dsh-context/docs/compatibility.md

Compatibility

dsh-context declares per-release compatibility with @deepseek-ai/dsh in its package manifest (dsh.compatibility.dshReleases). This page records what is actually verified for each declared release, and how.

Last verified: 2026-10-11 (plugin dsh-context@0.68.0 source tree, dsh 0.2.0-rc.2). The disposable-profile column below records its own, earlier manual dates.

Supported dsh releases

dsh release	Session log	Declared	Automated seam matrix	Disposable-profile install / uninstall
0.1.7-rc.2	V4	compatible	✅ baseline v0.1.7-rc.2	✅ install OK → uninstall OK (verified manually 2026-09-28)
0.2.0-rc.2	V4	compatible	✅ baseline v0.2.0-rc.2	✅ install OK → 1 composed row → uninstall OK → 0 rows (dsh-context@0.60.0, verified manually 2026-09-29)

The automated seam matrix runs for every row on every pnpm test in CI: both workflows fetch the pinned baseline tags and build the plugin first, and a missing prerequisite fails the run instead of skipping it. On a dev box with no checkout it skips with a printed reason. The disposable-profile column is a manual, per-release check: each release's CLI was installed from npm into a temporary DSH_HOME (the real ~/.dsh is never touched) — 0.1.7-rc.2 and 0.2.0-rc.2 on 2026-09-28 and 2026-09-29. The check pins the plugin build it installs: a bare-name install resolves through the profile's package manager, whose release-age policy can lag a just-published build (on 0.2.0-rc.2 the bare name resolved to 0.59.1, so the row records the explicit dsh-context@0.60.0). The only pnpm complaints there are peers the harness supplies from its own runtime module tree — none version-incompatible, and the install is never refused.

The supported range is the two lines above: the 0.1.7 line from rc.2 (the first release of that line with a complete npm dependency closure; rc.2 adds startsSeries to the first request/header of a resumed series and the developer/message tool-registry events; startsSeries is inert to the fold, while registry messages contribute to injected context), and the 0.2.0 line from rc.2 — the continuation of the V4 Config-form generation, every seam the plugin rides spelled as at 0.1.7-rc.2 (cordis 4.0.4, the same projection-registry contract, platform module table, client seats, settings transport, and durable-event vocabulary; still the V4 log generation, with no v4-to-v5 migration in the session-format tree). Everything older — the 0.1.5 line (supported through dsh-context@0.65.x), the 0.1.2/0.1.3 generations, the 0.1.1 line, and the alpha previews — is no longer in the support matrix; on such a harness the plugin's baseline gate composes the fallback units and shows the upgrade prompt instead of folding.

Harness-side compatibility admission (0.1.7+)

Every supported line enforces a plugin's own peerDependencies at startup and install: a bundle whose @deepseek-ai/dsh* peer ranges do not satisfy the running version (semver, includePrerelease: true) is skipped at composition, refused at install, and only loads with an exact-version exemption (dsh plugin allow-version). This plugin's dsh peers are all >=0.1.7-rc.2, which both supported lines satisfy, so the plugin composes normally with no exemption. The gate predates them (0.1.7 introduced it), so an older harness without it (the 0.1.5/0.1.6 lines) still loads the plugin — and the plugin's own baseline gate below is what composes the fallback units there.

Baseline gate

The declared floor is enforced at runtime, not only documented. At startup the host resolves the running harness's version — first from the module tree the plugin is bound to, then from the $DSH_HOME/profiles/node_modules mirror — and compares it against 0.1.7-rc.2 (channel order: release > rc > beta > alpha):

• Below the floor — fallback projection units that fold nothing and serve zeroed data: the Context and Fleet tabs keep their (empty) cards while a modal names both versions and urges the update.
• At or above the floor — the real projection units.
• Undetectable or unparseable — fails open into the real units, so a probe misfire never blanks a working deployment.

The mirror reflects whichever installation last booted a CLI profile, which need not be the running one (a packaged desktop client never refreshes it), so it never outranks the running module tree. Implementation: src/host/version.ts, src/host/fallback.ts, src/shared/version.ts.

