Skip to content

Project chat and planning updates

Project conversations work without a planning or specification document. Every chat surface—Feedback, Delivery, standalone Chat, Project Brief, specifications and planning—starts with Whole project as its read context. Open the Context button beside Ask and Update to choose a narrower scope in a side sheet. Expand the project explorer and select folders recursively or individual documents across directories. Choices stay with the conversation while you fill out feedback, navigate or reload the page.

Choose Selected files and folders to start a specific selection. Selecting a folder includes its descendants, including new files; deselect individual descendants to exclude them. Whole-project selection can also have exclusions. A partially selected folder is marked with a mixed checkbox. Excluded folders own all descendants; include the parent again or start a specific selection to choose within it. Use arrow keys to move through the tree and Space to select. Escape or Done closes the sheet and returns focus to its trigger.

Sending a question allows the selected provider to retrieve relevant content within the displayed scope. Narrowed project conversations also exclude unrelated planning artifacts. The explorer lists filenames locally; it does not load every selected file into the message. Ignored/private files, provider policy and retrieval budgets still apply. Up to 4,096 independent included paths and exclusions can be represented; recursive folders and Whole project avoid a document-count limit on their descendants. Incomplete directory listings are labelled, and selecting the directory includes eligible descendants beyond the displayed list.

Paste a public HTTP or HTTPS URL into your chat message to select that exact page for reading. You can also use Read URL in Chat context to select a page without including its URL in the message. Up to eight pages are selected for one request; the selection is separate from project file access and clears after sending. Turn Web search off beside Request details to limit external browsing to the page you supplied. The runtime checks each destination and redirect, blocks private network addresses, and supplies bounded readable HTML or plain text to the provider. Project context records the page title, URL, retrieval time and content fingerprint, or the reason a read failed. A URL citation opens the public page; it may have changed since retrieval. Pages that require sign-in or client-side rendering may have no readable text. Stop the response to end an in-progress URL read.

To search the public web, choose a backend under Settings → AI providers → Web search. Tavily uses DAY_SHIFT_TAVILY_API_KEY in the Day Shift runtime environment. SearXNG can connect to an existing public HTTPS or runtime-local loopback instance, or Set up local SearXNG with Docker or Podman already running in the runtime environment. The guided setup uses a pinned container image, a loopback-only port and a private generated secret. Settings shows setup progress while the image downloads, the container starts and Day Shift tests search. It also reports a retryable setup failure. Settings can test, start, stop, update or remove the Day Shift-owned local service. SearXNG JSON search must be enabled; the guided setup enables it automatically. The connection test sends a query for searxng. Removing the local service removes its Day Shift-generated configuration. The selected backend is fixed when a chat turn starts, and Day Shift does not silently switch to another service after a search failure.

If Day Shift runs in WSL 2, start Docker Desktop for Windows with its WSL 2 engine, then enable your Day Shift distribution under Docker Desktop → Settings → Resources → WSL Integration and apply the change. In Windows PowerShell, wsl --list --verbose confirms the distribution uses version 2. Back in that distribution, docker info should display a Server section. In Day Shift Settings, select Check for Docker, then Set up local SearXNG; Day Shift creates the container and tests JSON search. If docker is still unavailable, reopen the WSL terminal and check that Docker Desktop is running and the distribution integration is enabled. See the Docker Desktop WSL 2 guide.

The Web search switch beside Request details starts on in project chat. Turn it off before sending when you do not want search terms sent to the selected service. SearXNG forwards enabled searches to its configured search engines. Model-written terms can include details from the selected project context. In Chat context, choose Use an exact query to send only your own search terms for the next message. The switch remains at your chosen setting after sending; an exact query resets to model queries for the next message. Search permission is separate from project file access. The runtime permits up to three queries per turn and keeps at most five results per query, with timeout, rate and usage limits. Project context shows every exact query and its result status. Results provide titles, URLs, snippets and available dates as leads. A result becomes page evidence only when the runtime opens its exact URL through the guarded Read URL path. Failed reads and unread search results cannot support a page citation.

Context selection does not authorize edits. To update a planning document, use its separate Update target action in the explorer, then choose Update in chat. In planning and specification editors, the open document supplies the update target and remains included in the request. The same Context picker controls additional project access.

On Feedback, Prepare report context in chat stages a report, and Add feedback context to message adds its supplied text to your draft. Attached context shows the report title and version and can be removed. Report freshness is checked again before sending; stale context leaves the draft available for review. Adding report or delivery context does not select an update target. Creating follow-up work uses its separate preview and creation controls.

