Pi Posthorse synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Continue a task without retaining its whole active conversation
Pi Posthorse is a Pi extension for continuing long tasks across fresh model context windows. This account covers the complete extension at commit 307ef6bf05bc8bd8d718725761ffdf2f2ef72539, with applicability cutoff 2026-10-14. It is code-grounded: SRC-1 supplies implementation, declared design and attributed reports, but no supplied execution trace or causal experiment. Pi's model execution, scheduling, journal persistence and final compaction commitment remain external responsibilities. The shipped SDK callers and selected evaluators are included source, not observed passing runs. See the boundary and evidence and Source register.
The mechanism separates active input from retained material. A reset requests removal of the previous active conversation, leaving a small continuation payload. Caller-maintained notes remain in files; earlier conversation remains recoverable through the host journal. Neither store is automatically reloaded in full. This supports a continuation architecture, not a demonstrated claim that a model restores the right state or completes the task. The package's no-summary claim applies to the reached, supported custom automatic path, not every rollover: manual compaction and unsupported budgets leave native host behavior available. See RT-CLM-context-continuation, MEM-OBJ-notes and MEM-OBJ-session-journal.
Prepare and admit the next window
The extension installs request guidance, rollover hooks, capacity reporting, notes and history tools. The foreground model chooses task actions and explicit checkpoint content; Posthorse code controls callback admission, recovery selection and page delivery; Pi commits boundaries and schedules subsequent calls. The active model's exact version, parameter mutability and endpoint routing are not pinned by the extension revision. Installed capabilities do not establish deployed tool activation or permissions. See RT-CMP-active-model, RT-RTE-request-guidance and EPI-RTE-context-report.
Before model requests, Posthorse supplies current capacity and checkpoint guidance. Normal requests receive a named system section; forced prompts use request-local custom conversation content. Eligible working turns near the rollover line can receive a sparse reminder to save concise task state and request a fresh window. Final-answer turns do not receive that wakeup, and a large turn can cross the reminder band without one. Retained fingerprints suppress duplicate reminders and stale reminders are filtered from active input. These are code-controlled delivery decisions; following the prose remains model-interpreted behavior. See RT-RTE-request-guidance and RT-BAP-model-guidance.
Notes provide a separately retrievable checkpoint. Creation, replacement and append publish caller text without a mandatory human approval gate on the model path. Whole replacement can revise or clear a note; append does not reconcile contradictions. Shared Git-root resolution and migration protect particular storage locations, not semantic consistency. Available host editors are an expressly afforded alternative writer. Publication checks neither truth nor preservation of reasons, and lexical path checks do not create a filesystem sandbox. See RT-RTE-note-admission, MEM-OBJ-notes and MEM-RTE-notes-migration.
For an explicit reset, the model submits continuation prose through new_context. The callback checks capacity and records its own call ID; its success means requested, not committed. At turn end, a matching foreground result, sibling-error checks, current abort/error checks and recalculated pending-input-aware capacity determine whether Posthorse offers a retain-none compaction draft. Pi must then commit and continue. A foreign same-named callback cannot authorize this draft merely by returning success. Rejection leaves current context available but does not undo successful sibling writes or synchronize background work. Later-handler aborts and edits remain outside this enforcement point. See RT-OBJ-reset-requests, RT-RTE-explicit-reset, MEM-OBJ-explicit-handoff and RT-BAP-compaction-draft.
Recover automatically when the custom route is reached
Automatic recovery uses deterministic code rather than a summary-model call. A supported host compaction hook selects from branch entries and the live projection; a narrower threshold fallback applies only to eligible completed fresh-window turns. Construction failure cancels within the supported automatic hook rather than silently yielding to summarization. Manual or unsupported cases pass through, and host preparation, competing extensions and later edits can affect the final outcome. The eager threshold draft cannot incorporate arbitrary later same-phase edits. See RT-RTE-automatic-recovery.
Recovery honors projected omissions and replacements. It distinguishes direct owner inputs from visible coordination and treats an ask_question answer as owner input only with the required effective registration, unnamespaced projected call and successful linked result. This checks recorded provenance, not truth or present authorization. Tool receipts can include errors and orphan/legacy asynchronous results without certifying completed work. See EPI-RTE-recovery-provenance and EPI-OBJ-result-evidence.