Session-log generation

Every supported line is the V4 durable-log generation, and the plugin folds it from one shape-driven code path (src/host/logShapes.ts holds the vocabulary readers — firstTokenTimeOfStream, replaceRangeOf, isTokenChunk — and applyTimeline runs a single code path over them):

Seam	V4 (0.1.7-rc.2+, 0.2.0-rc.2+)
System prompt	system/message surface node (a request/header carrying header.system is rejected outright)
First token	embedded assistant/message.data.stream (also assistant/attempt.data.stream) — packed delta runs plus raw chunk records
Replacement endpoints	{ startSeq, endSeq }
Nested PTC dispatch	tool/ptc-dispatch
Tool result	first-class role: 'tool' message with lifted message.toolCallId/message.isError, direct content
Developer messages	developer/message tool-registry changes, nested under data.message

Legacy logs never reach the fold. The running harness migrates every historical log to V4 on EVERY read path the plugin rides — live open, persistence read handles, sessionQuery.observeSession (the detail/backfill cold reads) — in memory on read opens and as a published successor file on write opens; the fork-inherited prefix migrates with the rest (the V3→V4 stage lifts the old tool-result wrapper into the first-class message and rewrites sources). A log the migration chain cannot convert is refused entirely (SessionFormatUnsupportedError): the session is unavailable, never served raw. So on any supported harness the events the fold sees are V4-shaped by construction, and the pre-V4 spellings (the tool-result wrapper block and its isError, the event-level data.error mark, and further back the header.system envelopes, assistant/chunk floods, { start, end } endpoints, tool/code-dispatch) exist nowhere in the code.

The fold never branches on a detected harness version. The version probe is best-effort — it reports the module tree the embedding shell resolves plugin imports to, which a packaged client may redirect — so it decides which units register and nothing more.

Deltas of the current generation

• Developer messages: developer/message tool-registry changes are nested under data.message and fold into injected context, including request composition, live surface, and removed-node archives. They never count as human inputs; empty-content events keep their surface position at zero tokens, matching the harness's message projection (its deriveEventMessage projects an empty message to null — the fold's structural read keeps the node).
• Tool-result error mark: the event-level data.error metadata is admitted only alongside the message mark (the generation's own validation refuses a row whose data.error outlives message.isError), so the fold reads message.isError alone; a hostile row carrying only the envelope field degrades to unflagged, never to a throw.
• Image offload accounting (deferred): when a route rejects with IMAGE_OFFLOAD_REQUIRED, the harness offloads images out of the context through a message projection. The plugin's fold reads raw durable events, so an offloaded message keeps its pre-offload image count and price until compaction removes it — an over-count on a rare recovery path, not a crash. Revisit alongside the harness's own token-meter treatment if it shows up in practice.
• User preferences: plugin configuration lives on the Plugins page, derived from the loader entry's own Config schema (src/host/config.ts): cordis validates the config: block through the schema's Standard Schema face, and the SAME schema is what the settings describe projects — the namespace is the entry id (dsh-context) and the preference fields (.volatile()-marked) are the served, live-editable form; the fold bounds stay ordinary entry config (their edit remounts the entry, a preference edit commits volatile-only and never remounts). The browser half registers its card on the Plugins page's keyed plugins.bundle.config seat over the configForms transport (ctx.configForms.whileServed keeps the card alive only while the Host serves the namespace). The .volatile() modifier is feature-detected at schema build time (volatileField), so the same bundle still loads — preferences inert — on a below-floor harness whose schemastery predates it (the gate's fallback composition).
  • Preference values persisted on a retired line (the settings.yaml section dsh-context) are moved into the profile entry of the same name by the harness's own legacy import on first boot, so existing values carry into the current surface.
• System-prompt prefix/suffix injection: the promptPrefix/promptSuffix preferences are joined around the harness's assembled system prompt by two globally registered prompt sections (src/host/promptInjection.ts), not by string-editing session events. The seams this rides, probed per supported line in the compat matrix:
  • ctx.systemPrompt.section() — the ordered-section registry face (packages/core/system-prompt). The sections are registered through a deferred ctx.inject(['systemPrompt'], …), so a composition that starts the registry later still gets them, one without it degrades to a no-op, and the registration unwinds with this plugin's fiber. The orders anchor on the harness's own registry (HARNESS_IDENTITY - 1000, DEPLOYMENT_PERSONA_SUFFIX + 1000), keeping the prefix first and the suffix last even if the central order table is renumbered.
  • The volatile-only commit — vendor/loader/src/config/entry.ts rewrites the running config references on a preference edit without remounting the entry. The section text providers read those references per assembly, so an edit applies from the next model step; the harness's own system-prompt projection then updates the session log (system/message replace or append, per its routing), which is how a new session carries the text from its first turn.
  • Scope: global sections participate in every assembly — every session, subagents included (the same semantics as the harness's own deployment personaPrefix/personaSuffix). A deployment that registers a complete prompt section keeps ONLY that section, by the registry's design; the configured text is dropped there. The text is registered interpolate: false, so {{...}} in user text stays literal and can never trip the strict variable interpolation.