Temporary connection failures, selected rate-limit responses, and selected server failures can retry on the same provider and model up to three total attempts, with 100 ms and then 200 ms waits. Every dispatched attempt consumes the same cumulative input and turn budget; the operation deadline also covers retry waits. Cancellation stops the wait before another dispatch. Authentication denials, invalid requests, malformed responses, and unclassified failures require explicit recovery. A native tool call or partial answer draft stops automatic retries, and the protected finalization turn cannot be replenished by retrying. Durable receipts record every admitted request before delivery and keep actual token totals unknown when any attempt has no usage report.

Provider answers must finish successfully and match the complete requested response shape before Day Shift can use them. A truncated, refused, malformed or incomplete response cannot become a completed answer. Compatible providers use strict schema output when the selected model explicitly supports it; otherwise Day Shift sends the schema as bounded instructions and validates the answer locally. Generic JSON-format support does not establish strict schema support. Native document turns supply the response schema directly. Neither mechanism changes your write permissions or replenishes the operation budget.

Choose Ask for questions, suggestions and design discussions. It leaves documents unchanged. Update from this answer copies an answer into an Update request for you to review and send.

Choose Update to prepare changes to the selected documents. Review the changed text and full updated documents in chat. Apply changes saves the reviewed content after checking the current revisions; Discard leaves the documents unchanged. Pending reviews survive reloads. If a document or your editor draft changes, prepare a new update from the current version.

A successful application shows Saved and verified and offers Undo update. A selected document group is reviewed and applied together. Older Ask/Suggest/Edit history and proposal review links remain readable.

Specification decisions and accepted deferrals

Section titled “Specification decisions and accepted deferrals”

Specification review distinguishes unresolved decisions from resolved decisions and explicitly approved deferrals. Track them in the existing Open Questions section using the decision table below. A deferred value remains unanswered; its approved selection approach, owner, timing or dependency, rationale, and approval source explain why planning this scope can proceed before the value is selected.

## Open Questions
| ID | Status | Decision or question | Owner | Timing or dependency | Rationale | Approval source |
| --- | --- | --- | --- | --- | --- | --- |
| OQ-5 | deferred | Select and pin an exact supported stable framework release after compatibility checks; avoid prereleases and commit the lockfile. | Specification 2 | Implementation kickoff | Game rules do not depend on the exact version; version-specific work waits for selection. | Root specification, OQ-5: user approved selection at kickoff |
| OQ-8 | unresolved | Which hosting selection approach should apply? | Product owner | Before deployment planning | - | - |

Use stable, unique IDs and one of three statuses: unresolved, resolved, or deferred. Resolved rows need the adopted answer and its approval source. Deferred rows need all seven fields, including the full adopted selection approach and the rationale for planning before selection. An approval source identifies the actual user decision or source document and section; do not copy the example’s approval claim unless it describes your evidence. Unknown fields on unresolved rows may use -. A question without an approved answer or deferral remains unresolved.

The reviewer reports the decision ID and line requiring attention. Chat receives the same check for the current document or unsaved editor buffer when preparing an update. Existing prose and older tables remain readable, but ambiguous decision status continues to warn until an authorized refinement records the decisions in this format. Converting an existing approved deferral preserves its original approval; it does not require approving the same approach again. Keep supporting detail under the existing Scope or Constraints headings and preserve every unresolved choice in the register. Use None. only when there are no decisions left to track.

Saved and verified confirms a document update, not specification readiness. Draft feedback now reports remaining decision issues, and Review again checks the saved revision. A passing decision check does not authorize a planning-readiness disposition, create downstream work, or establish implementation or test completion. The checker validates recorded decision fields; it does not independently verify that an asserted approval source is truthful.

Select a phase and ask “Tell me about the parent overview.” The runtime resolves the registered parent identity and reads that document. Inspect Project context to see source paths, revisions and any exclusions. A source that is unavailable, stale or truncated cannot establish the missing requirements. Check the cited document before relying on an answer about those requirements.

Repository tools support directory listing, filename and text search, and file, section or line-range reads. Ask for an exact heading title or a one-based inclusive line range such as lines:4-8 to focus a read. Heading reads include their heading and nested subsections, ending at the next heading of the same or higher level; heading-like text inside fenced code is ignored. Duplicate or missing headings and invalid line ranges are reported without guessed content. Large documents use a bounded heading index tied to the current source revision, with bounded direct scanning when no usable index exists. Changed sources require a fresh selection. A partial scan can return a continuation before it supplies content; the provider continues with the same file and selector. Exact byte-range reads remain available as a fallback. Independent literal content phrases can share one scan in a batch of two to eight queries. Each query keeps its own identity, matched byte selections, status and continuation; completing one query does not complete the others. To resume a member, the provider replays the original full batch with that member’s cursor. If another member’s queued matches fill the bounded queue, drain those unfinished members before scanning farther. Search matches remain whole selections when a result page fills. Identical phrases with different query IDs retain separate progress and provenance. Reads remain subject to the project workspace, ignore rules, disclosure policy and retrieval budgets. Source text is untrusted input; instructions embedded in a file cannot authorize new writes. The runtime rechecks disclosure and revisions during a turn. It does not automatically send the entire repository to a provider.