The resulting record keeps selected owner anchors, budgeted input indices, bounded receipts and retrieval references. Consecutive automatic records are not nested wholesale; ancestry can carry earlier owner anchors or stop at a nonautomatic checkpoint. Thus deterministic construction does not exclude older model-authored interpretation. Excerpts can lose qualifiers, retain only headers or omit older evidence. Fixed warnings mark stale checkpoints and distinguish inputs from progress, but they do not mechanically prevent model reliance on stale text. See MEM-OBJ-recovery-record and EPI-RTE-recovery-reshape.
Restore selected evidence and deliver an answer
The next invocation receives an admitted handoff or recovery record automatically. Notes and detailed journal content require explicit retrieval. Notes use paths or literal text; history uses query, branch/project scope, cursors or exact entry IDs, with originals prioritized before recovery echoes. Project archive search can include nested subagent journals; it does not launch those agents, initialize their input or establish a completion barrier. Original and context-edit replacement IDs refer to different evidence versions. Normalized history is not byte-for-byte replay. See MEM-RTE-notes-recall and MEM-RTE-history-recall.
Notes and history share per-turn reservations, including sibling returns and stored images. Delivery is bounded using native estimates, character-based approximations, image estimates and margins; it is not provider-exact sizing and does not cover unrelated tools or raw transports. A partial page or refusal can require another call or a reset. Operator card expansion displays the returned page, not unseen pages. See RT-CMP-page-reservations, RT-RTE-retained-retrieval and MEM-OBJ-posthorse-display.
The representative shipped workflow is save a note → request reset → read notes/search history → return an answer. Its faux provider scripts these steps and the terminal response Recovered.; assertions check retained markers and journal reopening. A separate failed-sibling scenario scripts rejection followed by a new checkpoint request. It establishes a connected evaluator design, not autonomous checkpoint choice, live-state verification or observed recovery. Task verification is requested by guidance, not enforced by these checks. See the workflow account and EPI-RTE-reset-evaluator.
What this progression establishes
For a foreground task whose state can be checkpointed into concise prose and later explicitly retrieved, the implemented mechanism provides bounded continuation and recovery channels. Where a task depends on complete evidence, concurrent background completion or fresh external state, those channels do not establish completeness, synchronization or current correctness. Publication and reset admission are operational checks, not evidence-consuming acceptance of task propositions. Provenance selection preserves attribution of selected material while truncation degrades completeness; source warrant remains unknown. Capacity reporting derives arithmetic from host estimates, not independently warranted physical capacity. See MEM-BAP-recalled-content, EPI-RTE-recovery-reshape, EPI-RTE-context-report and the epistemic bounded conclusion.
The strongest supported learning-related contribution is an automatic trace-fed update: recorded conversation and tool experience become a retained continuation artifact with a later model-consumption channel. This meets the memory comparison's trace-learning meaning at wired strength. It does not establish improved capacity for future action; no outcome comparison was performed. Other note, handoff, editor and native-checkpoint update chains remain unresolved. The profile verification preserves this positive finding without upgrading it to observed learning. Its underlying support is MEM-OBJ-recovery-record, EPI-RTE-recovery-reshape and MEM-BAP-recalled-content.
No consumed-theory criticism and resulting capacity improvement is established. Theory-builder conditions 1–4 are not separately established by the accepted runtime records; task prose and fixed guidance cannot substitute for the required theory, consumption, content-directed criticism and iteration evidence. Reflection is not established, and computational authorship/admission does not establish autonomous theory building. Self-improvement is independently unestablished: continuation wiring does not show objective-bearing evaluation determining an improvement-directed organizational update or later operation causally depending on it. These are separate unresolved properties, not a bundled negative about the external model. See RT-RTE-request-guidance, RT-RTE-automatic-recovery and the epistemic bounded conclusion.