Parsing resilience [IMPORTANT!]

The plugin lives off data it does not own: the durable session log (event shapes vary across dsh versions, producers, and hand-edited replays), the conversation snapshot behind a client join, projection payloads on the wire, history RPC pages, and persisted stores. All of it is untrusted input at every layer, and two failure classes are designed out: parsing that blanks the page (the error card), and anything that leaves the page stuck on "loading".

• One bad record never takes down a view. A malformed node, event, tool entry, or file op degrades to zero rows for that item — the card, the tab, and the session keep working.
• Host-side projection folds are TOTAL. The harness projection registry drives apply straight off the session/event bus with no error boundary of its own: one throwing fold stalls that unit's cells and its session/projection push feed, and the browser waits on "loading" forever. Unknown event types return the state unchanged; per-event processing is isolated so a malformed event is dropped whole (all-or-nothing, no partial state); and no undefined-valued property is ever materialized into persisted state — the plain-JSON precondition makes one such property fail EVERY projection-cache write for the session, breaking sessions in unrelated, far-away places.
• Client-side parsing degrades visibly. Delivered projection payloads are sanitized at the boundary (the timelineOf pattern: collections re-proved, scalars zeroed, whole-value absence stays null → loading screen); per-item work in any fold over join/log data is isolated (per-item guards, or a bounded catch where a hostile object may throw on property access); every async fetch resolves to data or a visible retryable state, never an unhandled rejection that leaves a spinner.
• A session with no pushed contextTimeline value never stalls the view. The projection store's seed() clears a key its baseline omits, and a checkpoint row can be missing or version-stale on a cold-observed session — so the Context tab, the Fleet tab, or the /context modal can sit with no value at all and no push on the way. The source opens the detail channel itself in that case (rev 0): the plugin's own /api/dsh-context/detail route folds the durable log on demand and serves the slim head beside the collections, the next pushed head takes the channel back per its own revision, and a settled-without-data read surfaces the retryable failure note (the same DetailNote the detail cards use) instead of an eternal spinner. The Fleet tab reads that one source too, so its page carries the same guarantee, and every figure on it is derived by the same helper the Context tab's card uses (agentSelfOf), so the two mounts cannot disagree.
• A refold invalidates earlier detail reads. When the pushed revision decreases, the client resets its detail ledger and ignores any earlier read's data, absence, or failure. Once that read settles, the pending revision is fetched again; collections from the old fold never merge into the new head.
• Every field is re-proved at runtime. Structural narrowing over blind casts; optional chaining over non-null assertions; elements that fail the shape are skipped instead of throwing.
• Every parser carries hostile fixtures next to its happy path: wrong types, null/missing fields, null or primitive elements inside arrays, unpaired references, and objects that throw on property access. The repository's 100% per-file coverage bar applies to every guard branch — an untested guard is an unverified promise.