Use Context beside Ask and Update throughout the GUI to manage project read access. Choose Selected files and folders to browse the actual project locally, or Whole project to allow eligible content throughout the project. Browsing lists names without reading file bodies or sending them to the provider. The folder filter searches the current listing; incomplete listings are identified explicitly.

Folder selections are recursive and include new files. Exclusions take precedence within additional access. Review Selected paths and exclusions to remove a path or include an excluded path again. Exclusions do not revoke the open document or existing planning context in document chats. Ignore rules, private-path restrictions and the filesystem broker still apply.

Sending provides read-disclosure consent for the displayed scope to the selected provider for one hour. Changing context revokes the previous grant before applying the selection. Revoke analysis access withdraws the grant and clears the selection; choose context before sending again. Switching project, provider, model or runtime clears the active grant; another browser tab does not inherit it. The next message uses the displayed scope to create a fresh grant.

Local provider trust does not grant unrestricted access. Read access permits bounded evidence retrieval only: it does not authorize source edits, shell execution, validation, lifecycle acceptance, deployment or delivery actions.

For an overview with many phases, the runtime can query the declared children and their recorded metadata without loading every body or unrelated history. A phase document larger than a read page remains accessible through bounded byte, section or line selections. Recorded status is not accepted completion: acceptance needs the corresponding reviewed evidence.

Open Project context beside the model picker for analysis progress, sources and follow-up actions. It distinguishes paths discovered from content selections actually supplied. A finished provider response does not imply exhaustive project coverage. Excluded, denied, stale and uninspected evidence remain visible. Inspect a citation to see its source and selection. A matching bounded preview proves only matching preview text; an unseen range or changed source has unknown freshness until revalidated.

A turn has finite provider-call, read, elapsed-time, context and retained-state budgets. The standard document agent permits 16 provider calls; the broader analysis loop also enforces its own ceilings. Useful additional reads can continue beyond the earlier eight-call boundary, but a budget stop remains partial evidence. Continue remaining analysis prepares a message for review; it does not send or expand permission automatically. Narrow the remaining question and submit it explicitly.

Both supported provider protocols use application-managed context compaction. Codex uses declared Day Shift tools and an ephemeral thread within each operation; the thread is never reused across separate chat requests. OpenAI-compatible providers request evidence through brokered JSON completions. Compacted notes preserve bounded source identities and excerpts, not proof that omitted material was read. Current authorization and revisions are checked again before retained evidence is supplied.

Activity and Project context show the same current coverage receipt. Scoped coverage complete requires every current query to finish, no remaining source gaps, and support for every requested criterion. An unrelated completed read cannot hide another unfinished query. Coverage not assessed means there is no current observation; expired evidence loses its earlier completion claim. Earlier partial pages remain in retrieval history even after their query completes, and reconnecting preserves the current receipt without repeating the work. Coverage describes scoped evidence and satisfied criteria, not executed checks or lifecycle acceptance.

Current read-only comparisons and ordinary follow-up questions remain answers when Update was previously selected. Selected targets are recorded as read, and historical qualified memory cannot supply a new instruction or write authority. A new ordinary question clears obsolete review criteria while independent unfinished queries and source gaps remain visible. A mixed request containing a current correction can prepare changes only within the selected writable targets and the existing review/apply flow.

A review tracks each requested criterion as verified, contradicted, inferred, or Not assessed, with its claim scope and recovery when support is missing. For a general acceptance review, the selected document’s Acceptance Criteria supply the criterion IDs; requirement trace IDs and old conversation summaries do not add requested criteria. Bounded overflow remains explicitly Not assessed.

Review context follows criterion matches and declared relationships, including later milestones and tasks. Project context shows selected and omitted branches with reasons. An omitted branch is not absent and a bounded sample is not a complete review. Continue the recorded query or select a narrower branch when evidence remains unread.

Each material finding retains its own inspected supporting selections and conflicting selections. Retained evidence omitted from the actual provider input cannot support that answer. Changing a supporting source invalidates affected findings and criteria; evicting an unrelated source preserves independent findings. Source interpretation and recorded metadata do not establish executed validation or lifecycle acceptance.

