco#32192/anomalyco#19961). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

PR anomalyco#19961 moves system assembly out of LLMRequestPrep.prepare into the callers, so a caller passing raw parts now loses the agent prompt. This fixture predates the PR's base, so it was not updated with the others. The recorded fixtures are unchanged and still assert the agent prompt is present in the request body.

Keep the volatile current date out of the cached system prefix for DeepSeek models (both the V2 runner and V1 path) and instead append it to the trailing user turn, add SystemContext.omit, and rework DeepSeek reasoning handling so reasoning_content is only replayed across assistant tool turns (and caching hints are skipped) — maximizing DeepSeek prompt-prefix cache reuse. Ported from anomalyco#31867 (adapted for the Effect layer; the prepare() date test passes an explicit system prompt since anomalyco#19961 moved system assembly to callers). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add an experimental subagent-interrupt capability (behind OPENCODE_EXPERIMENTAL_SUBAGENT_INTERRUPT): an Interrupt service with pending/terminal state, task_steer/task_cancel/task_abort tools (gated by a new `interrupt` permission), a POST /session/:id/interrupt endpoint, and a turn-boundary consume in the prompt loop that injects an escaped steer/cancel frame plus a visible transcript marker. Cancel records a terminal reason and force-breaks after a grace window; abort writes a marker, records the terminal, and cancels the BackgroundJob immediately. The TUI gains an esc-with-reason flow, footer button, and distinct interrupt markers. Ported from anomalyco#32425 (adapted for the Effect layer; merged with the prior task.ts/registry.ts/prompt.ts changes from anomalyco#29447/anomalyco#32122/anomalyco#12520/anomalyco#32192/anomalyco#19961). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

PR anomalyco#19961 moves system assembly out of LLMRequestPrep.prepare into the callers, so a caller passing raw parts now loses the agent prompt. This fixture predates the PR's base, so it was not updated with the others. The recorded fixtures are unchanged and still assert the agent prompt is present in the request body.

…+ anomalyco#42045)

PR anomalyco#19961 moves system assembly out of LLMRequestPrep.prepare into the callers, so a caller passing raw parts now loses the agent prompt. This fixture predates the PR's base, so it was not updated with the others. The recorded fixtures are unchanged and still assert the agent prompt is present in the request body.

…quest prep The Aug-15 upstream merge dropped both firings of this hook: ualtinok's prompt.ts-side trigger from fcb023e (port of anomalyco#19961) conflicted away, and upstream's request.ts-side block never landed in the resolution. Every plugin registering the hook (magic-context guidance injection + prompt-hash cache-flush coordination, aft-opencode hints) has been silently inert since. Restore upstream v1.18.18 semantics: fire after system assembly with {sessionID, model}, then rejoin a mutated multi-part system into header+rest so prompt-prefix caching stays stable. Regression test proves prepare() now delivers plugin mutations into prepared.system.

PR anomalyco#19961 moves system assembly out of LLMRequestPrep.prepare into the callers, so a caller passing raw parts now loses the agent prompt. This fixture predates the PR's base, so it was not updated with the others. The recorded fixtures are unchanged and still assert the agent prompt is present in the request body.

Upstream PR anomalyco#19961 by @ualtinok, ported onto current dev. The design and the original implementation are theirs; this commit adapts them to dev's current session and runtime structure and updates the recorded-fixture caller that the reordering affects. Co-authored-by: ualtinok <ufukaltinok@gmail.com>

Carries five third-party upstream PRs, one commit each, with original authorship and Co-authored-by trailers preserved: anomalyco#32203 @ualtinok stabilize duplicate skill discovery anomalyco#31859 @ualtinok reject bootstrap client reentry anomalyco#19961 @ualtinok fire system transform before messages transform anomalyco#39930 @brauliobo bound compacted history hydration anomalyco#31867 @ChangedenCZD improve deepseek prompt cache reuse Reconciles anomalyco#19961's hook reordering with the fork's sparse-context and instruction-audience handling: sparse still branches before skills/MCP are built, and modelMsgs is computed after messages.transform so plugin mutations are picked up.

github-actions Bot added the contributor label Mar 29, 2026

github-actions Bot mentioned this pull request Mar 30, 2026

fix(opencode): keep dynamic user.system separate for prompt caching #20109

Closed

6 tasks

ualtinok force-pushed the fix/plugin-hook-ordering branch 3 times, most recently from 21ca0d6 to 9e0a1f2 Compare April 3, 2026 22:10

opencode-agent Bot force-pushed the dev branch from 2f2ed11 to 9f201d6 Compare April 16, 2026 21:55

ualtinok force-pushed the fix/plugin-hook-ordering branch from 9e0a1f2 to f99f9a6 Compare April 17, 2026 12:51