Low-level parity

The plugin's low-level logic mirrors the newest supported dsh implementation: token estimation re-fits dsh-token-meter's estimateSystemMessage (estimateSystemTokens / estimateSystemContent), and the TPS chip's first-token rule matches dsh-llm's assistantStreamFirstTokenTime. The Context card's occupancy percentage is the composer ring's own derivation — @deepseek-ai/dsh-client-ui-conversation's contextOccupancy (Math.min(100, Math.round(used / window * 100)), used = contextPressure.projectedTokens ?? pressureTokens) — read off the same token-meter projection and printed on a 0.1 grid, so the card and the ring can never disagree on the value and only the printed precision differs (the ring reads an integer, the card one decimal). The kernel's helper is internal to its browser bundle with no package export, so the plugin mirrors the formula (occupancyPercent in src/client/headline.ts) instead of importing it; where the meter's projection key is absent altogether (an older line), the card falls back to its own derived anchor. The timing card's wait/generation boundary deliberately diverges from that rule: it opens at the first OBSERVABLE instant — the first token, or an earlier block-start marker the token packing could not stamp (a redacted reasoning block leaves no chunk behind) — so the marker-tiled decode buckets and the wait/generation split read one window and the legend's shares never double-count it. When a supported line moves these implementations, the plugin's copies are updated to the newest one — the fold's statistics are differentially verified against the harness's own projection values on real logs (see "What each check means").

Web client seams

The browser half rides generation-specific seats, each reached through an optional seam so a deployment without the service simply goes without the capability. Capabilities only newer lines serve are reached through a deferred ctx.inject — never a hard module inject — so a composition without the service composes fully with the capability absent (no pending fiber, no throw): the service shape is re-proved before use, the registration is guarded, and the module is declared in dsh.client.inject. Every supported line ships the right Sidebar and the keyed main conversation panel; the plugin contributes to whichever face the running line serves:

Seam	V4 (0.1.7-rc.2, 0.2.0-rc.2)
Conversation panel root	keyed main panel's main.conversation (the plugin's Conversation-view tabs hang under it: Context (order 20), and Fleet (30) when defaultPlacement serves tabs and the fleetTab preference is not hidden)
Insight page (panel + entry)	the layout's keyed main slot under the dsh-context panel key, with the sidebar entry on the sidebar.panellist list under the same id — the shell owns the entry row and panel switching; both spellings probed per baseline (the insightPage probe)
Right Sidebar tab + guide entry	sidebarRightTabs registry, sidebar.right.pane.tab(.title) seats, guide capsule with order/title/description/icon (an optional commandId on guide entries and bindCommands on tab actions are additive and unused by the plugin)
Preferences card seat	plugins.bundle.config over the configForms transport
Session jump	uiWorkspace.openSession(target: SessionTarget) — the sidebar row click's own verb, its parameter re-spelled by the 0.1.6 selection refactor; the sessions service carries no selection verb any more

The guide capsule carries order, title, description, and icon (the guideEntry probe in tests/baselines.ts pins the fields of the supported generation).

The File Activity card's file-name affordance is an optional-column story: the plugin opens the file's preview tab through ctx.sidebarRight.openResource (the built-in Files sidebar's own path), building the dsh-resource://file/… address with the harness's browser-safe fileAddressFor (inlined by the client bundle, exactly as the Sidebar's own file types inline it). Where the preview face is absent (or refuses the address), the name keeps its original system-open behavior; the sidebar.nav probe pins the navigation face's spelling.

The session jump — a Context Insights session card or an Agent network node bringing its session forward — rides the harness's own selection verb (openSessionVia in src/client/services.ts): uiWorkspace.openSession, the verb the built-in sidebar's row click uses on every supported line. The face is re-proved per call and a deployment serving none (or a hostile face) swallows silently — the verb's own navigation returns the center column to the conversation (selectPanel(null)), so a dead jump simply stays on the page. Issue #90 was exactly the retired-spelling trap: the plugin once called the sessions service's own open(id), which the 0.1.6 selection refactor removed from the service (the click then did nothing on 0.1.7-rc.2), so the jump follows the view owner and the per-baseline sessionNav probe pins the verb's signature and the retired spelling's absence.