Day Shift supplies the same versioned conversational contract for project chat, document chat and ordinary inference operations. It tells the model to answer the current question directly, retain supplied decisions and approvals, distinguish deferred choices from unresolved decisions, cite available evidence and respect the current request’s scope. Previous answers and source documents cannot grant new write authority. Document updates add rules for adopting recommendations, recording suggestions, preserving unrelated content and describing proposed changes.

Tool invocation and response encoding are separate from these behavioral rules. Codex receives Day Shift instructions at developer level while retaining its native base instructions; OpenAI-compatible providers receive them as a system message. Codex receives a response schema; compatible-provider JSON is requested in the prompt and checked by Day Shift. A shared contract standardizes Day Shift’s instructions, but does not guarantee identical answers from different models. Runtime scope checks, candidate validation and save controls remain authoritative.

Use Chat mode to choose where a conversation belongs:

  • Page is the default on Project Brief, Specs, Planning, Feedback and Delivery. Each workspace keeps its own conversation. Navigating away backgrounds that chat and opens the destination workspace’s previous conversation or a fresh composer, without needing New conversation. Selecting another artifact within the same workspace keeps its page chat.
  • Global keeps the same conversation visible as you move between workspaces. Switching between Page and Global restores that mode’s conversation; it does not move or merge existing conversations.

Both modes keep requests running in the background. Returning to a workspace restores its chat and unsent draft during navigation. Use New conversation to run another chat on the same page. The header’s Chats list identifies each conversation’s owning workspace or Global mode; Open chat returns to that owner, and Stop cancels only the chosen request. Reloading the tab reconnects submitted requests and retains their ownership, but does not preserve unsent drafts.

Conversation history belongs to one project. History entries use the first message as their title, so conversations about feedback or general project questions remain identifiable without a planning artifact. Its operations retain their original target and revision identities when focus changes. Reconnecting to a restarted runtime requires its current identity; another project cannot inherit the prior project’s session authority.

Retired operation identities remain reserved to prevent redispatch. Deleting visible history is not permission to reuse an old operation ID. Preserve private conversation storage and transaction journals when diagnosing interrupted work; deleting those files can remove the evidence needed to reconcile a save.

Choose Update and state the intended planning documents explicitly. For example, request alignment of a phase and its direct milestones, or a Structured overview and its direct tasks. Reading related files is separate from authorizing edits. Ambiguous scope needs clarification; creation of new children and lifecycle changes require their own explicit commands and authority.

The runtime freezes the resolved target set, prepares every candidate, validates artifact shape and cross-document consistency, and waits for you to review the complete group. Apply changes checks current revisions before saving the group together. Drafts show proposed text. A successful save requires confirmed runtime results, not an assistant sentence claiming it saved something.

Handle conflicts and incomplete operations

Section titled “Handle conflicts and incomplete operations”

Keep unsaved editor text when a file changes on disk. Compare the current canonical version with the draft before selecting a replacement. A stale runtime identity or document revision must be refreshed before a new operation.

A verified rollback means the transaction restored its original contents. An incomplete transaction means recovery could not establish that result. Preserve its journal and inspect the affected files; do not retry blindly or describe the group as saved. Grouped undo checks every member against the saved versions and refuses intervening edits. It creates a compensating operation with its own identity.

Already-open editor and chat drafts remain mounted through disconnect and runtime restart. Connection warnings identify retained, potentially stale project data, and saving stays unavailable while verification is incomplete. Preserve or copy unsaved work before deliberately reloading the page: reconnect retention does not promise disk-persisted draft restoration.

Events tell connected clients to refresh canonical state. After a cursor gap or runtime restart, reconnect and reload snapshots instead of treating cached events as authoritative. An external rename or deletion can invalidate focus and draft targets.

The project-chat acceptance fixture exercises the serving HTTP runtime and the Codex app-server transport with a deterministic provider. It records tool kinds, target identities and result statuses without logging credentials or source contents. The optional configured provider smoke uses DAY_SHIFT_PROJECT_CHAT_CODEX_SMOKE=1 and requires an explicit DAY_SHIFT_PROJECT_CHAT_CODEX_MODEL; normal tests require no credentials.

The maintained evidence matrix is in Structured work 41’s task 09 implementation summary. Deterministic provider transport tests do not establish real-model quality. Browser fixtures, installed delivery, platform coverage and performance bounds require their own recorded results. A skipped browser or configured-provider lane remains unqualified; do not infer a pass from component or service tests.

Whole-project qualification and optional calibration

Section titled “Whole-project qualification and optional calibration”

The deterministic whole-project fixture exercises the built and extracted installed runtime and native browser with eight declared phases, a phase body above 100 KiB, at least 16 MB of unrelated history, and twelve useful content reads. It covers phase status, test coverage, architecture impact, contradictions and follow-up questions. Measurements distinguish physical analysis reads and directory traversal from unrelated GUI catalog work, and record provider inputs, retained state, memory and latency. Scripted transport results are integration evidence, not live-model reliability.

