'use strict';

/**
 * Transcript compaction — token-saving rewrites applied to the persisted
 * conversation history between turns.
 *
 * The `patch` tool's arguments embed entire file bodies: the `search` and
 * `replace` text of every operation. Once a patch has been applied those bodies
 * are pure dead weight — they are re-sent to the model on every later turn but
 * are never needed again. `compactPatchToolCalls` rewrites each `patch` tool
 * call in the transcript so its long text fields keep only a head/tail excerpt,
 * with a marker noting how many lines were elided.
 *
 * `invalidateStaleReadResults` attacks the same problem from the other side:
 * a `read_file` result carries a whole file body, and once a later `patch`
 * writes that file the body no longer matches what is on disk. Keeping it is
 * both a token waste and a correctness hazard (the model may reason from stale
 * contents). The stale result is replaced by a short "file was overwritten
 * meanwhile" marker that tells the model to re-read if it needs the file.
 */

import path from 'path';
import { ROOT_DIR } from './config.js';

// Lines kept from the top and bottom of each long text field.
const HEAD_LINES = 5;
const TAIL_LINES = 5;

// Explanatory text carried by the omission marker. The marker is inserted by
// the memory-management system when long tool-call text is elided between turns
// to save tokens. It is NOT written by the model, and the note says so
// explicitly so the model does not mistake it for its own earlier output.
const OMISSION_NOTE =
    'omitted from memory to save tokens — note to self: the Fido Agent did NOT ' +
    'write this note; the memory management system did';

/** Escape a string so it can be embedded literally inside a RegExp. */
const escapeRegExp = (s) => s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');

// Matches the omission marker line emitted below, so an already-compacted field
// can be recognised and left alone (keeping repeated passes idempotent). Built
// from OMISSION_NOTE so the pattern can never drift from the emitted marker.
const OMISSION_MARKER_RE = new RegExp(`^… \\[\\d+ lines? ${escapeRegExp(OMISSION_NOTE)}\\] …$`);

/**
 * Keep the first `head` and last `tail` lines of `text`, inserting a marker that
 * notes how many lines were omitted. Text that is already short enough
 * (<= head + tail lines), or that already carries an omission marker, is
 * returned unchanged.
 */
export function excerptLines(text, head = HEAD_LINES, tail = TAIL_LINES) {
    if (typeof text !== 'string') return text;
    const lines = text.split('\n');
    if (lines.length <= head + tail) return text;
    // Already compacted — don't re-truncate (the marker itself would otherwise
    // be counted as an extra line and trimmed away on the next pass).
    if (lines.some((l) => OMISSION_MARKER_RE.test(l))) return text;
    const omitted = lines.length - head - tail;
    return [
        ...lines.slice(0, head),
        `… [${omitted} line${omitted === 1 ? '' : 's'} ${OMISSION_NOTE}] …`,
        ...lines.slice(-tail),
    ].join('\n');
}

/**
 * Rewrite every `patch` tool call in `messages` in place, replacing the long
 * `search` and `replace` fields of each operation with a head/tail excerpt.
 * Malformed arguments are left untouched. Returns the number of tool calls that
 * were actually compacted.
 */
export function compactPatchToolCalls(messages) {
    if (!Array.isArray(messages)) return 0;
    let compacted = 0;
    for (const m of messages) {
        if (!m || m.role !== 'assistant' || !Array.isArray(m.tool_calls)) continue;
        for (const tc of m.tool_calls) {
            if (tc?.function?.name !== 'patch' || typeof tc.function.arguments !== 'string') continue;
            let args;
            try {
                args = JSON.parse(tc.function.arguments);
            } catch {
                continue; // leave malformed arguments untouched
            }
            if (!args || !Array.isArray(args.operations)) continue;
            let changed = false;
            for (const op of args.operations) {
                if (!op || typeof op !== 'object') continue;
                for (const field of ['search', 'replace']) {
                    const next = excerptLines(op[field]);
                    if (next !== op[field]) {
                        op[field] = next;
                        changed = true;
                    }
                }
            }
            if (changed) {
                tc.function.arguments = JSON.stringify(args);
                compacted++;
            }
        }
    }
    return compacted;
}

// ─── Stale read_file results ────────────────────────────────────────────────────

// Marker carried by a replaced read_file result. Detected verbatim so repeated
// passes over the transcript stay idempotent (an already-invalidated result is
// left as-is instead of being wrapped again).
const STALE_MARKER = 'file was overwritten meanwhile';