The Context Insights page's DeepSeek balance capsule is an all-optional stack, so it arms nowhere it cannot: the host reaches the settings and credentials services only inside a deferred ctx.inject (and re-reads them per request, so a key added while the host runs is picked up on the next open), mounts /api/dsh-context/balance through the same Connection exact-route registry as the detail and backfill routes, and the outbound /user/balance read carries the key only inside the host process. The provider's settings row is read off the Config-form describe() projection — the row is the entry whose served value declares the top-level apiKeyEnv credential ref, preferred under the settings ids the generations have served it as (llm-deepseek on product profiles, llm-deepseek-api-key elsewhere). The capsule appears only on a live, well-formed answer; a deployment whose DeepSeek API-key provider is not composed (no settings row), whose credentials service cannot resolve the key, whose connection service lacks the exact-route registry, or whose platform read fails (network, non-2xx, malformed payload, a third-party endpoint that does not serve the balance path) answers a typed null and the browser renders nothing — no pill, no spinner, no error state. The balance probe pins the settings face and the provider row's sources per baseline.

What each check means

• Automated seam matrix — part of this repository's pnpm test (the compat vitest project). For every baseline tag it stages the harness's REAL sources at that tag, boots the plugin's built host entry into that tag's actual SessionProjectionRegistry on the cordis release the line vendors, and probes the tag's client seams (slots, finalized-nodes seat, image loader, history face/envelope, markdown chrome, platform module table, that generation's durable-event vocabulary, the right Sidebar tab seam, its guide-entry contract and its resource-navigation face, the session-jump seam, the preferences card seat and settings transport). Definitions live in tests/baselines.ts — the one source of truth for the supported baselines and their per-version seam faces; both workflows fetch the pinned baseline tags before testing, and the compat project fails rather than skips when they are absent. Retired seams are asserted ABSENT (settings.register, the settings get(ns) read, the sessions service's open(id)), so a resurrection names the seam instead of silently shadowing the current face. Every event family src/host/fold.ts switches on is read back out of that source and asserted to be declared by every baseline, so a new fold case cannot go unprobed. To add a future dsh release: add one Baseline entry (tag, vendored cordis, faces) to tests/baselines.ts and its tag to both workflows' fetch step, then run pnpm test and re-fit whichever seam a failing probe names. The compat project boots the BUILT plugin (pnpm run build first) against the ACTUAL harness sources at each baseline tag and needs a dsh checkout (env DSH_REPO, default ~/dev/deepseek-harness).
• Statistics against the harness's own folds — the plugin's figures were verified differentially against the harness's OWN projection values (sessionStats, contextBreakdown, contextPressure, tokenUsage) over real session logs: system/tools/message tokens, per-request counts, turns/steps, TTFT, generation, tool time, and every billed cost bucket match exactly.
• Disposable-profile install / uninstall — for each release, that exact dsh CLI version was installed from npm into a temporary DSH_HOME (the real ~/.dsh is never touched), then:
  1. dsh plugin --profile <disposable> add dsh-context — install OK;
  2. dsh --profile <disposable> --dump-config — the bundle's - id: dsh-context row composes into the effective configuration (exactly 1 row);
  3. dsh plugin --profile <disposable> remove dsh-context — uninstall OK, dump-config back to 0 rows.

Scope

These checks prove source-level seam compatibility, statistical parity with the harness's own folds, and disposable-profile install/start-composition/uninstall per release. They are not a claim of full web-app runtime acceptance on a real Profile — visible UI behavior depends on the harness generation and the browser half, which the seam matrix approximates from the tag's sources.

Upgrading from an older plugin build

The skill-injection category (issue #66) bumps the timeline projection's stateVersion (18 → 19): skill content — the <available_skills> catalog digest, a user-explicit /name invocation's instructions message, and the content a skill-tool load returns — now folds into its own skill composition bucket instead of inject/tool. The re-bucketing changes the fold's per-category sums, so on upgrade every cached per-session row is invalidated and re-folded from the durable log, which rebuilds them under the new categories. The bump orphans the contextTimeline key for idle sessions, which have no refresh channel until they go live again — such a session re-folds when its log receives its next event. New sessions are exact from their first event.