Run the credential-free calibration contract check from the repository root:

Terminal window
node tests_v2/acceptance/project-analysis-calibration.cjs --check

Live calibration is opt-in and may send the generated fixture to the configured provider:

Terminal window
DAY_SHIFT_PROJECT_ANALYSIS_CALIBRATE=1 node tests_v2/acceptance/project-analysis-calibration.cjs --protocol codex-app-server --model YOUR_MODEL --seeds 46,47 --repetitions 1

The JSON report records the exact model, protocol, settings, fixture seeds, repetitions, answer/citation/coverage scores, failures and untested models. Seeds control fixture data; provider sampling seeds are unsupported. This entrypoint does not calibrate OpenAI-compatible providers. An offline check, a scripted answer or a report for another model cannot establish the selected live model’s reliability. Keep failed and incomplete results visible.

Ask why a requirement or decision exists while keeping the relevant specification selected. Project maps show the canonical sources and their checked revisions. Follow a source citation to inspect the document; explicit dependency links are distinguished from possible matches. A map with partial coverage does not establish that every downstream document was read. Changing a source invalidates the earlier map.

Ask for the impact of a requirement change before preparing an update. An impact preview is read-only. It identifies directly linked work, possible associations and coverage gaps. Select the affected items you intend to change, then use Prepare selected impact for review. Review the resulting diff and validation findings before Apply changes. The selected preview belongs to the current project and conversation; changed sources or dependencies require a fresh preview. Selecting impact does not grant permission to edit additional files.

Check a task and inspect an allowed local preview

Section titled “Check a task and inspect an allowed local preview”

Declared validation and app inspection controls are available when a task definition is focused and its runtime actions permit them. These controls do not turn a general conversation into an arbitrary command runner. Preview the exact declared check, its covered task and its effects, then explicitly confirm execution. A failing, cancelled, imported or stale record is not passing executed evidence. Even a current passing check does not accept or complete the task.

For app inspection, select an allowed local preview and review its origin and viewport before confirming. Inspection observes one page and a bounded read-only interaction, along with layout, accessibility and console findings. Unsafe origins and redirects are denied. Missing browser support, cancellation and time limits have explicit outcomes. Accessibility observations are heuristic, not a complete accessibility audit or complete product acceptance.

Use the task’s handoff and return controls for external implementation. Handoff prepares instructions for your chosen agent; it does not launch that agent. Returning work compares observed source changes with reported progress and distinguishes executed checks from imported claims. An external agent’s changes remain its own work. Out-of-scope or stale observations require correction or contract review. Acceptance still needs the owning review and lifecycle action.

Understand each action’s effects and grants

Section titled “Understand each action’s effects and grants”
Action What it permits What you control
Project reads, map and impact preview Bounded evidence retrieval; no source changes Current project read-disclosure consent, selected paths and exclusions
Web search or URL read External leads or a bounded page read Separate search or URL permission; a search result does not mean its page was read
Prepare and apply a document update A reviewed change to the selected documents Explicit update scope, current revisions and the separate apply action
Declared task validation Execution of the exact eligible declared check Previewed command and effects, followed by confirmation
Local app inspection One allowed page and a bounded read-only observation Previewed origin and viewport, followed by confirmation
External handoff and return Instructions and observed implementation evidence Your external agent choice; separate canonical review and acceptance

A provider cannot create new authority by requesting a tool. Revocation stops later reads; cancellation stops the selected operation. Source and web content are untrusted data and cannot authorize writes. Private content is filtered before model context and retained evidence. Read grants do not permit validation, browser inspection, deployment or lifecycle acceptance.

After a provider error, keep the current sources and narrow the next question. Limited coverage, large sources and exhausted context remain visible; continuation requires an explicit new request within the displayed scope. Reconnect restores operation identity and evidence rather than repeating an effect. Do not replay an expired proposal or validation plan: obtain a fresh preview and inspect what changed.

The verified SearXNG journey has deterministic integration evidence on both native Codex and OpenAI-compatible transports through the built runtime and actual GUI. It covers a large focused specification, retained conversation summary, decision citations, bounded impact, selected diff and apply, declared checks, available local observation and truthful external return. Local Chromium exercises the actual Settings search connection flow and a narrow viewport. The bounded sample-page inspection is separate from that Settings check.