github-actions Bot reviewed May 15, 2026

View reviewed changes

Comment thread packages/opencode/src/session/prompt.ts Outdated

github-actions Bot reviewed May 15, 2026

View reviewed changes

Comment thread packages/opencode/src/session/llm.ts Outdated

github-actions Bot reviewed May 16, 2026

View reviewed changes

ualtinok marked this pull request as draft May 16, 2026 12:54

github-actions Bot mentioned this pull request May 18, 2026

feat(plugin): add experimental.message.store.before hook #28224

Closed

6 tasks

ualtinok marked this pull request as ready for review May 19, 2026 05:24

ualtinok force-pushed the fix/plugin-hook-ordering branch 5 times, most recently from 8c371a9 to 1cb9c3e Compare June 6, 2026 09:45

ualtinok force-pushed the fix/plugin-hook-ordering branch 3 times, most recently from c97494d to 01faef0 Compare June 14, 2026 05:03

iceteaSA added a commit to iceteaSA/opencode that referenced this pull request Jun 17, 2026

fix(test): cast makeHttp layer + provider-options assertion after ano…

d479d83

…malyco#19961/anomalyco#31867 bundle merge

iceteaSA added a commit to iceteaSA/opencode that referenced this pull request Jun 17, 2026

fix(test): pass system input to DeepSeek date test after anomalyco#19961

9bbcd60

moved system assembly to caller

ualtinok force-pushed the fix/plugin-hook-ordering branch from 16d3672 to 26b60dc Compare June 28, 2026 10:23

ualtinok force-pushed the fix/plugin-hook-ordering branch from 26b60dc to 6e7412f Compare July 2, 2026 08:15

This was referenced Jul 6, 2026

📊 AI CLI 工具社区动态日报 2026-07-06 loxehate/loxehate.github.io#5

Closed

📊 AI CLI 工具社区动态日报 2026-07-06 loxehate/loxehate.github.io#8

Closed

ualtinok force-pushed the fix/plugin-hook-ordering branch from db13429 to 43afdf2 Compare July 7, 2026 08:56

ualtinok added 2 commits July 15, 2026 09:08

fix(session): fire system transform before messages transform

5d80969

fix(session): import v2 event helpers

5207442

ualtinok force-pushed the fix/plugin-hook-ordering branch from 43afdf2 to 5207442 Compare July 15, 2026 07:18

iceteaSA added a commit to iceteaSA/opencode that referenced this pull request Aug 15, 2026

merge: v1.18.18 into icetea_full (reconcile compaction anomalyco#19961 …

ce3bece

…+ anomalyco#42045)

github-actions Bot mentioned this pull request Aug 15, 2026

fix(opencode): coalesce system messages for OpenAI-compatible providers #42801

Open

3 tasks

                     })
+                    const finalizeFailedSummary = (error: unknown) =>
+                      Effect.gen(function* () {
+                        msg.error = MessageV2.fromError(error, { providerID: model.providerID })

fix(session): fire system.transform before messages.transform by ualtinok · Pull Request #19961 · anomalyco/opencode · GitHub

Uh oh!

fix(session): fire system.transform before messages.transform - #19961

fix(session): fire system.transform before messages.transform#19961
ualtinok wants to merge 2 commits into
anomalyco:devfrom
ualtinok:fix/plugin-hook-ordering

ualtinok commented Mar 29, 2026

Uh oh!

rekram1-node commented May 15, 2026

Uh oh!

Uh oh!

Uh oh!

rekram1-node commented May 16, 2026

Uh oh!

rekram1-node commented May 16, 2026

Uh oh!

github-actions Bot May 16, 2026

Uh oh!

github-actions Bot May 16, 2026

Uh oh!

ualtinok commented May 16, 2026

Uh oh!

ualtinok commented May 19, 2026

Uh oh!

elmagow commented May 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

fix(session): fire system.transform before messages.transform by ualtinok · Pull Request #19961 · anomalyco/opencode · GitHub

Uh oh!

fix(session): fire system.transform before messages.transform - #19961

fix(session): fire system.transform before messages.transform#19961ualtinok wants to merge 2 commits intoanomalyco:devanomalyco/opencode:devfrom ualtinok:fix/plugin-hook-orderingualtinok/opencode:fix/plugin-hook-orderingCopy head branch name to clipboard

Conversation

ualtinok commented Mar 29, 2026

Issue for this PR

Type of change

What does this PR do?

How did you verify your code works?

Checklist

Uh oh!

rekram1-node commented May 15, 2026

Uh oh!

Uh oh!

Uh oh!

rekram1-node commented May 16, 2026

Uh oh!

rekram1-node commented May 16, 2026

Uh oh!

github-actions Bot May 16, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot May 16, 2026

Choose a reason for hiding this comment

Uh oh!

ualtinok commented May 16, 2026

Uh oh!

ualtinok commented May 19, 2026

Uh oh!

elmagow commented May 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

fix(session): fire system.transform before messages.transform#19961
ualtinok wants to merge 2 commits into
anomalyco:devfrom
ualtinok:fix/plugin-hook-ordering