The Agent network card covers the orphaned relatives without waiting for a visit: a session listed without a contextTimeline row (composition absent, occupancy only) fetches its slim head from the plugin's /api/dsh-context/detail fetch route (mounted through Connection's exact-route registry, behind the same authenticated /api fence), which folds the durable log on demand — the node's composition bar fills in completely the first time the card opens (on the Fleet tab, or inline in the right Sidebar's Context panel). The same backfill serves sessions whose cache predates the plugin's installation entirely. Where the route is unavailable (the baseline gate, a connection service without the exact-route registry) — or the cold-read gate below skips the read under heap pressure — the card degrades to the pressure-only occupancy fill as before. Sibling order follows the parent row's own subagentCatalog projection (parent catalog event order — the same sequence the harness header's subagent popup lists) whenever the list block's projection cache carries it; a deployment or baseline without the catalog projection keeps the running-first/updatedAt heuristic, and children absent from a cached catalog sort after its entries under that heuristic. A card's work state rides the same signal the header popup's own state dot does: subagentTiming.lastTurnCompleted (folded since 0.1.7-alpha.1, so on every supported line) marks a subagent's last closed turn finished, and a running session is never done either way. The list block's row carries no completion flag at all — SessionSummary has never had one — and a top-level session never carries a subagentTiming projection, so a card that is not a subagent falls back to its own folded step count against a stopped session. That fallback is a proxy, not the harness's own statement: it cannot tell a turn that ended from one the user stopped, and a session with no steps yet draws no state glyph at all.

The session cards' last-user-message preview (lastUser on the contextTimeline head) bumps the timeline projection's stateVersion (19 → 20): a stale row can never gain the field while its session is idle (it only lands when a user message folds), so on upgrade every cached timeline row is invalidated and refolded from the durable log. To keep the bump from orphaning idle sessions' rows, the warm-up (src/host/backfill.ts) probes BOTH insight rows — contextActivity AND contextTimeline — and cold-refolds every session whose either row is missing or version-stale, so previews (and every other head figure) are exact for the whole session list the first time the page opens after the upgrade. A deployment without the cold-path services — or whose client cannot mount the warm-up's trigger route — degrades to the per-session refold on the session's next activity, the same as pre-warm-up upgrades.

The Context Insights page (the sidebar panel list's first-level main panel) reads its contextActivity projection — a per-day ledger of billed tokens and completed requests, stateVersion 1 — off every session-list row's projection column. A new key is purely additive: no existing cached row is invalidated, but sessions folded before the unit existed (and sessions that predate the plugin entirely, whose contextTimeline/contextHeaders rows are missing too) have nothing to serve until they next go live. The plugin therefore runs a one-pass warm-up on demand (src/host/backfill.ts): the page is the rows' only reader, so it POSTs the plugin's /api/dsh-context/backfill fetch route (the same authenticated /api fence as the detail route) the first time it mounts, and only then does the host walk the corpus — for every stored session missing the row it cold-reads the durable log once through the projection cache's own cold-read ladder (coldSnapshot), which seeds each unit from its cached rows, folds the remainder, and writes the refreshed checkpoint back. The pass runs once per host process (later opens and later POSTs are no-ops); live sessions are skipped (they fold for themselves); a legacy log the running harness refuses to migrate is permanent — its per-session detail logs at debug and the pass ends with one summary info line, so an unmigratable corpus costs one line per page-open process, not one warn per session per boot; and a deployment whose client never opens the page never cold-reads a thing. A deployment without the sessionQuery/sessionProjectionCache/sessionPersistence/sessions services — or whose connection service lacks the exact-route registry — simply never arms the pass (such a client could not open the page to read the rows anyway).

Every cold read the plugin makes — this pass's per-session fold and the detail route's cold rung alike — runs through one host-wide cold-read gate (src/host/coldRead.ts, issue #121): reads run one at a time (FIFO), and a read starts only while the process's V8 heap sits below two-thirds of its own ceiling (the persistence backend decodes a whole log per read, so an unbounded fan-out of 10k-event sessions stacks several-hundred-MB object graphs on a heap whose ceiling may not move — a capped Electron main). A read the gate skips under pressure takes the caller's designed no-data path rather than failing: the pass stops early with one info line (folded rows persist; the rest refold on their next live activity or a later process's pass), and the detail route answers its typed null.