Failure qualification covers current-source conflicts, revoked access, malformed requests, large and partial evidence, provider failures, context and operation budgets, private-content filtering, repeated retries, cancellation and restart. Some faults are transport-specific; this evidence does not claim every failure was reproduced on every provider. The maintained qualification record identifies each covered requirement and its evidence limits.

Scripted transports do not establish real-model answer quality or remote-provider reliability. The local SearXNG prerequisite can return useful JSON leads while some engines are unavailable. SearXNG engine-health diagnostics remain unqualified: chat evidence can omit engine warnings even when a search returns useful leads. Other operating systems, browser engines, destructive interactions and full accessibility auditing remain untested or outside this bounded capability. An unavailable lane is not a pass, and no recorded observation automatically accepts product work.

Final answers keep content citations only when the current operation inspected that source and supplied its exact content to the provider. Citations to unread files, stale revisions, omitted excerpts, metadata, or search leads are labeled unverified with a request to read the current content. Encoded paths and line or heading fragments must resolve inside the inspected selection. The same checks apply to Markdown links, plain URLs, autolinks, HTML anchors, and code-quoted project paths. A successful supplied web read can support its URL; a search result alone cannot. Uninspected web fragments are explicitly repaired to the read page URL, with the anchor marked Not assessed.

Statements that tests passed, changes were saved, computations were verified, or lifecycle work was accepted require matching host evidence. Source text and user reports remain reports, even when they say an operation succeeded. Unsupported model assertions are labeled unverified in answers, edit explanations, and conversation memory. Host-computed digest evidence establishes only the matching path, revision, and byte range; it does not establish test success or lifecycle acceptance. Actual saved outcomes continue to come from verified write receipts.

When context fills, Day Shift keeps active claim support, contradictory observations, and sources needed by unfinished criteria or queries ahead of unrelated and redundant excerpts. The provider receives the exact selections that still fit, not an assurance that all retained source text was delivered. If a required selection must be omitted, dependent findings are removed and affected criteria become Not assessed with a request to retrieve current evidence. Omitted selection identities and unresolved work remain visible in the bounded context record. If that qualified context cannot fit, the operation returns a supported partial result instead of presenting an unsupported conclusion.

Follow-ups retain bounded, versioned records with subject identities and original user wording. Decisions, alternatives, deferred choices, questions, and caveats remain separate. User statements are historical user reports; model summaries are unverified interpretations and cannot confirm a decision, prove a save, or grant approval or write authority.

Model notes carry the source revisions and access conditions they depended on. Before each response, Day Shift removes notes affected by changed sources, selected context, grant or policy changes, expiry, or context exclusions. It checks again after generation and discards a response whose retained factual context changed. Unrelated explicit user decisions remain historical context. An unscoped model note conservatively depends on all selected documents and retained analysis sources. Source-free alternatives and questions can remain explicitly unverified model proposals, separately from factual notes. They never establish a decision or current fact. Prior web content requires a fresh read; page metadata alone cannot restore a factual note.

The persisted format keeps at most 32 records within 24 KiB. Overflow is recorded as omitted context instead of merging subjects or quietly dropping dependencies. Legacy or unsupported summaries recover only original user wording and request fresh evidence for model claims. A valid user-edited summary replaces the inherited summary as explicitly user-authored, non-authoritative context. Removing operations, answers, or sources from the context window also removes their dependent memory; disabling previous analysis removes model notes. Raw legacy summaries and page metadata cannot bypass those controls.

Day Shift charges provider input before delivery, including instructions, delivered schemas and tool declarations, repeated context, and native tool results. Native retained context must fit the current provider allowance before each result is delivered. Admission estimates remain conservative; a provider report is kept separately, and cumulative actual usage stays unknown when any request has no report.

Exploration leaves one provider completion, 256 KiB of input, 256 KiB of evidence revalidation, and one minute within the operation limits for finalization. At a limit, the final assessment uses the available evidence without further tools or edits. Unfinished work remains Not assessed, with a narrower retry as recovery; finalization cannot establish missing coverage. Provider output tokens remain reserved independently for every request.

The configured provider timeout_ms covers the complete OpenAI-compatible request, including response-body reads and streams. A stalled native Codex turn uses the same setting as an inactivity limit: new response text, reasoning progress, plans, or scoped tool activity refresh that interval. Caller cancellation and timeouts cancel owned readers or native turns and close the affected session. A timeout is recorded as a limited result; it does not claim a saved update.

Chat operations also retain a cumulative 15-minute active-work limit across explicit recovery. Active time includes generation, retrieval, final source checks, and cleanup through final completion. Waiting between completed attempts and later reconnects do not consume active time. Recovery retains the original history, provider identity, and consumed request, tool, evidence, and context quotas. Before dispatch, admission is recorded durably. After a lost process, only the last persisted active checkpoint is known; unobserved time is not reconstructed from downtime. Legacy receipts preserve their known totals without inventing an active interval.