// Explanatory text carried by the replacement. Like the omission marker above,
// this is written by the memory-management system, not by the model — the note
// says so explicitly so the model does not mistake it for its own earlier
// output or for real file contents.
const STALE_NOTE =
    `${STALE_MARKER} — stale contents elided by the memory management system ` +
    'to save tokens; re-read with read_file for the current contents';

// `applyOperations` tags each change it makes as either an overwrite or a
// search/replace; both mean the file body a prior read returned is now stale.
const APPLIED_RE = /^(?:CREATE\/WRITE|PATCH in) (.+)$/;

/**
 * Canonical key for a workspace-relative path, so a path the model wrote as
 * "src/x.js" matches one written as "./src/x.js" or with different separators.
 */
function canonicalPath(p) {
    if (typeof p !== 'string' || !p) return null;
    try {
        return path.resolve(ROOT_DIR, p);
    } catch {
        return p;
    }
}

/** The replacement payload handed to the model in place of a stale read. */
function staleReadResult(filename) {
    return JSON.stringify({ filename, note: STALE_NOTE });
}

/**
 * Extract the files a `patch` result actually wrote. Both successful results
 * (`applied`) and partially-failed ones (`appliedSoFar`) list the mutations, so
 * files that were never touched are correctly left alone. Returns canonical
 * path keys; an unparseable result yields an empty list.
 */
function filesWrittenByPatch(resultContent) {
    let res = resultContent;
    if (typeof res === 'string') {
        try {
            res = JSON.parse(res);
        } catch {
            return [];
        }
    }
    if (!res || typeof res !== 'object') return [];
    const applied = Array.isArray(res.applied)
        ? res.applied
        : Array.isArray(res.appliedSoFar)
            ? res.appliedSoFar
            : [];
    const files = new Set();
    for (const entry of applied) {
        if (typeof entry !== 'string') continue;
        const hit = APPLIED_RE.exec(entry);
        if (hit) {
            const key = canonicalPath(hit[1]);
            if (key) files.add(key);
        }
    }
    return [...files];
}

/**
 * Replace `read_file` tool results in `messages` with a short "file was
 * overwritten meanwhile" marker whenever a later `patch` wrote the same file.
 *
 * A read result is stale only if the file changed *after* it was read, so we
 * compare transcript positions: a read is invalidated when a successful patch
 * write sits at a later index. Reads taken after their last patch (which still
 * reflect disk) are left untouched, as are reads for files no patch ever wrote.
 *
 * The rewrite only fires when it strictly shrinks the payload, so a tiny read
 * result is never traded for a longer marker. Repeated passes are idempotent.
 * Returns the number of results invalidated.
 */
export function invalidateStaleReadResults(messages) {
    if (!Array.isArray(messages)) return 0;

    // Index every tool call so a tool result can be mapped back to (name, args).
    const callInfo = new Map();
    for (const m of messages) {
        if (!m || m.role !== 'assistant' || !Array.isArray(m.tool_calls)) continue;
        for (const tc of m.tool_calls) {
            if (!tc || !tc.id) continue;
            let args = {};
            try {
                args = JSON.parse(tc.function?.arguments || '{}');
            } catch {
                // Malformed arguments: the call stays indexed but argless.
            }
            callInfo.set(tc.id, { name: tc.function?.name, args });
        }
    }

    // Every place a patch wrote a file, keyed by transcript position.
    const writes = [];
    for (let i = 0; i < messages.length; i++) {
        const m = messages[i];
        if (!m || m.role !== 'tool') continue;
        if (callInfo.get(m.tool_call_id)?.name !== 'patch') continue;
        const files = filesWrittenByPatch(m.content);
        if (files.length) writes.push({ index: i, files });
    }
    if (writes.length === 0) return 0;

    let invalidated = 0;
    for (let i = 0; i < messages.length; i++) {
        const m = messages[i];
        if (!m || m.role !== 'tool') continue;
        const info = callInfo.get(m.tool_call_id);
        if (info?.name !== 'read_file') continue;

        const key = canonicalPath(info.args?.filename);
        if (!key) continue;

        const content = typeof m.content === 'string' ? m.content : '';
        // Already replaced on an earlier pass — leave it alone.
        if (content.includes(STALE_MARKER)) continue;

        // Stale only if some patch wrote this file after the read happened.
        const stale = writes.some((w) => w.index > i && w.files.includes(key));
        if (!stale) continue;

        const replacement = staleReadResult(info.args.filename);
        // Never grow the payload: only swap when the marker is actually smaller.
        if (replacement.length >= content.length) continue;

        m.content = replacement;
        invalidated++;
    }
    return invalidated;
}