The page's per-day usage chart (the first row's right half) extends the contextActivity ledger with per-day pricing records — each day's billed buckets keyed provider → model → pricing period, the same SessionCostUsage raw material the timeline's session-cost totals carry, booked at each turn's initiation instant and priced client-side by the same book and estimator as the KPI band. Booking fees per (provider, model, period) changes the fold's state shape, so the unit's stateVersion bumps (1 → 2): on upgrade every cached contextActivity row reads version-stale, and the page's existing warm-up (above) cold-refolds the corpus from the durable logs, rebuilding every day's tokens AND fees exactly the first time the panel opens. On the wire the field is additive-optional: a day folded before it existed (or whose settlements carried no model header or no usage) serves without it, and the chart degrades to tokens-only bars for those days — the same sanitizer contract every other projection value rides.

The Insights page's skill card (the activity column's second card, under the heatmap: every skill loaded within the range — or the heatmap's pinned day — with its load tally and last-load instant, a row click pinning the session list to that skill's loaders) extends the contextActivity day entries with a per-name skill table. The fold reads the same two durable load gestures the timeline fold tags sub: 'skill' — a user-explicit /name invocation's skill-invocation source message and a skill-tool result's <skill_content name="…"> wrapper (the available-skills catalog digest carries no name and counts nothing) — and tallies them onto the load instant's local day, capped at 100 distinct names per day so a hostile replay cannot bloat the row. The tool-side gesture is trusted only under pairing: a result's wrapper counts solely when its callId matches a pending skill-tool call the fold tracked, so a read of source text merely QUOTING the wrapper (this file's own comments, say) never books a load; the pending list is bounded (100 ids) and consumed on pairing. Booking per-day tables changes the fold's state shape, so the unit's stateVersion bumps (3 → 4): on upgrade every cached contextActivity row reads version-stale and the page warm-up cold-refolds the corpus, rebuilding every day's tallies exactly. On the wire the field is additive-optional: a day folded before it existed (or without a load) serves without it, and the card reads the day as load-less — the same sanitizer contract as the pricing records.

The card's enrichment (what a loaded skill IS: its registry description, origin bucket, and on-disk path, joined by name onto the tallies — shown as a structured hover card per row and, while a skill is pinned, as the detail block under the rows whose path line copies on click) rides the plugin's own /api/dsh-context/skills fetch route, served off the host's skill registry rather than the harness's per-session skills/list Remote, whose wire shape strips the origin. The route mirrors that Remote's session-addressed resolution with re-proved faces — the session's observation yields its cwd and recorded preset, a live agent's preset-scoped registry wins, a cold session's preset rides a standing scope lease — because web compositions mount skill discovery on agent presets and the unscoped global registry lists nothing. Every rung degrades: the faces are read per request (an absent or hostile registry/preset/session service answers a typed null or the next-coarser view), the client re-proves every entry at its boundary (a failed proof drops that entry, a failed read renders the tallies unenriched), and a deployment without the connection fetch-route registry simply never arms it — no version gate, no state, nothing to backfill.

The timing card's session-time strip (the card's bottom band: every completed step's TTFT wait, decode blocks in stream order, tool-run windows, and in-step residue packed gapless in occurrence order — each slice's width is its share of the session's cumulative active time, with ticks at zero, the quartiles, and the full span) bumps the timeline projection's stateVersion (22 → 23): the strip's painted spans are a per-step collection the fold flushes at step/end, which later events cannot backfill for steps already folded (the timing totals' own v11 precedent), so on upgrade every cached timeline row is invalidated and refolded from the durable log, which rebuilds the spans for sessions started under older plugin builds; the v20 warm-up rebuilds idle sessions' rows the first time the page opens. On the wire the collection is additive-optional — it rides the detail channel (and the inline value on channel-less hosts), never the slim head — so a payload folded before it existed serves without it, and the timing card shows no strip then. The spans are bounded to the newest 2,000 records (a plain count cap), so an outsize session's strip paints its recent window; untrusted spans are re-proved per item at the client's render boundary, the same sanitizer contract every collection rides.