Read grants and continuation authority still expire by wall-clock time. Pausing active consumption does not renew permission, restore an old operation cursor, or authorize replay. Recovery needs current grants, revisions, and a fresh operation-scoped continuation.

Content search uses bounded 256-byte chunks, reducing a matched 32 KiB single-phrase scan from 256 reads to 128. Dense results and batched phrases further limit each chunk to the available result queue; continuation preserves exact UTF-8 selections. Each file page owns one descriptor, with authorization and file identity checked before and after reads. The descriptor closes at the page boundary, including cancellation, timeout and errors; disposal cancels active disclosure. These are local fixture measurements, not remote-provider or operating-system cold-cache latency guarantees.

Repeated references to the same planning document share one retained selection. An immediately verified snapshot can supply its metadata without a duplicate probe. Cached evidence still receives fresh source and permission checks before disclosure and on later requests; a changed or revoked source cannot support the answer.

Independent reads use a tested serial fallback with a cap of one. Local measurements of scheduled caps one, two and four over the synchronous broker did not create overlapping reads. Serial ownership preserves requested order, cumulative budgets, cancellation and each cursor. These fixture measurements do not establish remote-provider speedups.

Overlapping reads and related tool results share immutable cached evidence. Each reuse checks the current source, project, provider, grant, exclusions and disclosure policy. An excerpt removed from the analysis ledger cannot return through an older discovery, range, relationship or project-result cache; it requires a fresh authorized read. Heading indexes remain tied to their current source and disclosure conditions.

Cache retention has a per-operation bound. Diagnostic byte counts separate unique payload storage from index overhead and measure serialized UTF-8 bytes; they do not measure JavaScript heap size or establish faster provider responses. Removing the last index reference releases its cached payload.

Within each provider request or tool response, identical selected-document, prefetch, cached-result and ledger excerpts share one current source body. References preserve the source path, revision, exact UTF-8 byte range, section and trust classification. Unsaved edit text remains separate from different disk content. Host citation checks resolve only definitions actually supplied in that same frame.

Every frame is self-contained. After application compaction, omitted bodies and their references leave the disposable provider view together. Native history has no certified compaction lifetime, so a later request or tool response supplies its required body again and charges repeated disclosure. References never depend on an earlier message or native thread. Existing native session reuse remains available; this does not claim deduplication of inaccessible native history or faster provider responses. If current evidence cannot fit, narrow the request and continue from a supported partial result.

Expanded Activity distinguishes current observations from inherited budget consumption, shared cache retention and estimates. Repository file counts deduplicate the same path across chunks and pages; read operations count broker-owned fs.readSync and fs.readFileSync API invocations, including separately requested whole-source digest reads. Node internal calls and operating-system syscalls are not counted. Scanned bytes measure successful payload lengths at those owned boundaries, while evidence envelope bytes measure serialized UTF-8 retrieval results and provider-facing bytes measure admitted emitted frames. Repeated delivery is not another unique file. Metadata probe counts cover the instrumented repository metadata and revision checks; authorization has its own timing category.

Successful cache lookups are observed at shared-payload, planning-metadata and heading-index caches. Shared-payload lookups include inspection, pruning and range-candidate peeks; they do not measure delivered reuse or avoided reads. Last observed cache bytes separate immutable payloads from index overhead and are serialized UTF-8 measurements, not heap measurements. Lookup payload bytes count returned cache references across lookups; they are not new scanned bytes. Cache capacity and source invalidation remain unchanged. Recovery preserves cumulative budget consumption separately; cumulative unique resources across previous operations are unavailable rather than added together.

Exact provider-facing bytes come from transport callbacks. Adapters without those observations retain an explicitly estimated byte count, with exact bytes unknown. Token estimates describe the latest context and remain separate from actual provider usage. Legacy receipts and unobserved categories display unknown, while an observed zero remains zero.

Timing starts at operation admission and records queueing, provider work, retrieval, authorization, metadata probes, serialization, checkpoint writes, revalidation and reserved final generation. Provider wait measures observed transport fetch/body reads and native completion collection. Ordinary adapter preparation and response processing remain unattributed unless a host span observes them; custom adapters without a transport seam leave provider wait unknown. Reserved final generation is a parent stage whose exclusive time excludes nested provider wait and other child spans. Each observed interval belongs to the deepest active span; a later-started overlapping sibling takes precedence. Category times plus the explicitly unattributed time reconcile to the observed window. Span samples are bounded, but category totals continue after the sample limit. Correlation uses only operation and opaque query identifiers; timing receipts contain no source paths, source text, prompts or responses.