Inspectably linked real-model checkpoint/restoration traces would change the operation assessment. Objective-specific outcome comparisons could establish benefit; independently varying recovery selection would be needed to attribute a component effect. Audited host contracts and hook-ordering evidence would strengthen commitment claims. Full-scan comparisons across journal encodings would address recall equivalence. An evidence-consuming task-state acceptance route would change the current distinction between operational admission and semantic acceptance. None is supplied here; the report verification accepts the bounded mechanisms, not these stronger conclusions.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Recall equivalence and normalized replay are uncertified; carried report-verification limit | SRC-1; MEM-RTE-history-recall, RT-RTE-retained-retrieval; verification limit | Optional file/line prefilters, normalized matching and failure fallback in src/session-files.ts, src/history-search.ts, src/history-entries.ts |
Exhaustive-recall and exact-replay comparisons across arbitrary encodings and queries | Equivalence evidence against complete scans across the claimed encoding/query domain; an explicit normalization/replay contract |
| Selected evaluator coverage only; carried report-verification limit | SRC-1; EPI-RTE-reset-evaluator, workflow; verification limit | Selected reset/read/reopen and rejection/retry fixtures; other assertion bodies inventoried, not exhaustively audited; engineering checker plane excluded | Total test-coverage, checker-correctness and general recovery-reliability comparisons; passing execution or live task verification | Complete relevant evaluator audit and inspectable execution results; task-specific state checks for task reliability |
| External host, provider and deployed grants are not frozen | SRC-1; boundary, RT-CMP-active-model, RT-BAP-compaction-draft, RT-CMP-page-reservations | Extension-side drafts, estimates and file-consumption contracts; external persistence, scheduling, weights and transports excluded | Independent host qualification, pinned model behavior, deployment-independent commitment, isolation or provider-exact capacity guarantees | Frozen host/provider contracts and relevant qualification artifacts, deployment tool/grant configuration and physical payload measurements |
| Hook checks are local, not a transaction over all activity | SRC-1; RT-RTE-explicit-reset, RT-RTE-automatic-recovery | Foreground batch and reached handlers; eager fallback, later-handler countercases, sentinel append and competing-handler possibilities | Background synchronization, rollback of completed writes/all journal mutations, or unconditional abort/edit veto | Host-level ordering/commit contracts and synchronization/transaction mechanisms covering those alternate paths |
| Initial notes and external/native production remain unresolved | SRC-1; MEM-OBJ-notes, MEM-OBJ-session-journal, MEM-OBJ-recovery-record; profile verification | Subsequent note writes, external-editor affordance, consumed native/legacy checkpoints and stored media | Complete provenance, form, admission, authority, curation and trace-input inventories; universal human approval or model authorship | Content lineage, editor approval/implementation evidence, native checkpoint production traces and full operative-payload classification |
| Selection does not preserve complete semantic evidence or validate task truth | SRC-1; EPI-RTE-recovery-provenance, EPI-RTE-recovery-reshape, MEM-OBJ-notes, MEM-BAP-recalled-content | Inspected note/reset/recovery dispatches, bounded excerpts and model-interpreted warnings | Complete rationale, semantic equivalence, current-progress certification or accepted task claims from successful publication/retention alone | Retained full evidence and actual consumer traces; proposition-specific evidence-consuming acceptance with criterion, intended use and scope |
| No supplied operation or causal improvement evidence | SRC-1; RT-CLM-context-continuation, MEM-OBJ-recovery-record, EPI-RTE-reset-evaluator | Source implementation and scripted evaluator design; no target execution or controlled comparison | Observed restoration, model compliance, exercised learning, improved future capacity, production-cost benefit or component-causal attribution | Linked execution artifacts and objective-specific comparisons, with independent component variation for component claims |
| Theory criticism and improvement properties lack supporting chains | SRC-1; RT-RTE-request-guidance, RT-RTE-automatic-recovery, RT-CMP-active-model; epistemic conclusion | Fixed extension guidance, trace-fed continuation and opaque external model processing | Separate affirmative theory-builder conditions, criticism-induced capacity improvement, reflection, autonomous theory building or self-improvement attribution | Individuated theories, decisions, criticisms and retained next rounds; reflection's two causal directions; role-level autonomy evidence; objective-bearing updates and subsequent causally dependent operation for self-improvement |