The timing totals' wait/generation boundary moving from the first token to the first observable instant (above — an earlier block-start marker wins when the token packing could not stamp it, the same boundary the strip's spans already paint) bumps the timeline projection's stateVersion (23 → 24): rows folded under 23 charged the marker-tiled decode window to BOTH the wait and the generation split, so their legend rows sum past 100%, and the mis-split totals cannot be repaired by later events, so on upgrade every cached timeline row is invalidated and refolded from the durable log; the v20 warm-up rebuilds idle sessions' rows the first time the page opens. The wire shape is unchanged — the same ttftMs/genMs fields carry the corrected sums — and the TPS chip keeps the harness-parity first-token anchor, so only the card's wait/generation shares move.

Accounting for developer/message tool-registry changes bumps the timeline projection's stateVersion (24 → 25): earlier rows omitted these model-visible messages from their surface, request composition, and archives. On upgrade the derived rows refold from the durable log; the existing page warm-up also rebuilds idle sessions. The wire shape stays unchanged.

The trend chart's duration overlay (the card's Duration toggle: each request's step active time drawn as a curve over the bars, read off the right-hand quartile axis) bumps the timeline projection's stateVersion (25 → 26): the per-request activeMs is stamped at step/end fold time — the whole step window (the model call plus its tool runs) minus the step's user waits, those being approval decision intervals (approval/asked → approval/decided, a crash-dangling pair closing at the step's own end) and ask_user_question answer windows (the tool's body is the wait itself, in both the blocking and the timed schema's foreground phase) — so requests already folded can never gain it later (the v9 cacheRead precedent), and on upgrade every cached timeline row is invalidated and refolded from the durable log, while the v20 warm-up rebuilds idle sessions' rows the first time the page opens. On the wire the field is additive-optional: a request whose step is still open (the live tail) or that folded without one (an unpaired/hostile log) serves without it, and the overlay's curve simply breaks at that bar instead of faking a point — the same sanitizer contract every collection rides.

The session node's answers tally (answers on the contextTimeline head — one per assistant message carrying a non-blank text block; steps that only dispatched tool calls are work, not replies) bumps the timeline projection's stateVersion (26 → 27): it is a running total over the complete log that later events cannot backfill for sessions folded before the field existed (the v16 humanInputs precedent), so on upgrade every cached timeline row is invalidated and refolded from the durable log, while the v20 warm-up rebuilds idle sessions' rows the first time the page opens. On the wire the field is additive-optional: a row folded before it existed serves without it and the card degrades to a zero figure. The head's sibling counts.skills figure (the distinct loaded-skill names) needed no bump of its own — it is derived at view time over the retained events already living in every cached row's state (skill-tagged inject events since v19), so a pre-field row recomputes the exact tally the moment this build serves it.

The Fleet tab's per-agent detail (delegation prompt, descriptor composition, latest reply, inter-agent messages, and the lead's roster facts) rides the plugin's /api/dsh-context/fleet fetch route — the same authenticated /api fence, Connection exact-route registry, and cold-read gate as the detail route, folding nothing token-shaped of its own: it lifts subagent/descriptor, team/member, and user/message relay/mailbox/settled records out of one sessionQuery.observeSession read per session (live-preferred, so live and cold sessions share the path). Every seam degrades independently: a deployment without the route (or a skipped cold read) renders the inspector's core and hides the detail sections; a harness without the experimental Agent Teams plugin serves no agentTeam projection, so the members board and task board simply never render; the relay/mailbox/settled source kinds absent from a session's log yield an empty comms timeline, not an error. Sender-framing preambles ("Agent sent a message:", "Team message from :", "Team message from :") strip at extraction; the subagent-settled English boilerplate maps to localized variants at render, with unrecognized phrasings falling back to the raw text.

To force a full refold of an existing session sooner, delete its cached projection row — it is a derived cache and is rebuilt from the durable log:

rm ~/.dsh/storages/session_projcache/sessions/<session-id>.json