A persisted checkpoint cannot include the completed duration of its own write. It records its observation boundary and in-flight spans, with the tail after capture unknown. A live snapshot may include a write that completed after that durable snapshot; reconnect preserves the earlier boundary. No extra write is performed just to measure a previous write, and missing historical intervals are not reconstructed.

Chat debug exports include the runtime build identity captured at operation admission and a versioned, bounded diagnostic history. Older operations without these fields have unknown build and diagnostic coverage. Query records identify the logical query, returned status, bytes scanned, added selections, cursor fingerprints, progress and stagnation count. A section scan can legitimately return empty pages while advancing through large metadata; continue the same section request with its returned cursor.

Rejected tool arguments record structural facts, such as non-string, absolute, parent-traversing or out-of-grant paths, without retaining the rejected values. Provider failures record safe codes, retry decisions, output measurements and schema-validation keywords. Unknown failures retain a message fingerprint and bounded stack locations rather than raw error messages. Diagnostic totals and omitted-record counts make bounded-history loss explicit. These additional diagnostics contain no raw prompts, source bodies, credentials or cursor tokens; the existing full session export still contains the conversation and inspected evidence, so review it before sharing.

Ordinary progress, context measurements and analysis snapshots are coalesced over 100 milliseconds. Dispatch accounting, mutation receipts and terminal results remain synchronous. A crash may lose the latest unflushed progress details; it does not refund admitted dispatch budgets. Each source-validation pass observes a unique retained file once, with fresh checks at the next disclosure boundary. Completed section requests can reuse the exact current result after those checks; continuation tokens always go back to the cursor service. Repeated reads remain bounded. Static tool guidance already present in native instructions is not resent; source excerpts remain independently resolvable.

Diagnostic totals continue accumulating after individual measurement samples are evicted. CHAT_CHECKPOINT_MEASURED separates serialization, atomic-write and index-update milliseconds and records serialized bytes. Durable snapshots retain the existing observation boundary: a checkpoint cannot contain the completed measurement of its own write. Budget admissions use compact, synchronously durable records bound to the current full snapshot generation. Restart merges their consumed quotas before recovery; an older generation cannot override a newer snapshot. CHAT_ADMISSION_MEASURED measures these writes separately. Ordinary progress snapshots coalesce over 500 ms and omit derived history/index updates. Admission of a new operation, effect records and terminal state still persist full snapshots and update the index synchronously.

PROVIDER_FRAME_MEASURED reports actual delivered serialized frame bytes, new and repeated logical source-body bytes, and the remaining non-body bytes (including metadata, framing and string escaping). These are distinct from the conservative cumulative provider-context budget, which repeatedly charges retained history and does not measure newly transmitted bytes. TOOL_VALIDATION_MEASURED records pre-tool and final disclosure-check durations. Measurements occur at actual delivery; rejected candidate frames do not become supplied evidence.

Claim diagnostics include the classification rule, Markdown-normalized text fingerprint, scope, cell/row citation association, bounded evidence keys and cited byte ranges, with omitted-reference counts. Table cells and normalized clauses are classified separately; “Verified: draft” describes recorded status, while a later execution claim in the same row still requires qualification. They do not contain the claim text. An unattributed statement such as “Phase 05 remains only partially accepted” still requires qualification; prefer explicit document attribution such as “Phase 05 documents report acceptance of some scoped work; full completion remains unsupported.”

Native criterion results include bounded conclusions with separate support and conflict references. Reference capacity scales with the output budget up to eight references per support/conflict list. A governed per-phase assessment must include its own overview and inspected sections from each retained local milestone reconciliation. A foreign handoff or header-only reconciliation cannot substitute for those sections; insufficient support remains Not assessed. These checks establish evidence coverage, not semantic truth or lifecycle acceptance. Explicit numbered per-phase requests have separate phase dispositions and a separate next-gate disposition. Adjacent identical citations collapse for display; distinct exact byte ranges keep separate links. Visible labels use compact identities such as P04/M01 reconciliation; adjacent identical labels for different ranges receive numbered markers, while exact ranges remain in link targets and source details. Large provenance records remain intact: section scans use bounded 16 KiB chunks and reuse a revision-bound heading index after the first complete scan, avoiding a migration of historical lifecycle evidence. Missing-section results include up to twelve current heading hints, with an explicit truncation flag. Heading guidance distinguishes implementation summaries from phase overviews and reconciliations. Repeated missing selectors reuse revision-current feedback without rescanning; source changes still require refresh. Retrieval errors retain query and selector fingerprints. Exhausted exploration latches once: further tool calls return the same stop feedback without new budget admissions or finalization events.